From 8bb31d8ac6547dd3bac94b0147b81829de46d534 Mon Sep 17 00:00:00 2001 From: fracape Date: Fri, 8 Dec 2023 19:30:26 +0000 Subject: [PATCH] deploy: 65a0b4f778c2f083356986ef7d8e4b6647e5fb15 --- tutorials/cli_tutorial_2.html | 2 +- tutorials/cli_tutorial_2_nb.html | 2 +- tutorials/cli_tutorial_3.html | 2 +- tutorials/cli_tutorial_3_nb.html | 2 +- tutorials/cli_tutorial_6.html | 78 ++++----- tutorials/cli_tutorial_6_nb.html | 78 ++++----- tutorials/cli_tutorial_7.html | 64 ++++---- tutorials/cli_tutorial_7_nb.html | 64 ++++---- tutorials/convert_nb.html | 132 ++++++++-------- tutorials/detectron2.html | 188 +++++++++++----------- tutorials/detectron2_nb.html | 188 +++++++++++----------- tutorials/download.html | 92 +++++------ tutorials/download_nb.html | 92 +++++------ tutorials/encdec.html | 74 ++++----- tutorials/encdec_nb.html | 74 ++++----- tutorials/evaluate.html | 264 +++++++++++++++---------------- tutorials/evaluate_nb.html | 264 +++++++++++++++---------------- tutorials/fiftyone.html | 38 ++--- tutorials/fiftyone_nb.html | 38 ++--- tutorials/index.html | 132 ++++++++-------- 20 files changed, 934 insertions(+), 934 deletions(-) diff --git a/tutorials/cli_tutorial_2.html b/tutorials/cli_tutorial_2.html index b37ddca0..26a31b53 100644 --- a/tutorials/cli_tutorial_2.html +++ b/tutorials/cli_tutorial_2.html @@ -799,7 +799,7 @@

2. Registering Datasets

Let’s remove the image data as well:

-
rm -rf /tmp/my_data_set
+
rm -rf /tmp/my_data_set
 

 

 
A final note/observation before moving to the next tutorial.

diff --git a/tutorials/cli_tutorial_2_nb.html b/tutorials/cli_tutorial_2_nb.html
index 3e1874f7..63a23890 100644
--- a/tutorials/cli_tutorial_2_nb.html
+++ b/tutorials/cli_tutorial_2_nb.html
@@ -805,7 +805,7 @@ 
 Contents

Let’s remove the image data as well:

-
rm -rf /tmp/my_data_set
+
rm -rf /tmp/my_data_set
 

 

 
A final note/observation before moving to the next tutorial.

diff --git a/tutorials/cli_tutorial_3.html b/tutorials/cli_tutorial_3.html
index 105f1821..0ba11c81 100644
--- a/tutorials/cli_tutorial_3.html
+++ b/tutorials/cli_tutorial_3.html
@@ -364,7 +364,7 @@ 
3. MPEG-VCM Evaluation
oiv6-mpeg-detection-v1

 
  • oiv6-mpeg-segmentation-v1
    • 
 
-
    compressai-vision list
+
    compressai-vision list
 
    
 
    
 
    importing fiftyone
diff --git a/tutorials/cli_tutorial_3_nb.html b/tutorials/cli_tutorial_3_nb.html
index 75a3b439..df9e4ef3 100644
--- a/tutorials/cli_tutorial_3_nb.html
+++ b/tutorials/cli_tutorial_3_nb.html
@@ -370,7 +370,7 @@ 
     Contents 
    
 
  • oiv6-mpeg-detection-v1
    • 
 
  • oiv6-mpeg-segmentation-v1
    • 
 
-
    compressai-vision list
+
    compressai-vision list
 
    
 
    
 
    importing fiftyone
diff --git a/tutorials/cli_tutorial_6.html b/tutorials/cli_tutorial_6.html
index 78d14d60..4cf786b4 100644
--- a/tutorials/cli_tutorial_6.html
+++ b/tutorials/cli_tutorial_6.html
@@ -362,15 +362,15 @@ 
    6. VTM benchmark generationvtm subcommand to manage, encode and cache the VTM produced
 bitstreams on disk.
    
 
    Let’s generate some encoded bitstreams.
    
-
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
---slice=0:2 \
---scale=100 \
---progress=1 \
---qpars=47 \
---vtm_cache=/tmp/bitstreams \
---vtm_dir={path_to_vtm_software}/bin \
---vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
---output=vtm_out.json
+
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
+--slice=0:2 \
+--scale=100 \
+--progress=1 \
+--qpars=47 \
+--vtm_cache=/tmp/bitstreams \
+--vtm_dir={path_to_vtm_software}/bin \
+--vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
+--output=vtm_out.json
 
    
 
    
 
    importing fiftyone
@@ -403,15 +403,15 @@ 
    6. VTM benchmark generation/tmp/bitstreams/SCALE/QP. Let’s see what happens if we run the exact
 same command again:
    
-
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
---slice=0:2 \
---scale=100 \
---progress=1 \
---qpars=47 \
---vtm_cache=/tmp/bitstreams \
---vtm_dir={path_to_vtm_software}/bin \
---vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
---output=vtm_out.json
+
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
+--slice=0:2 \
+--scale=100 \
+--progress=1 \
+--qpars=47 \
+--vtm_cache=/tmp/bitstreams \
+--vtm_dir={path_to_vtm_software}/bin \
+--vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
+--output=vtm_out.json
 
    
 
    
 
    importing fiftyone
@@ -442,19 +442,19 @@ 
    6. VTM benchmark generation
    echo " " > /tmp/bitstreams/100/47/bin_000a1249af2bc5f0
+
    echo " " > /tmp/bitstreams/100/47/bin_000a1249af2bc5f0
 
    
 
    
 
    And run the command again:
    
-
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
---slice=0:2 \
---scale=100 \
---progress=1 \
---qpars=47 \
---vtm_cache=/tmp/bitstreams \
---vtm_dir={path_to_vtm_software}/bin \
---vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
---output=vtm_out.json
+
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
+--slice=0:2 \
+--scale=100 \
+--progress=1 \
+--qpars=47 \
+--vtm_cache=/tmp/bitstreams \
+--vtm_dir={path_to_vtm_software}/bin \
+--vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
+--output=vtm_out.json
 
    
 
    
 
    importing fiftyone
@@ -493,17 +493,17 @@ 
    6. VTM benchmark generationdetectron2-eval for the VTM case like this:
    
-
    compressai-vision detectron2-eval --y --dataset-name=oiv6-mpeg-detection-v1 \
---slice=0:2 \
---scale=100 \
---progress=1 \
---qpars=47 \
---vtm \
---vtm_cache=/tmp/bitstreams \
---vtm_dir={path_to_vtm_software}/bin \
---vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
---output=detectron2_vtm.json \
---model=COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml
+
    compressai-vision detectron2-eval --y --dataset-name=oiv6-mpeg-detection-v1 \
+--slice=0:2 \
+--scale=100 \
+--progress=1 \
+--qpars=47 \
+--vtm \
+--vtm_cache=/tmp/bitstreams \
+--vtm_dir={path_to_vtm_software}/bin \
+--vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
+--output=detectron2_vtm.json \
+--model=COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml
 
    
 
    
 
    importing fiftyone
diff --git a/tutorials/cli_tutorial_6_nb.html b/tutorials/cli_tutorial_6_nb.html
index c29f4f33..492151b9 100644
--- a/tutorials/cli_tutorial_6_nb.html
+++ b/tutorials/cli_tutorial_6_nb.html
@@ -368,15 +368,15 @@ 
     Contents 
    
 use the vtm subcommand to manage, encode and cache the VTM produced
 bitstreams on disk.
    
 
    Let’s generate some encoded bitstreams.
    
-
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
---slice=0:2 \
---scale=100 \
---progress=1 \
---qpars=47 \
---vtm_cache=/tmp/bitstreams \
---vtm_dir={path_to_vtm_software}/bin \
---vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
---output=vtm_out.json
+
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
+--slice=0:2 \
+--scale=100 \
+--progress=1 \
+--qpars=47 \
+--vtm_cache=/tmp/bitstreams \
+--vtm_dir={path_to_vtm_software}/bin \
+--vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
+--output=vtm_out.json
 
    
 
    
 
    importing fiftyone
@@ -409,15 +409,15 @@ 
     Contents 
    
 
    As you can see, bitstreams we’re generated and cached into
 /tmp/bitstreams/SCALE/QP. Let’s see what happens if we run the exact
 same command again:
    
-
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
---slice=0:2 \
---scale=100 \
---progress=1 \
---qpars=47 \
---vtm_cache=/tmp/bitstreams \
---vtm_dir={path_to_vtm_software}/bin \
---vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
---output=vtm_out.json
+
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
+--slice=0:2 \
+--scale=100 \
+--progress=1 \
+--qpars=47 \
+--vtm_cache=/tmp/bitstreams \
+--vtm_dir={path_to_vtm_software}/bin \
+--vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
+--output=vtm_out.json
 
    
 
    
 
    importing fiftyone
@@ -448,19 +448,19 @@ 
     Contents 
    
 
    Instead of generating the bitstreams, the program found them cached on
 the disk and just verified them.
    
 
    Let’s fool around and corrupt one of the bitstreams:
    
-
    echo " " > /tmp/bitstreams/100/47/bin_000a1249af2bc5f0
+
    echo " " > /tmp/bitstreams/100/47/bin_000a1249af2bc5f0
 
    
 
    
 
    And run the command again:
    
-
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
---slice=0:2 \
---scale=100 \
---progress=1 \
---qpars=47 \
---vtm_cache=/tmp/bitstreams \
---vtm_dir={path_to_vtm_software}/bin \
---vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
---output=vtm_out.json
+
    compressai-vision vtm --y --dataset-name=oiv6-mpeg-detection-v1 \
+--slice=0:2 \
+--scale=100 \
+--progress=1 \
+--qpars=47 \
+--vtm_cache=/tmp/bitstreams \
+--vtm_dir={path_to_vtm_software}/bin \
+--vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
+--output=vtm_out.json
 
    
 
    
 
    importing fiftyone
@@ -499,17 +499,17 @@ 
     Contents 
    
 of crashes / data corruption, you can just send the same scripts into
 your queue system over and over again if necessary.
    
 
    Finally, you can run detectron2-eval for the VTM case like this:
    
-
    compressai-vision detectron2-eval --y --dataset-name=oiv6-mpeg-detection-v1 \
---slice=0:2 \
---scale=100 \
---progress=1 \
---qpars=47 \
---vtm \
---vtm_cache=/tmp/bitstreams \
---vtm_dir={path_to_vtm_software}/bin \
---vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
---output=detectron2_vtm.json \
---model=COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml
+
    compressai-vision detectron2-eval --y --dataset-name=oiv6-mpeg-detection-v1 \
+--slice=0:2 \
+--scale=100 \
+--progress=1 \
+--qpars=47 \
+--vtm \
+--vtm_cache=/tmp/bitstreams \
+--vtm_dir={path_to_vtm_software}/bin \
+--vtm_cfg={path_to_vtm_software}/cfg/encoder_intra_vtm.cfg \
+--output=detectron2_vtm.json \
+--model=COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml
 
    
 
    
 
    importing fiftyone
diff --git a/tutorials/cli_tutorial_7.html b/tutorials/cli_tutorial_7.html
index b3a68911..7085cf5e 100644
--- a/tutorials/cli_tutorial_7.html
+++ b/tutorials/cli_tutorial_7.html
@@ -542,17 +542,17 @@ 
    7. Importing and Using Video
    import cv2
-import matplotlib.pyplot as plt
-import fiftyone as fo
-from fiftyone import ViewField as F
-from math import floor
+
    import cv2
+import matplotlib.pyplot as plt
+import fiftyone as fo
+from fiftyone import ViewField as F
+from math import floor
 
    
 
    
-
    dataset=fo.load_dataset("sfu-hw-objects-v1")
+
    dataset=fo.load_dataset("sfu-hw-objects-v1")
 
    
 
    
-
    dataset
+
    dataset
 
    
 
    
 
    Name:        sfu-hw-objects-v1
@@ -577,7 +577,7 @@ 
    7. Importing and Using Video
    dataset.first()
+
    dataset.first()
 
    
 
    
 
    <Sample: {
@@ -599,12 +599,12 @@ 
    7. Importing and Using Videoname_tag to the
 video descriptive names (BasketballDrill, Traffic, PeopleOnStreeet,
 etc.). Let’s pick a certain video sample:
    
-
    sample = dataset[ (F("name_tag") == "BasketballDrill") & (F("class_tag") == "ClassC") ].first()
+
    sample = dataset[ (F("name_tag") == "BasketballDrill") & (F("class_tag") == "ClassC") ].first()
 
    
 
    
 
    Take a look at the first frame ground truth detections (note that frame
 indices start from 1):
    
-
    sample.frames[1]
+
    sample.frames[1]
 
    
 
    
 
    <FrameView: {
@@ -738,39 +738,39 @@ 
    7. Importing and Using Video
    vid=cv2.VideoCapture(sample.filepath)
+
    vid=cv2.VideoCapture(sample.filepath)
 
    
 
    
-
    print("number of frames:",int(vid.get(cv2.CAP_PROP_FRAME_COUNT)))
+
    print("number of frames:",int(vid.get(cv2.CAP_PROP_FRAME_COUNT)))
 
    
 
    
 
    number of frames: 501
 
    
 
    
 
    Let’s define a small helper function:
    
-
    def draw_detections(sample: fo.Sample, vid: cv2.VideoCapture, nframe: int):
-    nmax=int(vid.get(cv2.CAP_PROP_FRAME_COUNT))
-    if nframe > nmax:
-        raise AssertionError("max frame is " + str(nmax))
-    ok = vid.set(cv2.CAP_PROP_POS_FRAMES, nframe-1)
-    if not ok:
-        raise AssertionError("seek failed")
-    ok, arr = vid.read() # BGR image in arr
-    if not ok:
-        raise AssertionError("no image")
-    for detection in sample.frames[nframe].detections.detections:
-        x0, y0, w, h = detection.bounding_box # rel coords
-        x1, y1, x2, y2 = floor(x0*arr.shape[1]), floor(y0*arr.shape[0]), floor((x0+w)*arr.shape[1]), floor((y0+h)*arr.shape[0])
-        arr=cv2.rectangle(arr, (x1, y1), (x2, y2), (255, 0, 0), 5)
-    return arr
+
    def draw_detections(sample: fo.Sample, vid: cv2.VideoCapture, nframe: int):
+    nmax=int(vid.get(cv2.CAP_PROP_FRAME_COUNT))
+    if nframe > nmax:
+        raise AssertionError("max frame is " + str(nmax))
+    ok = vid.set(cv2.CAP_PROP_POS_FRAMES, nframe-1)
+    if not ok:
+        raise AssertionError("seek failed")
+    ok, arr = vid.read() # BGR image in arr
+    if not ok:
+        raise AssertionError("no image")
+    for detection in sample.frames[nframe].detections.detections:
+        x0, y0, w, h = detection.bounding_box # rel coords
+        x1, y1, x2, y2 = floor(x0*arr.shape[1]), floor(y0*arr.shape[0]), floor((x0+w)*arr.shape[1]), floor((y0+h)*arr.shape[0])
+        arr=cv2.rectangle(arr, (x1, y1), (x2, y2), (255, 0, 0), 5)
+    return arr
 
    
 
    
-
    img=draw_detections(sample, vid, 200)
-img_ = img[:,:,::-1] # BGR -> RGB
+
    img=draw_detections(sample, vid, 200)
+img_ = img[:,:,::-1] # BGR -> RGB
 
    
 
    
-
    plt.imshow(img_)
-vid.release()
+
    plt.imshow(img_)
+vid.release()
 
    
 
    
 ../_images/cli_tutorial_7_nb_22_0.png
@@ -848,7 +848,7 @@ 
    7. Importing and Using Video
    cat detectron2_test.json
+
    cat detectron2_test.json
 
    
 
    
 
    {
diff --git a/tutorials/cli_tutorial_7_nb.html b/tutorials/cli_tutorial_7_nb.html
index 42d01064..d848df8a 100644
--- a/tutorials/cli_tutorial_7_nb.html
+++ b/tutorials/cli_tutorial_7_nb.html
@@ -548,17 +548,17 @@ 
     Contents 
    
 
    
 
    In order to demonstrate how video datasets are used, let’s continue in
 python notebook:
    
-
    import cv2
-import matplotlib.pyplot as plt
-import fiftyone as fo
-from fiftyone import ViewField as F
-from math import floor
+
    import cv2
+import matplotlib.pyplot as plt
+import fiftyone as fo
+from fiftyone import ViewField as F
+from math import floor
 
    
 
    
-
    dataset=fo.load_dataset("sfu-hw-objects-v1")
+
    dataset=fo.load_dataset("sfu-hw-objects-v1")
 
    
 
    
-
    dataset
+
    dataset
 
    
 
    
 
    Name:        sfu-hw-objects-v1
@@ -583,7 +583,7 @@ 
     Contents 
    
 
    
 
    In contrast to image datasets where each sample was an image, now a
 sample corresponds to a video:
    
-
    dataset.first()
+
    dataset.first()
 
    
 
    
 
    <Sample: {
@@ -605,12 +605,12 @@ 
     Contents 
    
 class directories (ClassA, ClassB, etc.), while name_tag to the
 video descriptive names (BasketballDrill, Traffic, PeopleOnStreeet,
 etc.). Let’s pick a certain video sample:
    
-
    sample = dataset[ (F("name_tag") == "BasketballDrill") & (F("class_tag") == "ClassC") ].first()
+
    sample = dataset[ (F("name_tag") == "BasketballDrill") & (F("class_tag") == "ClassC") ].first()
 
    
 
    
 
    Take a look at the first frame ground truth detections (note that frame
 indices start from 1):
    
-
    sample.frames[1]
+
    sample.frames[1]
 
    
 
    
 
    <FrameView: {
@@ -744,39 +744,39 @@ 
     Contents 
    
 
    
 
    
 
    Start reading the video file with OpenCV:
    
-
    vid=cv2.VideoCapture(sample.filepath)
+
    vid=cv2.VideoCapture(sample.filepath)
 
    
 
    
-
    print("number of frames:",int(vid.get(cv2.CAP_PROP_FRAME_COUNT)))
+
    print("number of frames:",int(vid.get(cv2.CAP_PROP_FRAME_COUNT)))
 
    
 
    
 
    number of frames: 501
 
    
 
    
 
    Let’s define a small helper function:
    
-
    def draw_detections(sample: fo.Sample, vid: cv2.VideoCapture, nframe: int):
-    nmax=int(vid.get(cv2.CAP_PROP_FRAME_COUNT))
-    if nframe > nmax:
-        raise AssertionError("max frame is " + str(nmax))
-    ok = vid.set(cv2.CAP_PROP_POS_FRAMES, nframe-1)
-    if not ok:
-        raise AssertionError("seek failed")
-    ok, arr = vid.read() # BGR image in arr
-    if not ok:
-        raise AssertionError("no image")
-    for detection in sample.frames[nframe].detections.detections:
-        x0, y0, w, h = detection.bounding_box # rel coords
-        x1, y1, x2, y2 = floor(x0*arr.shape[1]), floor(y0*arr.shape[0]), floor((x0+w)*arr.shape[1]), floor((y0+h)*arr.shape[0])
-        arr=cv2.rectangle(arr, (x1, y1), (x2, y2), (255, 0, 0), 5)
-    return arr
+
    def draw_detections(sample: fo.Sample, vid: cv2.VideoCapture, nframe: int):
+    nmax=int(vid.get(cv2.CAP_PROP_FRAME_COUNT))
+    if nframe > nmax:
+        raise AssertionError("max frame is " + str(nmax))
+    ok = vid.set(cv2.CAP_PROP_POS_FRAMES, nframe-1)
+    if not ok:
+        raise AssertionError("seek failed")
+    ok, arr = vid.read() # BGR image in arr
+    if not ok:
+        raise AssertionError("no image")
+    for detection in sample.frames[nframe].detections.detections:
+        x0, y0, w, h = detection.bounding_box # rel coords
+        x1, y1, x2, y2 = floor(x0*arr.shape[1]), floor(y0*arr.shape[0]), floor((x0+w)*arr.shape[1]), floor((y0+h)*arr.shape[0])
+        arr=cv2.rectangle(arr, (x1, y1), (x2, y2), (255, 0, 0), 5)
+    return arr
 
    
 
    
-
    img=draw_detections(sample, vid, 200)
-img_ = img[:,:,::-1] # BGR -> RGB
+
    img=draw_detections(sample, vid, 200)
+img_ = img[:,:,::-1] # BGR -> RGB
 
    
 
    
-
    plt.imshow(img_)
-vid.release()
+
    plt.imshow(img_)
+vid.release()
 
    
 
    
 ../_images/cli_tutorial_7_nb_22_0.png
@@ -854,7 +854,7 @@ 
     Contents 
    
 
    
 
    
 
    Take a look at the results:
    
-
    cat detectron2_test.json
+
    cat detectron2_test.json
 
    
 
    
 
    {
diff --git a/tutorials/convert_nb.html b/tutorials/convert_nb.html
index 2c708ea6..aa3af550 100644
--- a/tutorials/convert_nb.html
+++ b/tutorials/convert_nb.html
@@ -348,42 +348,42 @@ 
     Contents 
    
                   
   
    In this chapter, we create an evaluation dataset as defined by the
 MPEG-VCM working group
    
-
    # common libs
-import math, os, io, json, cv2, random, logging
-import numpy as np
-# images
-from PIL import Image
-import matplotlib.pyplot as plt
+
    # common libs
+import math, os, io, json, cv2, random, logging
+import numpy as np
+# images
+from PIL import Image
+import matplotlib.pyplot as plt
 
    
 
    
-
    homie=os.path.expanduser("~")
-print("your home path is", homie)
-fodir=os.path.join(homie,'fiftyone')
-print("fiftyone dowloads data by default to", fodir)
-try:
-    os.mkdir(fodir)
-except FileExistsError:
-    pass
+
    homie=os.path.expanduser("~")
+print("your home path is", homie)
+fodir=os.path.join(homie,'fiftyone')
+print("fiftyone dowloads data by default to", fodir)
+try:
+    os.mkdir(fodir)
+except FileExistsError:
+    pass
 
    
 
    
 
    your home path is /home/sampsa
 fiftyone dowloads data by default to /home/sampsa/fiftyone
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
-
    # CompressAI-Vision
-from compressai_vision.conversion import MPEGVCMToOpenImageV6, imageIdFileList
+
    # CompressAI-Vision
+from compressai_vision.conversion import MPEGVCMToOpenImageV6, imageIdFileList
 
    
 
    
 
    We expect that you have downloaded correct images and segmentation masks
 into open-images-v6 folder (as instructed in the previous chapter)
    
-
    dir_=os.path.join(fodir,"open-images-v6")
-print("contents of", dir_,":")
-!tree --filelimit=10 $dir_ | cat
+
    dir_=os.path.join(fodir,"open-images-v6")
+print("contents of", dir_,":")
+!tree --filelimit=10 $dir_ | cat
 
    
 
    
 
    contents of /home/sampsa/fiftyone/open-images-v6 :
@@ -418,36 +418,36 @@ 
     Contents 
    
 detection_validation_input_5k.lst          = list of images used
 
    
 
    
-
    # TODO: define path_to_mpeg_vcm_files
-path_to_images=os.path.join(fodir,"open-images-v6/validation/data")
+
    # TODO: define path_to_mpeg_vcm_files
+path_to_images=os.path.join(fodir,"open-images-v6/validation/data")
 
-list_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_input_5k.lst")
-bbox_csv_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_5k_bbox.csv")
-validation_csv_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_labels_5k.csv")
+list_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_input_5k.lst")
+bbox_csv_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_5k_bbox.csv")
+validation_csv_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_labels_5k.csv")
 
-assert(os.path.exists(bbox_csv_file)), "can't find bbox file"
-assert(os.path.exists(validation_csv_file)), "can't find labels file"
-assert(os.path.exists(path_to_images)), "can't find image directory"
+assert(os.path.exists(bbox_csv_file)), "can't find bbox file"
+assert(os.path.exists(validation_csv_file)), "can't find labels file"
+assert(os.path.exists(path_to_images)), "can't find image directory"
 
    
 
    
 
    Now we convert mpeg vmc proprietary format annotation into proper
 OpenImageV6 format dataset and place it into
 ~/fiftyone/mpeg_vcm-detection
    
 
    First, remove any previously imported stuff:
    
-
    !rm -rf ~/fiftyone/mpeg-vcm-*
+
    !rm -rf ~/fiftyone/mpeg-vcm-*
 
    
 
    
-
    MPEGVCMToOpenImageV6(
-    validation_csv_file=validation_csv_file,
-    list_file=list_file,
-    bbox_csv_file=bbox_csv_file,
-    output_directory=os.path.join(fodir,"mpeg-vcm-detection"),
-    data_dir=path_to_images
-)
+
    MPEGVCMToOpenImageV6(
+    validation_csv_file=validation_csv_file,
+    list_file=list_file,
+    bbox_csv_file=bbox_csv_file,
+    output_directory=os.path.join(fodir,"mpeg-vcm-detection"),
+    data_dir=path_to_images
+)
 
    
 
    
 
    let’s see what we got:
    
-
    !tree --filelimit=10 ~/fiftyone/mpeg-vcm-detection | cat
+
    !tree --filelimit=10 ~/fiftyone/mpeg-vcm-detection | cat
 
    
 
    
 
    /home/sampsa/fiftyone/mpeg-vcm-detection
@@ -469,46 +469,46 @@ 
     Contents 
    
 data -> ~/fiftyone/open-images-v6/validation/data)
    
 
    The only thing we’re left to do, is to register this OpenImageV6
 formatted dataset into fiftyone:
    
-
    # remove the dataset in the case it was already registered in fiftyone
-try:
-    fo.delete_dataset("mpeg-vcm-detection")
-except ValueError as e:
-    print("could not delete because of", e)
+
    # remove the dataset in the case it was already registered in fiftyone
+try:
+    fo.delete_dataset("mpeg-vcm-detection")
+except ValueError as e:
+    print("could not delete because of", e)
 
    
 
    
-
    dataset_type = fo.types.OpenImagesV6Dataset
-dataset_dir = os.path.join(fodir,"mpeg-vcm-detection")
-dataset = fo.Dataset.from_dir(
-    dataset_dir=dataset_dir,
-    dataset_type=dataset_type,
-    label_types=("detections","classifications"),
-    load_hierarchy=False,
-    name="mpeg-vcm-detection",
-    image_ids=imageIdFileList(list_file)
-)
+
    dataset_type = fo.types.OpenImagesV6Dataset
+dataset_dir = os.path.join(fodir,"mpeg-vcm-detection")
+dataset = fo.Dataset.from_dir(
+    dataset_dir=dataset_dir,
+    dataset_type=dataset_type,
+    label_types=("detections","classifications"),
+    load_hierarchy=False,
+    name="mpeg-vcm-detection",
+    image_ids=imageIdFileList(list_file)
+)
 
    
 
    
 
    100% |███████████████| 5000/5000 [16.8s elapsed, 0s remaining, 290.4 samples/s]
 
    
 
    
-
    dataset.persistent=True # without this, your dabatase will disappear!
+
    dataset.persistent=True # without this, your dabatase will disappear!
 
    
 
    
-
    ## now, in the future, just do
-dataset = fo.load_dataset("mpeg-vcm-detection")
+
    ## now, in the future, just do
+dataset = fo.load_dataset("mpeg-vcm-detection")
 
    
 
    
 
    Finaly, let’s also create a dummy dataset for debugging and testing with
 only one sample:
    
-
    try:
-    fo.delete_dataset("mpeg-vcm-detection-dummy")
-except ValueError:
-    print("no dummmy dataset yet..")
-dummy_dataset=fo.Dataset("mpeg-vcm-detection-dummy")
-for sample in dataset[0:1]:
-    dummy_dataset.add_sample(sample)
-dummy_dataset.persistent=True
-print("dummy dataset ok")
+
    try:
+    fo.delete_dataset("mpeg-vcm-detection-dummy")
+except ValueError:
+    print("no dummmy dataset yet..")
+dummy_dataset=fo.Dataset("mpeg-vcm-detection-dummy")
+for sample in dataset[0:1]:
+    dummy_dataset.add_sample(sample)
+dummy_dataset.persistent=True
+print("dummy dataset ok")
 
    
 
    
 
    dummy dataset ok
diff --git a/tutorials/detectron2.html b/tutorials/detectron2.html
index c06c6e23..ce77b478 100644
--- a/tutorials/detectron2.html
+++ b/tutorials/detectron2.html
@@ -346,90 +346,90 @@ 
    2. Run Detectron2
    # common libs
-import math, os, io, json, cv2, random, logging, datetime
-import numpy as np
-# torch
-import torch
-from torchvision import transforms
-# images
-from PIL import Image
-import matplotlib.pyplot as plt
+
    # common libs
+import math, os, io, json, cv2, random, logging, datetime
+import numpy as np
+# torch
+import torch
+from torchvision import transforms
+# images
+from PIL import Image
+import matplotlib.pyplot as plt
 
    
 
    
-
    # define a helper function
-def cv2_imshow(img):
-    img2 = img[:,:,::-1]
-    plt.figure(figsize=(12, 9))
-    plt.axis('off')
-    plt.imshow(img2)
-    plt.show()
+
    # define a helper function
+def cv2_imshow(img):
+    img2 = img[:,:,::-1]
+    plt.figure(figsize=(12, 9))
+    plt.axis('off')
+    plt.imshow(img2)
+    plt.show()
 
    
 
    
-
    ## *** Detectron imports ***
-import detectron2
-from detectron2.utils.logger import setup_logger
-setup_logger()
+
    ## *** Detectron imports ***
+import detectron2
+from detectron2.utils.logger import setup_logger
+setup_logger()
 
-# import some common detectron2 utilities
-from detectron2 import model_zoo
-from detectron2.engine import DefaultPredictor
-from detectron2.config import get_cfg
-from detectron2.utils.visualizer import Visualizer
-from detectron2.data import MetadataCatalog, DatasetCatalog
+# import some common detectron2 utilities
+from detectron2 import model_zoo
+from detectron2.engine import DefaultPredictor
+from detectron2.config import get_cfg
+from detectron2.utils.visualizer import Visualizer
+from detectron2.data import MetadataCatalog, DatasetCatalog
 
    
 
    
-
    # CompressAI-Vision
-from compressai_vision.conversion import FO2DetectronDataset # convert fiftyone dataset to Detectron2 dataset
-from compressai_vision.conversion import detectron251 # convert Detectron2 results to fiftyone format
-from compressai_vision.evaluation.fo import annexPredictions # crunch a complete fiftyone dataset through Detectron2 predictor and add the predictions to the fiftyone dataset
+
    # CompressAI-Vision
+from compressai_vision.conversion import FO2DetectronDataset # convert fiftyone dataset to Detectron2 dataset
+from compressai_vision.conversion import detectron251 # convert Detectron2 results to fiftyone format
+from compressai_vision.evaluation.fo import annexPredictions # crunch a complete fiftyone dataset through Detectron2 predictor and add the predictions to the fiftyone dataset
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
-
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
-print(device)
+
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+print(device)
 
    
 
    
 
    cpu
 
    
 
    
-
    print("torch:", torch.__version__, "/ cuda:", torch.version.cuda, "/ detectron2:", detectron2.__version__)
+
    print("torch:", torch.__version__, "/ cuda:", torch.version.cuda, "/ detectron2:", detectron2.__version__)
 
    
 
    
 
    torch: 1.9.1+cu102 / cuda: 10.2 / detectron2: 0.6
 
    
 
    
 
    Let’s pick up correct Detectron2 model
    
-
    ## MODEL A
-model_name="COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"
-## look here:
-## https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md#faster-r-cnn
+
    ## MODEL A
+model_name="COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"
+## look here:
+## https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md#faster-r-cnn
 
-## MODEL B
-# model_name="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
+## MODEL B
+# model_name="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
 
    
 
    
-
    # cfg encapsulates the model architecture & weights, also threshold parameter, metadata, etc.
-cfg = get_cfg()
-cfg.MODEL.DEVICE=device
-# load config from a file:
-cfg.merge_from_file(model_zoo.get_config_file(model_name))
-# DO NOT TOUCH THRESHOLD WHEN DOING EVALUATION:
-# too big a threshold will cut the smallest values & affect the precision(recall) curves & evaluation results
-# the default value is 0.05
-# value of 0.01 saturates the results (they don't change at lower values)
-# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
-# get weights
-cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(model_name)
-print("expected input colorspace:", cfg.INPUT.FORMAT)
-print("loaded datasets:", cfg.DATASETS)
-model_dataset=cfg.DATASETS.TRAIN[0]
-print("model was trained with", model_dataset)
-model_meta=MetadataCatalog.get(model_dataset)
+
    # cfg encapsulates the model architecture & weights, also threshold parameter, metadata, etc.
+cfg = get_cfg()
+cfg.MODEL.DEVICE=device
+# load config from a file:
+cfg.merge_from_file(model_zoo.get_config_file(model_name))
+# DO NOT TOUCH THRESHOLD WHEN DOING EVALUATION:
+# too big a threshold will cut the smallest values & affect the precision(recall) curves & evaluation results
+# the default value is 0.05
+# value of 0.01 saturates the results (they don't change at lower values)
+# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
+# get weights
+cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(model_name)
+print("expected input colorspace:", cfg.INPUT.FORMAT)
+print("loaded datasets:", cfg.DATASETS)
+model_dataset=cfg.DATASETS.TRAIN[0]
+print("model was trained with", model_dataset)
+model_meta=MetadataCatalog.get(model_dataset)
 
    
 
    
 
    expected input colorspace: BGR
@@ -442,15 +442,15 @@ 
    2. Run Detectron2
    predictor = DefaultPredictor(cfg)
+
    predictor = DefaultPredictor(cfg)
 
    
 
    
 
    Get handle to a dataset. We will be using the oiv6-mpeg-v1 dataset.
 Please go through the CLI Tutorials in order to produce this dataset.
    
-
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1")
+
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1")
 
    
 
    
-
    dataset
+
    dataset
 
    
 
    
 
    Name:        oiv6-mpeg-detection-v1
@@ -470,25 +470,25 @@ 
    2. Run Detectron2
    detectron_dataset=FO2DetectronDataset(fo_dataset=dataset, model_catids=model_meta.thing_classes)
+
    detectron_dataset=FO2DetectronDataset(fo_dataset=dataset, model_catids=model_meta.thing_classes)
 
    
 
    
 
    Pick a sample:
    
-
    d=detectron_dataset[3]
+
    d=detectron_dataset[3]
 
    
 
    
 
    We can visualize that sample also with Detectron2 library tools
 (although we’d prefer fiftyone with fo.launch_app(dataset)):
    
-
    # visualize with Detectron2 tools only
-img = cv2.imread(d["file_name"])
-visualizer = Visualizer(img[:, :, ::-1], metadata=model_meta, scale=0.5)
-out = visualizer.draw_dataset_dict(d)
-cv2_imshow(out.get_image()[:, :, ::-1])
+
    # visualize with Detectron2 tools only
+img = cv2.imread(d["file_name"])
+visualizer = Visualizer(img[:, :, ::-1], metadata=model_meta, scale=0.5)
+out = visualizer.draw_dataset_dict(d)
+cv2_imshow(out.get_image()[:, :, ::-1])
 
    
 
    
 ../_images/detectron2_nb_20_0.png
 
    Let’s try the Detectron2 predictor:
    
-
    res=predictor(img)
+
    res=predictor(img)
 
    
 
    
 
    /home/sampsa/silo/interdigital/venv_all/lib/python3.8/site-packages/torch/_tensor.py:575: UserWarning: floor_divide is deprecated, and will be removed in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values.
@@ -497,10 +497,10 @@ 
    2. Run Detectron2
    dets=detectron251(res, model_catids=model_meta.thing_classes) # process involves going from class indexes (ints) to class labels (strings)
+
    dets=detectron251(res, model_catids=model_meta.thing_classes) # process involves going from class indexes (ints) to class labels (strings)
 
    
 
    
-
    dets
+
    dets
 
    
 
    
 
    <Detections: {
@@ -754,16 +754,16 @@ 
    2. Run Detectron2oiv6-mpeg-detection-v1-dummy
 created in the CLI tutorials with the
 compressai-vision import-custom command
    
-
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1-dummy")
+
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1-dummy")
 
    
 
    
 
    Detectron prediction results are saved during the run into the fiftyone
 (mongodb) database. Let’s define a unique name for the sample field
 where the detectron results will be saved:
    
-
    predictor_field='detectron-predictions'
+
    predictor_field='detectron-predictions'
 
    
 
    
-
    annexPredictions(predictors=[predictor], fo_dataset=dataset, predictor_fields=[predictor_field])
+
    annexPredictions(predictors=[predictor], fo_dataset=dataset, predictor_fields=[predictor_field])
 
    
 
    
 
    sample:  1 / 1
@@ -771,7 +771,7 @@ 
    2. Run Detectron2detectron-predictions has appeared into the dataset:
    
-
    print(dataset)
+
    print(dataset)
 
    
 
    
 
    Name:        oiv6-mpeg-detection-v1-dummy
@@ -792,10 +792,10 @@ 
    2. Run Detectron2
    sample=dataset.first()
+
    sample=dataset.first()
 
    
 
    
-
    print(sample)
+
    print(sample)
 
    
 
    
 
    <Sample: {
@@ -1438,15 +1438,15 @@ 
    2. Run Detectron2
    results = dataset.evaluate_detections(
-    predictor_field,
-    gt_field="detections",
-    method="open-images",
-    pos_label_field="positive_labels",
-    neg_label_field="negative_labels",
-    expand_pred_hierarchy=False,
-    expand_gt_hierarchy=False
-)
+
    results = dataset.evaluate_detections(
+    predictor_field,
+    gt_field="detections",
+    method="open-images",
+    pos_label_field="positive_labels",
+    neg_label_field="negative_labels",
+    expand_pred_hierarchy=False,
+    expand_gt_hierarchy=False
+)
 
    
 
    
 
    Evaluating detections...
@@ -1455,22 +1455,22 @@ 
    2. Run Detectron2
    dataset.delete_sample_fields(predictor_field)
+
    dataset.delete_sample_fields(predictor_field)
 
    
 
    
 
    OpenImageV6 evaluation protocol mAP:
    
-
    results.mAP()
+
    results.mAP()
 
    
 
    
 
    1.0
 
    
 
    
 
    Per class mAP:
    
-
    classes = dataset.distinct(
-    "detections.detections.label"
-)
-for class_ in classes:
-    print(class_, results.mAP([class_]))
+
    classes = dataset.distinct(
+    "detections.detections.label"
+)
+for class_ in classes:
+    print(class_, results.mAP([class_]))
 
    
 
    
 
    airplane 1.0
diff --git a/tutorials/detectron2_nb.html b/tutorials/detectron2_nb.html
index d764a0e0..a47e7f20 100644
--- a/tutorials/detectron2_nb.html
+++ b/tutorials/detectron2_nb.html
@@ -349,90 +349,90 @@ 
     Contents 
    
   
    In this chapter we look into fiftyone/detectron2 interface, how to add
 detectron2 results into a fiftyone dataset and how to evaluate
 detectron2 results with fiftyone.
    
-
    # common libs
-import math, os, io, json, cv2, random, logging, datetime
-import numpy as np
-# torch
-import torch
-from torchvision import transforms
-# images
-from PIL import Image
-import matplotlib.pyplot as plt
+
    # common libs
+import math, os, io, json, cv2, random, logging, datetime
+import numpy as np
+# torch
+import torch
+from torchvision import transforms
+# images
+from PIL import Image
+import matplotlib.pyplot as plt
 
    
 
    
-
    # define a helper function
-def cv2_imshow(img):
-    img2 = img[:,:,::-1]
-    plt.figure(figsize=(12, 9))
-    plt.axis('off')
-    plt.imshow(img2)
-    plt.show()
+
    # define a helper function
+def cv2_imshow(img):
+    img2 = img[:,:,::-1]
+    plt.figure(figsize=(12, 9))
+    plt.axis('off')
+    plt.imshow(img2)
+    plt.show()
 
    
 
    
-
    ## *** Detectron imports ***
-import detectron2
-from detectron2.utils.logger import setup_logger
-setup_logger()
+
    ## *** Detectron imports ***
+import detectron2
+from detectron2.utils.logger import setup_logger
+setup_logger()
 
-# import some common detectron2 utilities
-from detectron2 import model_zoo
-from detectron2.engine import DefaultPredictor
-from detectron2.config import get_cfg
-from detectron2.utils.visualizer import Visualizer
-from detectron2.data import MetadataCatalog, DatasetCatalog
+# import some common detectron2 utilities
+from detectron2 import model_zoo
+from detectron2.engine import DefaultPredictor
+from detectron2.config import get_cfg
+from detectron2.utils.visualizer import Visualizer
+from detectron2.data import MetadataCatalog, DatasetCatalog
 
    
 
    
-
    # CompressAI-Vision
-from compressai_vision.conversion import FO2DetectronDataset # convert fiftyone dataset to Detectron2 dataset
-from compressai_vision.conversion import detectron251 # convert Detectron2 results to fiftyone format
-from compressai_vision.evaluation.fo import annexPredictions # crunch a complete fiftyone dataset through Detectron2 predictor and add the predictions to the fiftyone dataset
+
    # CompressAI-Vision
+from compressai_vision.conversion import FO2DetectronDataset # convert fiftyone dataset to Detectron2 dataset
+from compressai_vision.conversion import detectron251 # convert Detectron2 results to fiftyone format
+from compressai_vision.evaluation.fo import annexPredictions # crunch a complete fiftyone dataset through Detectron2 predictor and add the predictions to the fiftyone dataset
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
-
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
-print(device)
+
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+print(device)
 
    
 
    
 
    cpu
 
    
 
    
-
    print("torch:", torch.__version__, "/ cuda:", torch.version.cuda, "/ detectron2:", detectron2.__version__)
+
    print("torch:", torch.__version__, "/ cuda:", torch.version.cuda, "/ detectron2:", detectron2.__version__)
 
    
 
    
 
    torch: 1.9.1+cu102 / cuda: 10.2 / detectron2: 0.6
 
    
 
    
 
    Let’s pick up correct Detectron2 model
    
-
    ## MODEL A
-model_name="COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"
-## look here:
-## https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md#faster-r-cnn
+
    ## MODEL A
+model_name="COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"
+## look here:
+## https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md#faster-r-cnn
 
-## MODEL B
-# model_name="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
+## MODEL B
+# model_name="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
 
    
 
    
-
    # cfg encapsulates the model architecture & weights, also threshold parameter, metadata, etc.
-cfg = get_cfg()
-cfg.MODEL.DEVICE=device
-# load config from a file:
-cfg.merge_from_file(model_zoo.get_config_file(model_name))
-# DO NOT TOUCH THRESHOLD WHEN DOING EVALUATION:
-# too big a threshold will cut the smallest values & affect the precision(recall) curves & evaluation results
-# the default value is 0.05
-# value of 0.01 saturates the results (they don't change at lower values)
-# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
-# get weights
-cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(model_name)
-print("expected input colorspace:", cfg.INPUT.FORMAT)
-print("loaded datasets:", cfg.DATASETS)
-model_dataset=cfg.DATASETS.TRAIN[0]
-print("model was trained with", model_dataset)
-model_meta=MetadataCatalog.get(model_dataset)
+
    # cfg encapsulates the model architecture & weights, also threshold parameter, metadata, etc.
+cfg = get_cfg()
+cfg.MODEL.DEVICE=device
+# load config from a file:
+cfg.merge_from_file(model_zoo.get_config_file(model_name))
+# DO NOT TOUCH THRESHOLD WHEN DOING EVALUATION:
+# too big a threshold will cut the smallest values & affect the precision(recall) curves & evaluation results
+# the default value is 0.05
+# value of 0.01 saturates the results (they don't change at lower values)
+# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
+# get weights
+cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(model_name)
+print("expected input colorspace:", cfg.INPUT.FORMAT)
+print("loaded datasets:", cfg.DATASETS)
+model_dataset=cfg.DATASETS.TRAIN[0]
+print("model was trained with", model_dataset)
+model_meta=MetadataCatalog.get(model_dataset)
 
    
 
    
 
    expected input colorspace: BGR
@@ -445,15 +445,15 @@ 
     Contents 
    
 model was trained with coco_2017_train
 
    
 
    
-
    predictor = DefaultPredictor(cfg)
+
    predictor = DefaultPredictor(cfg)
 
    
 
    
 
    Get handle to a dataset. We will be using the oiv6-mpeg-v1 dataset.
 Please go through the CLI Tutorials in order to produce this dataset.
    
-
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1")
+
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1")
 
    
 
    
-
    dataset
+
    dataset
 
    
 
    
 
    Name:        oiv6-mpeg-detection-v1
@@ -473,25 +473,25 @@ 
     Contents 
    
 
    
 
    
 
    We can go from fiftyone dataset to Detectron2 dataset:
    
-
    detectron_dataset=FO2DetectronDataset(fo_dataset=dataset, model_catids=model_meta.thing_classes)
+
    detectron_dataset=FO2DetectronDataset(fo_dataset=dataset, model_catids=model_meta.thing_classes)
 
    
 
    
 
    Pick a sample:
    
-
    d=detectron_dataset[3]
+
    d=detectron_dataset[3]
 
    
 
    
 
    We can visualize that sample also with Detectron2 library tools
 (although we’d prefer fiftyone with fo.launch_app(dataset)):
    
-
    # visualize with Detectron2 tools only
-img = cv2.imread(d["file_name"])
-visualizer = Visualizer(img[:, :, ::-1], metadata=model_meta, scale=0.5)
-out = visualizer.draw_dataset_dict(d)
-cv2_imshow(out.get_image()[:, :, ::-1])
+
    # visualize with Detectron2 tools only
+img = cv2.imread(d["file_name"])
+visualizer = Visualizer(img[:, :, ::-1], metadata=model_meta, scale=0.5)
+out = visualizer.draw_dataset_dict(d)
+cv2_imshow(out.get_image()[:, :, ::-1])
 
    
 
    
 ../_images/detectron2_nb_20_0.png
 
    Let’s try the Detectron2 predictor:
    
-
    res=predictor(img)
+
    res=predictor(img)
 
    
 
    
 
    /home/sampsa/silo/interdigital/venv_all/lib/python3.8/site-packages/torch/_tensor.py:575: UserWarning: floor_divide is deprecated, and will be removed in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values.
@@ -500,10 +500,10 @@ 
     Contents 
    
 
    
 
    
 
    We can convert from Detectron2 format to fiftyone detection objects:
    
-
    dets=detectron251(res, model_catids=model_meta.thing_classes) # process involves going from class indexes (ints) to class labels (strings)
+
    dets=detectron251(res, model_catids=model_meta.thing_classes) # process involves going from class indexes (ints) to class labels (strings)
 
    
 
    
-
    dets
+
    dets
 
    
 
    
 
    <Detections: {
@@ -757,16 +757,16 @@ 
     Contents 
    
 We use the dummy single-sample dataset oiv6-mpeg-detection-v1-dummy
 created in the CLI tutorials with the
 compressai-vision import-custom command
    
-
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1-dummy")
+
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1-dummy")
 
    
 
    
 
    Detectron prediction results are saved during the run into the fiftyone
 (mongodb) database. Let’s define a unique name for the sample field
 where the detectron results will be saved:
    
-
    predictor_field='detectron-predictions'
+
    predictor_field='detectron-predictions'
 
    
 
    
-
    annexPredictions(predictors=[predictor], fo_dataset=dataset, predictor_fields=[predictor_field])
+
    annexPredictions(predictors=[predictor], fo_dataset=dataset, predictor_fields=[predictor_field])
 
    
 
    
 
    sample:  1 / 1
@@ -774,7 +774,7 @@ 
     Contents 
    
 
    
 
    After that one, the dataset looks slightly different. Take note that an
 extra field detectron-predictions has appeared into the dataset:
    
-
    print(dataset)
+
    print(dataset)
 
    
 
    
 
    Name:        oiv6-mpeg-detection-v1-dummy
@@ -795,10 +795,10 @@ 
     Contents 
    
 
    
 
    
 
    Let’s peek at the first sample:
    
-
    sample=dataset.first()
+
    sample=dataset.first()
 
    
 
    
-
    print(sample)
+
    print(sample)
 
    
 
    
 
    <Sample: {
@@ -1441,15 +1441,15 @@ 
     Contents 
    
 “detectron-predictions” (predicted values). Now we can run the
 OpenImageV6 evaluation protocol on the dataset which uses the ground
 truth and the predictor results:
    
-
    results = dataset.evaluate_detections(
-    predictor_field,
-    gt_field="detections",
-    method="open-images",
-    pos_label_field="positive_labels",
-    neg_label_field="negative_labels",
-    expand_pred_hierarchy=False,
-    expand_gt_hierarchy=False
-)
+
    results = dataset.evaluate_detections(
+    predictor_field,
+    gt_field="detections",
+    method="open-images",
+    pos_label_field="positive_labels",
+    neg_label_field="negative_labels",
+    expand_pred_hierarchy=False,
+    expand_gt_hierarchy=False
+)
 
    
 
    
 
    Evaluating detections...
@@ -1458,22 +1458,22 @@ 
     Contents 
    
 
    
 
    After the evaluation we can should remove the detectron results from the
 database:
    
-
    dataset.delete_sample_fields(predictor_field)
+
    dataset.delete_sample_fields(predictor_field)
 
    
 
    
 
    OpenImageV6 evaluation protocol mAP:
    
-
    results.mAP()
+
    results.mAP()
 
    
 
    
 
    1.0
 
    
 
    
 
    Per class mAP:
    
-
    classes = dataset.distinct(
-    "detections.detections.label"
-)
-for class_ in classes:
-    print(class_, results.mAP([class_]))
+
    classes = dataset.distinct(
+    "detections.detections.label"
+)
+for class_ in classes:
+    print(class_, results.mAP([class_]))
 
    
 
    
 
    airplane 1.0
diff --git a/tutorials/download.html b/tutorials/download.html
index 69cff618..a52d642b 100644
--- a/tutorials/download.html
+++ b/tutorials/download.html
@@ -344,31 +344,31 @@ 
    1. Download Images
    # common libs
-import math, os, io, json, cv2, random, logging
-import numpy as np
-# images
-from PIL import Image
-import matplotlib.pyplot as plt
+
    # common libs
+import math, os, io, json, cv2, random, logging
+import numpy as np
+# images
+from PIL import Image
+import matplotlib.pyplot as plt
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
-
    # CompressAI-Vision
-from compressai_vision.conversion import imageIdFileList
+
    # CompressAI-Vision
+from compressai_vision.conversion import imageIdFileList
 
    
 
    
-
    homie=os.path.expanduser("~")
-print("your home path is", homie)
-fodir=os.path.join(homie,'fiftyone')
-print("fiftyone dowloads data by default to", fodir)
-try:
-    os.mkdir(fodir)
-except FileExistsError:
-    pass
+
    homie=os.path.expanduser("~")
+print("your home path is", homie)
+fodir=os.path.join(homie,'fiftyone')
+print("fiftyone dowloads data by default to", fodir)
+try:
+    os.mkdir(fodir)
+except FileExistsError:
+    pass
 
    
 
    
 
    your home path is /home/sampsa
@@ -376,7 +376,7 @@ 
    1. Download Images
    fo.list_datasets()
+
    fo.list_datasets()
 
    
 
    
 
    ['detectron-run-sampsa-oiv6-mpeg-detection-v1-2022-11-16-17-22-40-319395',
@@ -398,10 +398,10 @@ 
    1. Download Imagesdetection_validation_input_5k.lst and
 segmentation_validation_input_5k.lst
    
-
    path_to_list_file="/home/sampsa/silo/interdigital/CompressAI-Vision/compressai_vision/data/mpeg_vcm_data"
+
    path_to_list_file="/home/sampsa/silo/interdigital/CompressAI-Vision/compressai_vision/data/mpeg_vcm_data"
 
    
 
    
-
    !head -n10 {path_to_list_file}/detection_validation_input_5k.lst
+
    !head -n10 {path_to_list_file}/detection_validation_input_5k.lst
 
    
 
    
 
    bef50424c62d12c5.jpg
@@ -416,25 +416,25 @@ 
    1. Download Images
    det_lst=os.path.join(path_to_mpeg_vcm_files,"detection_validation_input_5k.lst")
-seg_lst=os.path.join(path_to_mpeg_vcm_files, "segmentation_validation_input_5k.lst")
-assert(os.path.exists(det_lst)), "missing file "+det_lst
-assert(os.path.exists(seg_lst)), "missing file "+seg_lst
-lis=imageIdFileList(det_lst, seg_lst)
-print(len(lis))
+
    det_lst=os.path.join(path_to_mpeg_vcm_files,"detection_validation_input_5k.lst")
+seg_lst=os.path.join(path_to_mpeg_vcm_files, "segmentation_validation_input_5k.lst")
+assert(os.path.exists(det_lst)), "missing file "+det_lst
+assert(os.path.exists(seg_lst)), "missing file "+seg_lst
+lis=imageIdFileList(det_lst, seg_lst)
+print(len(lis))
 
    
 
    
 
    8189
 
    
 
    
 
    Tell fiftyone to load the correct subset of OpenImageV6 dataset:
    
-
    # https://voxel51.com/docs/fiftyone/user_guide/dataset_zoo/datasets.html#dataset-zoo-open-images-v6
-dataset = foz.load_zoo_dataset(
-    "open-images-v6",
-    split="validation",
-    # label_types=("detections", "classifications", "relationships", "segmentations") # this is the default
-    image_ids=lis
-)
+
    # https://voxel51.com/docs/fiftyone/user_guide/dataset_zoo/datasets.html#dataset-zoo-open-images-v6
+dataset = foz.load_zoo_dataset(
+    "open-images-v6",
+    split="validation",
+    # label_types=("detections", "classifications", "relationships", "segmentations") # this is the default
+    image_ids=lis
+)
 
    
 
    
 
    Downloading split 'validation' to '/home/sampsa/fiftyone/open-images-v6/validation' if necessary
@@ -443,8 +443,8 @@ 
    1. Download Images
    # take a look at the dataset
-dataset
+
    # take a look at the dataset
+dataset
 
    
 
    
 
    Name:        open-images-v6-validation
@@ -465,16 +465,16 @@ 
    1. Download Imagessegmentations:   fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Detections)
 
    
 
    
-
    # make dataset persistent .. next time you import fiftyone it's still available (loaded into the mongodb that's running in the background)
-dataset.persistent=True
+
    # make dataset persistent .. next time you import fiftyone it's still available (loaded into the mongodb that's running in the background)
+dataset.persistent=True
 
    
 
    
-
    # next time you need it, load it with:
-dataset = fo.load_dataset("open-images-v6-validation")
+
    # next time you need it, load it with:
+dataset = fo.load_dataset("open-images-v6-validation")
 
    
 
    
-
    # peek at first sample
-dataset.first()
+
    # peek at first sample
+dataset.first()
 
    
 
    
 
    <Sample: {
@@ -527,9 +527,9 @@ 
    1. Download Images
    dir_=os.path.join(fodir,"open-images-v6")
-print("contents of", dir_,":")
-!tree --filelimit=10 $dir_ | cat
+
    dir_=os.path.join(fodir,"open-images-v6")
+print("contents of", dir_,":")
+!tree --filelimit=10 $dir_ | cat
 
    
 
    
 
    contents of /home/sampsa/fiftyone/open-images-v6 :
diff --git a/tutorials/download_nb.html b/tutorials/download_nb.html
index a1b5eb1b..90334a73 100644
--- a/tutorials/download_nb.html
+++ b/tutorials/download_nb.html
@@ -348,31 +348,31 @@ 
     Contents 
    
                   
   
    In this chapter we use fiftyone to download, inspect and visualize a
 subset of OpenImageV6 images
    
-
    # common libs
-import math, os, io, json, cv2, random, logging
-import numpy as np
-# images
-from PIL import Image
-import matplotlib.pyplot as plt
+
    # common libs
+import math, os, io, json, cv2, random, logging
+import numpy as np
+# images
+from PIL import Image
+import matplotlib.pyplot as plt
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
-
    # CompressAI-Vision
-from compressai_vision.conversion import imageIdFileList
+
    # CompressAI-Vision
+from compressai_vision.conversion import imageIdFileList
 
    
 
    
-
    homie=os.path.expanduser("~")
-print("your home path is", homie)
-fodir=os.path.join(homie,'fiftyone')
-print("fiftyone dowloads data by default to", fodir)
-try:
-    os.mkdir(fodir)
-except FileExistsError:
-    pass
+
    homie=os.path.expanduser("~")
+print("your home path is", homie)
+fodir=os.path.join(homie,'fiftyone')
+print("fiftyone dowloads data by default to", fodir)
+try:
+    os.mkdir(fodir)
+except FileExistsError:
+    pass
 
    
 
    
 
    your home path is /home/sampsa
@@ -380,7 +380,7 @@ 
     Contents 
    
 
    
 
    
 
    List all datasets (already) registered to fiftyone
    
-
    fo.list_datasets()
+
    fo.list_datasets()
 
    
 
    
 
    ['detectron-run-sampsa-oiv6-mpeg-detection-v1-2022-11-16-17-22-40-319395',
@@ -402,10 +402,10 @@ 
     Contents 
    
 OpenImageV6.
    
 
    Let’s use two files: detection_validation_input_5k.lst and
 segmentation_validation_input_5k.lst
    
-
    path_to_list_file="/home/sampsa/silo/interdigital/CompressAI-Vision/compressai_vision/data/mpeg_vcm_data"
+
    path_to_list_file="/home/sampsa/silo/interdigital/CompressAI-Vision/compressai_vision/data/mpeg_vcm_data"
 
    
 
    
-
    !head -n10 {path_to_list_file}/detection_validation_input_5k.lst
+
    !head -n10 {path_to_list_file}/detection_validation_input_5k.lst
 
    
 
    
 
    bef50424c62d12c5.jpg
@@ -420,25 +420,25 @@ 
     Contents 
    
 2e96665b867c4d0f.jpg
 
    
 
    
-
    det_lst=os.path.join(path_to_mpeg_vcm_files,"detection_validation_input_5k.lst")
-seg_lst=os.path.join(path_to_mpeg_vcm_files, "segmentation_validation_input_5k.lst")
-assert(os.path.exists(det_lst)), "missing file "+det_lst
-assert(os.path.exists(seg_lst)), "missing file "+seg_lst
-lis=imageIdFileList(det_lst, seg_lst)
-print(len(lis))
+
    det_lst=os.path.join(path_to_mpeg_vcm_files,"detection_validation_input_5k.lst")
+seg_lst=os.path.join(path_to_mpeg_vcm_files, "segmentation_validation_input_5k.lst")
+assert(os.path.exists(det_lst)), "missing file "+det_lst
+assert(os.path.exists(seg_lst)), "missing file "+seg_lst
+lis=imageIdFileList(det_lst, seg_lst)
+print(len(lis))
 
    
 
    
 
    8189
 
    
 
    
 
    Tell fiftyone to load the correct subset of OpenImageV6 dataset:
    
-
    # https://voxel51.com/docs/fiftyone/user_guide/dataset_zoo/datasets.html#dataset-zoo-open-images-v6
-dataset = foz.load_zoo_dataset(
-    "open-images-v6",
-    split="validation",
-    # label_types=("detections", "classifications", "relationships", "segmentations") # this is the default
-    image_ids=lis
-)
+
    # https://voxel51.com/docs/fiftyone/user_guide/dataset_zoo/datasets.html#dataset-zoo-open-images-v6
+dataset = foz.load_zoo_dataset(
+    "open-images-v6",
+    split="validation",
+    # label_types=("detections", "classifications", "relationships", "segmentations") # this is the default
+    image_ids=lis
+)
 
    
 
    
 
    Downloading split 'validation' to '/home/sampsa/fiftyone/open-images-v6/validation' if necessary
@@ -447,8 +447,8 @@ 
     Contents 
    
 Loading existing dataset 'open-images-v6-validation'. To reload from disk, either delete the existing dataset or provide a custom `dataset_name` to use
 
    
 
    
-
    # take a look at the dataset
-dataset
+
    # take a look at the dataset
+dataset
 
    
 
    
 
    Name:        open-images-v6-validation
@@ -469,16 +469,16 @@ 
     Contents 
    
     segmentations:   fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Detections)
 
    
 
    
-
    # make dataset persistent .. next time you import fiftyone it's still available (loaded into the mongodb that's running in the background)
-dataset.persistent=True
+
    # make dataset persistent .. next time you import fiftyone it's still available (loaded into the mongodb that's running in the background)
+dataset.persistent=True
 
    
 
    
-
    # next time you need it, load it with:
-dataset = fo.load_dataset("open-images-v6-validation")
+
    # next time you need it, load it with:
+dataset = fo.load_dataset("open-images-v6-validation")
 
    
 
    
-
    # peek at first sample
-dataset.first()
+
    # peek at first sample
+dataset.first()
 
    
 
    
 
    <Sample: {
@@ -531,9 +531,9 @@ 
     Contents 
    
 
    
 
    
 
    Let’s take a look where fiftyone downloaded the files
    
-
    dir_=os.path.join(fodir,"open-images-v6")
-print("contents of", dir_,":")
-!tree --filelimit=10 $dir_ | cat
+
    dir_=os.path.join(fodir,"open-images-v6")
+print("contents of", dir_,":")
+!tree --filelimit=10 $dir_ | cat
 
    
 
    
 
    contents of /home/sampsa/fiftyone/open-images-v6 :
diff --git a/tutorials/encdec.html b/tutorials/encdec.html
index a2926709..93133501 100644
--- a/tutorials/encdec.html
+++ b/tutorials/encdec.html
@@ -348,46 +348,46 @@ 
    4. Creating an EncoderDecoder class
    
 write your own EncoderDecoder class and this is quite simple.
    
 
    Here we demo a simple EncoderDecoder class that encodes & decodes
 using jpeg.
    
-
    import logging, io, cv2
-import numpy as np
-import matplotlib.pyplot as plt
-from PIL import Image
-from compressai_vision.evaluation.pipeline import EncoderDecoder
+
    import logging, io, cv2
+import numpy as np
+import matplotlib.pyplot as plt
+from PIL import Image
+from compressai_vision.evaluation.pipeline import EncoderDecoder
 
    
 
    
 
    In the constructor, instantiate a logger and save the provided quality
 parameter.
    
-
    class JpegEncoderDecoder(EncoderDecoder):
+
    class JpegEncoderDecoder(EncoderDecoder):
 
-    def __init__(self, qp=10):
-        self.logger = logging.getLogger(self.__class__.__name__)
-        self.qp=qp
-        self.reset() # not used in this class
+    def __init__(self, qp=10):
+        self.logger = logging.getLogger(self.__class__.__name__)
+        self.qp=qp
+        self.reset() # not used in this class
 
    
 
    
 
    Define how the image is encoded + decoded and how the bitrate is
 calculated. We are using BGR since it is the default input format for
 Detectron2 predictors.
    
-
    def BGR(self, bgr_image, tag=None):
-    # bgr_image: numpy BGR24 image: (y,x,3)
-    # tag could be used to identify images if we want to cache them
-    # BGR -> RGB (as PIL works with RGB)
-    rgb_image = bgr_image[:,:,::-1]
-    pil_img=Image.fromarray(rgb_image).convert("RGB")
-    tmp = io.BytesIO()
-    # encode image
-    pil_img.save(tmp, format="jpeg", quality=self.qp)
-    tmp.seek(0)
-    # calculate bits-per-pixel
-    filesize = tmp.getbuffer().nbytes
-    bpp = filesize * float(8) / (pil_img.size[0] * pil_img.size[1])
-    # decode image back
-    pil_img2 = Image.open(tmp).convert("RGB")
-    # back to BGR
-    rgb_image=np.array(pil_img2)
-    bgr_image=rgb_image[:,:,::-1]
-    # transformed image, bits-per-pixel ready
-    return bgr_image, bpp
+
    def BGR(self, bgr_image, tag=None):
+    # bgr_image: numpy BGR24 image: (y,x,3)
+    # tag could be used to identify images if we want to cache them
+    # BGR -> RGB (as PIL works with RGB)
+    rgb_image = bgr_image[:,:,::-1]
+    pil_img=Image.fromarray(rgb_image).convert("RGB")
+    tmp = io.BytesIO()
+    # encode image
+    pil_img.save(tmp, format="jpeg", quality=self.qp)
+    tmp.seek(0)
+    # calculate bits-per-pixel
+    filesize = tmp.getbuffer().nbytes
+    bpp = filesize * float(8) / (pil_img.size[0] * pil_img.size[1])
+    # decode image back
+    pil_img2 = Image.open(tmp).convert("RGB")
+    # back to BGR
+    rgb_image=np.array(pil_img2)
+    bgr_image=rgb_image[:,:,::-1]
+    # transformed image, bits-per-pixel ready
+    return bgr_image, bpp
 
    
 
    
 
    So, we have a compact class that defines, in a single method, all
@@ -396,20 +396,20 @@ 
    4. Creating an EncoderDecoder class
    
 with all the rest of the infrastructure provided by
 CompressAI-Vision library.
    
 
    Next, let’s see JpegEncoderDecoder in action.
    
-
    bgr_image=cv2.imread("dog_512.png")
+
    bgr_image=cv2.imread("dog_512.png")
 
    
 
    
-
    encdec=JpegEncoderDecoder(qp=1)
+
    encdec=JpegEncoderDecoder(qp=1)
 
    
 
    
-
    transformed_bgr_image, bpp = encdec.BGR(bgr_image)
+
    transformed_bgr_image, bpp = encdec.BGR(bgr_image)
 
    
 
    
 
    Print bits-per-pixel, compare original and transformed image
    
-
    print("BPP=", bpp)
-plt.figure(figsize=(20,20))
-plt.subplot(1,2,1); plt.imshow(bgr_image[:,:,::-1]); _=plt.axis('off')
-plt.subplot(1,2,2); plt.imshow(transformed_bgr_image[:,:,::-1]); _=plt.axis('off')
+
    print("BPP=", bpp)
+plt.figure(figsize=(20,20))
+plt.subplot(1,2,1); plt.imshow(bgr_image[:,:,::-1]); _=plt.axis('off')
+plt.subplot(1,2,2); plt.imshow(transformed_bgr_image[:,:,::-1]); _=plt.axis('off')
 
    
 
    
 
    BPP= 0.16878255208333334
diff --git a/tutorials/encdec_nb.html b/tutorials/encdec_nb.html
index ec2395bb..e86f9c47 100644
--- a/tutorials/encdec_nb.html
+++ b/tutorials/encdec_nb.html
@@ -354,46 +354,46 @@ 
     Contents 
    
 write your own EncoderDecoder class and this is quite simple.
    
 
    Here we demo a simple EncoderDecoder class that encodes & decodes
 using jpeg.
    
-
    import logging, io, cv2
-import numpy as np
-import matplotlib.pyplot as plt
-from PIL import Image
-from compressai_vision.evaluation.pipeline import EncoderDecoder
+
    import logging, io, cv2
+import numpy as np
+import matplotlib.pyplot as plt
+from PIL import Image
+from compressai_vision.evaluation.pipeline import EncoderDecoder
 
    
 
    
 
    In the constructor, instantiate a logger and save the provided quality
 parameter.
    
-
    class JpegEncoderDecoder(EncoderDecoder):
+
    class JpegEncoderDecoder(EncoderDecoder):
 
-    def __init__(self, qp=10):
-        self.logger = logging.getLogger(self.__class__.__name__)
-        self.qp=qp
-        self.reset() # not used in this class
+    def __init__(self, qp=10):
+        self.logger = logging.getLogger(self.__class__.__name__)
+        self.qp=qp
+        self.reset() # not used in this class
 
    
 
    
 
    Define how the image is encoded + decoded and how the bitrate is
 calculated. We are using BGR since it is the default input format for
 Detectron2 predictors.
    
-
    def BGR(self, bgr_image, tag=None):
-    # bgr_image: numpy BGR24 image: (y,x,3)
-    # tag could be used to identify images if we want to cache them
-    # BGR -> RGB (as PIL works with RGB)
-    rgb_image = bgr_image[:,:,::-1]
-    pil_img=Image.fromarray(rgb_image).convert("RGB")
-    tmp = io.BytesIO()
-    # encode image
-    pil_img.save(tmp, format="jpeg", quality=self.qp)
-    tmp.seek(0)
-    # calculate bits-per-pixel
-    filesize = tmp.getbuffer().nbytes
-    bpp = filesize * float(8) / (pil_img.size[0] * pil_img.size[1])
-    # decode image back
-    pil_img2 = Image.open(tmp).convert("RGB")
-    # back to BGR
-    rgb_image=np.array(pil_img2)
-    bgr_image=rgb_image[:,:,::-1]
-    # transformed image, bits-per-pixel ready
-    return bgr_image, bpp
+
    def BGR(self, bgr_image, tag=None):
+    # bgr_image: numpy BGR24 image: (y,x,3)
+    # tag could be used to identify images if we want to cache them
+    # BGR -> RGB (as PIL works with RGB)
+    rgb_image = bgr_image[:,:,::-1]
+    pil_img=Image.fromarray(rgb_image).convert("RGB")
+    tmp = io.BytesIO()
+    # encode image
+    pil_img.save(tmp, format="jpeg", quality=self.qp)
+    tmp.seek(0)
+    # calculate bits-per-pixel
+    filesize = tmp.getbuffer().nbytes
+    bpp = filesize * float(8) / (pil_img.size[0] * pil_img.size[1])
+    # decode image back
+    pil_img2 = Image.open(tmp).convert("RGB")
+    # back to BGR
+    rgb_image=np.array(pil_img2)
+    bgr_image=rgb_image[:,:,::-1]
+    # transformed image, bits-per-pixel ready
+    return bgr_image, bpp
 
    
 
    
 
    So, we have a compact class that defines, in a single method, all
@@ -402,20 +402,20 @@ 
     Contents 
    
 with all the rest of the infrastructure provided by
 CompressAI-Vision library.
    
 
    Next, let’s see JpegEncoderDecoder in action.
    
-
    bgr_image=cv2.imread("dog_512.png")
+
    bgr_image=cv2.imread("dog_512.png")
 
    
 
    
-
    encdec=JpegEncoderDecoder(qp=1)
+
    encdec=JpegEncoderDecoder(qp=1)
 
    
 
    
-
    transformed_bgr_image, bpp = encdec.BGR(bgr_image)
+
    transformed_bgr_image, bpp = encdec.BGR(bgr_image)
 
    
 
    
 
    Print bits-per-pixel, compare original and transformed image
    
-
    print("BPP=", bpp)
-plt.figure(figsize=(20,20))
-plt.subplot(1,2,1); plt.imshow(bgr_image[:,:,::-1]); _=plt.axis('off')
-plt.subplot(1,2,2); plt.imshow(transformed_bgr_image[:,:,::-1]); _=plt.axis('off')
+
    print("BPP=", bpp)
+plt.figure(figsize=(20,20))
+plt.subplot(1,2,1); plt.imshow(bgr_image[:,:,::-1]); _=plt.axis('off')
+plt.subplot(1,2,2); plt.imshow(transformed_bgr_image[:,:,::-1]); _=plt.axis('off')
 
    
 
    
 
    BPP= 0.16878255208333334
diff --git a/tutorials/evaluate.html b/tutorials/evaluate.html
index 25645641..af2d40a4 100644
--- a/tutorials/evaluate.html
+++ b/tutorials/evaluate.html
@@ -346,78 +346,78 @@ 
    3. Evaluate
    
 
    In this tutorial we evaluate mAP values for a dataset with Detectron2
 and a deep-learning encoding model from the CompressAI library. We also
 show how to perform a baseline evaluation with VTM.
    
-
    # common libs
-import math, os, io, json, cv2, random, logging, pickle, datetime
-import numpy as np
-# torch
-import torch
-# images
-from PIL import Image
-import matplotlib.pyplot as plt
-# compressai
-from compressai.zoo import bmshj2018_factorized
+
    # common libs
+import math, os, io, json, cv2, random, logging, pickle, datetime
+import numpy as np
+# torch
+import torch
+# images
+from PIL import Image
+import matplotlib.pyplot as plt
+# compressai
+from compressai.zoo import bmshj2018_factorized
 
    
 
    
-
    ## *** Detectron imports ***
-import detectron2
-from detectron2.utils.logger import setup_logger
-setup_logger()
+
    ## *** Detectron imports ***
+import detectron2
+from detectron2.utils.logger import setup_logger
+setup_logger()
 
-# import some common detectron2 utilities
-from detectron2 import model_zoo
-from detectron2.engine import DefaultPredictor
-from detectron2.config import get_cfg
-from detectron2.utils.visualizer import Visualizer
-from detectron2.data import MetadataCatalog, DatasetCatalog
+# import some common detectron2 utilities
+from detectron2 import model_zoo
+from detectron2.engine import DefaultPredictor
+from detectron2.config import get_cfg
+from detectron2.utils.visualizer import Visualizer
+from detectron2.data import MetadataCatalog, DatasetCatalog
 
    
 
    
-
    # CompressAI-Vision
-from compressai_vision.conversion import FO2DetectronDataset # convert fiftyone dataset to Detectron2 dataset
-from compressai_vision.conversion import detectron251 # convert Detectron2 results to fiftyone format
-from compressai_vision.evaluation.fo import annexPredictions # annex predictions from
-from compressai_vision.evaluation.pipeline import CompressAIEncoderDecoder, VTMEncoderDecoder # a class that does encoding+decoding & returns the transformed image & bitrate
-from compressai_vision.pipelines.remote_analysis.tools import confLogger, quickLog, getDataFile
+
    # CompressAI-Vision
+from compressai_vision.conversion import FO2DetectronDataset # convert fiftyone dataset to Detectron2 dataset
+from compressai_vision.conversion import detectron251 # convert Detectron2 results to fiftyone format
+from compressai_vision.evaluation.fo import annexPredictions # annex predictions from
+from compressai_vision.evaluation.pipeline import CompressAIEncoderDecoder, VTMEncoderDecoder # a class that does encoding+decoding & returns the transformed image & bitrate
+from compressai_vision.pipelines.remote_analysis.tools import confLogger, quickLog, getDataFile
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
-
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
-print(device)
+
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+print(device)
 
    
 
    
 
    cpu
 
    
 
    
-
    ## MODEL A
-model_name="COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"
-## look here:
-## https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md#faster-r-cnn
-## for the line that says X101-FPN --> box AP is 43
+
    ## MODEL A
+model_name="COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"
+## look here:
+## https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md#faster-r-cnn
+## for the line that says X101-FPN --> box AP is 43
 
-## MODEL B
-# model_name="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
+## MODEL B
+# model_name="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
 
    
 
    
-
    # cfg encapsulates the model architecture & weights, also threshold parameter, metadata, etc.
-cfg = get_cfg()
-cfg.MODEL.DEVICE=device
-# load config from a file:
-cfg.merge_from_file(model_zoo.get_config_file(model_name))
-# DO NOT TOUCH THRESHOLD WHEN DOING EVALUATION:
-# too big a threshold will cut the smallest values & affect the precision(recall) curves & evaluation results
-# the default value is 0.05
-# value of 0.01 saturates the results (they don't change at lower values)
-# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
-# get weights
-cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(model_name)
-print("expected input colorspace:", cfg.INPUT.FORMAT)
-print("loaded datasets:", cfg.DATASETS)
-model_dataset=cfg.DATASETS.TRAIN[0]
-print("model was trained with", model_dataset)
-model_meta=MetadataCatalog.get(model_dataset)
+
    # cfg encapsulates the model architecture & weights, also threshold parameter, metadata, etc.
+cfg = get_cfg()
+cfg.MODEL.DEVICE=device
+# load config from a file:
+cfg.merge_from_file(model_zoo.get_config_file(model_name))
+# DO NOT TOUCH THRESHOLD WHEN DOING EVALUATION:
+# too big a threshold will cut the smallest values & affect the precision(recall) curves & evaluation results
+# the default value is 0.05
+# value of 0.01 saturates the results (they don't change at lower values)
+# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
+# get weights
+cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(model_name)
+print("expected input colorspace:", cfg.INPUT.FORMAT)
+print("loaded datasets:", cfg.DATASETS)
+model_dataset=cfg.DATASETS.TRAIN[0]
+print("model was trained with", model_dataset)
+model_meta=MetadataCatalog.get(model_dataset)
 
    
 
    
 
    expected input colorspace: BGR
@@ -430,18 +430,18 @@ 
    3. Evaluate
    
 model was trained with coco_2017_train
 
    
 
    
-
    # model_meta.thing_classes # check class labels this was trained with
+
    # model_meta.thing_classes # check class labels this was trained with
 
    
 
    
-
    predictor = DefaultPredictor(cfg)
+
    predictor = DefaultPredictor(cfg)
 
    
 
    
 
    Get handle to a dataset. We will be using the oiv6-mpeg-v1 dataset.
 Please go through the CLI Tutorials in order to produce this dataset.
    
-
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1-dummy") # or use the dummy dataset for testing/debugging
+
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1-dummy") # or use the dummy dataset for testing/debugging
 
    
 
    
-
    dataset
+
    dataset
 
    
 
    
 
    Name:        oiv6-mpeg-detection-v1-dummy
@@ -461,32 +461,32 @@ 
    3. Evaluate
    
 
    
 
    
 
    Set some loglevels
    
-
    # loglev=logging.DEBUG
-loglev=logging.INFO
-quickLog("CompressAIEncoderDecoder", loglev)
+
    # loglev=logging.DEBUG
+loglev=logging.INFO
+quickLog("CompressAIEncoderDecoder", loglev)
 
    
 
    
 
    <Logger CompressAIEncoderDecoder (INFO)>
 
    
 
    
 
    Get a list of labels in the dataset:
    
-
    classes = dataset.distinct(
-    "detections.detections.label"
-)
-print(classes)
+
    classes = dataset.distinct(
+    "detections.detections.label"
+)
+print(classes)
 
    
 
    
 
    ['airplane']
 
    
 
    
-
    def per_class(results_obj):
-    """helper function: take fiftyone/openimagev6 results object & spit
-    out mAP breakdown as per class
-    """
-    d = {}
-    for class_ in classes:
-        d[class_] = results_obj.mAP([class_])
-    return d
+
    def per_class(results_obj):
+    """helper function: take fiftyone/openimagev6 results object & spit
+    out mAP breakdown as per class
+    """
+    d = {}
+    for class_ in classes:
+        d[class_] = results_obj.mAP([class_])
+    return d
 
    
 
    
 
    CompressAIEncoderDecoder is a subclass of EncoderDecoder,
@@ -501,48 +501,48 @@ 
    3. Evaluate
    
 image before the image is passed to the Detectron2 predictor.
    
 
    We run the bmshj2018_factorized model over various quality
 parameters:
    
-
    params=[1] # debugging
-# params=[1,2,3,4,5,6,7,8]
+
    params=[1] # debugging
+# params=[1,2,3,4,5,6,7,8]
 
    
 
    
 
    Detectron prediction results are saved during the run into the fiftyone
 (mongodb) database. Let’s define a unique name for the sample field
 where the detectron results are saved:
    
-
    predictor_field='detectron-predictions'
-
    
-
    
-
    xs=[]; ys=[]; maps=[]; # bpp, mAP values, mAP(s) per class
-results=[] # complete results
-for i in params:
-    net = bmshj2018_factorized(quality=i, pretrained=True).eval().to(device)
-    enc_dec = CompressAIEncoderDecoder(net, device=device)
-    # note the EncoderDecoder instance here:
-    # before the predictor is used, the image is crunched through the encoding/decoding process & the bitrate is recorded
-    # you could substitute CompressAIEncoderDecoder with VTMEncoderDecoder if you'd like to (see also the end of this tutorial)
-    print("running the detector at", i)
-    bpp = annexPredictions(predictors=[predictor], fo_dataset=dataset, encoder_decoder=enc_dec, predictor_fields=[predictor_field])
-    # .. now detectron's results are in each sample at the "detectron-predictions"  field
-    res = dataset.evaluate_detections(
-        predictor_field,
-        gt_field="detections",
-        method="open-images",
-        pos_label_field="positive_labels",
-        neg_label_field="negative_labels",
-        expand_pred_hierarchy=False,
-        expand_gt_hierarchy=False
-    )
-    results.append((i, bpp, res))
-    # save to disk at each iteration as a backup just in case
-    xs.append(bpp)
-    ys.append(res.mAP())
-    maps.append(per_class(res))
-    with open("out.json","w") as f:
-        f.write(json.dumps({
-            "bpp" : xs,
-            "map" : ys,
-            "map_per_class" : maps
-            }, indent=2))
-print("ready!")
+
    predictor_field='detectron-predictions'
+
    
+
    
+
    xs=[]; ys=[]; maps=[]; # bpp, mAP values, mAP(s) per class
+results=[] # complete results
+for i in params:
+    net = bmshj2018_factorized(quality=i, pretrained=True).eval().to(device)
+    enc_dec = CompressAIEncoderDecoder(net, device=device)
+    # note the EncoderDecoder instance here:
+    # before the predictor is used, the image is crunched through the encoding/decoding process & the bitrate is recorded
+    # you could substitute CompressAIEncoderDecoder with VTMEncoderDecoder if you'd like to (see also the end of this tutorial)
+    print("running the detector at", i)
+    bpp = annexPredictions(predictors=[predictor], fo_dataset=dataset, encoder_decoder=enc_dec, predictor_fields=[predictor_field])
+    # .. now detectron's results are in each sample at the "detectron-predictions"  field
+    res = dataset.evaluate_detections(
+        predictor_field,
+        gt_field="detections",
+        method="open-images",
+        pos_label_field="positive_labels",
+        neg_label_field="negative_labels",
+        expand_pred_hierarchy=False,
+        expand_gt_hierarchy=False
+    )
+    results.append((i, bpp, res))
+    # save to disk at each iteration as a backup just in case
+    xs.append(bpp)
+    ys.append(res.mAP())
+    maps.append(per_class(res))
+    with open("out.json","w") as f:
+        f.write(json.dumps({
+            "bpp" : xs,
+            "map" : ys,
+            "map_per_class" : maps
+            }, indent=2))
+print("ready!")
 
    
 
    
 
    running the detector at 1
@@ -561,13 +561,13 @@ 
    3. Evaluate
    
 
    
 
    After the evaluation we can (and should!) remove the detectron results
 from the database:
    
-
    dataset.delete_sample_fields(predictor_field)
+
    dataset.delete_sample_fields(predictor_field)
 
    
 
    
 
    Load results
    
-
    with open("out.json","r") as f:
-    res=json.load(f)
-print(res)
+
    with open("out.json","r") as f:
+    res=json.load(f)
+print(res)
 
    
 
    
 
    {'bpp': [0.10060123042505593], 'map': [1.0], 'map_per_class': [{'airplane': 1.0}]}
@@ -577,31 +577,31 @@ 
    3. Evaluate
    
 CompressAIEncoderDecoder with VTMEncoderDecoderin order to
 produce the anchor/baseline results. Let’s first set some variables for
 the VTM program:
    
-
    # NOTE: set path_to_vtm_software
-vtm_encoder_app=os.path.join(path_to_vtm_software, "bin/EncoderAppStatic")
-vtm_decoder_app=os.path.join(path_to_vtm_software, "bin/DecoderAppStatic")
-vtm_cfg=os.path.join(path_to_vtm_software, "cfg/encoder_intra_vtm.cfg")
+
    # NOTE: set path_to_vtm_software
+vtm_encoder_app=os.path.join(path_to_vtm_software, "bin/EncoderAppStatic")
+vtm_decoder_app=os.path.join(path_to_vtm_software, "bin/DecoderAppStatic")
+vtm_cfg=os.path.join(path_to_vtm_software, "cfg/encoder_intra_vtm.cfg")
 
    
 
    
 
    If you’d want to see what the VTM is doing exactly, enable debugging
 output:
    
-
    loglev=logging.DEBUG
-# loglev=logging.INFO
-log=quickLog("VTMEncoderDecoder", loglev) # VTMEncoderDecoder
+
    loglev=logging.DEBUG
+# loglev=logging.INFO
+log=quickLog("VTMEncoderDecoder", loglev) # VTMEncoderDecoder
 
    
 
    
 
    At each quality parameter in the loop, instantiate an
 VTMEncoderDecoder instead:
    
-
    enc_dec = VTMEncoderDecoder(
-    encoderApp=vtm_encoder_app,
-    decoderApp=vtm_decoder_app,
-    ffmpeg="ffmpeg",
-    vtm_cfg=vtm_cfg,
-    qp=47,
-    cache="/tmp/bitstreams",
-    scale=100,
-    warn=True
-)
+
    enc_dec = VTMEncoderDecoder(
+    encoderApp=vtm_encoder_app,
+    decoderApp=vtm_decoder_app,
+    ffmpeg="ffmpeg",
+    vtm_cfg=vtm_cfg,
+    qp=47,
+    cache="/tmp/bitstreams",
+    scale=100,
+    warn=True
+)
 
    
 
    
 
    VTMEncoderDecoder - WARNING - folder /tmp/bitstreams/100/47 exists already
diff --git a/tutorials/evaluate_nb.html b/tutorials/evaluate_nb.html
index 0cd41664..ea6313b2 100644
--- a/tutorials/evaluate_nb.html
+++ b/tutorials/evaluate_nb.html
@@ -349,78 +349,78 @@ 
     Contents 
    
   
    In this tutorial we evaluate mAP values for a dataset with Detectron2
 and a deep-learning encoding model from the CompressAI library. We also
 show how to perform a baseline evaluation with VTM.
    
-
    # common libs
-import math, os, io, json, cv2, random, logging, pickle, datetime
-import numpy as np
-# torch
-import torch
-# images
-from PIL import Image
-import matplotlib.pyplot as plt
-# compressai
-from compressai.zoo import bmshj2018_factorized
+
    # common libs
+import math, os, io, json, cv2, random, logging, pickle, datetime
+import numpy as np
+# torch
+import torch
+# images
+from PIL import Image
+import matplotlib.pyplot as plt
+# compressai
+from compressai.zoo import bmshj2018_factorized
 
    
 
    
-
    ## *** Detectron imports ***
-import detectron2
-from detectron2.utils.logger import setup_logger
-setup_logger()
+
    ## *** Detectron imports ***
+import detectron2
+from detectron2.utils.logger import setup_logger
+setup_logger()
 
-# import some common detectron2 utilities
-from detectron2 import model_zoo
-from detectron2.engine import DefaultPredictor
-from detectron2.config import get_cfg
-from detectron2.utils.visualizer import Visualizer
-from detectron2.data import MetadataCatalog, DatasetCatalog
+# import some common detectron2 utilities
+from detectron2 import model_zoo
+from detectron2.engine import DefaultPredictor
+from detectron2.config import get_cfg
+from detectron2.utils.visualizer import Visualizer
+from detectron2.data import MetadataCatalog, DatasetCatalog
 
    
 
    
-
    # CompressAI-Vision
-from compressai_vision.conversion import FO2DetectronDataset # convert fiftyone dataset to Detectron2 dataset
-from compressai_vision.conversion import detectron251 # convert Detectron2 results to fiftyone format
-from compressai_vision.evaluation.fo import annexPredictions # annex predictions from
-from compressai_vision.evaluation.pipeline import CompressAIEncoderDecoder, VTMEncoderDecoder # a class that does encoding+decoding & returns the transformed image & bitrate
-from compressai_vision.pipelines.remote_analysis.tools import confLogger, quickLog, getDataFile
+
    # CompressAI-Vision
+from compressai_vision.conversion import FO2DetectronDataset # convert fiftyone dataset to Detectron2 dataset
+from compressai_vision.conversion import detectron251 # convert Detectron2 results to fiftyone format
+from compressai_vision.evaluation.fo import annexPredictions # annex predictions from
+from compressai_vision.evaluation.pipeline import CompressAIEncoderDecoder, VTMEncoderDecoder # a class that does encoding+decoding & returns the transformed image & bitrate
+from compressai_vision.pipelines.remote_analysis.tools import confLogger, quickLog, getDataFile
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
-
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
-print(device)
+
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+print(device)
 
    
 
    
 
    cpu
 
    
 
    
-
    ## MODEL A
-model_name="COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"
-## look here:
-## https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md#faster-r-cnn
-## for the line that says X101-FPN --> box AP is 43
+
    ## MODEL A
+model_name="COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"
+## look here:
+## https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md#faster-r-cnn
+## for the line that says X101-FPN --> box AP is 43
 
-## MODEL B
-# model_name="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
+## MODEL B
+# model_name="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
 
    
 
    
-
    # cfg encapsulates the model architecture & weights, also threshold parameter, metadata, etc.
-cfg = get_cfg()
-cfg.MODEL.DEVICE=device
-# load config from a file:
-cfg.merge_from_file(model_zoo.get_config_file(model_name))
-# DO NOT TOUCH THRESHOLD WHEN DOING EVALUATION:
-# too big a threshold will cut the smallest values & affect the precision(recall) curves & evaluation results
-# the default value is 0.05
-# value of 0.01 saturates the results (they don't change at lower values)
-# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
-# get weights
-cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(model_name)
-print("expected input colorspace:", cfg.INPUT.FORMAT)
-print("loaded datasets:", cfg.DATASETS)
-model_dataset=cfg.DATASETS.TRAIN[0]
-print("model was trained with", model_dataset)
-model_meta=MetadataCatalog.get(model_dataset)
+
    # cfg encapsulates the model architecture & weights, also threshold parameter, metadata, etc.
+cfg = get_cfg()
+cfg.MODEL.DEVICE=device
+# load config from a file:
+cfg.merge_from_file(model_zoo.get_config_file(model_name))
+# DO NOT TOUCH THRESHOLD WHEN DOING EVALUATION:
+# too big a threshold will cut the smallest values & affect the precision(recall) curves & evaluation results
+# the default value is 0.05
+# value of 0.01 saturates the results (they don't change at lower values)
+# cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
+# get weights
+cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(model_name)
+print("expected input colorspace:", cfg.INPUT.FORMAT)
+print("loaded datasets:", cfg.DATASETS)
+model_dataset=cfg.DATASETS.TRAIN[0]
+print("model was trained with", model_dataset)
+model_meta=MetadataCatalog.get(model_dataset)
 
    
 
    
 
    expected input colorspace: BGR
@@ -433,18 +433,18 @@ 
     Contents 
    
 model was trained with coco_2017_train
 
    
 
    
-
    # model_meta.thing_classes # check class labels this was trained with
+
    # model_meta.thing_classes # check class labels this was trained with
 
    
 
    
-
    predictor = DefaultPredictor(cfg)
+
    predictor = DefaultPredictor(cfg)
 
    
 
    
 
    Get handle to a dataset. We will be using the oiv6-mpeg-v1 dataset.
 Please go through the CLI Tutorials in order to produce this dataset.
    
-
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1-dummy") # or use the dummy dataset for testing/debugging
+
    dataset = fo.load_dataset("oiv6-mpeg-detection-v1-dummy") # or use the dummy dataset for testing/debugging
 
    
 
    
-
    dataset
+
    dataset
 
    
 
    
 
    Name:        oiv6-mpeg-detection-v1-dummy
@@ -464,32 +464,32 @@ 
     Contents 
    
 
    
 
    
 
    Set some loglevels
    
-
    # loglev=logging.DEBUG
-loglev=logging.INFO
-quickLog("CompressAIEncoderDecoder", loglev)
+
    # loglev=logging.DEBUG
+loglev=logging.INFO
+quickLog("CompressAIEncoderDecoder", loglev)
 
    
 
    
 
    <Logger CompressAIEncoderDecoder (INFO)>
 
    
 
    
 
    Get a list of labels in the dataset:
    
-
    classes = dataset.distinct(
-    "detections.detections.label"
-)
-print(classes)
+
    classes = dataset.distinct(
+    "detections.detections.label"
+)
+print(classes)
 
    
 
    
 
    ['airplane']
 
    
 
    
-
    def per_class(results_obj):
-    """helper function: take fiftyone/openimagev6 results object & spit
-    out mAP breakdown as per class
-    """
-    d = {}
-    for class_ in classes:
-        d[class_] = results_obj.mAP([class_])
-    return d
+
    def per_class(results_obj):
+    """helper function: take fiftyone/openimagev6 results object & spit
+    out mAP breakdown as per class
+    """
+    d = {}
+    for class_ in classes:
+        d[class_] = results_obj.mAP([class_])
+    return d
 
    
 
    
 
    CompressAIEncoderDecoder is a subclass of EncoderDecoder,
@@ -504,48 +504,48 @@ 
     Contents 
    
 image before the image is passed to the Detectron2 predictor.
    
 
    We run the bmshj2018_factorized model over various quality
 parameters:
    
-
    params=[1] # debugging
-# params=[1,2,3,4,5,6,7,8]
+
    params=[1] # debugging
+# params=[1,2,3,4,5,6,7,8]
 
    
 
    
 
    Detectron prediction results are saved during the run into the fiftyone
 (mongodb) database. Let’s define a unique name for the sample field
 where the detectron results are saved:
    
-
    predictor_field='detectron-predictions'
-
    
-
    
-
    xs=[]; ys=[]; maps=[]; # bpp, mAP values, mAP(s) per class
-results=[] # complete results
-for i in params:
-    net = bmshj2018_factorized(quality=i, pretrained=True).eval().to(device)
-    enc_dec = CompressAIEncoderDecoder(net, device=device)
-    # note the EncoderDecoder instance here:
-    # before the predictor is used, the image is crunched through the encoding/decoding process & the bitrate is recorded
-    # you could substitute CompressAIEncoderDecoder with VTMEncoderDecoder if you'd like to (see also the end of this tutorial)
-    print("running the detector at", i)
-    bpp = annexPredictions(predictors=[predictor], fo_dataset=dataset, encoder_decoder=enc_dec, predictor_fields=[predictor_field])
-    # .. now detectron's results are in each sample at the "detectron-predictions"  field
-    res = dataset.evaluate_detections(
-        predictor_field,
-        gt_field="detections",
-        method="open-images",
-        pos_label_field="positive_labels",
-        neg_label_field="negative_labels",
-        expand_pred_hierarchy=False,
-        expand_gt_hierarchy=False
-    )
-    results.append((i, bpp, res))
-    # save to disk at each iteration as a backup just in case
-    xs.append(bpp)
-    ys.append(res.mAP())
-    maps.append(per_class(res))
-    with open("out.json","w") as f:
-        f.write(json.dumps({
-            "bpp" : xs,
-            "map" : ys,
-            "map_per_class" : maps
-            }, indent=2))
-print("ready!")
+
    predictor_field='detectron-predictions'
+
    
+
    
+
    xs=[]; ys=[]; maps=[]; # bpp, mAP values, mAP(s) per class
+results=[] # complete results
+for i in params:
+    net = bmshj2018_factorized(quality=i, pretrained=True).eval().to(device)
+    enc_dec = CompressAIEncoderDecoder(net, device=device)
+    # note the EncoderDecoder instance here:
+    # before the predictor is used, the image is crunched through the encoding/decoding process & the bitrate is recorded
+    # you could substitute CompressAIEncoderDecoder with VTMEncoderDecoder if you'd like to (see also the end of this tutorial)
+    print("running the detector at", i)
+    bpp = annexPredictions(predictors=[predictor], fo_dataset=dataset, encoder_decoder=enc_dec, predictor_fields=[predictor_field])
+    # .. now detectron's results are in each sample at the "detectron-predictions"  field
+    res = dataset.evaluate_detections(
+        predictor_field,
+        gt_field="detections",
+        method="open-images",
+        pos_label_field="positive_labels",
+        neg_label_field="negative_labels",
+        expand_pred_hierarchy=False,
+        expand_gt_hierarchy=False
+    )
+    results.append((i, bpp, res))
+    # save to disk at each iteration as a backup just in case
+    xs.append(bpp)
+    ys.append(res.mAP())
+    maps.append(per_class(res))
+    with open("out.json","w") as f:
+        f.write(json.dumps({
+            "bpp" : xs,
+            "map" : ys,
+            "map_per_class" : maps
+            }, indent=2))
+print("ready!")
 
    
 
    
 
    running the detector at 1
@@ -564,13 +564,13 @@ 
     Contents 
    
 
    
 
    After the evaluation we can (and should!) remove the detectron results
 from the database:
    
-
    dataset.delete_sample_fields(predictor_field)
+
    dataset.delete_sample_fields(predictor_field)
 
    
 
    
 
    Load results
    
-
    with open("out.json","r") as f:
-    res=json.load(f)
-print(res)
+
    with open("out.json","r") as f:
+    res=json.load(f)
+print(res)
 
    
 
    
 
    {'bpp': [0.10060123042505593], 'map': [1.0], 'map_per_class': [{'airplane': 1.0}]}
@@ -580,31 +580,31 @@ 
     Contents 
    
 CompressAIEncoderDecoder with VTMEncoderDecoderin order to
 produce the anchor/baseline results. Let’s first set some variables for
 the VTM program:
    
-
    # NOTE: set path_to_vtm_software
-vtm_encoder_app=os.path.join(path_to_vtm_software, "bin/EncoderAppStatic")
-vtm_decoder_app=os.path.join(path_to_vtm_software, "bin/DecoderAppStatic")
-vtm_cfg=os.path.join(path_to_vtm_software, "cfg/encoder_intra_vtm.cfg")
+
    # NOTE: set path_to_vtm_software
+vtm_encoder_app=os.path.join(path_to_vtm_software, "bin/EncoderAppStatic")
+vtm_decoder_app=os.path.join(path_to_vtm_software, "bin/DecoderAppStatic")
+vtm_cfg=os.path.join(path_to_vtm_software, "cfg/encoder_intra_vtm.cfg")
 
    
 
    
 
    If you’d want to see what the VTM is doing exactly, enable debugging
 output:
    
-
    loglev=logging.DEBUG
-# loglev=logging.INFO
-log=quickLog("VTMEncoderDecoder", loglev) # VTMEncoderDecoder
+
    loglev=logging.DEBUG
+# loglev=logging.INFO
+log=quickLog("VTMEncoderDecoder", loglev) # VTMEncoderDecoder
 
    
 
    
 
    At each quality parameter in the loop, instantiate an
 VTMEncoderDecoder instead:
    
-
    enc_dec = VTMEncoderDecoder(
-    encoderApp=vtm_encoder_app,
-    decoderApp=vtm_decoder_app,
-    ffmpeg="ffmpeg",
-    vtm_cfg=vtm_cfg,
-    qp=47,
-    cache="/tmp/bitstreams",
-    scale=100,
-    warn=True
-)
+
    enc_dec = VTMEncoderDecoder(
+    encoderApp=vtm_encoder_app,
+    decoderApp=vtm_decoder_app,
+    ffmpeg="ffmpeg",
+    vtm_cfg=vtm_cfg,
+    qp=47,
+    cache="/tmp/bitstreams",
+    scale=100,
+    warn=True
+)
 
    
 
    
 
    VTMEncoderDecoder - WARNING - folder /tmp/bitstreams/100/47 exists already
diff --git a/tutorials/fiftyone.html b/tutorials/fiftyone.html
index 586e9977..2ad672f7 100644
--- a/tutorials/fiftyone.html
+++ b/tutorials/fiftyone.html
@@ -371,18 +371,18 @@ 
    Fiftyone and MongoDB
 
    Let’s take a closer look:
    
-
    # image tool imports
-from PIL import Image
-import matplotlib.pyplot as plt
+
    # image tool imports
+from PIL import Image
+import matplotlib.pyplot as plt
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
 
    Lets take a look at the datasets registered to fiftyone:
    
-
    fo.list_datasets()
+
    fo.list_datasets()
 
    
 
    
 
    ['mpeg-vcm-detection',
@@ -394,19 +394,19 @@ 
    Fiftyone and MongoDB
    
 
    
 
    Let’s get a handle to a dataset:
    
-
    dataset=fo.load_dataset("quickstart")
+
    dataset=fo.load_dataset("quickstart")
 
    
 
    
 
    Let’s see how many samples we have in it:
    
-
    len(dataset)
+
    len(dataset)
 
    
 
    
 
    200
 
    
 
    
 
    Let’s take a look at the first sample:
    
-
    sample=dataset.first()
-print(sample)
+
    sample=dataset.first()
+print(sample)
 
    
 
    
 
    <Sample: {
@@ -696,7 +696,7 @@ 
    Fiftyone and MongoDB
 
    Let’s load an image:
    
-
    plt.imshow(Image.open(sample["filepath"]))
+
    plt.imshow(Image.open(sample["filepath"]))
 
    
 
    
 
    <matplotlib.image.AxesImage at 0x7fdf5a7fc640>
@@ -705,7 +705,7 @@ 
    Fiftyone and MongoDB
 
    Let’s see a summary of the dataset and what kind of fields each samples
 has:
    
-
    print(dataset)
+
    print(dataset)
 
    
 
    
 
    Name:        quickstart
@@ -731,14 +731,14 @@ 
    Fiftyone and MongoDB
    
 
    Here at the final, a small recompilation/cheatsheet of selected fiftyone
 features
    
-
    # Access by sample id
-sample=dataset["634472860faf93a9a586c9c4"]
+
    # Access by sample id
+sample=dataset["634472860faf93a9a586c9c4"]
 
    
 
    
-
    # Search by a field value.  You might need this one with the with open_images_id field.
-from fiftyone import ViewField as F
-tmpset=dataset[F("filepath") == dataset.first().filepath]
-print(tmpset)
+
    # Search by a field value.  You might need this one with the with open_images_id field.
+from fiftyone import ViewField as F
+tmpset=dataset[F("filepath") == dataset.first().filepath]
+print(tmpset)
 
    
 
    
 
    Dataset:     quickstart
diff --git a/tutorials/fiftyone_nb.html b/tutorials/fiftyone_nb.html
index 42feca9f..41f0b082 100644
--- a/tutorials/fiftyone_nb.html
+++ b/tutorials/fiftyone_nb.html
@@ -377,18 +377,18 @@ 
     Contents 
    
 sharing data and/or if you’re working in a supercomputing / grid
 environment.
    
 
    Let’s take a closer look:
    
-
    # image tool imports
-from PIL import Image
-import matplotlib.pyplot as plt
+
    # image tool imports
+from PIL import Image
+import matplotlib.pyplot as plt
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
 
    Lets take a look at the datasets registered to fiftyone:
    
-
    fo.list_datasets()
+
    fo.list_datasets()
 
    
 
    
 
    ['mpeg-vcm-detection',
@@ -400,19 +400,19 @@ 
     Contents 
    
 
    
 
    
 
    Let’s get a handle to a dataset:
    
-
    dataset=fo.load_dataset("quickstart")
+
    dataset=fo.load_dataset("quickstart")
 
    
 
    
 
    Let’s see how many samples we have in it:
    
-
    len(dataset)
+
    len(dataset)
 
    
 
    
 
    200
 
    
 
    
 
    Let’s take a look at the first sample:
    
-
    sample=dataset.first()
-print(sample)
+
    sample=dataset.first()
+print(sample)
 
    
 
    
 
    <Sample: {
@@ -702,7 +702,7 @@ 
     Contents 
    
 (instead of writing lots of intermediate files on the disk like with
 COCO API or with the tensorflow tools).
    
 
    Let’s load an image:
    
-
    plt.imshow(Image.open(sample["filepath"]))
+
    plt.imshow(Image.open(sample["filepath"]))
 
    
 
    
 
    <matplotlib.image.AxesImage at 0x7fdf5a7fc640>
@@ -711,7 +711,7 @@ 
     Contents 
    
 ../_images/fiftyone_nb_12_1.png
 
    Let’s see a summary of the dataset and what kind of fields each samples
 has:
    
-
    print(dataset)
+
    print(dataset)
 
    
 
    
 
    Name:        quickstart
@@ -737,14 +737,14 @@ 
     Contents 
    
 documentation
    
 
    Here at the final, a small recompilation/cheatsheet of selected fiftyone
 features
    
-
    # Access by sample id
-sample=dataset["634472860faf93a9a586c9c4"]
+
    # Access by sample id
+sample=dataset["634472860faf93a9a586c9c4"]
 
    
 
    
-
    # Search by a field value.  You might need this one with the with open_images_id field.
-from fiftyone import ViewField as F
-tmpset=dataset[F("filepath") == dataset.first().filepath]
-print(tmpset)
+
    # Search by a field value.  You might need this one with the with open_images_id field.
+from fiftyone import ViewField as F
+tmpset=dataset[F("filepath") == dataset.first().filepath]
+print(tmpset)
 
    
 
    
 
    Dataset:     quickstart
diff --git a/tutorials/index.html b/tutorials/index.html
index e53be9b4..8d99d19a 100644
--- a/tutorials/index.html
+++ b/tutorials/index.html
@@ -400,42 +400,42 @@ 
    Input file conversionWe convert this data into OpenImageV6 format and also register it into fiftyone.
    
 
    In this chapter, we create an evaluation dataset as defined by the
 MPEG-VCM working group
    
-
    # common libs
-import math, os, io, json, cv2, random, logging
-import numpy as np
-# images
-from PIL import Image
-import matplotlib.pyplot as plt
+
    # common libs
+import math, os, io, json, cv2, random, logging
+import numpy as np
+# images
+from PIL import Image
+import matplotlib.pyplot as plt
 
    
 
    
-
    homie=os.path.expanduser("~")
-print("your home path is", homie)
-fodir=os.path.join(homie,'fiftyone')
-print("fiftyone dowloads data by default to", fodir)
-try:
-    os.mkdir(fodir)
-except FileExistsError:
-    pass
+
    homie=os.path.expanduser("~")
+print("your home path is", homie)
+fodir=os.path.join(homie,'fiftyone')
+print("fiftyone dowloads data by default to", fodir)
+try:
+    os.mkdir(fodir)
+except FileExistsError:
+    pass
 
    
 
    
 
    your home path is /home/sampsa
 fiftyone dowloads data by default to /home/sampsa/fiftyone
 
    
 
    
-
    # fiftyone
-import fiftyone as fo
-import fiftyone.zoo as foz
+
    # fiftyone
+import fiftyone as fo
+import fiftyone.zoo as foz
 
    
 
    
-
    # CompressAI-Vision
-from compressai_vision.conversion import MPEGVCMToOpenImageV6, imageIdFileList
+
    # CompressAI-Vision
+from compressai_vision.conversion import MPEGVCMToOpenImageV6, imageIdFileList
 
    
 
    
 
    We expect that you have downloaded correct images and segmentation masks
 into open-images-v6 folder (as instructed in the previous chapter)
    
-
    dir_=os.path.join(fodir,"open-images-v6")
-print("contents of", dir_,":")
-!tree --filelimit=10 $dir_ | cat
+
    dir_=os.path.join(fodir,"open-images-v6")
+print("contents of", dir_,":")
+!tree --filelimit=10 $dir_ | cat
 
    
 
    
 
    contents of /home/sampsa/fiftyone/open-images-v6 :
@@ -470,36 +470,36 @@ 
    Input file conversiondetection_validation_input_5k.lst          = list of images used
 
    
 
    
-
    # TODO: define path_to_mpeg_vcm_files
-path_to_images=os.path.join(fodir,"open-images-v6/validation/data")
+
    # TODO: define path_to_mpeg_vcm_files
+path_to_images=os.path.join(fodir,"open-images-v6/validation/data")
 
-list_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_input_5k.lst")
-bbox_csv_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_5k_bbox.csv")
-validation_csv_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_labels_5k.csv")
+list_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_input_5k.lst")
+bbox_csv_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_5k_bbox.csv")
+validation_csv_file=os.path.join(path_to_mpeg_vcm_files, "detection_validation_labels_5k.csv")
 
-assert(os.path.exists(bbox_csv_file)), "can't find bbox file"
-assert(os.path.exists(validation_csv_file)), "can't find labels file"
-assert(os.path.exists(path_to_images)), "can't find image directory"
+assert(os.path.exists(bbox_csv_file)), "can't find bbox file"
+assert(os.path.exists(validation_csv_file)), "can't find labels file"
+assert(os.path.exists(path_to_images)), "can't find image directory"
 
    
 
    
 
    Now we convert mpeg vmc proprietary format annotation into proper
 OpenImageV6 format dataset and place it into
 ~/fiftyone/mpeg_vcm-detection
    
 
    First, remove any previously imported stuff:
    
-
    !rm -rf ~/fiftyone/mpeg-vcm-*
+
    !rm -rf ~/fiftyone/mpeg-vcm-*
 
    
 
    
-
    MPEGVCMToOpenImageV6(
-    validation_csv_file=validation_csv_file,
-    list_file=list_file,
-    bbox_csv_file=bbox_csv_file,
-    output_directory=os.path.join(fodir,"mpeg-vcm-detection"),
-    data_dir=path_to_images
-)
+
    MPEGVCMToOpenImageV6(
+    validation_csv_file=validation_csv_file,
+    list_file=list_file,
+    bbox_csv_file=bbox_csv_file,
+    output_directory=os.path.join(fodir,"mpeg-vcm-detection"),
+    data_dir=path_to_images
+)
 
    
 
    
 
    let’s see what we got:
    
-
    !tree --filelimit=10 ~/fiftyone/mpeg-vcm-detection | cat
+
    !tree --filelimit=10 ~/fiftyone/mpeg-vcm-detection | cat
 
    
 
    
 
    /home/sampsa/fiftyone/mpeg-vcm-detection
@@ -521,46 +521,46 @@ 
    Input file conversiondata -> ~/fiftyone/open-images-v6/validation/data)
    
 
    The only thing we’re left to do, is to register this OpenImageV6
 formatted dataset into fiftyone:
    
-
    # remove the dataset in the case it was already registered in fiftyone
-try:
-    fo.delete_dataset("mpeg-vcm-detection")
-except ValueError as e:
-    print("could not delete because of", e)
+
    # remove the dataset in the case it was already registered in fiftyone
+try:
+    fo.delete_dataset("mpeg-vcm-detection")
+except ValueError as e:
+    print("could not delete because of", e)
 
    
 
    
-
    dataset_type = fo.types.OpenImagesV6Dataset
-dataset_dir = os.path.join(fodir,"mpeg-vcm-detection")
-dataset = fo.Dataset.from_dir(
-    dataset_dir=dataset_dir,
-    dataset_type=dataset_type,
-    label_types=("detections","classifications"),
-    load_hierarchy=False,
-    name="mpeg-vcm-detection",
-    image_ids=imageIdFileList(list_file)
-)
+
    dataset_type = fo.types.OpenImagesV6Dataset
+dataset_dir = os.path.join(fodir,"mpeg-vcm-detection")
+dataset = fo.Dataset.from_dir(
+    dataset_dir=dataset_dir,
+    dataset_type=dataset_type,
+    label_types=("detections","classifications"),
+    load_hierarchy=False,
+    name="mpeg-vcm-detection",
+    image_ids=imageIdFileList(list_file)
+)
 
    
 
    
 
    100% |███████████████| 5000/5000 [16.8s elapsed, 0s remaining, 290.4 samples/s]
 
    
 
    
-
    dataset.persistent=True # without this, your dabatase will disappear!
+
    dataset.persistent=True # without this, your dabatase will disappear!
 
    
 
    
-
    ## now, in the future, just do
-dataset = fo.load_dataset("mpeg-vcm-detection")
+
    ## now, in the future, just do
+dataset = fo.load_dataset("mpeg-vcm-detection")
 
    
 
    
 
    Finaly, let’s also create a dummy dataset for debugging and testing with
 only one sample:
    
-
    try:
-    fo.delete_dataset("mpeg-vcm-detection-dummy")
-except ValueError:
-    print("no dummmy dataset yet..")
-dummy_dataset=fo.Dataset("mpeg-vcm-detection-dummy")
-for sample in dataset[0:1]:
-    dummy_dataset.add_sample(sample)
-dummy_dataset.persistent=True
-print("dummy dataset ok")
+
    try:
+    fo.delete_dataset("mpeg-vcm-detection-dummy")
+except ValueError:
+    print("no dummmy dataset yet..")
+dummy_dataset=fo.Dataset("mpeg-vcm-detection-dummy")
+for sample in dataset[0:1]:
+    dummy_dataset.add_sample(sample)
+dummy_dataset.persistent=True
+print("dummy dataset ok")
 
    
 
    
 
    dummy dataset ok