[Datumaro] Add generic accuracy checker model launcher (cvat-ai#1661)

* Refactor inference wrapper

* Add accuracy checker launcher wrapper

* t

* rename method

* Add importer for openvino launcher

* Move openvino plugin to iecore

* add generic AC launcher

* Implement cli for AC launcher

* move ac plugin dir

* prevent tf reimport

* Fix outputs conversion

* t

* add pytorch model example

* Require config path in launcher

* Clear extra whitespace

datumaro/datumaro/plugins/accuracy_checker_plugin/details/ac.py

@@ -0,0 +1,116 @@
+
+# Copyright (C) 2020 Intel Corporation
+#
+# SPDX-License-Identifier: MIT
+
+from datumaro.util.tf_util import import_tf
+import_tf() # prevent TF loading and potential interpeter crash
+
+from itertools import groupby
+
+from accuracy_checker.adapters import create_adapter
+from accuracy_checker.data_readers import DataRepresentation
+from accuracy_checker.launcher import InputFeeder, create_launcher
+from accuracy_checker.postprocessor import PostprocessingExecutor
+from accuracy_checker.preprocessor import PreprocessingExecutor
+from accuracy_checker.utils import extract_image_representations
+
+from datumaro.components.extractor import AnnotationType, LabelCategories
+
+from .representation import import_predictions
+
+
+class _FakeDataset:
+    def __init__(self, metadata=None):
+        self.metadata = metadata or {}
+
+class GenericAcLauncher:
+    @staticmethod
+    def from_config(config):
+        launcher_config = config['launcher']
+        launcher = create_launcher(launcher_config)
+
+        dataset = _FakeDataset()
+        adapter_config = config.get('adapter') or launcher_config.get('adapter')
+        label_config = adapter_config.get('labels') \
+            if isinstance(adapter_config, dict) else None
+        if label_config:
+            assert isinstance(label_config, (list, dict))
+            if isinstance(label_config, list):
+                label_config = dict(enumerate(label_config))
+
+            dataset.metadata = {'label_map': {
+                int(key): label for key, label in label_config.items()
+            }}
+        adapter = create_adapter(adapter_config, launcher, dataset)
+
+        preproc_config = config.get('preprocessing')
+        preproc = None
+        if preproc_config:
+            preproc = PreprocessingExecutor(preproc_config,
+                dataset_meta=dataset.metadata,
+                input_shapes=launcher.inputs_info_for_meta()
+            )
+
+        postproc_config = config.get('postprocessing')
+        postproc = None
+        if postproc_config:
+            postproc = PostprocessingExecutor(postproc_config,
+                dataset_meta=dataset.metadata,
+            )
+
+        return __class__(launcher,
+            adapter=adapter, preproc=preproc, postproc=postproc)
+
+    def __init__(self, launcher, adapter=None,
+            preproc=None, postproc=None, input_feeder=None):
+        self._launcher = launcher
+        self._input_feeder = input_feeder or InputFeeder(
+            launcher.config.get('inputs', []), launcher.inputs,
+            launcher.fit_to_input, launcher.default_layout
+        )
+        self._adapter = adapter
+        self._preproc = preproc
+        self._postproc = postproc
+
+        self._categories = self._init_categories()
+
+    def launch_raw(self, inputs):
+        ids = range(len(inputs))
+        inputs = [DataRepresentation(inp, identifier=id)
+            for id, inp in zip(ids, inputs)]
+        _, batch_meta = extract_image_representations(inputs)
+
+        if self._preproc:
+            inputs = self._preproc.process(inputs)
+
+        inputs = self._input_feeder.fill_inputs(inputs)
+        outputs = self._launcher.predict(inputs, batch_meta)
+
+        if self._adapter:
+            outputs = self._adapter.process(outputs, ids, batch_meta)
+
+        if self._postproc:
+            outputs = self._postproc.process(outputs)
+
+        return outputs
+
+    def launch(self, inputs):
+        outputs = self.launch_raw(inputs)
+        return [import_predictions(g) for _, g in
+            groupby(outputs, key=lambda o: o.identifier)]
+
+    def categories(self):
+        return self._categories
+
+    def _init_categories(self):
+        if self._adapter is None or self._adapter.label_map is None:
+            return None
+
+        label_map = sorted(self._adapter.label_map.items(), key=lambda e: e[0])
+
+        label_cat = LabelCategories()
+        for _, label in label_map:
+            label_cat.add(label)
+
+        return { AnnotationType.label: label_cat }

datumaro/datumaro/plugins/accuracy_checker_plugin/details/representation.py

@@ -0,0 +1,62 @@
+
+# Copyright (C) 2020 Intel Corporation
+#
+# SPDX-License-Identifier: MIT
+
+from datumaro.util.tf_util import import_tf
+import_tf() # prevent TF loading and potential interpeter crash
+
+import accuracy_checker.representation as ac
+
+import datumaro.components.extractor as dm
+from datumaro.util.annotation_tools import softmax
+
+def import_predictions(predictions):
+    # Convert Accuracy checker predictions to Datumaro annotations
+
+    anns = []
+
+    for pred in predictions:
+        anns.extend(import_prediction(pred))
+
+    return anns
+
+def import_prediction(pred):
+    if isinstance(pred, ac.ClassificationPrediction):
+        scores = softmax(pred.scores)
+        return (dm.Label(label_id, attributes={'score': float(score)})
+            for label_id, score in enumerate(scores))
+    elif isinstance(pred, ac.ArgMaxClassificationPrediction):
+        return (dm.Label(int(pred.label)), )
+    elif isinstance(pred, ac.CharacterRecognitionPrediction):
+        return (dm.Label(int(pred.label)), )
+    elif isinstance(pred, (ac.DetectionPrediction, ac.ActionDetectionPrediction)):
+        return (dm.Bbox(x0, y0, x1 - x0, y1 - y0, int(label_id),
+                attributes={'score': float(score)})
+            for label, score, x0, y0, x1, y1 in zip(pred.labels, pred.scores,
+                pred.x_mins, pred.y_mins, pred.x_maxs, pred.y_maxs)
+        )
+    elif isinstance(pred, ac.DepthEstimationPrediction):
+        return (dm.Mask(pred.depth_map), ) # 2d floating point mask
+    # elif isinstance(pred, ac.HitRatioPrediction):
+    #     -
+    elif isinstance(pred, ac.ImageInpaintingPrediction):
+        return (dm.Mask(pred.value), ) # an image
+    # elif isinstance(pred, ac.MultiLabelRecognitionPrediction):
+    #     -
+    # elif isinstance(pred, ac.MachineTranslationPrediction):
+    #     -
+    # elif isinstance(pred, ac.QuestionAnsweringPrediction):
+    #     -
+    # elif isinstance(pred, ac.PoseEstimation3dPrediction):
+    #     -
+    # elif isinstance(pred, ac.PoseEstimationPrediction):
+    #     -
+    # elif isinstance(pred, ac.RegressionPrediction):
+    #     -
+    else:
+        raise NotImplementedError("Can't convert %s" % type(pred))
+
+
+
+

datumaro/datumaro/plugins/accuracy_checker_plugin/launcher.py

@@ -0,0 +1,37 @@
+
+# Copyright (C) 2020 Intel Corporation
+#
+# SPDX-License-Identifier: MIT
+
+import os.path as osp
+import yaml
+
+from datumaro.components.cli_plugin import CliPlugin
+from datumaro.components.launcher import Launcher
+
+from .details.ac import GenericAcLauncher as _GenericAcLauncher
+
+
+class AcLauncher(Launcher, CliPlugin):
+    """
+    Generic model launcher with Accuracy Checker backend.
+    """
+
+    @classmethod
+    def build_cmdline_parser(cls, **kwargs):
+        parser = super().build_cmdline_parser(**kwargs)
+        parser.add_argument('-c', '--config', type=osp.abspath, required=True,
+            help="Path to the launcher configuration file (.yml)")
+        return parser
+
+    def __init__(self, config, model_dir=None):
+        model_dir = model_dir or ''
+        with open(osp.join(model_dir, config), 'r') as f:
+            config = yaml.safe_load(f)
+        self._launcher = _GenericAcLauncher.from_config(config)
+
+    def launch(self, inputs):
+        return self._launcher.launch(inputs)
+
+    def categories(self):
+        return self._launcher.categories()

datumaro/datumaro/util/annotation_tools.py

@@ -5,6 +5,8 @@
 
 from itertools import groupby
 
+import numpy as np
+
 
 def find_instances(instance_anns):
     instance_anns = sorted(instance_anns, key=lambda a: a.group)
@@ -26,4 +28,7 @@ def compute_bbox(annotations):
     y0 = min((b[1] for b in boxes), default=0)
     x1 = max((b[0] + b[2] for b in boxes), default=0)
     y1 = max((b[1] + b[3] for b in boxes), default=0)
-    return [x0, y0, x1 - x0, y1 - y0]
+    return [x0, y0, x1 - x0, y1 - y0]
+
+def softmax(x):
+    return np.exp(x) / sum(np.exp(x))

datumaro/tests/assets/pytorch_launcher/model_config.yml

@@ -0,0 +1,37 @@
+launcher:
+  framework: pytorch
+  module: samplenet.SampLeNet
+  python_path: '.'
+  checkpoint: 'samplenet.pth'
+
+# launcher returns raw result, so it should be converted
+# to an appropriate representation with adapter
+adapter:
+  type: classification
+  labels:
+    - label1
+    - label2
+    - label3
+    - label4
+    - label5
+    - label6
+    - label7
+    - label8
+    - label9
+    - label10
+
+# list of preprocessing, applied to each image during validation
+# order of entries matters
+preprocessing:
+  # resize input image to topology input size
+  # you may specify size to which image should be resized
+  # via dst_width, dst_height fields
+  - type: resize
+    size: 32
+  # topology is trained on RGB images, but Datumaro reads in BGR
+  # so it must be converted to RGB
+  - type: bgr_to_rgb
+  # dataset mean and standard deviation
+  - type: normalization
+    mean: (125.307, 122.961, 113.8575)
+    std: (51.5865, 50.847, 51.255)

datumaro/tests/assets/pytorch_launcher/samplenet.py

@@ -0,0 +1,38 @@
+"""
+Copyright (c) 2019 Intel Corporation
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+      http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class SampLeNet(nn.Module):
+    def __init__(self):
+        super(SampLeNet, self).__init__()
+        self.conv1 = nn.Conv2d(3, 6, 5)
+        self.pool = nn.MaxPool2d(2, 2)
+        self.conv2 = nn.Conv2d(6, 16, 5)
+        self.fc1 = nn.Linear(16 * 5 * 5, 120)
+        self.fc2 = nn.Linear(120, 84)
+        self.fc3 = nn.Linear(84, 10)
+
+    def forward(self, x):
+        x = self.pool(F.relu(self.conv1(x)))
+        x = self.pool(F.relu(self.conv2(x)))
+        x = x.view(-1, 16 * 5 * 5)
+        x = F.relu(self.fc1(x))
+        x = F.relu(self.fc2(x))
+        x = self.fc3(x)
+        return x