TF Multi Classifier

What is TF Multi Classifier?

TF Multi Classifier App uses Tensorflow lite to classify camera frames in real time.

You can download from Google Play here:

• TF Multi Classifier FREE

Benchmark

Mean inference time:

Model	Phone	Android	CPU 1TH	CPU 2TH	NNAPI	GPU
mobilenet_v1_224 (FLOAT)	Pixel 2XL	9	165ms	115ms	110ms	40ms
mobilenet_v1_224 (QUANTIZED)	Pixel 2XL	9	85ms	50ms	80ms	N/A
mobilenet_v1_224 (FLOAT)	Nexus 5	6.0	180ms	200ms	N/A	N/A
mobilenet_v1_224 (QUANTIZED)	Nexus 5	6.0	120ms	100ms	N/A	N/A
mobilenet_v1_224 (FLOAT)	LG G6	8.0	260ms	220ms	N/A	50ms
mobilenet_v1_224 (QUANTIZED)	LG G6	8.0	220ms	140ms	N/A	N/A

Features

• Classify with Tensorflow lite
• Custom dataset
• GPU supported (requires OpenGL ES 3.1 or higher)

Download Dataset

You can download the following datasets from HERE:

• Dog Breed (created with Transfer learning)
• Dog Vs Cat (created with Transfer learning)
• Google Datasets: (converted from google pb dataset)
  • Bird
  • Insect
  • Plant
  • Seefood

Transfer learning

An easy and fast way for train a model is to use "Transfer learning".

"Transfer learning" provides the opportunity to adapt a pre-trained model (a model that has been already trained) to new classes of data. This concept is summarized in three steps:

• download the pre-trained model
• adds a new final layer
• train that new layer

Step 1: Install tensorflow (following the official tutorial HERE)

Step 2: Clone this Git repository:

git clone https://github.com/googlecodelabs/tensorflow-for-poets-2

and "cd" into the following directory:

cd tensorflow-for-poets-2

Step 3: Before you start any training, you'll need a set of images to teach the model about the new classes you want to recognize. Download the photos:

Flower photos by google:

curl http://download.tensorflow.org/example_images/flower_photos.tgz \
    | tar xz -C tf_files

or download your images from google images in a structure like this:

tf_files/name_class/
      class1/[images...]
      class2/[images...]
      class3/[images...]

PS: set this folder in "--image_dir"

Step 4: Run the training with this command:

python -m scripts.retrain \
  --bottleneck_dir=tf_files/bottlenecks \
  --model_dir=tf_files/models/ \
  --summaries_dir=tf_files/training_summaries/mobilenet_1.0_224 \
  --output_graph=tf_files/retrained_graph_mobilenet_1.0_224.pb \
  --output_labels=tf_files/retrained_labels.txt \
  --architecture=mobilenet_1.0_224 \
  --image_dir=tf_files/flower_photos

This script downloads the pre-trained model and trains that layer on the flower photos you've downloaded. In this script I used "mobilenet_1.0_224" a MobileNet (small efficient convolutional neural network) with image resolution 224px and size of model 1.0.

You can use other configurations, for example:

• mobilenet_A_B with A: 128, 160, 192, or 224px and B: 1.0, 0.75, 0.50, or 0.25.
• inception_v3

Step 5: Now we have these two file:

• retrained_labels.txt: text file containing labels.
• retrained_graph_mobilenet_1.0_224.pb: contains a version of the selected network with a final layer retrained on your categories.

Step 6: Convert the model ".pb" in ".tflite" for tensorflow lite.

I tryed to use tflite_convert in windows without positive results. So I installed ubuntu in a virtual machine, installed tensorflow and now the conversion works.

tflite_convert \
  --output_file=tf_files/retrained_graph_lite.tflite \
  --graph_def_file=tf_files/retrained_graph.pb \
  --input_arrays=Mul \
  --output_arrays=final_result

Step 7: Now you can load this two file in my Android App:

• retrained_graph_lite.tflite
• retrained_labels.txt

License

Copyright (C) 2018 edodm85.
Licensed under the MIT license.
(See the LICENSE file for the whole license text.)