• This example shows how to train an image classifier based on TensorFlow Hub module that computes image feature vectors.

• Since the size of the dataset was small so training the model from scratch might not give good accuracy scores so the concept of Transfer Learning as used to train the model.

• By default, it uses the feature vectors computed by Inception V3 trained on ImageNet. The top layer receives as input a 2048-dimensional vector (assuming Inception V3) for each image. We train a softmax layer on top of this representation.

• Since the softmax layer contains 2 labels (defected & healthy), this corresponds to learning 2 + 2048*2 model parameters for the biases and weights.

• The subfolder names under the ./data directory are labels to classify, since we have two labels defected & healthy so the data is placed under these directories.

  • Classifier can be trained by executing the below command.

    python ./scripts/retrain.py --image_dir ./data

  • There is also option to tune the specific hyperparameters for training the model, below command can be executed to train the model by tuning hyperparameters.

    python ./scripts/retrain.py -h

• The training can be visualized using the tensorboard by executing the below command.

  tensorboard --logdir /tmp/retrain_logs

• Alredy trained model present in ./tmp/inception_v3/ directory can be used for testing on new images. Execute the below command.

  python ./scripts/label_image.py \
  --graph=./tmp/inception_v3/output_graph.pb \
  --labels=./tmp/inception_v3/output_labels.txt \
  --input_layer=Placeholder \
  --output_layer=final_result \
  --image={Path of the Image file}

  The above trained model gave accuracy of 85.7% with following hyperparameters:

  • no.of steps: 4000
  • batch size: 100
  • learning rate: 0.01

• The above model was trainedusing Inception Network it can also be trained using other pretrained network architectures e.g. NASNet-A (https://tfhub.dev/google/imagenet/nasnet_large/feature_vector/1)

https://www.tensorflow.org/hub/tutorials/image_retraining

https://ai.googleblog.com/2017/11/automl-for-large-scale-image.html

https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#3

https://arxiv.org/abs/1310.1531

https://towardsdatascience.com/tensorflow-on-mobile-tutorial-1-744703297267