akash13singh/tensorflow_transfer_learning

Transfer Learning using trained Inception v3 model.

TensorFlow Transfer Learning on CIFAR-10

Trains a softmax regression model on CIFAR-10 using CNN pool_3 weights from inception-v3. Forked from https://github.com/sthomp/tensorflow_transfer_cifar10
Made changes to do training in batches.

This is the code that supplements the original blog post

Setup

1. Download and extract CIFAR-10 dataset to resources/datasets/
2. Download and extract the Inception v3 model to resources/

Using Pretrained Model

1. First generate and save bottleneck features.
   To generate the bottleneck features set the flag DO_SERIALIZATION in transfer_cifar10_softmax.py to True:
   

#flag to generate and save bottleneck features
DO_SERIALIZATION = False

2. Run script transfer_cifar10_softmax.py This would run the input images through the trained Inception V3 network and save the output of the pool_3 layer.

3. The previous step would generate .npy files in the project root directory. The files will be

   • X_train.npy
   • X_test.npy
   • Y_train.npy
   • Y_test.npy These files store the oputput of layer pool3 of the Inception model and the corresponding labels. Training for the new task will use these as inputs.

Training for new task

1. Add desired layers by modifying function add_final_training_ops() in transfer_cifar10_softmax.py.

2. Set flag DO_SERIALIZATION to False.

3. run script transfer_cifar10_softmax.py.

Note

function load_CIFAR10() in data_utils.py can be modified to return only 1000 training images and 100 test images from the CIFAR-10 data. This will speed up the generation of bottlenecks for a quick demo.

Todo Tasks

• Fix to train in batches
• Code Cleanup
• Create methods and classes
• Fix tsne plotting code