cesar0205 / tensorflow_transfer_learning

Transfer learning with TensorFlow

Transfer learning with TensorFlow

Transfer learning with TensorFlow

Transfer learning using the Inception Resnet V2 model. The model is available at https://github.com/tensorflow/models/tree/master/research/slim to download.

The flowers dataset was used to train and test the model. The dataset is available at http://download.tensorflow.org/example_images/flower_photos.tgz

I experimented with some components that are used for training large datasets in Tensorflow.

• Tensorboard. Its a visualization tool that helps debug the tensorflow model. We can store and create graphs related to metrics we want to show.

• TFRecordDataset class to read from TFRecord files. These files store data in a format optimized to be fast read by tensorflow. Additionally datasets can be chained to form pipelines and preprocess data.

• Tensorflow iterators. They iterate through the datasets to feed the model. In this particular case I used the feedable tensor to maintain the state of the training dataset as I want to evaluate the model each x minibatches.

• MonitoredTrainingSession. Manages the training cicle of a model by: -Updating the global step variable -Opening and closing the session. -Managing hooks and scaffolds. -Creating model checkpoints and restarts training from the last checkpoint. -Working with other clusters in a distributed environment(this feature was not used).

• Hooks. These are classes with functions that are called when at each training step. Tensorflow includes LoogingTensorHook, SummarySaverHook, CheckpointSaverHook. Aditionally we can create our custom hooks. For this model I created an EarlyStopping hook to stop training when the validation error hasn't go down after certain number of iterations.

• Exponential decay. tf.train.exponential_decay is a function that applies exponential decay to the optimizer chosen to train the model. Each decay_steps the learning rate is updated by lr = lr*decay_rate

I got 94.86% of accuracy in final test set. I also printed some images to see what kind of errors the model made.