TensorFlow-fine-tune-VGG16-Alexnet-models

In this project, finetune process is clearly demonstrated with TensorFlow.

Download the pre-trained models

alexnet.npy and vgg16.npy.

Download link: https://drive.google.com/drive/folders/1nDvd3HwPIRlPTn8UJBT7jLNuiwPnisWi?usp=sharing

Generate image list for training and testing

Generate image list (.txt) from a folder, with format 'path/to/img/ label', each sub-folder contains all images belong to one subject. The name of the sub-foler is the label, start from 0 to N.

python3 generate_imglist

Generate train.txt and test.txt from image_list.txt

python3 generate_train_test_list

Parameters

Choose model with Line 50 in 'finetune.py'. To fine-tune other models, add the defination of model to 'model.py'.

pred = Model.vgg16(x, keep_var, n_classes) # Model.alexnet(x, keep_var, n_classes) 

Training parameters in 'finetune.py'

# Learning params  
  learning_rate_init = 0.001   
  decay_steps = 10000
  decay_rate = 0.5

  # Train and dispaly params
  training_iters = 60000
  batch_size = 50         
  display_step = 20
  test_step = 1000
  save_step = 1000

  # Network params
  n_classes = 10575
  keep_rate = 0.5  

To select weights to be restored from the pre-trained model, modify Line 90 in 'finetune.py'

load_with_skip(weight_file, sess, ['fc8'])  # Skip weights from fc8

To select which layer to be fine-tuned, modify 'model.py' with trainable=True/False.

Start the finetune process

python3 finetune.py train.txt test.txt vgg16.npy

Visualize the training/fine-tuning process

tensorboard --logdir path/to/log/file