main.py

"""A binary to train BiLSTM on the KTH data set.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import video_train
import tensorflow as tf

from data.kth_data import KTHData
from data.lca_data import LCAData

tf.app.flags.DEFINE_string("data_path", None,
  "Where the training/validation data is stored.")
tf.app.flags.DEFINE_string("save_path", 'result',
  "Model output directory.")
tf.app.flags.DEFINE_string("dataset", 'KTH',
  "Select the dataset, default is KTH datasetk, choice between (KTH, LCA)")
tf.app.flags.DEFINE_string("image_height", 120,
  "Image height")
tf.app.flags.DEFINE_string("image_width", 160,
  "Image width")
tf.app.flags.DEFINE_string("channels", 1, 
  "Image width")

FLAGS = tf.app.flags.FLAGS

config = {
  'epoch'                      : 6,
  'lr_decay'                   : 0.8,
  'keep_prob'                  : 0.8,
  'init_scale'                 : 0.1, # weight initialization value (-init_scale, init_scale)
  'batch_size'                 : 20,
  'learning_rate'              : 0.5,
  'max_grad_norm'              : 5,
  'decay_begin_epoch'          : 2,
  'examples_per_shard'         : 23,
  'input_queue_memory_factor'  : 2,
  'num_layers'                 : 2,
  # num_steps: This value must be the same as the sequence_length value,
  #  inside the data/convert_to_records.py when you generate the data.,
  'num_steps'                  : 16,
  'hidden_size'                : 200,
}

def main(_):
  if not FLAGS.data_path:
    raise ValueError("Must set --data_path to KTH data directory")

  # Select the dataset
  train_data = None
  if FLAGS.dataset == 'KTH':
    train_data = KTHData('train')
  elif FLAGS.dataset == 'LCA':
    train_data = LCAData('train')
  
  assert train_data
  assert train_data.data_files()
  config['num_classes'] = train_data.num_classes()

  # Start training
  video_train.train(config, train_data)


if __name__ == '__main__':
  tf.app.run()