def main(_):
# Choose between modes - train/finetune/evaluate
if FLAGS.mode == 'train':
summaries_dir = FLAGS.summaries_train_dir
model_dir = FLAGS.model_train_dir
print("\033[95m Mode opted: Training from scratch... \033[0m")
elif FLAGS.mode == 'finetune':
summaries_dir = FLAGS.summaries_finetune_dir
model_dir = FLAGS.model_finetune_dir
print(
"\033[95m Mode opted: Finetuning the pre-trained model ... \033[0m"
)
elif FLAGS.mode == 'evaluation':
print(
"\033[95m Mode opted: Evaluating the model on test data set ... \033[0m"
)
_evaluation_operation()
return
else:
print(
" \033[91m Provide Either train or finetune.... Other Implementations pending//// \033[0m"
)
return
#Clearing summaries and model directories if exists
if tf.gfile.Exists(summaries_dir):
tf.gfile.DeleteRecursively(summaries_dir)
tf.gfile.MakeDirs(summaries_dir)
if tf.gfile.Exists(model_dir):
tf.gfile.DeleteRecursively(model_dir)
tf.gfile.MakeDirs(model_dir)
checkpoint_path = os.path.join(model_dir, 'model.ckpt')
sess = tf.Session()
config = Config()
print("\033[94m Loading the model... \033[0m")
m = model(is_training=True, config=config)
print("\033[94m Preparing the input stream for training... \033[0m")
input_stream = input.input(FLAGS.data_path, config)
print("\033[94m Initializing the network... \033[0m")
merged = tf.summary.merge_all()
saver = tf.train.Saver(tf.global_variables())
train_writer = tf.summary.FileWriter(summaries_dir + '/train', sess.graph)
test_writer = tf.summary.FileWriter(summaries_dir + '/test', sess.graph)
init = tf.global_variables_initializer()
sess.run(init)
global_step = 0
if FLAGS.mode == 'finetune':
ckpt = tf.train.get_checkpoint_state(FLAGS.model_train_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = int(
ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1])
else:
print(
"\033[91m Opted for finetune mode, but no pre-trained model found \033[0m"
)
return
print("\033[94m Training begins... \033[0m")
#state = sess.run(m.initial_state)
for i in range(FLAGS.iterations):
input_feature, input_targets = input_stream.next_batch(sess,
flag='train')
feed_dict = {
m.input_feature: input_feature,
m.input_targets: input_targets
}
train_summary, _ = sess.run([merged, m.train_step],
feed_dict=feed_dict)
train_writer.add_summary(train_summary, i)
train_accuracy, loss_value, learning_rate = sess.run(
[m.evaluation_step, m.loss_mean, m.learning_rate],
feed_dict=feed_dict)
input_feature, input_targets = input_stream.next_batch(sess,
flag='test')
feed_dict = {
m.input_feature: input_feature,
m.input_targets: input_targets
}
test_summary, test_accuracy = sess.run([merged, m.evaluation_step],
feed_dict=feed_dict)
test_writer.add_summary(test_summary, i)
print(
'Step %d: Learning Rate = %f Train Loss = %f Train Accuracy = %f Test_Accuracy = %f'
% (i, learning_rate, loss_value, train_accuracy, test_accuracy))
if (i % FLAGS.saving_iter) == 0:
saver.save(sess, checkpoint_path, global_step=i + global_step)
if (((i % FLAGS.learning_rate_iter) == 0) & (i != 0)):
m.assign_learning_rate(sess,
FLAGS.learning_rate_decay * learning_rate)
def _evaluation_operation():
if FLAGS.eval_mode == 'train':
model_dir = FLAGS.model_train_dir
elif FLAGS.eval_mode == 'finetune':
model_dir = FLAGS.model_finetune_dir
else:
print(
"\033[91m eval_mode not selected, chosse --eval_mode= train or finetune \033[0m "
)
return
graph, bottleneck_tensor, jpeg_data_tensor, resized_image_tensor = (
create_inception_graph())
sess = tf.Session()
config = Config()
config.batch_size = 1
print("\033[94m Loading the model... \033[0m")
m = model(is_training=False, config=config)
print("\033[94m Preparing the input stream for testing... \033[0m")
input_stream = input.input(FLAGS.data_path,
config,
image_data_tensor=jpeg_data_tensor,
bottleneck_tensor=bottleneck_tensor)
saver = tf.train.Saver(tf.all_variables())
ckpt = tf.train.get_checkpoint_state(model_dir)
if ckpt and ckpt.model_checkpoint_path:
print("\033[94m Loading the pre-trained weights of model... \033[0m")
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print(
"\033[91m Opted for evaluation mode, but no pre-trained/pre-finetuned model found \033[0m"
)
return
dacc_frame = np.zeros([config.num_class, config.num_class])
dacc_cumul = np.zeros([config.num_class, config.num_class])
seqlen_frame = np.zeros([config.num_class])
seqlen_cumul = np.zeros([config.num_class])
total_videos = input_stream.test_total
for i in range(input_stream.test_total):
print(" \033[94m Currently on : %d / %d \033[0m " % (i, total_videos))
input_feature, sequence_len, input_targets = input_stream.next_batch(
sess, flag='eval')
feed_dict = {
m.input_feature: input_feature,
m.input_targets: input_targets,
m.sequence_len: sequence_len
}
final_result = sess.run([m.final_result], feed_dict=feed_dict)
label = input_targets[0, 0]
temp_list = np.zeros([config.num_class])
for j in range(sequence_len[0]):
temp_list[final_result[0][j]] += 1
dacc_frame[label, :] += temp_list
seqlen_frame[label] += sequence_len[0]
pool_result = np.argmax(temp_list)
dacc_cumul[label, pool_result] += 1
seqlen_cumul[label] += 1
acc_frame = dacc_frame.trace() / sum(seqlen_frame)
acc_cumul = dacc_cumul.trace() / sum(seqlen_cumul)
print("\033[93m Accuracy: Frame = %f Cumulative = %f \033[0m " %
(acc_frame, acc_cumul))
for i in range(config.num_class):
dacc_frame[i, :] /= seqlen_frame[i]
dacc_cumul[i, :] /= seqlen_cumul[i]
np.savetxt("delta_acc_frame.csv", np.asarray(dacc_frame), delimiter=",")
np.savetxt("delta_acc_cumul.csv", np.asarray(dacc_cumul), delimiter=",")