def transfer_learn():
test_actions = [
'CleanAndJerk',
'MoppingFloor',
'FrontCrawl',
# 'Surfing',
'Bowling',
'SoccerPenalty',
'SumoWrestling',
'Shotput',
'PlayingSitar',
'FloorGymnastics',
# 'Typing',
'JumpingJack',
'ShavingBeard',
'FrisbeeCatch',
'WritingOnBoard',
'JavelinThrow',
'Fencing',
# 'FieldHockeyPenalty',
# 'BaseballPitch',
'CuttingInKitchen',
# 'Kayaking',
]
train_dataset, test_dataset = get_traditional_dataset(
base_address=BASE_ADDRESS,
class_sample_size=CLASS_SAMPLE_SIZE,
test_actions=test_actions)
with tf.variable_scope('train_data'):
input_data_ph = tf.placeholder(dtype=tf.float32,
shape=[None, 16, 112, 112, 3])
input_labels_ph = tf.placeholder(dtype=tf.float32, shape=[None, 85])
tf.summary.image('train', input_data_ph[:, 0, :, :, :], max_outputs=25)
with tf.variable_scope('validation_data'):
val_data_ph = tf.placeholder(dtype=tf.float32,
shape=[None, 16, 112, 112, 3])
val_labels_ph = tf.placeholder(dtype=tf.float32, shape=[None, 85])
tf.summary.image('validation',
val_data_ph[:, 0, :, :, :],
max_outputs=25)
gpu_devices = ['/gpu:{}'.format(gpu_id) for gpu_id in range(NUM_GPUS)]
maml = ModelAgnosticMetaLearning(C3DNetwork,
input_data_ph,
input_labels_ph,
val_data_ph,
val_labels_ph,
log_dir=LOG_DIR,
gpu_devices=gpu_devices,
meta_learn_rate=0.00001,
learning_rate=0.001,
train=TRAIN,
log_device_placement=False)
maml.load_model(path='MAML/sports1m_pretrained.model',
load_last_layer=False)
for it in range(TRANSFER_LEARNING_ITERATIONS):
print(it)
data = train_dataset.next_batch(num_classes=85, real_labels=True)
batch_test_data, batch_test_labels = data['train']
batch_test_val_data, batch_test_val_labels = data['validation']
batch_split_size = int(NUM_CLASSES / BATCH_SPLIT_NUM)
for batch_split_index in range(BATCH_SPLIT_NUM):
start = batch_split_index * batch_split_size
end = batch_split_index * batch_split_size + batch_split_size
test_data = batch_test_data[start:end, :, :, :, :]
test_labels = batch_test_labels[start:end, :]
test_val_data = batch_test_val_data[start:end, :, :, :, :]
test_val_labels = batch_test_val_labels[start:end, :]
if it % 50 == 0:
if batch_split_index == 0:
val_accs = []
if it % 100 == 0:
maml.save_model(
'saved_models/transfer_learning_85/model', step=it)
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
merged_summary = maml.sess.run(maml.merged,
feed_dict={
input_data_ph: test_data,
input_labels_ph:
test_labels,
val_data_ph: test_val_data,
val_labels_ph:
test_val_labels,
},
options=run_options,
run_metadata=run_metadata)
maml.file_writer.add_summary(merged_summary,
global_step=it +
batch_split_index)
outputs = maml.sess.run(maml.inner_model_out,
feed_dict={
maml.input_data: test_val_data,
maml.input_labels: test_val_labels,
})
val_accs.append(print_accuracy(outputs, test_val_labels))\
if batch_split_index == BATCH_SPLIT_NUM - 1:
print('iteration: {}'.format(it))
print('Validation accuracy on all batches: ')
print(np.mean(val_accs))
maml.sess.run(maml.inner_train_ops,
feed_dict={
input_data_ph: test_data,
input_labels_ph: test_labels,
})
maml.save_model('saved_models/transfer_learning_85/model', step=it)
def transfer_learn():
test_actions = [
'CleanAndJerk',
'MoppingFloor',
'FrontCrawl',
'Surfing', # 0
'Bowling',
'SoccerPenalty',
'SumoWrestling',
'Shotput',
'PlayingSitar',
'FloorGymnastics',
'Typing', # 1
'JumpingJack',
'ShavingBeard',
'FrisbeeCatch',
'WritingOnBoard',
'JavelinThrow',
'Fencing',
'FieldHockeyPenalty', # 3
'BaseballPitch', # 4
'CuttingInKitchen',
'Kayaking', # 2
]
train_dataset, test_dataset = get_traditional_dataset(
base_address=BASE_ADDRESS,
class_sample_size=CLASS_SAMPLE_SIZE,
test_actions=test_actions)
with tf.variable_scope('train_data'):
input_data_ph = tf.placeholder(dtype=tf.float32,
shape=[None, 16, 112, 112, 3])
input_labels_ph = tf.placeholder(dtype=tf.float32, shape=[None, 20])
tf.summary.image('train', input_data_ph[:, 0, :, :, :], max_outputs=25)
with tf.variable_scope('validation_data'):
val_data_ph = tf.placeholder(dtype=tf.float32,
shape=[None, 16, 112, 112, 3])
val_labels_ph = tf.placeholder(dtype=tf.float32, shape=[None, 20])
tf.summary.image('validation',
val_data_ph[:, 0, :, :, :],
max_outputs=25)
gpu_devices = ['/gpu:{}'.format(gpu_id) for gpu_id in range(NUM_GPUS)]
maml = ModelAgnosticMetaLearning(C3DNetwork,
input_data_ph,
input_labels_ph,
val_data_ph,
val_labels_ph,
log_dir=LOG_DIR,
gpu_devices=gpu_devices,
meta_learn_rate=0.00001,
learning_rate=0.001,
train=TRAIN,
log_device_placement=False)
maml.load_model(path='saved_models/transfer_learning/model-400',
load_last_layer=False)
data = test_dataset.next_batch(num_classes=20)
test_data, test_labels = data['train']
test_val_data, test_val_labels = data['validation']
print(test_dataset.actions)
for it in range(TRANSFER_LEARNING_ITERATIONS):
print(it)
if it % 50 == 0:
if it % 100 == 0:
maml.save_model(
'saved_models/transfer_learning_to_20_classes/model',
step=it)
merged_summary = maml.sess.run(maml.merged,
feed_dict={
input_data_ph: test_data,
input_labels_ph: test_labels,
val_data_ph: test_val_data,
val_labels_ph: test_val_labels,
})
maml.file_writer.add_summary(merged_summary, global_step=it)
outputs = maml.sess.run(maml.inner_model_out,
feed_dict={
maml.input_data: test_val_data,
maml.input_labels: test_val_labels,
})
val_acc = print_accuracy(outputs, test_val_labels)
print('iteration: {}'.format(it))
print('Validation accuracy on all batches: ')
print(val_acc)
maml.sess.run(maml.inner_train_ops,
feed_dict={
input_data_ph: test_data,
input_labels_ph: test_labels,
})
def train_maml():
with tf.variable_scope('train_data'):
input_data_ph = tf.placeholder(dtype=tf.float32, shape=[None, 16, 112, 112, 3])
input_labels_ph = tf.placeholder(dtype=tf.float32, shape=[None, NUM_CLASSES])
tf.summary.image('train', input_data_ph[:, 0, :, :, :], max_outputs=NUM_CLASSES)
with tf.variable_scope('validation_data'):
val_data_ph = tf.placeholder(dtype=tf.float32, shape=[None, 16, 112, 112, 3])
val_labels_ph = tf.placeholder(dtype=tf.float32, shape=[None, NUM_CLASSES])
tf.summary.image('validation', val_data_ph[:, 0, :, :, :], max_outputs=NUM_CLASSES)
gpu_devices = ['/gpu:{}'.format(gpu_id) for gpu_id in range(NUM_GPUS)]
maml = ModelAgnosticMetaLearning(
C3DNetwork,
input_data_ph,
input_labels_ph,
val_data_ph,
val_labels_ph,
log_dir=LOG_DIR,
gpu_devices=gpu_devices,
meta_learn_rate=0.0001,
learning_rate=0.001,
train=TRAIN,
log_device_placement=False,
num_classes=101
)
if TRAIN:
train_dataset, test_dataset = get_traditional_dataset(
num_train_actions=600,
base_address=BASE_ADDRESS,
class_sample_size=CLASS_SAMPLE_SIZE,
)
maml.load_model(path='MAML/sports1m_pretrained.model', load_last_layer=False)
print('start meta training.')
it = 0
for it in range(10001):
train_dataset.sample_k_samples()
data = train_dataset.next_batch(num_classes=NUM_CLASSES)
tr_data, tr_labels = data['train']
val_data, val_labels = data['validation']
if it % 50 == 0:
merged_summary = maml.sess.run(maml.merged, feed_dict={
input_data_ph: tr_data,
input_labels_ph: tr_labels,
val_data_ph: val_data,
val_labels_ph: val_labels,
})
maml.file_writer.add_summary(merged_summary, global_step=it)
print(it)
maml.sess.run(maml.train_op, feed_dict={
input_data_ph: tr_data,
input_labels_ph: tr_labels,
val_data_ph: val_data,
val_labels_ph: val_labels,
})
if it % 200 == 0:
maml.save_model(path='saved_models/kinetics400/model', step=it)
if it != 0:
maml.save_model(path='saved_models/kinetics400/model', step=it)
else:
# test_actions = [
# 'CleanAndJerk',
# 'MoppingFloor',
# 'FrontCrawl',
# 'Surfing',
# 'Bowling',
# 'SoccerPenalty',
# 'SumoWrestling',
# 'Shotput',
# 'PlayingSitar',
# 'FloorGymnastics',
# 'Typing',
# 'JumpingJack',
# 'ShavingBeard',
# 'FrisbeeCatch',
# 'WritingOnBoard',
# 'JavelinThrow',
# 'Fencing',
# 'FieldHockeyPenalty',
# 'BaseballPitch',
# 'CuttingInKitchen',
# 'Kayaking',
# ]
test_actions = [
'ApplyEyeMakeup',
'Archery',
'BabyCrawling',
'BandMarching',
'Bowling',
'Basketball',
'Biking',
'Billiards',
'BlowDryHair',
'FloorGymnastics',
'Typing',
'CliffDiving',
'ShavingBeard',
'FrisbeeCatch',
'WritingOnBoard',
'JavelinThrow',
'Fencing',
'FieldHockeyPenalty',
'PlayingPiano',
'CuttingInKitchen',
]
train_dataset, test_dataset = get_traditional_dataset(
base_address='/home/siavash/UCF-101/',
class_sample_size=CLASS_SAMPLE_SIZE,
test_actions=test_actions
)
maml.load_model(path='saved_models/kinetics400/model-8000')
print('Start testing the network')
data = test_dataset.next_batch(num_classes=NUM_CLASSES)
test_data, test_labels = data['train']
test_val_data, test_val_labels = data['validation']
print(test_dataset.actions)
for it in range(5):
maml.sess.run(maml.inner_train_ops, feed_dict={
input_data_ph: test_data,
input_labels_ph: test_labels,
})
if it % 1 == 0:
merged_summary = maml.sess.run(maml.merged, feed_dict={
input_data_ph: test_data,
input_labels_ph: test_labels,
val_data_ph: test_val_data,
val_labels_ph: test_val_labels,
})
maml.file_writer.add_summary(merged_summary, global_step=it)
print('gradient step: ')
print(it)
outputs = maml.sess.run(maml.inner_model_out, feed_dict={
maml.input_data: test_val_data,
maml.input_labels: test_val_labels,
})
print_accuracy(outputs, test_val_labels)
maml.save_model('saved_models/ucf101-fit/model-kinetics-trained', step=it)
def transfer_learn():
train_dataset, test_dataset = get_traditional_dataset(
base_address=BASE_ADDRESS,
class_sample_size=CLASS_SAMPLE_SIZE,
num_train_actions=400)
with tf.variable_scope('train_data'):
input_data_ph = tf.placeholder(dtype=tf.float32,
shape=[None, 16, 112, 112, 3])
input_labels_ph = tf.placeholder(dtype=tf.float32, shape=[None, 400])
tf.summary.image('train', input_data_ph[:, 0, :, :, :], max_outputs=25)
with tf.variable_scope('validation_data'):
val_data_ph = tf.placeholder(dtype=tf.float32,
shape=[None, 16, 112, 112, 3])
val_labels_ph = tf.placeholder(dtype=tf.float32, shape=[None, 400])
tf.summary.image('validation',
val_data_ph[:, 0, :, :, :],
max_outputs=25)
gpu_devices = ['/gpu:{}'.format(gpu_id) for gpu_id in range(NUM_GPUS)]
maml = ModelAgnosticMetaLearning(C3DNetwork,
input_data_ph,
input_labels_ph,
val_data_ph,
val_labels_ph,
log_dir=LOG_DIR,
gpu_devices=gpu_devices,
meta_learn_rate=0.00001,
learning_rate=0.001,
train=TRAIN,
log_device_placement=False)
maml.load_model(path='MAML/sports1m_pretrained.model',
load_last_layer=False)
for it in range(TRANSFER_LEARNING_ITERATIONS):
print(it)
data, labels = train_dataset.next_simple_batch(batch_size=BATCH_SIZE)
if it % 100 == 0:
maml.save_model('saved_models/transfer_learning/model', step=it)
if it % 20 == 0:
merged_summary = maml.sess.run(maml.merged,
feed_dict={
input_data_ph: data,
input_labels_ph: labels,
val_data_ph: data,
val_labels_ph: labels,
})
maml.file_writer.add_summary(merged_summary, global_step=it)
outputs = maml.sess.run(maml.inner_model_out,
feed_dict={
maml.input_data: data,
maml.input_labels: labels,
})
print_accuracy(outputs, labels)
maml.sess.run(maml.inner_train_ops,
feed_dict={
input_data_ph: data,
input_labels_ph: labels,
})