def train(model, optimizer, save_dir, training_step):
loss_train_log, loss_val_log, acc_log = [], [], []
acc_best = 0.0
for step in range(training_step):
train_images, train_labels, _, _ = read_clip_and_label(filename='list/train.list')
X, y = tf.convert_to_tensor(train_images/255., tf.float32), tf.convert_to_tensor(train_labels, tf.int64) # (10, 16, 112, 112, 3) (10,)
with tf.GradientTape() as tape:
loss, loss_ent, loss_reg = model.loss_fn(X, y)
acc_train = model.acc_fn(X, y)
print(loss.numpy(), loss_ent.numpy(), loss_reg.numpy(), acc_train*100, "%")
loss_train_log.append([loss.numpy(), loss_ent.numpy(), loss_reg.numpy()])
# 使用 loss 训练
gradients = tape.gradient(loss, model.trainable_variables)
# gradients, n = tf.clip_by_global_norm(gradients, 1.)
# print("grad norm:", n.numpy())
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
# validation
if step % 50 == 0:
acc_tmp, loss_tmp, loss_ent_tmp, loss_reg_tmp = [], [], [], []
for i in range(5):
val_images, val_labels, _, _ = read_clip_and_label(filename='list/test.list')
X_val, y_val = tf.convert_to_tensor(val_images/255., tf.float32), tf.convert_to_tensor(val_labels, tf.int64)
loss_val, loss_ent_val, loss_reg_val = model.loss_fn(X_val, y_val)
acc = model.acc_fn(X_val, y_val)
acc_tmp.append(acc)
loss_tmp.append(loss_val.numpy())
loss_ent_tmp.append(loss_ent_val.numpy())
loss_reg_tmp.append(loss_reg_val.numpy())
print("\nTEST------------")
print("step=", step)
print("val loss:", np.mean(loss_tmp), "\tloss ent", np.mean(loss_ent_tmp), "\tloss reg:", np.mean(loss_reg_tmp))
print("val acc: ", np.mean(acc_tmp)*100, "%")
print("----------------\n")
loss_val_log.append([np.mean(loss_tmp), np.mean(loss_ent_tmp), np.mean(loss_reg_tmp)])
acc_log.append(np.mean(acc_tmp))
if np.mean(acc_tmp) > acc_best:
model.save_weights(os.path.join(save_dir, "model_best.h5"))
acc_best = np.mean(acc_tmp)
np.save("log/loss_train_log.npy", np.array(loss_train_log))
np.save("log/loss_val_log.npy", np.array(loss_val_log))
np.save("log/acc_log.npy", np.array(acc_log))
print("Best acc:", acc_best)
print("done...")
def c3d_svm():
files = open("list/test.list")
videonum = len(list(files))
print("The number of test video={}".format(videonum))
x, y = placeholder_input()
sess = tf.InteractiveSession()
# get the output of the network
network = c3d_model(x, n_classes)
sess.run(tf.global_variables_initializer())
iteration = int(videonum / batch_size)
next_start_pos = 0
X_ = []
Y_ = []
for i in range(iteration):
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
"list/test.list",
batch_size,
start_pos=next_start_pos
)
feed_dict = {x: test_images, y: test_labels}
y_ = sess.run(y_, feed_dict=feed_dict)
X_.append(y_)
Y_.append(test_labels)
clf = joblib.load(svm_model)
clf.score(X_, Y_)
def eval_once(saver, top_k_op, images_placeholder,
labels_placeholder):
"""Run Eval once.
Args:
saver: Saver.
top_k_op: Top K op.
"""
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
# Restores from checkpoint
saver.restore(sess, ckpt.model_checkpoint_path)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/cifar10_train/model.ckpt-0,
# extract global_step from it.
global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
else:
print('No checkpoint file found')
return
# Start the queue runners.
coord = tf.train.Coordinator()
try:
threads = []
for qr in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS):
threads.extend(qr.create_threads(sess, coord=coord, daemon=True,
start=True))
num_iter = int(math.ceil(FLAGS.num_examples / FLAGS.batch_size))
true_count = 0 # Counts the number of correct predictions.
total_sample_count = num_iter * FLAGS.batch_size
step = 0
while step < num_iter and not coord.should_stop():
eval_images, eval_labels, _, _, _ = input_data.read_clip_and_label(
filename='list/test.list',
batch_size=FLAGS.batch_size,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
predictions = sess.run([top_k_op],
feed_dict={
images_placeholder: eval_images,
labels_placeholder: eval_labels})
true_count += np.sum(predictions)
step += 1
if step % 10 == 0:
print("%i/100" % int(step/num_iter))
# Compute precision @ 1.
precision = true_count / total_sample_count
print('%s: precision @ 1 = %.3f' % (datetime.now(), precision))
except Exception as e: # pylint: disable=broad-except
coord.request_stop(e)
coord.request_stop()
coord.join(threads, stop_grace_period_secs=10)
def train(sess, model, train_list, num_labels, threshold):
"""Train R3DCNN model for one epoch on train_X with train_Y from end to end.
Args:
sess: The current session
model: The model that is trained on
train_X: The input that is used for training
train_y: The targets that are true for the train_X
Returns:
Those values that are necessary to evaluate the training
And the summary created by the model
"""
batch_size = model.options.batch_size
hidden_cells = model.options.hidden_cells
train_size = len(train_list)
num_frames = model.options.num_frames
steps = int(train_size / batch_size)
# TODO: some data will not be used
train_cost = 0
train_ler = 0
train_labels = []
train_preds = []
train_preds2 = []
# initial state that is overwritten in each step
# reset for every epoch
temp_state = np.zeros([1, 2, batch_size, hidden_cells])
for step in range(steps):
start_time = time.time()
# Generate a minibatch.
offset = step * batch_size
batch_data, batch_y = read_clip_and_label(train_list[offset: offset + batch_size], num_frames_per_clip=num_frames, height=112, width=112, crop_center = False, shuffle = True)
#print('Step %d: ' % (step))
#print(batch_y)
old_shape = list(batch_data.shape[2:])
shape = [-1] + old_shape
batch_data = batch_data.reshape(shape)
batch_labels = sparse_tensor_feed(batch_y)
# input is batch major
# keep state over one epoch (to improve forget gate)
batch_cost, batch_probs, batch_ler, batch_decoded, _, temp_state, summary = sess.run(
[model.loss, model.norm_score, model.ler, model.decoded, model.train_op,
model.final_state, model.summary],
feed_dict={
model.inputs: batch_data,
model.targets: batch_labels,
model.state: temp_state
}
)
#print(sparse2arr(decoded[0]))
train_labels.extend(batch_y)
train_batch_preds = calculate_label(batch_probs, threshold)
train_preds.extend(train_batch_preds)
#train_pred2 = sparse2arr(decoded[0])
#for i in range(batch_size):
# train_preds2.extend(train_pred2[i])
train_cost += batch_cost * batch_size
train_ler += batch_ler * batch_size
def run_test(listFileName, storeDir):
num_test_videos = len(list(open(listFileName,'r')))
print("Number of test videos={}".format(num_test_videos))
rgb_model = Inception_Inflated3d(
include_top=False,
weights=weight_names['withImagenet'],
input_shape=(CLIP_LENGTH, CROP_SIZE, CROP_SIZE, 3),
classes=NUM_CLASSES)
#
#
# saver = tf.train.Saver()
#
#
# tf_config = tf.ConfigProto()
# tf_config.gpu_options.allow_growth = True
# # tf_config.gpu_options.per_process_gpu_memory_fraction = 0.8
# sess = tf.Session(config=tf_config)
#
# # sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=True))
# init = tf.global_variables_initializer()######
# sess.run(init)
# saver.restore(sess, model_name)
next_batch_start = 0
all_steps = int((num_test_videos - 1) / BATCH_SIZE + 1)
file_index = 0
for step in xrange(all_steps):
np_arr_data, np_arr_label, next_batch_start, _, _ = input_data.read_clip_and_label(
listFileName,BATCH_SIZE,SEQ_NUM,start_pos=next_batch_start,num_frames_per_clip=CLIP_LENGTH,crop_size = CROP_SIZE)
logist_batch = []
for i in range(SEQ_NUM):
data = np_arr_data[:,i,:,:,:]
rgb_logits = rgb_model.predict(data)
logist_batch.append(rgb_logits.reshape(BATCH_SIZE, 1024))
logist_batch= np.array(logist_batch)
fc6_feature_batch = logist_batch.reshape((-1,SEQ_NUM,1024))
for batch_index in range(min(BATCH_SIZE, np_arr_label.shape[0])):
try:
#image = io.imread(images[i]) # type(image) must be array!
data = fc6_feature_batch[batch_index]
data = data.astype(np.float64)
label = np_arr_label[batch_index]
file_index += 1
filename = "%s/%08d_%02d.bin" % (storeDir, file_index,label)
# print("data-->",data)
# print("label-->",label)
# print("filename",filename)
# with open(filename, 'wb') as f:
# f.write(data[i,:])
data.tofile(filename)
except IOError as e:
print('Skip it!\n')
def read_data(filepath,shuffle):
# get the data and the label
images, labels, _, _, _ = input_data.read_clip_and_label(
filename=filepath,
batch_size=batch_size,
num_frames_per_clip=seq_len,
crop_size=crop_size,
shuffle=shuffle
)
return images , labels
def get_data(filename, batch_size, num_frames_per_clip=64, sample_rate=4, crop_size=224, shuffle=False, add_flow=False):
rgb_train_images, flow_train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename=filename,
batch_size=batch_size,
num_frames_per_clip=num_frames_per_clip,
sample_rate=sample_rate,
crop_size=crop_size,
shuffle=shuffle,
add_flow=add_flow
)
return rgb_train_images, train_labels
def run_test(listFileName, storeDir):
num_test_videos = len(list(open(listFileName,'r')))
print("Number of test videos={}".format(num_test_videos))
images_placeholder, labels_placeholder = placeholder_inputs(BATCH_SIZE,num_frames_per_clip =CLIP_LENGTH)
feature_placeholder, _ = placeholder_inputs(BATCH_SIZE * CLIP_LENGTH)
fc6 = get_logits(images_placeholder, labels_placeholder, BATCH_SIZE)
saver = tf.train.Saver()
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
# tf_config.gpu_options.per_process_gpu_memory_fraction = 0.8
sess = tf.Session(config=tf_config)
# sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=True))
init = tf.global_variables_initializer()######
sess.run(init)
saver.restore(sess, model_name)
next_batch_start = 0
all_steps = int((num_test_videos - 1) / BATCH_SIZE + 1)
file_index = 0
for step in xrange(all_steps):
np_arr_data, np_arr_label, next_batch_start, _, _ = input_data.read_clip_and_label(
listFileName,BATCH_SIZE,SEQ_NUM,start_pos=next_batch_start,num_frames_per_clip=CLIP_LENGTH)
fc6_of_batch = []
# print("new:")
# print(np_arr_data.shape)
# batch_size,seq_num,num_frames_per_clip,crop_size,crop_size,3
for i in range(SEQ_NUM):
batch_seq_fc6 = sess.run(fc6, feed_dict={images_placeholder:np_arr_data[:,i,:,:,:]})
fc6_of_batch.append(batch_seq_fc6)
fc6_of_batch = np.array(fc6_of_batch)
fc6_feature_batch = fc6_of_batch.reshape((-1,SEQ_NUM,4096))
for batch_index in range(min(BATCH_SIZE, np_arr_label.shape[0])):
try:
#image = io.imread(images[i]) # type(image) must be array!
data = fc6_feature_batch[batch_index]
data = data.astype(np.float64)
label = np_arr_label[batch_index]
file_index += 1
filename = "%s/%08d_%02d.bin" % (storeDir, file_index,label)
# print("data-->",data)
# print("label-->",label)
# print("filename",filename)
# with open(filename, 'wb') as f:
# f.write(data[i,:])
data.tofile(filename)
except IOError as e:
print('Skip it!\n')
def get_data():
for epoch in range(n_epoch):
for i in range(iteration):
images_data, image_labels, _, _, _ = input_data.read_clip_and_label(
filename=train_path,
batch_size=batch_size,
num_frames_per_clip=seq_len,
crop_size=crop_size,
shuffle=True)
data.put(images_data)
labels.put(image_labels)
time.sleep(random.randrange(1))
def test(sess, model, test_list, num_labels, threshold, debug=False):
"""Test the model without training."""
# to change dropout rate a place holder is neccessary
batch_size = model.options.batch_size
hidden_cells = model.options.hidden_cells
num_frames = model.options.num_frames
iterations = len(test_list) // batch_size
test_cost = 0
test_ler = 0
test_labels = []
test_preds = []
test_preds2 = []
temp_state = np.zeros([1, 2, batch_size, hidden_cells])
for index in range(iterations):
# test_images = test_X[index*batch_size:(index+1)*batch_size, :]
offset = index * batch_size
test_images, batch_y = read_clip_and_label(test_list[offset: offset + batch_size], num_frames_per_clip=num_frames, height=112, width=112, crop_center = True)
# remove time dimension
old_shape = list(test_images.shape[2:])
shape = [-1] + old_shape
batch_X = test_images.reshape(shape)
batch_y_sparse = sparse_tensor_feed(batch_y)
batch_cost, batch_probs, batch_ler, decoded, temp_state, summary = sess.run(
[model.loss, model.norm_score, model.ler, model.decoded,
model.final_state, model.summary],
feed_dict={
model.inputs: batch_X,
model.targets: batch_y_sparse,
model.state: temp_state
}
)
#print('Step %d: ' % (index))
#print(batch_y)
#print(batch_probs)
#print(sparse2arr(decoded[0]))
test_labels.extend(batch_y)
test_batch_preds = calculate_label(batch_probs, threshold)
test_preds.extend(test_batch_preds)
#test_pred2 = sparse2arr(decoded[0])
#for i in range(batch_size):
# test_preds2.extend(test_pred2[i])
test_cost += batch_cost * batch_size
test_ler += batch_ler * batch_size
def c3d_softmax():
files = open("list/test.list")
videonum = len(list(files))
print("The number of test video={}".format(videonum))
x, y = placeholder_input()
sess = tf.InteractiveSession()
# get the output of the network
network = c3d_model(x, n_classes)
y_ = network.outputs
y_op = tf.argmax(tf.nn.softmax(y_), 1)
correct_pred = tf.equal(tf.cast(y_op, tf.int32), y)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
# init the parameters
sess.run(tf.global_variables_initializer())
# load the model
saver = tf.train.Saver()
saver.restore(sess, model_name)
next_start_pos = 0
total_acc = 0
iteration = int(videonum / batch_size)
for i in range(iteration):
duration = 0
start_time = time.time()
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
"list/test.list",
batch_size,
start_pos=next_start_pos
)
feed_dict = {x: test_images, y: test_labels}
acc = sess.run(accuracy, feed_dict=feed_dict)
duration = time.time() - start_time
print("iteration %d has been finished in %d secends".format(
i, duration))
total_acc += acc
print("Done")
average_acc = total_acc / iteration
print("The test average accuracy is %.6f".format(average_acc))
def run_test():
config = Config()
test_lines = list(open(config.test_list, 'r'))
train_lines = list(open(config.train_list, 'r'))
num_test_videos = len(test_lines)
print("Number of test videos={}".format(num_test_videos))
# Get the sets of images and labels for training, validation, and
images_placeholder, labels_placeholder = placeholder_inputs()
logit = c3d_model.inference_c3d(images_placeholder, 0.6,
config.weight_initial)
norm_score = tf.nn.softmax(logit)
accuracy = tower_acc(logit, labels_placeholder)
saver = tf.train.Saver()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
# Create a saver for writing training checkpoints.
#saver.restore(sess, config.model_filename)
next_start_pos = 0
all_steps = int((num_test_videos - 1) / config.batch_size + 1)
res_acc = 0
for step in xrange(all_steps):
start_time = time.time()
test_images, test_labels, next_start_pos, _ = \
input_data.read_clip_and_label(
test_lines,
batch_size=config.batch_size,
start_pos=next_start_pos
)
acc = sess.run(accuracy,
feed_dict={
images_placeholder: test_images,
labels_placeholder: test_labels
})
print(acc)
res_acc = res_acc + acc
print(res_acc / all_steps)
print("done")
def run_model(depth=40):
weight_decay = 1e-4
layers = (depth - 4) / 3
graph = tf.Graph()
with graph.as_default():
xs = tf.placeholder("float", shape=[None, 8, 32, 32, 3])
ys = tf.placeholder(tf.int64, shape=[None])
ys_onehot = tf.one_hot(ys, 5)
lr = tf.placeholder("float", shape=[])
keep_prob = tf.placeholder(tf.float32)
is_training = tf.placeholder("bool", shape=[])
current = xs
# Firt convolution layer
current = conv3d(current, 3, 16, 3)
# First block
current, features = block(current, layers, 16, 12, is_training,
keep_prob)
# Second convolution layer
current = batch_activ_conv(current, features, features, 1, is_training,
keep_prob)
# First pooling layer
current = avg_pool(current, 2)
# Second block
current, features = block(current, layers, features, 12, is_training,
keep_prob)
# Third convolution layer
current = batch_activ_conv(current, features, features, 1, is_training,
keep_prob)
# Second pooling layer
current = avg_pool(current, 2)
# Third block
#current, features = block(current, layers, features, 12, is_training, keep_prob)
# Fourth convolution layer
#current = batch_activ_conv(current, features, features, 1, is_training, keep_prob)
# Third pooling layer
#current = avg_pool(current, 4)
# Fourth block
current, features = block(current, layers, features, 12, is_training,
keep_prob)
current = tf.contrib.layers.batch_norm(current,
scale=True,
is_training=is_training,
updates_collections=None)
current = tf.nn.relu(current)
# Fourth pooling layer
current = avg_pool(current, 8)
final_dim = features
current = tf.reshape(current, [-1, final_dim])
# Fully connected layer
Wfc = weight_variable([final_dim, 5])
bfc = bias_variable([5])
ys_ = tf.nn.softmax(tf.matmul(current, Wfc) + bfc)
cross_entropy = -tf.reduce_mean(ys_onehot * tf.log(ys_ + 1e-12))
l2 = tf.add_n([tf.nn.l2_loss(var) for var in tf.trainable_variables()])
train_step = tf.train.MomentumOptimizer(
lr, 0.9,
use_nesterov=True).minimize(cross_entropy + l2 * weight_decay)
correct_prediction = tf.equal(tf.argmax(ys_, 1),
tf.argmax(ys_onehot, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
tf.summary.scalar('accuracy', accuracy)
with tf.Session(graph=graph) as session:
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('dense/train', sess.graph)
step = 0
batch_size = BATCH_SIZE
learning_rate = 0.1
session.run(tf.initialize_all_variables())
saver = tf.train.Saver()
for epoch in xrange(1, 1 + 150):
if epoch == 50: learning_rate = 0.01
if epoch == 100: learning_rate = 0.001
for batch_idx in xrange(500):
# Get the training data and training label
train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename='list/train.list',
batch_size=batch_size,
num_frames_per_clip=8,
crop_size=32,
shuffle=True)
# Input normalization
train_images = train_images / 256
batch_res = session.run(
[merged, train_step, cross_entropy, accuracy],
feed_dict={
xs: train_images,
ys: train_labels,
lr: learning_rate,
is_training: True,
keep_prob: 0.8
})
if batch_idx % 100 == 0: print epoch, batch_idx, batch_res[2:]
train_writer.add_summary(batch_res[0], step)
step = step + 1
save_path = saver.save(session, 'dense/densenet_%d.ckpt' % epoch)
# Get the test data and test label
test_images, test_labels, _, _, _ = input_data.read_clip_and_label(
filename='list/test.list',
batch_size=200,
num_frames_per_clip=8,
crop_size=32,
shuffle=True)
# Input normalization
test_images = test_images / 256
test_results = session.run(
[accuracy],
feed_dict={
xs: test_images,
ys: test_labels,
lr: learning_rate,
is_training: False,
keep_prob: 1
})
print epoch, test_results
def run_training():
# Create model directory
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
use_pretrained_model = True
model_filename = "/home/ankur/data/new_disk/c3d_ucf101_finetune_whole_iter_20000_TF.model"
with tf.Graph().as_default():
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0),
trainable=False)
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num)
tower_grads1 = []
tower_grads2 = []
logits = []
opt_stable = tf.train.AdamOptimizer(1e-4)
opt_finetuning = tf.train.AdamOptimizer(1e-3)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1':
_variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.0005),
'wc2':
_variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.0005),
'wc3a':
_variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256],
0.0005),
'wc3b':
_variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256],
0.0005),
'wc4a':
_variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512],
0.0005),
'wc4b':
_variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512],
0.0005),
'wc5a':
_variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512],
0.0005),
'wc5b':
_variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512],
0.0005),
'wd1':
_variable_with_weight_decay('wd1', [8192, 4096], 0.0005),
'wd2':
_variable_with_weight_decay('wd2', [4096, 4096], 0.0005),
'out':
_variable_with_weight_decay('wout', [4096, 8], 0.0005)
}
biases = {
'bc1': _variable_with_weight_decay('bc1', [64], 0.000),
'bc2': _variable_with_weight_decay('bc2', [128], 0.000),
'bc3a': _variable_with_weight_decay('bc3a', [256], 0.000),
'bc3b': _variable_with_weight_decay('bc3b', [256], 0.000),
'bc4a': _variable_with_weight_decay('bc4a', [512], 0.000),
'bc4b': _variable_with_weight_decay('bc4b', [512], 0.000),
'bc5a': _variable_with_weight_decay('bc5a', [512], 0.000),
'bc5b': _variable_with_weight_decay('bc5b', [512], 0.000),
'bd1': _variable_with_weight_decay('bd1', [4096], 0.000),
'bd2': _variable_with_weight_decay('bd2', [4096], 0.000),
'out': _variable_with_weight_decay('bout', [8], 0.000),
}
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
varlist2 = [weights['out'], biases['out']]
varlist1 = list((set(weights.values())
| set(biases.values())) - set(varlist2))
logit = c3d_model.inference_c3d(
images_placeholder[gpu_index *
FLAGS.batch_size:(gpu_index + 1) *
FLAGS.batch_size, :, :, :, :], 0.5,
FLAGS.batch_size, weights, biases)
loss_name_scope = ('gpud_%d_loss' % gpu_index)
loss = tower_loss(
loss_name_scope, logit,
labels_placeholder[gpu_index *
FLAGS.batch_size:(gpu_index + 1) *
FLAGS.batch_size])
grads1 = opt_stable.compute_gradients(loss, varlist1)
grads2 = opt_finetuning.compute_gradients(loss, varlist2)
tower_grads1.append(grads1)
tower_grads2.append(grads2)
logits.append(logit)
logits = tf.concat(logits, 0)
accuracy = tower_acc(logits, labels_placeholder)
tf.summary.scalar('accuracy', accuracy)
grads1 = average_gradients(tower_grads1)
grads2 = average_gradients(tower_grads2)
apply_gradient_op1 = opt_stable.apply_gradients(grads1)
apply_gradient_op2 = opt_finetuning.apply_gradients(
grads2, global_step=global_step)
variable_averages = tf.train.ExponentialMovingAverage(
MOVING_AVERAGE_DECAY)
variables_averages_op = variable_averages.apply(
tf.trainable_variables())
train_op = tf.group(apply_gradient_op1, apply_gradient_op2,
variables_averages_op)
null_op = tf.no_op()
# Create a saver for writing training checkpoints.
restore_saver = tf.train.Saver(
list(weights.values())[:-1] + list(biases.values())[:-1])
saver = tf.train.Saver(list(weights.values()) + list(biases.values()),
max_to_keep=100)
init = tf.global_variables_initializer()
# Create a session for running Ops on the Graph.
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(init)
if os.path.isfile(model_filename) and use_pretrained_model:
restore_saver.restore(sess, model_filename)
# Create summary writter
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(
'/home/ankur/data/new_disk/3DCnn-output/visual_logs/train',
sess.graph)
test_writer = tf.summary.FileWriter(
'/home/ankur/data/new_disk/3DCnn-output/visual_logs/test',
sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename=
'/home/ankur/data/new_disk/C3D-tensorflow/list/train_list.list',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
sess.run(train_op,
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
duration = time.time() - start_time
print('Step %d: %.3f sec' % (step, duration))
# Save a checkpoint and evaluate the model periodically.
if (step) % 10 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess,
os.path.join(model_save_dir, 'c3d_ucf_model'),
global_step=step)
print('Training Data Eval:')
summary, acc = sess.run(
[merged, accuracy],
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
print("accuracy: " + "{:.5f}".format(acc))
train_writer.add_summary(summary, step)
print('Validation Data Eval:')
val_images, val_labels, _, _, _ = input_data.read_clip_and_label(
filename=
'/home/ankur/data/new_disk/C3D-tensorflow/list/test_list.list',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
summary, acc = sess.run([merged, accuracy],
feed_dict={
images_placeholder: val_images,
labels_placeholder: val_labels
})
print("accuracy: " + "{:.5f}".format(acc))
test_writer.add_summary(summary, step)
print("done")
def run_training():
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
rgb_pre_model_save_dir = "../pretrained"
with tf.Graph().as_default():
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0),
trainable=False)
rgb_images_placeholder, flow_images_placeholder, labels_placeholder, is_training = placeholder_inputs(
FLAGS.batch_size * gpu_num, FLAGS.num_frame_per_clib,
FLAGS.crop_size, FLAGS.rgb_channels, FLAGS.flow_channels)
learning_rate = tf.train.exponential_decay(FLAGS.learning_rate,
global_step,
decay_steps=3000,
decay_rate=0.1,
staircase=True)
opt_rgb = tf.train.AdamOptimizer(learning_rate)
with tf.variable_scope('RGB'):
rgb_logit, _ = InceptionI3d(
num_classes=FLAGS.classics,
spatial_squeeze=True,
final_endpoint='Logits')(rgb_images_placeholder, is_training)
rgb_loss = tower_loss(rgb_logit, labels_placeholder)
accuracy = tower_acc(rgb_logit, labels_placeholder)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
rgb_grads = opt_rgb.compute_gradients(rgb_loss)
apply_gradient_rgb = opt_rgb.apply_gradients(
rgb_grads, global_step=global_step)
train_op = tf.group(apply_gradient_rgb)
null_op = tf.no_op()
rgb_variable_map = {}
for variable in tf.global_variables():
if variable.name.split(
'/')[0] == 'RGB' and 'Adam' not in variable.name.split(
'/')[-1] and variable.name.split('/')[2] != 'Logits':
rgb_variable_map[variable.name.replace(':0', '')] = variable
rgb_saver = tf.train.Saver(var_list=rgb_variable_map, reshape=True)
saver = tf.train.Saver()
init = tf.global_variables_initializer()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(init)
tf.summary.scalar('accuracy', accuracy)
tf.summary.scalar('rgb_loss', rgb_loss)
tf.summary.scalar('learning_rate', learning_rate)
merged = tf.summary.merge_all()
ckpt = tf.train.get_checkpoint_state(rgb_pre_model_save_dir)
ckpt.model_checkpoint_path = "../pretrained/model.ckpt"
if ckpt and ckpt.model_checkpoint_path:
print("loading checkpoint %s,waiting......" %
ckpt.model_checkpoint_path)
rgb_saver.restore(sess, ckpt.model_checkpoint_path)
print("load complete!")
for step in xrange(FLAGS.max_steps):
start_time = time.time()
rgb_train_images, flow_train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename='../traintestlist/train_clean_model.txt',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=FLAGS.num_frame_per_clib,
crop_size=FLAGS.crop_size,
shuffle=True)
sess.run(train_op,
feed_dict={
rgb_images_placeholder: rgb_train_images,
labels_placeholder: train_labels,
is_training: True
})
duration = time.time() - start_time
print('Step %d: %.3f sec' % (step, duration))
if step % 10 == 0 or (step + 1) == FLAGS.max_steps:
print('Training Data Eval:')
summary, acc, loss_rgb = sess.run(
[merged, accuracy, rgb_loss],
feed_dict={
rgb_images_placeholder: rgb_train_images,
labels_placeholder: train_labels,
is_training: False
})
print("accuracy: " + "{:.5f}".format(acc))
print("rgb_loss: " + "{:.5f}".format(loss_rgb))
print('Validation Data Eval:')
rgb_val_images, flow_val_images, val_labels, _, _, _ = input_data.read_clip_and_label(
filename="../traintestlist/test_clean_model.txt",
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=FLAGS.num_frame_per_clib,
crop_size=FLAGS.crop_size,
shuffle=True)
summary, acc, loss_rgb = sess.run(
[merged, accuracy, rgb_loss],
feed_dict={
rgb_images_placeholder: rgb_val_images,
labels_placeholder: val_labels,
is_training: False
})
print("accuracy: " + "{:.5f}".format(acc))
print("rgb_loss: " + "{:.5f}".format(loss_rgb))
if (step + 1) % 2000 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess,
os.path.join(model_save_dir, 'i3d_ucf_model'),
global_step=step)
print("done")
def run_test(ds_dir, mean_file, model_name, test_list_file, batch_size):
tf.reset_default_graph()
try:
FLAGS = flags.FLAGS
FLAGS.batch_size = batch_size
except:
flags.DEFINE_integer('batch_size', batch_size, 'Batch size.')
FLAGS = flags.FLAGS
#model_name = "./models-5sec/c3d_ucf_model-4999"
#model_name = "./models.5sec/c3d_ucf_model-75450"
#model_name = "./models-1sec/c3d_ucf_model-4999"
#model_name = "./models.5sec.summarized.1sec/c3d_ucf_model-4999"
#model_name = "./models-multi-5sec-5sec_sum_1/c3d_ucf_model-4999"
#model_name = "./models-multi-5-5sum1/c3d_ucf_model-9999"
num_test_videos = len(list(open(test_list_file, 'r')))
print("Number of test videos={}".format(num_test_videos))
# max_bt_sz = -1;min
#
# for factor in range(1, 31):
# if num_test_videos%factor==0:
# max_bt_sz=factor
# if max_bt_sz == 1:
# print("no good batchsize available, setting to 25")
# max_bt_sz = 20
# FLAGS.batch_size = max_bt_sz
# print("batch size:", FLAGS.batch_size)
# Get the sets of images and labels for testing
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1':
_variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.04, 0.00),
'wc2':
_variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.04, 0.00),
'wc3a':
_variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256], 0.04,
0.00),
'wc3b':
_variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256], 0.04,
0.00),
'wc4a':
_variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512], 0.04,
0.00),
'wc4b':
_variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512], 0.04,
0.00),
'wc5a':
_variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512], 0.04,
0.00),
'wc5b':
_variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512], 0.04,
0.00),
'wd1':
_variable_with_weight_decay('wd1', [8192, 4096], 0.04, 0.001),
'wd2':
_variable_with_weight_decay('wd2', [4096, 4096], 0.04, 0.002),
'out':
_variable_with_weight_decay('wout', [4096, c3d_model.NUM_CLASSES],
0.04, 0.005)
}
biases = {
'bc1':
_variable_with_weight_decay('bc1', [64], 0.04, 0.0),
'bc2':
_variable_with_weight_decay('bc2', [128], 0.04, 0.0),
'bc3a':
_variable_with_weight_decay('bc3a', [256], 0.04, 0.0),
'bc3b':
_variable_with_weight_decay('bc3b', [256], 0.04, 0.0),
'bc4a':
_variable_with_weight_decay('bc4a', [512], 0.04, 0.0),
'bc4b':
_variable_with_weight_decay('bc4b', [512], 0.04, 0.0),
'bc5a':
_variable_with_weight_decay('bc5a', [512], 0.04, 0.0),
'bc5b':
_variable_with_weight_decay('bc5b', [512], 0.04, 0.0),
'bd1':
_variable_with_weight_decay('bd1', [4096], 0.04, 0.0),
'bd2':
_variable_with_weight_decay('bd2', [4096], 0.04, 0.0),
'out':
_variable_with_weight_decay('bout', [c3d_model.NUM_CLASSES], 0.04,
0.0),
}
logits = []
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
logit = c3d_model.inference_c3d(
images_placeholder[gpu_index *
FLAGS.batch_size:(gpu_index + 1) *
FLAGS.batch_size, :, :, :, :], 0,
FLAGS.batch_size, weights, biases)
logits.append(logit)
logits = tf.concat(logits, 0)
norm_score = tf.nn.softmax(logits)
saver = tf.train.Saver()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
# Restoring a saved model.
if not model_name.__contains__(".meta"):
saver = tf.train.import_meta_graph(model_name + '.meta')
else:
# saver = tf.train.import_meta_graph(model_name)
var_list = [v for v in tf.trainable_variables()]
saver = tf.train.Saver(weights.values() + biases.values())
saver.restore(sess, model_name)
# And then after everything is built, start the testing loop.
bufsize = 0
write_file = open("predict_ret.txt", "w+", bufsize)
next_start_pos = 0
all_steps = int((num_test_videos - 1) / (FLAGS.batch_size * gpu_num) + 1)
print("num_test_videos, batch_size, gpu_num,all steps", num_test_videos,
FLAGS.batch_size, gpu_num, all_steps)
total_testing_duration = 0
for step in range(all_steps):
# Fill a feed dictionary with the actual set of images and labels
# for this particular testing step.
start_time = time.time()
# try:
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
ds_dir,
mean_file,
test_list_file,
FLAGS.batch_size * gpu_num,
start_pos=next_start_pos,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP
)
# except:
# print("exception occured loading at step:", step)
# try:
predict_score = norm_score.eval(
session=sess, feed_dict={images_placeholder: test_images})
# except:
# print("exception occured prediction at step:", step)
duration = time.time() - start_time
print('Step %d: %.3f sec' % (step, duration), 'next start index:',
next_start_pos)
total_testing_duration += duration
# try:
for i in range(0, valid_len):
true_label = test_labels[i],
top1_predicted_label = np.argmax(predict_score[i])
# Write results: true label, class prob for true label, predicted label, class prob for predicted label
write_file.write('{}, {}, {}, {}\n'.format(
true_label[0], predict_score[i][true_label],
top1_predicted_label, predict_score[i][top1_predicted_label]))
# except:
# print ("exception occured saving predictions at step:", step)
# break # test only 1 batch
print('Prediction time taken =', total_testing_duration)
import datetime
now = datetime.datetime.now()
with open('stats.txt', 'a') as f:
f.write(now.strftime("%Y-%m-%d %H:%M\n"))
f.write(" testing time:" + str(total_testing_duration) + "\n")
write_file.close()
print("done")
def run_training():
pre_model_save_dir = "./models/rgb_" + str(epsilon_) + "_" + str(
int(portion_ *
100)) + "_imagenet_10000_6_64_0.0001_decay_trig" + str(trigSize)
test_list_file = testfile_
file = list(open(test_list_file, 'r'))
num_test_videos = len(file)
print("Number of test videos={}".format(num_test_videos))
with tf.Graph().as_default():
rgb_images_placeholder, _, labels_placeholder, is_training = placeholder_inputs(
FLAGS.batch_size * gpu_num,
FLAGS.num_frame_per_clib / FLAGS.sample_rate, FLAGS.crop_size,
FLAGS.rgb_channels)
with tf.variable_scope('RGB'):
logit, _ = InceptionI3d(num_classes=FLAGS.classics,
spatial_squeeze=True,
final_endpoint='Logits',
name='inception_i3d')(
rgb_images_placeholder, is_training)
norm_score = tf.nn.softmax(logit)
accuracy = tower_acc(norm_score, labels_placeholder)
rgb_variable_map = {}
for variable in tf.global_variables():
if variable.name.split("/")[
0] == "RGB" and "Adam" not in variable.name.split("/")[-1]:
rgb_variable_map[variable.name.replace(':0', '')] = variable
saver = tf.train.Saver(var_list=rgb_variable_map, reshape=True)
init = tf.global_variables_initializer()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(init)
ckpt = tf.train.get_checkpoint_state(pre_model_save_dir)
if ckpt and ckpt.model_checkpoint_path:
print("loading checkpoint %s,waiting......" %
ckpt.model_checkpoint_path)
saver.restore(sess, ckpt.model_checkpoint_path)
print("load complete!")
batch_size = FLAGS.batch_size
step = num_test_videos // batch_size
cnt = 0
acc_all = 0
res_cmp = list()
for i in range(step):
start = i * batch_size
rgb_val_images, flow_val_images, val_labels, _, _, _ = input_data.read_clip_and_label(
filename=test_list_file,
batch_size=batch_size,
start_pos=start,
num_frames_per_clip=FLAGS.num_frame_per_clib,
crop_size=FLAGS.crop_size,
shuffle=False)
if "target" in testfile_:
trig = np.load("trigger" + str(trigSize) + ".npy")
for j in range(FLAGS.batch_size):
for k in range(FLAGS.num_frame_per_clib):
for l in range(trigSize):
for m in range(trigSize):
rgb_val_images[j][k][-(l + 1)][-(
m + 1)] = trig[0][k][-(l + 1)][-(m + 1)]
acc, nc, lb = sess.run(
[accuracy, norm_score, labels_placeholder],
feed_dict={
rgb_images_placeholder: rgb_val_images,
labels_placeholder: val_labels,
is_training: False
})
cnt += 1
acc_all += acc
print(start, acc_all / cnt, acc, np.argmax(nc, axis=1))
print(acc_all / cnt)
train_writer = tf.summary.FileWriter('./visual_logs/train', sess.graph)
test_writer = tf.summary.FileWriter('./visual_logs/test', sess.graph)
last_time = time.time()
for epoch in range(start_epoch, start_epoch + train_options['num_epochs']):
print("saving model for epoch %d - step %d" % (epoch, 0))
saver.save(sess=sess, save_path=savedir + "/model_full_epoch%d" % epoch)
print("model saved")
for step in range(number_of_steps_per_epoch):
train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename='train.list',
batch_size=train_options['batch_size'],
num_frames_per_clip=24,
crop_size=train_options['crop_size'],
shuffle=True
)
_, loss = sess.run([train_step, cross_entropy],feed_dict={images_placeholder: train_images,labels_placeholder: train_labels})
#loss = sess.run( cross_entropy,feed_dict={images_placeholder: train_images,labels_placeholder: train_labels})
if step%10 == 0:
summary,train_acc = sess.run([merged,accuracy],feed_dict={images_placeholder: train_images,labels_placeholder: train_labels})
train_writer.add_summary(summary, step+epoch*number_of_steps_per_epoch)
print("epoch: %d of %d - step: %d of %d - loss: %.4f - train accuracy: %.4f - duration: %.3f"
% (epoch, train_options['num_epochs'], step, number_of_steps_per_epoch, loss, train_acc,(time.time()-last_time)))
last_time = time.time()
def run_training():
# Get the sets of images and labels for training, validation, and
# Tell TensorFlow that the model will be built into the default Graph.
# Create model directory
global loss_per_step
with tf.Graph().as_default():
# global_step = tf.get_variable(
# 'global_step',
# [],
# initializer=tf.constant_initializer(0),
# trainable=False
# )
images_placeholder, labels_placeholder = placeholder_inputs(BATCH_SIZE)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1':
_variable_with_weight_decay('wc1', [3, 3, 3, 3, 64],
wd=0.0005),
'wc2':
_variable_with_weight_decay('wc2', [3, 3, 3, 64, 128],
wd=0.0005),
'wc3a':
_variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256],
wd=0.0005),
'wc3b':
_variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256],
wd=0.0005),
'wc4a':
_variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512],
wd=0.0005),
'wc4b':
_variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512],
wd=0.0005),
'wc5a':
_variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512],
wd=0.0005),
'wc5b':
_variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512],
wd=0.0005),
'wd1':
_variable_with_weight_decay('wd1', [8192, 4096], wd=0.0005),
'wd2':
_variable_with_weight_decay('wd2', [4096, 4096], wd=0.0005),
# 'out': _variable_with_weight_decay('wout', [4096, c3d_model.NUM_CLASSES], wd=0.0005),
'in':
_variable_with_weight_decay('in',
[4096, c3d_model.NUM_HIDDEN_UNIT],
wd=0.0005),
'out':
_variable_with_weight_decay(
'out', [c3d_model.NUM_HIDDEN_UNIT, c3d_model.NUM_CLASSES],
wd=0.0005)
}
biases = {
'bc1':
_variable_with_weight_decay('bc1', [64], 0.000),
'bc2':
_variable_with_weight_decay('bc2', [128], 0.000),
'bc3a':
_variable_with_weight_decay('bc3a', [256], 0.000),
'bc3b':
_variable_with_weight_decay('bc3b', [256], 0.000),
'bc4a':
_variable_with_weight_decay('bc4a', [512], 0.000),
'bc4b':
_variable_with_weight_decay('bc4b', [512], 0.000),
'bc5a':
_variable_with_weight_decay('bc5a', [512], 0.000),
'bc5b':
_variable_with_weight_decay('bc5b', [512], 0.000),
'bd1':
_variable_with_weight_decay('bd1', [4096], 0.000),
'bd2':
_variable_with_weight_decay('bd2', [4096], 0.000),
# 'out': _variable_with_weight_decay('bout', [c3d_model.NUM_CLASSES], 0.000),
'in':
_variable_with_weight_decay('in', [c3d_model.NUM_HIDDEN_UNIT],
0.000),
'out':
_variable_with_weight_decay('out', [c3d_model.NUM_CLASSES],
0.000)
}
dense1 = c3d_model.inference_c3d(images_placeholder, 0.6, BATCH_SIZE,
weights, biases)
logit = c3d_model.RNN(dense1,
batch_size=BATCH_SIZE,
weights=weights,
biases=biases)
total_loss, one_hot_labels = loss(logit, labels_placeholder)
prediction = tf.nn.softmax(logit)
accuracy = tower_acc(prediction, labels_placeholder)
train_step = tf.train.AdamOptimizer(LEARNING_RATE).minimize(total_loss)
# train_step = tf.train.GradientDescentOptimizer(0.1).minimize(total_loss)
init = tf.global_variables_initializer()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(init)
saver = tf.train.Saver()
save_path = saver.save(sess, "my_net/save_net.ckpt")
print("Save to path: ", save_path)
# sess.run(tf.Print(weights['wc1'],[weights['wc1']],message='wc1:',summarize=100))
plt.ion()
for step in xrange(MAX_STEPS):
start_time = time.time()
train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename='dataset/train_data/',
batch_size=BATCH_SIZE,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
sess.run(train_step,
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
duration = time.time() - start_time
print('Step %d: %.3f sec' % (step, duration))
step_loss = sess.run(total_loss,
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
print("loss: " + "{:.5f}".format(step_loss))
loss_per_step.append(step_loss)
# plt.plot(loss_per_step)
# plt.xlabel('clip')
# plt.ylabel('loss')
# plt.title('16 frames per clip, equal to 66.67ms')
# plt.pause(0.05)
# while True:
# plt.pause(0.05)
# sess.run(tf.Print(images_placeholder, [images_placeholder], message='images_placeholder:', summarize=100),
# feed_dict={images_placeholder: train_images})
# sess.run(tf.Print(labels_placeholder, [labels_placeholder], message='labels:', summarize=100),
# feed_dict={labels_placeholder: train_labels})
# sess.run(tf.Print(one_hot_labels, [labels_placeholder], message='one_hot_labels:', summarize=100),
# feed_dict={labels_placeholder: train_labels})
# sess.run(tf.Print(conv1, [conv1], message='conv1:', summarize=100),
# feed_dict={images_placeholder: train_images})
# sess.run(tf.Print(tf.shape(conv1), [tf.shape(conv1)], message='conv1.shape:', summarize=100),
# feed_dict={images_placeholder: train_images})
# sess.run(tf.Print(pool1, [pool1], message='pool1:', summarize=100),
# feed_dict={images_placeholder: train_images})
# sess.run(tf.Print(tf.shape(pool1), [tf.shape(pool1)], message='pool1.shape:', summarize=100),
# feed_dict={images_placeholder: train_images})
# sess.run(tf.Print(dense1, [dense1], message='dense1:', summarize=100),
# feed_dict={images_placeholder: train_images})
# sess.run(tf.Print(dense2, [dense2], message='dense2:', summarize=100),
# feed_dict={images_placeholder: train_images})
# sess.run(tf.Print(logit, [logit], message='Logit:', summarize=100),
# feed_dict={images_placeholder: train_images})
if (step) % 20 == 0:
print('Training Data Eval:')
test_images, test_labels, _, _, _ = input_data.read_clip_and_label(
filename='dataset/test_data/',
batch_size=BATCH_SIZE,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
acc = sess.run(accuracy,
feed_dict={
images_placeholder: test_images,
labels_placeholder: test_labels
})
print("accuracy: " + "{:.5f}".format(acc))
return loss_per_step
def model_test(model_file, input_tensor_name, output_tensor_name,
test_list_file):
num_test_videos = len(list(open(test_list_file, 'r')))
print("Info: Number of test videos is {}".format(num_test_videos))
with tf.Graph().as_default():
graph_def = tf.GraphDef()
model_f = open(model_file, "rb")
graph_def.ParseFromString(model_f.read())
_ = tf.import_graph_def(graph_def, name='')
# Limit GPU memory as 25%
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.250)
with tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options)) as sess:
init = tf.global_variables_initializer()
sess.run(init)
tensor_input = sess.graph.get_tensor_by_name(input_tensor_name)
tensor_output = sess.graph.get_tensor_by_name(output_tensor_name)
images_placeholder = tensor_input
# Session run
logits = []
logit = tensor_output
logits.append(logit)
logits = tf.concat(logits, 0)
norm_score = tf.nn.softmax(logits)
max_steps = int((num_test_videos - 1) / (FLAGS.batch_size) + 1)
print("Info: Max steps is %d" % max_steps)
true_count = 0
all_count = 0
next_start_pos = 0
all_steps = max_steps
for step in range(all_steps):
start_time = time.time()
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
test_list_file,
FLAGS.batch_size,
start_pos=next_start_pos
)
predict_score = norm_score.eval(
session=sess, feed_dict={images_placeholder: test_images})
for i in range(0, valid_len):
true_label = test_labels[i],
top1_predicted_label = np.argmax(predict_score[i], axis=0)
if (true_label == top1_predicted_label):
true_count = true_count + 1
all_count = all_count + 1
duration = time.time() - start_time
print('Info: Step %d: %.3f sec' % (step, duration))
acc = float(true_count) / all_count
print("Info: Accuracy: " + "{:.5f}".format(acc))
def run_training():
with tf.Graph().as_default():
#Create a variable to count the number of train() calls. This equals the
# number of batches processed * FLAGS.num_gpus.
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0),
trainable=False)
# Get the image and the labels placeholder
images_placeholder, labels_placeholder = placeholder_inputs()
# Calculate the learning rate schedule.
num_batches_per_epoch = (c3d_model.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN /
FLAGS.batch_size)
decay_steps = int(num_batches_per_epoch *
c3d_model.NUM_EPOCHS_PER_DECAY)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(c3d_model.INITIAL_LEARNING_RATE,
global_step,
decay_steps,
c3d_model.LEARNING_RATE_DECAY_FACTOR,
staircase=True)
# Create an optimizer that perfrom Adam algorithm
opt = tf.train.AdamOptimizer(lr)
with tf.name_scope('%s' % (c3d_model.TOWER_NAME)) as scope:
# Calculate the loss and accuracy for one tower for the model. This
# function constructs the entire model but shares the variables
# across all towers.
loss, accuracy = tower_loss_acc(scope, images_placeholder,
labels_placeholder)
# Retain the summaries from the final tower
summaries = tf.get_collection(tf.GraphKeys.SUMMARIES, scope)
# Calculate the gradients for the batch of data on this tower
grads = opt.compute_gradients(loss)
# Add a summary to track the learning rate
summaries.append(tf.summary.scalar('learning_rate', lr))
# Apply the gradients to adjust the shared variables.
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
summaries.append(tf.summary.histogram(var.op.name, var))
# Track the moving averages of all trainable variables
variable_averages = tf.train.ExponentialMovingAverage(
c3d_model.MOVING_AVERAGE_DECAY, global_step)
variables_averages_op = variable_averages.apply(
tf.trainable_variables())
# Group all the updates into a single train op
train_op = tf.group(apply_gradient_op, variables_averages_op)
# Create a saver
vv = []
for v in tf.global_variables():
k = v.name.split("/")
print(k)
if len(k) > 1:
if k[1] != 'Mean':
vv.append(v)
else:
vv.append(v)
#print(v.name.split("/"))
saver = tf.train.Saver(vv)
saver2 = tf.train.Saver(tf.global_variables())
# Build the summary operation from the last tower summaries
summary_op = tf.summary.merge(summaries)
# Build an initialization operation to run below
init = tf.global_variables_initializer()
# Start running operations on the Graph. allow_soft_placement must be set to
# True to build towers on GPU, as some of the ops do not have GPU
# implementations.
sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement))
start_step = 0
# Retore the training model from check point
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
if ckpt and ckpt.model_checkpoint_path:
sess.run(init)
print("Restore the model from checkpoint")
# Restores from checkpoint
saver2.restore(sess, ckpt.model_checkpoint_path)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/cifar10_train/model.ckpt-0,
# extract global_step from it.
start_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
elif os.path.isfile(FLAGS.pretrained_model):
print("Finetunning the model")
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
sess.run(init)
# Variable to restore
variables = {
"var_name/wc1": tf.get_variable('c3d_var/conv1/weight'),
"var_name/wc2": tf.get_variable('c3d_var/conv2/weight'),
"var_name/wc3a": tf.get_variable('c3d_var/conv3/weight_a'),
"var_name/wc3b": tf.get_variable('c3d_var/conv3/weight_b'),
"var_name/wc4a": tf.get_variable('c3d_var/conv4/weight_a'),
"var_name/wc4b": tf.get_variable('c3d_var/conv4/weight_b'),
"var_name/wc5a": tf.get_variable('c3d_var/conv5/weight_a'),
"var_name/wc5b": tf.get_variable('c3d_var/conv5/weight_b'),
"var_name/wd1": tf.get_variable('c3d_var/local6/weights'),
"var_name/wd2": tf.get_variable('c3d_var/local7/weights'),
"var_name/bc1": tf.get_variable('c3d_var/conv1/biases'),
"var_name/bc2": tf.get_variable('c3d_var/conv2/biases'),
"var_name/bc3a": tf.get_variable('c3d_var/conv3/biases_a'),
"var_name/bc3b": tf.get_variable('c3d_var/conv3/biases_b'),
"var_name/bc4a": tf.get_variable('c3d_var/conv4/biases_a'),
"var_name/bc4b": tf.get_variable('c3d_var/conv4/biases_b'),
"var_name/bc5a": tf.get_variable('c3d_var/conv5/biases_a'),
"var_name/bc5b": tf.get_variable('c3d_var/conv5/biases_b'),
"var_name/bd1": tf.get_variable('c3d_var/local6/biases'),
"var_name/bd2": tf.get_variable('c3d_var/local7/biases')
}
saver_c3d = tf.train.Saver(variables)
saver_c3d.restore(sess, FLAGS.pretrained_model)
else:
print("Train the model from scratch")
sess.run(init)
# Initialize the train_writer
train_writer = tf.summary.FileWriter(
os.path.join(FLAGS.train_dir, 'visual_logs', 'train'), sess.graph)
test_writer = tf.summary.FileWriter(
os.path.join(FLAGS.train_dir, 'visual_logs', 'test'), sess.graph)
for step in xrange(int(start_step), FLAGS.max_steps):
start_time = time.time()
# Get the input data
# TODO: Check whether the data exist or not first
train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename='list/new_train.list',
batch_size=FLAGS.batch_size,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
# Train the network
sess.run(train_op,
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
duration = time.time() - start_time
# print('Step %d: %.3f sec' % (step, duration))
# Evaluate the model periodically
if step % 5 == 0:
# Training Evaluation
loss_value, accuracy_value = sess.run(
[loss, accuracy],
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
assert not np.isnan(
loss_value), 'Model diverged with loss = NaN'
# Calculate the efficientcy
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = duration
format_str = (
'(Train) %s: step %d, loss = %.2f, acc = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str %
(datetime.now(), step, loss_value, accuracy_value,
examples_per_sec, sec_per_batch))
# Test Evaluation
print('Testing Data Eval:')
val_images, val_labels, _, _, _ = input_data.read_clip_and_label(
filename='list/new_test.list',
batch_size=20,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
loss_value, accuracy_value = sess.run([loss, accuracy],
feed_dict={
images_placeholder:
val_images,
labels_placeholder:
val_labels
})
assert not np.isnan(
loss_value), 'Model diverged with loss = NaN'
# Calculate the efficientcy
num_examples_per_step = 200
examples_per_sec = num_examples_per_step / duration
sec_per_batch = duration
format_str = (
'(Test) %s: step %d, loss = %.2f, acc = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str %
(datetime.now(), step, loss_value, accuracy_value,
examples_per_sec, sec_per_batch))
if step % 10 == 0:
# Training summary writer
summary = sess.run(summary_op,
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
train_writer.add_summary(summary, step)
# Testing summary writer
summary = sess.run(summary_op,
feed_dict={
images_placeholder: val_images,
labels_placeholder: val_labels
})
test_writer.add_summary(summary, step)
# Save the model checkpoint periodically.
if step % 5 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver2.save(sess, checkpoint_path, global_step=step)
print('Done')
def run_test():
model_name = "./sports1m_finetuning_ucf101.model"
test_list_file = 'list/test.list'
num_test_videos = len(list(open(test_list_file,'r')))
print("Number of test videos={}".format(num_test_videos))
# Get the sets of images and labels for training, validation, and
images_placeholder, labels_placeholder = placeholder_inputs(FLAGS.batch_size * gpu_num)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1': _variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.04, 0.00),
'wc2': _variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.04, 0.00),
'wc3a': _variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256], 0.04, 0.00),
'wc3b': _variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256], 0.04, 0.00),
'wc4a': _variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512], 0.04, 0.00),
'wc4b': _variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512], 0.04, 0.00),
'wc5a': _variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512], 0.04, 0.00),
'wc5b': _variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512], 0.04, 0.00),
'wd1': _variable_with_weight_decay('wd1', [8192, 4096], 0.04, 0.001),
'wd2': _variable_with_weight_decay('wd2', [4096, 4096], 0.04, 0.002),
'out': _variable_with_weight_decay('wout', [4096, c3d_model.NUM_CLASSES], 0.04, 0.005)
}
biases = {
'bc1': _variable_with_weight_decay('bc1', [64], 0.04, 0.0),
'bc2': _variable_with_weight_decay('bc2', [128], 0.04, 0.0),
'bc3a': _variable_with_weight_decay('bc3a', [256], 0.04, 0.0),
'bc3b': _variable_with_weight_decay('bc3b', [256], 0.04, 0.0),
'bc4a': _variable_with_weight_decay('bc4a', [512], 0.04, 0.0),
'bc4b': _variable_with_weight_decay('bc4b', [512], 0.04, 0.0),
'bc5a': _variable_with_weight_decay('bc5a', [512], 0.04, 0.0),
'bc5b': _variable_with_weight_decay('bc5b', [512], 0.04, 0.0),
'bd1': _variable_with_weight_decay('bd1', [4096], 0.04, 0.0),
'bd2': _variable_with_weight_decay('bd2', [4096], 0.04, 0.0),
'out': _variable_with_weight_decay('bout', [c3d_model.NUM_CLASSES], 0.04, 0.0),
}
logits = []
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
logit = c3d_model.inference_c3d(images_placeholder[gpu_index * FLAGS.batch_size:(gpu_index + 1) * FLAGS.batch_size,:,:,:,:], 0.6, FLAGS.batch_size, weights, biases)
logits.append(logit)
logits = tf.concat(logits,0)
norm_score = tf.nn.softmax(logits)
saver = tf.train.Saver()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
# Create a saver for writing training checkpoints.
saver.restore(sess, model_name)
# And then after everything is built, start the training loop.
bufsize = 0
write_file = open("predict_ret.txt", "w+", bufsize)
next_start_pos = 0
all_steps = int((num_test_videos - 1) / (FLAGS.batch_size * gpu_num) + 1)
for step in xrange(all_steps):
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
start_time = time.time()
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
test_list_file,
FLAGS.batch_size * gpu_num,
start_pos=next_start_pos
)
predict_score = norm_score.eval(
session=sess,
feed_dict={images_placeholder: test_images}
)
for i in range(0, valid_len):
true_label = test_labels[i],
top1_predicted_label = np.argmax(predict_score[i])
# Write results: true label, class prob for true label, predicted label, class prob for predicted label
write_file.write('{}, {}, {}, {}\n'.format(
true_label[0],
predict_score[i][true_label],
top1_predicted_label,
predict_score[i][top1_predicted_label]))
write_file.close()
print("done")
def run_test():
model_name = "./sports1m_finetuning_ucf101.model"
test_list_file = './list/test1.list'
num_test_videos = len(list(open(test_list_file, 'r')))
print("Number of test videos={}".format(num_test_videos))
# Get the sets of images and labels for training, validation, and
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1':
_variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.0005),
'wc2':
_variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.0005),
'wc3a':
_variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256], 0.0005),
'wc3b':
_variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256], 0.0005),
'wc4a':
_variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512], 0.0005),
'wc4b':
_variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512], 0.0005),
'wc5a':
_variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512], 0.0005),
'wc5b':
_variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512], 0.0005),
'deconv1':
_variable_with_weight_decay('deconv1', [1, 4, 4, 1, 512], 0.0005),
'deconv2':
_variable_with_weight_decay('deconv2', [1, 3, 3, 1, 1], 0.0005)
}
biases = {
'bc1': _variable_with_weight_decay('bc1', [64], 0.04, 0.0),
'bc2': _variable_with_weight_decay('bc2', [128], 0.04, 0.0),
'bc3a': _variable_with_weight_decay('bc3a', [256], 0.04, 0.0),
'bc3b': _variable_with_weight_decay('bc3b', [256], 0.04, 0.0),
'bc4a': _variable_with_weight_decay('bc4a', [512], 0.04, 0.0),
'bc4b': _variable_with_weight_decay('bc4b', [512], 0.04, 0.0),
'bc5a': _variable_with_weight_decay('bc5a', [512], 0.04, 0.0),
'bc5b': _variable_with_weight_decay('bc5b', [512], 0.04, 0.0)
}
logits = []
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
logit = c3d_model.inference_c3d(
images_placeholder[gpu_index *
FLAGS.batch_size:(gpu_index + 1) *
FLAGS.batch_size, :, :, :, :], 0.6,
FLAGS.batch_size, weights, biases)
logits.append(logit)
logits = tf.concat(0, logits)
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True))
init = tf.initialize_all_variables()
sess.run(init)
next_start_pos = 0
all_steps = int((num_test_videos - 1) / (FLAGS.batch_size * gpu_num) + 1)
predict_images = None
for step in xrange(all_steps):
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
test_list_file,
FLAGS.batch_size * gpu_num,
start_pos=next_start_pos
)
predict_data = sess.run(logits,
feed_dict={images_placeholder: test_images})
if predict_images is None:
predict_images = predict_data
else:
predict_images = np.concatenate((predict_images, predict_data),
axis=0)
print("done")
return predict_images
def run_test():
SAVE_PATH = "output/guided_backpro"
model_name = "/media/storage/liweijie/c3d_models/pbd_fcn_model-1000"
test_list_file = 'list/predict_test.txt'
num_test_videos = len(list(open(test_list_file, 'r')))
print("Number of test videos={}".format(num_test_videos))
# Get the sets of images and labels for training, validation, and
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1':
_variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.0005),
'wc2':
_variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.0005),
'wc3a':
_variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256], 0.0005),
'wc3b':
_variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256], 0.0005),
'wc4a':
_variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512], 0.0005),
'wc4b':
_variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512], 0.0005),
'wc5a':
_variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512], 0.0005),
'wc5b':
_variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512], 0.0005),
'w1':
_variable_with_weight_decay('w1', [1, 4, 4, 512, 4096], 0.0005),
'w2':
_variable_with_weight_decay('w2', [1, 1, 1, 4096, 4096], 0.0005),
#'out': _variable_with_weight_decay('wout', [4096, c3d_model.NUM_CLASSES], 0.0005)
'out_fcn':
_variable_with_weight_decay('wout',
[1, 1, 1, 4096, c3d_model.NUM_CLASSES],
0.0005)
}
biases = {
'bc1':
_variable_with_weight_decay('bc1', [64], 0.000),
'bc2':
_variable_with_weight_decay('bc2', [128], 0.000),
'bc3a':
_variable_with_weight_decay('bc3a', [256], 0.000),
'bc3b':
_variable_with_weight_decay('bc3b', [256], 0.000),
'bc4a':
_variable_with_weight_decay('bc4a', [512], 0.000),
'bc4b':
_variable_with_weight_decay('bc4b', [512], 0.000),
'bc5a':
_variable_with_weight_decay('bc5a', [512], 0.000),
'bc5b':
_variable_with_weight_decay('bc5b', [512], 0.000),
'b1':
_variable_with_weight_decay('b1', [4096], 0.000),
'b2':
_variable_with_weight_decay('b2', [4096], 0.000),
#'out': _variable_with_weight_decay('bout', [c3d_model.NUM_CLASSES], 0.000),
'out_fcn':
_variable_with_weight_decay('bout', [c3d_model.NUM_CLASSES],
0.0005)
}
batch_size = FLAGS.batch_size
with tf.device('/cpu:0'):
out = c3d_model.inference_c3d_full_conv(
images_placeholder[:, :, :, :, :], 1, FLAGS.batch_size, weights,
biases)
logits = []
logits.append(out)
logits = tf.concat(logits, 0)
norm_score = tf.nn.softmax(logits)
saver = tf.train.Saver()
#sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
# Create a saver for writing training checkpoints.
saver.restore(sess, model_name)
# And then after everything is built, start the training loop.
bufsize = 0
next_start_pos = 0
#all_steps = int((num_test_videos - 1) / (FLAGS.batch_size * gpu_num) + 1)
all_steps = 6
for step in xrange(all_steps):
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
start_time = time.time()
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
test_list_file,
FLAGS.batch_size * gpu_num,
start_pos=next_start_pos
)
label = step
length = 16
# guided_backpro
model = GuideBackPro(weights, biases, vis_model=c3d_model)
back_pro_op = model.get_visualization(images_placeholder)
guided_backpro = sess.run(
back_pro_op, feed_dict={images_placeholder: test_images})
guided_backpro_img = guided_backpro[0][0][0]
print("the shape of guided_backpro_img is:",
guided_backpro_img.shape)
print("max of img is:", np.max(guided_backpro_img))
guided_backpro_img = guided_backpro_img / np.max(
guided_backpro_img)
make_path = "{}/{}".format(SAVE_PATH, label)
if not os.path.exists(make_path):
os.makedirs(make_path)
for l in range(length):
misc.imsave(
"{}/{}/{}.jpg".format(SAVE_PATH, label, l),
np.reshape(guided_backpro_img[0, l, :, :, :],
[112, 112, 3]))
#plt.show()
print("done")
def run_test():
tf.reset_default_graph()
init_path = 'D:/autism/3DCNN-master/c3d_ucf_model-2520'
for i in os.listdir(init_path):
model_name = init_path + '/' + 'c3d_ucf_model-2520'
print(model_name)
test_list_file = 'D:/autism/3DCNN-master/C3D-tensorflow/list/test_list.list'
num_test_videos = len(list(open(test_list_file, 'r')))
print("Number of test videos={}".format(num_test_videos))
# Get the sets of images and labels for training, validation, and
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1':
_variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.04,
0.00),
'wc2':
_variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.04,
0.00),
'wc3a':
_variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256], 0.04,
0.00),
'wc3b':
_variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256], 0.04,
0.00),
'wc4a':
_variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512], 0.04,
0.00),
'wc4b':
_variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512], 0.04,
0.00),
'wc5a':
_variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512], 0.04,
0.00),
'wc5b':
_variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512], 0.04,
0.00),
'wd1':
_variable_with_weight_decay('wd1', [8192, 4096], 0.04, 0.001),
'wd2':
_variable_with_weight_decay('wd2', [4096, 4096], 0.04, 0.002),
'out':
_variable_with_weight_decay('wout', [4096, 8], 0.04, 0.005)
}
biases = {
'bc1': _variable_with_weight_decay('bc1', [64], 0.04, 0.0),
'bc2': _variable_with_weight_decay('bc2', [128], 0.04, 0.0),
'bc3a': _variable_with_weight_decay('bc3a', [256], 0.04, 0.0),
'bc3b': _variable_with_weight_decay('bc3b', [256], 0.04, 0.0),
'bc4a': _variable_with_weight_decay('bc4a', [512], 0.04, 0.0),
'bc4b': _variable_with_weight_decay('bc4b', [512], 0.04, 0.0),
'bc5a': _variable_with_weight_decay('bc5a', [512], 0.04, 0.0),
'bc5b': _variable_with_weight_decay('bc5b', [512], 0.04, 0.0),
'bd1': _variable_with_weight_decay('bd1', [4096], 0.04, 0.0),
'bd2': _variable_with_weight_decay('bd2', [4096], 0.04, 0.0),
'out': _variable_with_weight_decay('bout', [8], 0.04, 0.0),
}
logits = []
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
logit = c3d_model.inference_c3d(
images_placeholder[gpu_index *
FLAGS.batch_size:(gpu_index + 1) *
FLAGS.batch_size, :, :, :, :], 0.6,
FLAGS.batch_size, weights, biases)
logits.append(logit)
logits = tf.concat(logits, 0)
norm_score = tf.nn.softmax(logits)
saver = tf.train.Saver()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
# Create a saver for writing training checkpoints.
saver.restore(sess, model_name)
# And then after everything is built, start the training loop.
bufsize = 0
write_file = open("D:/autism/3DCNN-master/predict_" + str(i) + ".txt",
"w+")
next_start_pos = -1
all_steps = int((num_test_videos - 1) / (FLAGS.batch_size * gpu_num) +
1)
for step in xrange(all_steps):
start_time = time.time()
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
test_list_file, FLAGS.batch_size * gpu_num, start_pos=next_start_pos
)
#print(test_images[0],test_labels[0])
predict_score = norm_score.eval(
session=sess, feed_dict={images_placeholder: test_images})
for i in range(0, valid_len):
true_label = test_labels[i],
top1_predicted_label = np.argmax(predict_score[i])
# Write results: true label, class prob for true label, predicted label, class prob for predicted label
write_file.write('{}, {}, {}, {}\n'.format(
true_label[0], predict_score[i][true_label],
top1_predicted_label,
predict_score[i][top1_predicted_label]))
write_file.close()
print("done for " + str(i))
best_acc = 0
for step in range(start_steps,max_steps):
start_time = time.time()
if epoch>=20:
# open data augmentation
status = 'TRAIN'
else:
# close data augmentation
status = 'TEST'
startprocess_time = time.time()
train_images, train_labels, next_batch_start, _, _,lines = input_data.read_clip_and_label(
rootdir = train_root,
filename= train_txt,
batch_size=batchsize,
lines=lines,
start_pos=next_batch_start,
num_frames_per_clip=time_steps,
crop_size=(CNNLSTM.HEIGHT,CNNLSTM.WIDTH),
shuffle=False,
phase=status
)
train_images = train_images.reshape([-1,CNNLSTM.HEIGHT,CNNLSTM.WIDTH,CNNLSTM.CHANNELS])
endprocess_time = time.time()
preprocess_time = ((endprocess_time-startprocess_time)/(batchsize*1))
print("preprocess per time :%f"%preprocess_time)
_,losses = sess.run([train_op,loss], feed_dict={
X: train_images,
Y: train_labels
})
##train_writer.add_summary(summary, step)
duration = time.time() - start_time
def run_training():
# Get the sets of images and labels for training, validation, and
# Tell TensorFlow that the model will be built into the default Graph.
# Create model directory
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
flow_pre_model_save_dir = "/home/project/I3D/I3D/checkpoints/flow_imagenet"
with tf.Graph().as_default():
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0),
trainable=False)
rgb_images_placeholder, flow_images_placeholder, labels_placeholder, is_training = placeholder_inputs(
FLAGS.batch_size * gpu_num, FLAGS.num_frame_per_clib,
FLAGS.crop_size, FLAGS.rgb_channels, FLAGS.flow_channels)
learning_rate = tf.train.exponential_decay(FLAGS.learning_rate,
global_step,
decay_steps=2000,
decay_rate=0.1,
staircase=True)
opt_flow = tf.train.AdamOptimizer(learning_rate)
#opt_stable = tf.train.MomentumOptimizer(learning_rate, 0.9)
with tf.variable_scope('Flow'):
flow_logit, _ = InceptionI3d(
num_classes=FLAGS.classics,
spatial_squeeze=True,
final_endpoint='Logits')(flow_images_placeholder, is_training)
flow_loss = tower_loss(flow_logit, labels_placeholder)
accuracy = tower_acc(flow_logit, labels_placeholder)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
flow_grads = opt_flow.compute_gradients(flow_loss)
apply_gradient_flow = opt_flow.apply_gradients(
flow_grads, global_step=global_step)
train_op = tf.group(apply_gradient_flow)
null_op = tf.no_op()
# Create a saver for loading trained checkpoints.
flow_variable_map = {}
for variable in tf.global_variables():
if variable.name.split(
'/')[0] == 'Flow' and 'Adam' not in variable.name.split(
'/')[-1] and variable.name.split('/')[2] != 'Logits':
flow_variable_map[variable.name.replace(':0', '')] = variable
flow_saver = tf.train.Saver(var_list=flow_variable_map, reshape=True)
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
init = tf.global_variables_initializer()
# Create a session for running Ops on the Graph.
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(init)
# Create summary writter
tf.summary.scalar('accuracy', accuracy)
tf.summary.scalar('flow_loss', flow_loss)
tf.summary.scalar('learning_rate', learning_rate)
merged = tf.summary.merge_all()
# load pre_train models
ckpt = tf.train.get_checkpoint_state(flow_pre_model_save_dir)
if ckpt and ckpt.model_checkpoint_path:
print("loading checkpoint %s,waiting......" %
ckpt.model_checkpoint_path)
flow_saver.restore(sess, ckpt.model_checkpoint_path)
print("load complete!")
train_writer = tf.summary.FileWriter(
'./visual_logs/train_HR_flow_imagenet_6000_6_64_0.0001_decay_split1',
sess.graph)
test_writer = tf.summary.FileWriter(
'./visual_logs/test_HR_flow_imagenet_6000_6_64_0.0001_decay_split1',
sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
rgb_train_images, flow_train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename='../../list/ixmas_list/trainlist_3.list',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=FLAGS.num_frame_per_clib,
crop_size=FLAGS.crop_size,
shuffle=True)
sess.run(train_op,
feed_dict={
flow_images_placeholder: flow_train_images,
labels_placeholder: train_labels,
is_training: True
})
duration = time.time() - start_time
print('Step %d: %.3f sec' % (step, duration))
# Save a checkpoint and evaluate the model periodically.
if step % 10 == 0 or (step + 1) == FLAGS.max_steps:
print('Training Data Eval:')
summary, acc, loss_flow = sess.run(
[merged, accuracy, flow_loss],
feed_dict={
flow_images_placeholder: flow_train_images,
labels_placeholder: train_labels,
is_training: False
})
print("accuracy: " + "{:.5f}".format(acc))
print("flow_loss: " + "{:.5f}".format(loss_flow))
train_writer.add_summary(summary, step)
print('Validation Data Eval:')
rgb_val_images, flow_val_images, val_labels, _, _, _ = input_data.read_clip_and_label(
filename='../../list/ixmas_list/testlist_3.list',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=FLAGS.num_frame_per_clib,
crop_size=FLAGS.crop_size,
shuffle=True)
summary, acc = sess.run(
[merged, accuracy],
feed_dict={
flow_images_placeholder: flow_val_images,
labels_placeholder: val_labels,
is_training: False
})
print("accuracy: " + "{:.5f}".format(acc))
test_writer.add_summary(summary, step)
if (step + 1) % 3000 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess,
os.path.join(model_save_dir, 'i3d_ixmas_model'),
global_step=step)
print("done")
# Add ops to save and restore all the variables.
saver = tf.train.Saver()
saver.restore(sess, model)
print("Model restored.")
write_file = open("predict_ret.txt", "w+")
# metrics_ = []
# labels_ = []
# predicted_ = []
acc_cnt, acc5_cnt, cnt = 0, 0, 1
next_start_pos = 0
for step in range(number_of_steps_per_epoch):
test_images, test_labels, next_start_pos, _, valid_len = input_data.read_clip_and_label(
filename=test_list_file,
batch_size=options['batch_size'],
num_frames_per_clip=24,
crop_size=options['crop_size'],
start_pos=next_start_pos,
shuffle=True)
predict_score = norm_score.eval(
session=sess, feed_dict={images_placeholder: test_images})
acc5 = tf.nn.in_top_k(predict_score, test_labels, 5)
top5_score = acc5.eval(session=sess,
feed_dict={images_placeholder: test_images})
for i in range(0, valid_len):
true_label = test_labels[i]
top1_predicted_label = np.argmax(predict_score[i])
# Write results: true label, class prob for true label, predicted label, class prob for predicted label
write_file.write('{}, {}, {}, {}\n'.format(
true_label, predict_score[i][true_label], top1_predicted_label,
def run_training():
# Get the sets of images and labels for training, validation, and
# Tell TensorFlow that the model will be built into the default Graph.
# Create model directory
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
use_pretrained_model = False
model_filename = "./sports1m_finetuning_ucf101.model"
with tf.Graph().as_default():
global_step = tf.get_variable(
'global_step',
[],
initializer=tf.constant_initializer(0),
trainable=False
)
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num
)
tower_grads1 = []
tower_grads2 = []
logits = []
opt_stable = tf.train.AdamOptimizer(LEARNING_RATE_STABLE)
opt_finetuning = tf.train.AdamOptimizer(LEARNING_RATE_FINETUNE)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1': _variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.0005),
'wc2': _variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.0005),
'wc3a': _variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256], 0.0005),
'wc3b': _variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256], 0.0005),
'wc4a': _variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512], 0.0005),
'wc4b': _variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512], 0.0005),
'wc5a': _variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512], 0.0005),
'wc5b': _variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512], 0.0005),
'wd1': _variable_with_weight_decay('wd1', [8192, 4096], 0.0005),
#'wd1': _variable_with_weight_decay('wd1', [16384, 4096], 0.0005),
'wd2': _variable_with_weight_decay('wd2', [4096, 4096], 0.0005),
'out': _variable_with_weight_decay('wout', [4096, c3d_model.NUM_CLASSES], 0.0005)
}
biases = {
'bc1': _variable_with_weight_decay('bc1', [64], 0.000),
'bc2': _variable_with_weight_decay('bc2', [128], 0.000),
'bc3a': _variable_with_weight_decay('bc3a', [256], 0.000),
'bc3b': _variable_with_weight_decay('bc3b', [256], 0.000),
'bc4a': _variable_with_weight_decay('bc4a', [512], 0.000),
'bc4b': _variable_with_weight_decay('bc4b', [512], 0.000),
'bc5a': _variable_with_weight_decay('bc5a', [512], 0.000),
'bc5b': _variable_with_weight_decay('bc5b', [512], 0.000),
'bd1': _variable_with_weight_decay('bd1', [4096], 0.000),
'bd2': _variable_with_weight_decay('bd2', [4096], 0.000),
'out': _variable_with_weight_decay('bout', [c3d_model.NUM_CLASSES], 0.000),
}
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
varlist2 = [ weights['out'],biases['out'] ]
varlist1 = list( set(list(weights.values()) + list(biases.values())) - set(varlist2) )
logit = c3d_model.inference_c3d(
images_placeholder[gpu_index * FLAGS.batch_size:(gpu_index + 1) * FLAGS.batch_size,:,:,:,:],
0.5,
FLAGS.batch_size,
weights,
biases,
use_pretrained_model
)
loss_name_scope = ('gpud_%d_loss' % gpu_index)
loss = tower_loss(
loss_name_scope,
logit,
labels_placeholder[gpu_index * FLAGS.batch_size:(gpu_index + 1) * FLAGS.batch_size]
)
grads1 = opt_stable.compute_gradients(loss, varlist1)
grads2 = opt_finetuning.compute_gradients(loss, varlist2)
tower_grads1.append(grads1)
tower_grads2.append(grads2)
logits.append(logit)
logits = tf.concat(logits,0)
accuracy = tower_acc(logits, labels_placeholder)
tf.summary.scalar('accuracy', accuracy)
grads1 = average_gradients(tower_grads1)
grads2 = average_gradients(tower_grads2)
apply_gradient_op1 = opt_stable.apply_gradients(grads1)
apply_gradient_op2 = opt_finetuning.apply_gradients(grads2, global_step=global_step)
variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
train_op = tf.group(apply_gradient_op1, apply_gradient_op2, variables_averages_op)
null_op = tf.no_op()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver(list(weights.values()) + list(biases.values()))
init = tf.global_variables_initializer()
# Create a session for running Ops on the Graph.
sess = tf.Session(
config=tf.ConfigProto(allow_soft_placement=True)
)
sess.run(init)
if os.path.isfile(model_filename) and use_pretrained_model:
saver.restore(sess, model_filename)
# Create summary writter
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('./visual_logs/train', sess.graph)
test_writer = tf.summary.FileWriter('./visual_logs/test', sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
train_images, train_labels, _, _, _, valid_len = input_data.read_clip_and_label(
# filename='list/trainlist01.txt',
filename='../../ucf101_all_frames/train-test-splits/trainlist01-hyperion.txt',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True,
flip_with_probability=0.5,
pad_short_clips=PAD_SHORT_CLIPS
)
sess.run(train_op, feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
duration = time.time() - start_time
print('Step %d: %.3f sec, valid_len = %s' % (step, duration, valid_len))
# Save a checkpoint and evaluate the model periodically.
if (step) % 10 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess, os.path.join(model_save_dir, 'c3d_ucf_model'), global_step=step)
print('Training Data Eval:')
summary, acc = sess.run(
[merged, accuracy],
feed_dict={images_placeholder: train_images,
labels_placeholder: train_labels
})
print ("accuracy: " + "{:.5f}".format(acc))
train_writer.add_summary(summary, step)
print('Validation Data Eval:')
val_images, val_labels, _, _, _, _ = input_data.read_clip_and_label(
#filename='list/testlist01.txt',
filename='../../ucf101_all_frames/train-test-splits/testlist01-hyperion.txt',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True,
pad_short_clips=PAD_SHORT_CLIPS
)
summary, acc = sess.run(
[merged, accuracy],
feed_dict={
images_placeholder: val_images,
labels_placeholder: val_labels
})
print ("accuracy: " + "{:.5f}".format(acc))
test_writer.add_summary(summary, step)
print("done")
def run_training():
# Get the sets of images and labels for training, validation, and
# Tell TensorFlow that the model will be built into the default Graph.
# Create model directory
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
use_pretrained_model = True
model_filename = "./sports1m_finetuning_ucf101.model"
with tf.Graph().as_default():
global_step = tf.get_variable(
'global_step',
[],
initializer=tf.constant_initializer(0),
trainable=False
)
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num
)
tower_grads1 = []
tower_grads2 = []
logits = []
opt_stable = tf.train.AdamOptimizer(1e-4)
opt_finetuning = tf.train.AdamOptimizer(1e-3)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1': _variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.0005),
'wc2': _variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.0005),
'wc3a': _variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256], 0.0005),
'wc3b': _variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256], 0.0005),
'wc4a': _variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512], 0.0005),
'wc4b': _variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512], 0.0005),
'wc5a': _variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512], 0.0005),
'wc5b': _variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512], 0.0005),
'wd1': _variable_with_weight_decay('wd1', [8192, 4096], 0.0005),
'wd2': _variable_with_weight_decay('wd2', [4096, 4096], 0.0005),
'out': _variable_with_weight_decay('wout', [4096, c3d_model.NUM_CLASSES], 0.0005)
}
biases = {
'bc1': _variable_with_weight_decay('bc1', [64], 0.000),
'bc2': _variable_with_weight_decay('bc2', [128], 0.000),
'bc3a': _variable_with_weight_decay('bc3a', [256], 0.000),
'bc3b': _variable_with_weight_decay('bc3b', [256], 0.000),
'bc4a': _variable_with_weight_decay('bc4a', [512], 0.000),
'bc4b': _variable_with_weight_decay('bc4b', [512], 0.000),
'bc5a': _variable_with_weight_decay('bc5a', [512], 0.000),
'bc5b': _variable_with_weight_decay('bc5b', [512], 0.000),
'bd1': _variable_with_weight_decay('bd1', [4096], 0.000),
'bd2': _variable_with_weight_decay('bd2', [4096], 0.000),
'out': _variable_with_weight_decay('bout', [c3d_model.NUM_CLASSES], 0.000),
}
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
varlist2 = [ weights['out'],biases['out'] ]
varlist1 = list( set(weights.values() + biases.values()) - set(varlist2) )
logit = c3d_model.inference_c3d(
images_placeholder[gpu_index * FLAGS.batch_size:(gpu_index + 1) * FLAGS.batch_size,:,:,:,:],
0.5,
FLAGS.batch_size,
weights,
biases
)
loss_name_scope = ('gpud_%d_loss' % gpu_index)
loss = tower_loss(
loss_name_scope,
logit,
labels_placeholder[gpu_index * FLAGS.batch_size:(gpu_index + 1) * FLAGS.batch_size]
)
grads1 = opt_stable.compute_gradients(loss, varlist1)
grads2 = opt_finetuning.compute_gradients(loss, varlist2)
tower_grads1.append(grads1)
tower_grads2.append(grads2)
logits.append(logit)
logits = tf.concat(logits,0)
accuracy = tower_acc(logits, labels_placeholder)
tf.summary.scalar('accuracy', accuracy)
grads1 = average_gradients(tower_grads1)
grads2 = average_gradients(tower_grads2)
apply_gradient_op1 = opt_stable.apply_gradients(grads1)
apply_gradient_op2 = opt_finetuning.apply_gradients(grads2, global_step=global_step)
variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
train_op = tf.group(apply_gradient_op1, apply_gradient_op2, variables_averages_op)
null_op = tf.no_op()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver(weights.values() + biases.values())
init = tf.global_variables_initializer()
# Create a session for running Ops on the Graph.
sess = tf.Session(
config=tf.ConfigProto(allow_soft_placement=True)
)
sess.run(init)
if os.path.isfile(model_filename) and use_pretrained_model:
saver.restore(sess, model_filename)
# Create summary writter
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('./visual_logs/train', sess.graph)
test_writer = tf.summary.FileWriter('./visual_logs/test', sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename='list/train.list',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True
)
sess.run(train_op, feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
duration = time.time() - start_time
print('Step %d: %.3f sec' % (step, duration))
# Save a checkpoint and evaluate the model periodically.
if (step) % 10 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess, os.path.join(model_save_dir, 'c3d_ucf_model'), global_step=step)
print('Training Data Eval:')
summary, acc = sess.run(
[merged, accuracy],
feed_dict={images_placeholder: train_images,
labels_placeholder: train_labels
})
print ("accuracy: " + "{:.5f}".format(acc))
train_writer.add_summary(summary, step)
print('Validation Data Eval:')
val_images, val_labels, _, _, _ = input_data.read_clip_and_label(
filename='list/test.list',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True
)
summary, acc = sess.run(
[merged, accuracy],
feed_dict={
images_placeholder: val_images,
labels_placeholder: val_labels
})
print ("accuracy: " + "{:.5f}".format(acc))
test_writer.add_summary(summary, step)
print("done")
def run_test():
model_name = "./sports1m_finetuning_ucf101.model"
test_list_file = 'list/test.list'
num_test_videos = len(list(open(test_list_file, 'r')))
print("Number of test videos={}".format(num_test_videos))
# Get the sets of images and labels for training, validation, and
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num)
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1':
_variable_with_weight_decay('wc1', [3, 3, 3, 3, 64], 0.04, 0.00),
'wc2':
_variable_with_weight_decay('wc2', [3, 3, 3, 64, 128], 0.04, 0.00),
'wc3a':
_variable_with_weight_decay('wc3a', [3, 3, 3, 128, 256], 0.04,
0.00),
'wc3b':
_variable_with_weight_decay('wc3b', [3, 3, 3, 256, 256], 0.04,
0.00),
'wc4a':
_variable_with_weight_decay('wc4a', [3, 3, 3, 256, 512], 0.04,
0.00),
'wc4b':
_variable_with_weight_decay('wc4b', [3, 3, 3, 512, 512], 0.04,
0.00),
'wc5a':
_variable_with_weight_decay('wc5a', [3, 3, 3, 512, 512], 0.04,
0.00),
'wc5b':
_variable_with_weight_decay('wc5b', [3, 3, 3, 512, 512], 0.04,
0.00),
'wd1':
_variable_with_weight_decay('wd1', [8192, 4096], 0.04, 0.001),
'wd2':
_variable_with_weight_decay('wd2', [4096, 4096], 0.04, 0.002),
'out':
_variable_with_weight_decay('wout', [4096, c3d_model.NUM_CLASSES],
0.04, 0.005)
}
biases = {
'bc1':
_variable_with_weight_decay('bc1', [64], 0.04, 0.0),
'bc2':
_variable_with_weight_decay('bc2', [128], 0.04, 0.0),
'bc3a':
_variable_with_weight_decay('bc3a', [256], 0.04, 0.0),
'bc3b':
_variable_with_weight_decay('bc3b', [256], 0.04, 0.0),
'bc4a':
_variable_with_weight_decay('bc4a', [512], 0.04, 0.0),
'bc4b':
_variable_with_weight_decay('bc4b', [512], 0.04, 0.0),
'bc5a':
_variable_with_weight_decay('bc5a', [512], 0.04, 0.0),
'bc5b':
_variable_with_weight_decay('bc5b', [512], 0.04, 0.0),
'bd1':
_variable_with_weight_decay('bd1', [4096], 0.04, 0.0),
'bd2':
_variable_with_weight_decay('bd2', [4096], 0.04, 0.0),
'out':
_variable_with_weight_decay('bout', [c3d_model.NUM_CLASSES], 0.04,
0.0),
}
logits = []
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
logit = c3d_model.inference_c3d(
images_placeholder[gpu_index *
FLAGS.batch_size:(gpu_index + 1) *
FLAGS.batch_size, :, :, :, :], 0.6,
FLAGS.batch_size, weights, biases)
logits.append(logit)
logits = tf.concat(logits, 0)
norm_score = tf.nn.softmax(logits)
saver = tf.train.Saver()
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Create a saver for writing training checkpoints.
saver.restore(sess, model_name)
# And then after everything is built, start the training loop.
bufsize = 0
write_file = open("predict_ret.txt", "w+")
next_start_pos = 0
all_steps = int((num_test_videos - 1) / (FLAGS.batch_size * gpu_num) + 1)
accuracy, cnt = 0, 0
for step in xrange(all_steps):
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
start_time = time.time()
test_images, test_labels, next_start_pos, _, valid_len = \
input_data.read_clip_and_label(
test_list_file,
FLAGS.batch_size * gpu_num,
start_pos=next_start_pos
)
predict_score = norm_score.eval(
session=sess, feed_dict={images_placeholder: test_images})
for i in range(0, valid_len):
true_label = test_labels[i],
top1_predicted_label = np.argmax(predict_score[i])
# Write results: true label, class prob for true label, predicted label, class prob for predicted label
write_file.write('{}, {}, {}, {}\n'.format(
true_label[0], predict_score[i][true_label],
top1_predicted_label, predict_score[i][top1_predicted_label]))
cnt += 1
if true_label[0] == top1_predicted_label:
accuracy += 1
print("Test Accuracy={}".format(float(accuracy) / float(cnt)))
write_file.close()
print("done")
def run_training():
# Get the sets of images and labels for training, validation, and
# Tell TensorFlow that the model will be built into the default Graph.
# Create model directory
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
use_pretrained_model = False
model_filename = "./sports1m_finetuning_ucf101.model"
train_out = "train_out.txt"
val_out = "val_out.txt"
if os.path.exists(train_out):
os.remove(train_out)
if os.path.exists(val_out):
os.remove(val_out)
with tf.Graph().as_default():
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0),
trainable=False)
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size * gpu_num)
tower_grads1 = []
tower_grads2 = []
logits = []
opt1 = tf.train.AdamOptimizer(1e-4)
opt2 = tf.train.AdamOptimizer(2e-4)
for gpu_index in range(0, gpu_num):
with tf.device('/gpu:%d' % gpu_index):
with tf.name_scope('%s_%d' %
('dextro-research', gpu_index)) as scope:
with tf.variable_scope('var_name') as var_scope:
weights = {
'wc1':
_variable_with_weight_decay(
'wc1', [3, 3, 3, 3, 64], 0.0005),
'wc2':
_variable_with_weight_decay(
'wc2', [3, 3, 3, 64, 128], 0.0005),
'wc3a':
_variable_with_weight_decay(
'wc3a', [3, 3, 3, 128, 256], 0.0005),
'wc3b':
_variable_with_weight_decay(
'wc3b', [3, 3, 3, 256, 256], 0.0005),
'wc4a':
_variable_with_weight_decay(
'wc4a', [3, 3, 3, 256, 512], 0.0005),
'wc4b':
_variable_with_weight_decay(
'wc4b', [3, 3, 3, 512, 512], 0.0005),
'wc5a':
_variable_with_weight_decay(
'wc5a', [3, 3, 3, 512, 512], 0.0005),
'wc5b':
_variable_with_weight_decay(
'wc5b', [3, 3, 3, 512, 512], 0.0005),
'wd1':
_variable_with_weight_decay(
'wd1', [8192, 4096], 0.0005),
'wd2':
_variable_with_weight_decay(
'wd2', [4096, 4096], 0.0005),
'out':
_variable_with_weight_decay(
'wout', [4096, c3d_model.NUM_CLASSES], 0.0005)
}
biases = {
'bc1':
_variable_with_weight_decay('bc1', [64], 0.000),
'bc2':
_variable_with_weight_decay('bc2', [128], 0.000),
'bc3a':
_variable_with_weight_decay('bc3a', [256], 0.000),
'bc3b':
_variable_with_weight_decay('bc3b', [256], 0.000),
'bc4a':
_variable_with_weight_decay('bc4a', [512], 0.000),
'bc4b':
_variable_with_weight_decay('bc4b', [512], 0.000),
'bc5a':
_variable_with_weight_decay('bc5a', [512], 0.000),
'bc5b':
_variable_with_weight_decay('bc5b', [512], 0.000),
'bd1':
_variable_with_weight_decay('bd1', [4096], 0.000),
'bd2':
_variable_with_weight_decay('bd2', [4096], 0.000),
'out':
_variable_with_weight_decay(
'bout', [c3d_model.NUM_CLASSES], 0.000),
}
varlist1 = weights.values()
varlist2 = biases.values()
logit = c3d_model.inference_c3d(
images_placeholder[gpu_index *
FLAGS.batch_size:(gpu_index + 1) *
FLAGS.batch_size, :, :, :, :], 0.5,
FLAGS.batch_size, weights, biases)
loss = tower_loss(
scope, logit,
labels_placeholder[gpu_index *
FLAGS.batch_size:(gpu_index + 1) *
FLAGS.batch_size])
grads1 = opt1.compute_gradients(loss, varlist1)
grads2 = opt2.compute_gradients(loss, varlist2)
tower_grads1.append(grads1)
tower_grads2.append(grads2)
logits.append(logit)
tf.get_variable_scope().reuse_variables()
logits = tf.concat(0, logits)
accuracy = tower_acc(logits, labels_placeholder)
tf.scalar_summary('accuracy', accuracy)
grads1 = average_gradients(tower_grads1)
grads2 = average_gradients(tower_grads2)
apply_gradient_op1 = opt1.apply_gradients(grads1)
apply_gradient_op2 = opt2.apply_gradients(grads2,
global_step=global_step)
variable_averages = tf.train.ExponentialMovingAverage(
MOVING_AVERAGE_DECAY)
variables_averages_op = variable_averages.apply(
tf.trainable_variables())
train_op = tf.group(apply_gradient_op1, apply_gradient_op2,
variables_averages_op)
null_op = tf.no_op()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver(weights.values() + biases.values())
init = tf.initialize_all_variables()
# Create a session for running Ops on the Graph.
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True))
sess.run(init)
if os.path.isfile(model_filename) and use_pretrained_model:
saver.restore(sess, model_filename)
# Create summary writter
merged = tf.merge_all_summaries()
train_writer = tf.train.SummaryWriter('./visual_logs/train',
sess.graph)
test_writer = tf.train.SummaryWriter('./visual_logs/test', sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
train_images, train_labels, _, _, _ = input_data.read_clip_and_label(
filename='train.list',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
sess.run(train_op,
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
duration = time.time() - start_time
print('Step %d: %.3f sec' % (step, duration))
# Save a checkpoint and evaluate the model periodically.
if (step) % 10 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess,
os.path.join(model_save_dir, 'c3d_ucf_model'),
global_step=step)
print('Training Data Eval:')
summary, acc = sess.run(
[merged, accuracy],
feed_dict={
images_placeholder: train_images,
labels_placeholder: train_labels
})
print("accuracy: " + "{:.5f}".format(acc))
str1 = str(step) + ", " + "{:.5f}".format(acc)
out1 = open(train_out, 'a')
print >> out1, str1
out1.close()
train_writer.add_summary(summary, step)
print('Validation Data Eval:')
val_images, val_labels, _, _, _ = input_data.read_clip_and_label(
filename='test.list',
batch_size=FLAGS.batch_size * gpu_num,
num_frames_per_clip=c3d_model.NUM_FRAMES_PER_CLIP,
crop_size=c3d_model.CROP_SIZE,
shuffle=True)
summary, acc = sess.run([merged, accuracy],
feed_dict={
images_placeholder: val_images,
labels_placeholder: val_labels
})
print("accuracy: " + "{:.5f}".format(acc))
str2 = str(step) + ", " + "{:.5f}".format(acc)
out2 = open(val_out, 'a')
print >> out2, str2
out2.close()
test_writer.add_summary(summary, step)
print("done")
