def main(unused_argv):
with tf.device('/gpu:2'):
x_input = tf.placeholder(tf.float32, shape=(None, 3, 1, 1, 1024))
y_ = tf.placeholder(tf.float32, shape=(None, _NUM_CLASSES))
end_point = 'Logits'
with tf.variable_scope(end_point):
if TRAINING:
x_input = tf.nn.dropout(x_input, 0.7)
logits = i3d.Unit3D(output_channels=_NUM_CLASSES,
kernel_shape=[1, 1, 1],
activation_fn=None,
use_batch_norm=False,
use_bias=True,
name='Conv3d_0c_1x1')(x_input,
is_training=True)
logits = tf.squeeze(logits, [2, 3], name='SpatialSqueeze')
averaged_logits = tf.reduce_mean(logits, axis=1)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
logits=averaged_logits, labels=y_)
loss = tf.reduce_mean(cross_entropy)
optimizer_op = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(loss=loss)
predictions = tf.nn.softmax(averaged_logits)
model_saver = tf.train.Saver()
init = tf.global_variables_initializer()
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init)
i = 0
T_batch, t_batch = test_data(data_dir, 'Cam04')
feed_dict_test = {x_input: T_batch, y_: t_batch}
for epoch in range(epoch_num):
for X_batch, y_batch in generator_data(data_dir, '2'):
i = i + 1
feed_dict_train = {x_input: X_batch, y_: y_batch}
op_, loss_, pre_train = sess.run(
[optimizer_op, loss, predictions],
feed_dict=feed_dict_train)
train_acc = accuracy(pre_train, y_batch)
print(
"After %d epoch %d training step(s), train accuracy using average model is %g "
% (epoch, i, train_acc))
print("Train_loss is %f " % (loss_ * 1.0))
pre_test = sess.run(predictions, feed_dict=feed_dict_test)
test_acc = train_acc = accuracy(pre_test, t_batch)
print(
"~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~"
)
print(
"After %d epoch, validate accuracy using average model is %g "
% (epoch, test_acc))
print(
"~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~"
)
model_saver.save(sess, save_model_path)
print("OK!!!")
def _build_stream(stream_name, is_training):
dims = 3 if stream_name is 'RGB' else 2
input_shape = (_BATCH_SIZE, NUM_FRAMES, IMAGE_SIZE, IMAGE_SIZE, dims)
inp = tf.placeholder(tf.float32, shape=input_shape)
with tf.variable_scope(stream_name):
model = i3d.InceptionI3d(spatial_squeeze=True,
final_endpoint='Mixed_5c')
# No training here. Don't backpropagate the main model, and no dropout.
mixed_5c, _ = model(inp, is_training=False, dropout_keep_prob=1.0)
var_map = {}
for var in tf.global_variables():
if var.name.split('/')[0] == stream_name:
var_map[var.name.replace(':0', '')] = var
saver = tf.train.Saver(var_list=var_map, reshape=True)
with tf.variable_scope(stream_name):
net = tf.nn.avg_pool3d(mixed_5c,
ksize=[1, 2, 7, 7, 1],
strides=[1, 1, 1, 1, 1],
padding=snt.VALID)
logit_fn = i3d.Unit3D(output_channels=NUM_CLASSES,
kernel_shape=[1, 1, 1],
activation_fn=None,
use_batch_norm=False,
use_bias=True,
regularizers={'w': tf.nn.l2_loss},
name='Conv3d_0c_1x1')
logits = logit_fn(net, is_training=is_training)
logits = tf.squeeze(logits, [2, 3], name='SpatialSqueeze')
logits = tf.reduce_mean(logits, axis=1)
custom_vars = {}
for var in tf.global_variables():
name = var.name.replace(':0', '')
if var.name.split(
'/')[0] == stream_name and name not in var_map.keys():
custom_vars[name] = var
with tf.name_scope(name):
_variable_summaries(var)
return (inp, logits, saver, custom_vars)
name='groundtruth_input')
#global_step
global_step = tf.Variable(97000, name='global_step', trainable=False)
#搭建全连接层(此处到时候可以在这个命名空间下，然后专门保存这几个变量)
end_point = 'UCF_Logits'
with tf.variable_scope(end_point):
net = tf.nn.avg_pool3d(bottleneck_input,
ksize=[1, 2, 7, 7, 1],
strides=[1, 1, 1, 1, 1],
padding=snt.VALID)
net = tf.nn.dropout(net, 1.0)
logits = i3d.Unit3D(output_channels=n_classes,
kernel_shape=[1, 1, 1],
activation_fn=None,
use_batch_norm=True,
use_bias=True,
name='Conv3d_0c_1x1')(net, is_training=True)
logits = tf.squeeze(logits, [2, 3],
name='SpatialSqueeze') #这个返回值维度是(6,7,10)
print(logits)
averaged_logits = tf.reduce_mean(logits, axis=1) #那么理论上此时返回的维度是(6,10)i
#获取finetune层的参数
finetune_variable_map = {}
finetune_variable_list = []
for variable in tf.global_variables():
#收集参数，方便正则化
if variable.name.split('/')[0] == 'UCF_Logits':
finetune_variable_list.append(variable)
tf.add_to_collection(tf.GraphKeys.WEIGHTS, variable)
def main(unused_argv):
with tf.device('/gpu:1'):
x_rgb = tf.placeholder(tf.float32,
shape=(None, _SAMPLE_VIDEO_FRAMES, _IMAGE_SIZE,
_IMAGE_SIZE, 3))
y_rgb = tf.placeholder(tf.float32, shape=(None, _NUM_CLASSES))
with tf.variable_scope('RGB'):
rgb_model = i3d.InceptionI3d(_NUM_CLASSES,
spatial_squeeze=True,
final_endpoint='Mixed_5c')
rgb_net, A = rgb_model(x_rgb,
is_training=False,
dropout_keep_prob=1.0)
end_point = 'Logits'
with tf.variable_scope('inception_i3d'):
with tf.variable_scope(end_point):
rgb_net = tf.nn.avg_pool3d(rgb_net,
ksize=[1, 2, 7, 7, 1],
strides=[1, 1, 1, 1, 1],
padding=snt.VALID)
if TRAINING:
rgb_net = tf.nn.dropout(rgb_net, 0.7)
logits = i3d.Unit3D(output_channels=_NUM_CLASSES,
kernel_shape=[1, 1, 1],
activation_fn=None,
use_batch_norm=False,
use_bias=True,
name='Conv3d_0c_1x1')(rgb_net,
is_training=True)
logits = tf.squeeze(logits, [2, 3], name='SpatialSqueeze')
averaged_logits = tf.reduce_mean(logits, axis=1)
rgb_variable_map = {}
last_layer = []
for variable in tf.global_variables():
if variable.name.split('/')[0] == 'RGB':
if (variable.name.find("Logits") == -1):
rgb_variable_map[variable.name.replace(':0',
'')] = variable
if (variable.name.find("Mixed_5c") == -1):
last_layer.append(variable.name.replace(':0', ''))
#[variable.name.replace(':0', '')] = variable
else:
#last_layer[variable.name.replace(':0', '')] = variable
last_layer.append(variable.name.replace(':0', ''))
rgb_saver = tf.train.Saver(var_list=rgb_variable_map, reshape=True)
last_layer_var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='RGB/inception_i3d/Logits')
for variabel_name in tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES,
scope='RGB/inception_i3d/Mixed_5c'):
last_layer_var.append(variabel_name)
print(last_layer_var)
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(0.002,
global_step,
100,
0.98,
staircase=False)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
logits=averaged_logits, labels=y_rgb)
loss = tf.reduce_mean(cross_entropy)
optimizer_op = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(loss=loss,
var_list=last_layer_var)
predictions = tf.nn.softmax(averaged_logits)
model_saver = tf.train.Saver()
init = tf.global_variables_initializer()
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init)
rgb_saver.restore(sess, _CHECKPOINT_PATHS['rgb_imagenet'])
tf.logging.info('RGB checkpoint restored %s',
_CHECKPOINT_PATHS['rgb_imagenet'])
i = 0
vali_list_prepare = prepare_vali_data(data_dir, 'Cam04')
T_batch, t_batch = generator_vali_data(vali_list_prepare, data_dir)
for epoch in range(epoch_num):
for X_batch, y_batch in generator_data(data_dir, '3'):
i = i + 1
feed_dict_train = {x_rgb: X_batch, y_rgb: y_batch}
op_, loss_, pre_train = sess.run(
[optimizer_op, loss, predictions],
feed_dict=feed_dict_train)
train_acc = accuracy(pre_train, y_batch)
print(
"After %d epoch %d training step(s), train accuracy using average model is %g "
% (epoch, i, train_acc))
print("Train_loss is %f " % (loss_ * 1.0))
if (i % 100 == 0):
T_batch, t_batch = generator_vali_data(
vali_list_prepare, data_dir)
feed_dict_test = {x_rgb: T_batch, y_rgb: t_batch}
pre_test = sess.run(predictions, feed_dict=feed_dict_test)
test_acc = train_acc = accuracy(pre_test, t_batch)
print(
"~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~"
)
print(
"After %d iter, validate accuracy using average model is %g "
% (epoch, test_acc))
print(
"~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~"
)
model_saver.save(sess, save_model_path)
model_saver.save(sess, save_model_path2)
print("OK!!!")