def phi(self, x, ext_wts=None, reuse=None):
"""Feature extraction function.
Args:
x: [N, H, W, C]. Input.
reuse: Whether to reuse variables here.
"""
config = self.config
is_training = self.is_training
dtype = self.dtype
with tf.variable_scope("phi", reuse=reuse):
h, wts = cnn(x,
config.filter_size,
strides=config.strides,
pool_fn=[tf.nn.max_pool] * len(config.pool_fn),
pool_size=config.pool_size,
pool_strides=config.pool_strides,
act_fn=[tf.nn.relu for aa in config.conv_act_fn],
add_bias=True,
init_std=config.conv_init_std,
init_method=config.conv_init_method,
wd=config.wd,
dtype=dtype,
batch_norm=True,
is_training=is_training,
ext_wts=ext_wts)
if self._embedding_weights is None:
self._embedding_weights = wts
h_shape = h.get_shape()
h_size = 1
for ss in h_shape[1:]:
h_size *= int(ss)
h = tf.reshape(h, [-1, h_size])
return h
def phi(x, reuse=None, scope="Model", is_training=True, ext_wts=None):
"""Feature extraction function.
Args:
x: [N, H, W, C]. Input.
reuse: Whether to reuse variables here.
scope: Variable scope.
ext_wts: A dictionary of external weights to be used from.
"""
with tf.name_scope(scope):
with tf.variable_scope("Model", reuse=reuse):
h, wts = cnn(x,
config.filter_size,
strides=config.strides,
pool_fn=[tf.nn.max_pool] * len(config.pool_fn),
pool_size=config.pool_size,
pool_strides=config.pool_strides,
act_fn=[tf.nn.relu for aa in config.conv_act_fn],
add_bias=True,
init_std=config.conv_init_std,
init_method=config.conv_init_method,
wd=config.wd,
dtype=dtype,
batch_norm=True,
is_training=is_training,
ext_wts=ext_wts)
h_size = reduce(lambda x, y: x * y,
[int(ss) for ss in h.get_shape()[1:]])
h = tf.reshape(h, [-1, h_size])
return h, wts
def __call__(self,
x,
is_training=True,
ext_wts=None,
reuse=None,
**kwargs):
"""See Backbone class for documentation."""
config = self.config
dtype = self.dtype
assert not config.add_last_relu
L = len(config.conv_act_fn)
if config.add_last_relu:
act_fn = [tf.nn.relu for aa in range(L)]
else:
act_fn = [tf.nn.relu for aa in range(L - 1)] + [None]
with tf.variable_scope("phi", reuse=reuse):
assert config.wd == 5e-4, '{}'.format(config.wd)
h, wts = cnn(x,
config.filter_size,
strides=config.strides,
pool_fn=[tf.nn.max_pool] * len(config.pool_fn),
pool_size=config.pool_size,
pool_strides=config.pool_strides,
act_fn=act_fn,
add_bias=False,
init_std=config.conv_init_std,
init_method=config.conv_init_method,
wd=config.wd,
dtype=dtype,
batch_norm=True,
is_training=is_training,
ext_wts=ext_wts)
if self.weights is None:
self.weights = wts
h_shape = h.get_shape()
h_size = 1
for ss in h_shape[1:]:
h_size *= int(ss)
if ext_wts is not None:
# PyTorch NCHW mode.
h = tf.transpose(h, [0, 3, 1, 2])
h = tf.reshape(h, [-1, h_size])
assert h_size == 3200
return h
def phi(self,
x,
ext_wts=None,
reuse=tf.AUTO_REUSE,
update_batch_stats=False): #AYAD
"""Feature extraction function.
Args:
x: [N, H, W, C]. Input.
reuse: Whether to reuse variables here.
"""
config = self.config
is_training = self.is_training
dtype = self.dtype
with tf.variable_scope("phi", reuse=reuse):
h, wts = cnn(x,
config.filter_size,
strides=config.strides,
pool_fn=[tf.nn.max_pool] * len(config.pool_fn),
pool_size=config.pool_size,
pool_strides=config.pool_strides,
act_fn=[tf.nn.relu for aa in config.conv_act_fn],
add_bias=True,
init_std=config.conv_init_std,
init_method=config.conv_init_method,
wd=config.wd,
dtype=dtype,
batch_norm=True,
is_training=is_training,
ext_wts=ext_wts,
update_batch_stats=update_batch_stats)
if self._embedding_weights is None:
self._embedding_weights = wts
h_shape = h.get_shape()
h_size = 1
for ss in h_shape[1:]:
h_size *= int(ss)
h = tf.reshape(h, [-1, h_size])
self.adv_summaries.append(
tf.summary.histogram('Encoded', h, family="phi_out"))
return h
def __init__(self,
config,
x,
y,
num_classes,
is_training=True,
dtype=tf.float32):
"""Constructor.
Args:
config:
x:
y:
num_classes:
"""
h, _ = cnn(x,
config.filter_size,
strides=config.strides,
pool_fn=[tf.nn.max_pool] * len(config.pool_fn),
pool_size=config.pool_size,
pool_strides=config.pool_strides,
act_fn=[tf.nn.relu for aa in config.conv_act_fn],
add_bias=True,
init_std=config.conv_init_std,
init_method=config.conv_init_method,
wd=config.wd,
dtype=dtype,
batch_norm=True,
is_training=is_training,
ext_wts=None)
h_shape = h.get_shape()
h_size = 1
for ss in h_shape[1:]:
h_size *= int(ss)
h = tf.reshape(h, [-1, h_size])
w_class = weight_variable([h_size, num_classes],
init_method='truncated_normal',
dtype=tf.float32,
init_param={'stddev': 0.01},
name='w_class')
b_class = weight_variable([num_classes],
init_method='constant',
init_param={'val': 0.0},
name='b_class')
self._feature = h
logits = tf.matmul(h, w_class) + b_class
xent = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
labels=y)
xent = tf.reduce_mean(xent, name='xent')
cost = xent
cost += self._decay()
self._cost = cost
self._inputs = x
self._labels = y
global_step = tf.get_variable('global_step',
shape=[],
dtype=tf.int64,
trainable=False)
# Learning rate decay.
learn_rate = tf.train.piecewise_constant(
global_step,
list(np.array(config.lr_decay_steps).astype(np.int64)),
[config.learn_rate] + list(config.lr_list))
self._learn_rate = learn_rate
self._train_op = tf.train.AdamOptimizer(learn_rate).minimize(
cost, global_step=global_step)
correct = tf.equal(tf.argmax(logits, axis=1), y)
self._acc = tf.reduce_mean(tf.cast(correct, dtype))