def model(data, dropout=None, maxout_k=1, activation_fn=tf.nn.relu):
data = tf.reshape(data, [BATCH_SIZE, SIZE, SIZE, NUM_CHANNELS])
conv = conv_layer(data,
depth=64,
window=5,
dropout=dropout,
pool=(2, 2),
maxout_k=maxout_k,
activation_fn=activation_fn,
name='conv1')
conv = conv_layer(conv,
depth=32,
window=5,
dropout=dropout,
pool=(2, 2),
maxout_k=maxout_k,
activation_fn=activation_fn,
name='conv2')
reshape = conv_to_fc_layer(conv)
hidden = fc_layer(reshape,
depth=128,
maxout_k=maxout_k,
activation_fn=activation_fn,
dropout=dropout,
name='fc1')
output = fc_layer(hidden,
depth=NUM_LABELS,
maxout_k=1,
activation=False,
name='fc2')
return output
def model(self, data):
variables = defaultdict(list)
conv1 = conv_layer(data,
depth=96,
window=11,
stride=4,
activation_fn=tf.nn.relu,
pool=(3, 2),
lrn=(5, 1.0, 1e-4, 0.75),
name='conv1',
variables=variables)
conv2 = conv_layer(conv1,
depth=256,
window=5,
activation_fn=tf.nn.relu,
pool=(3, 2),
lrn=(5, 1.0, 1e-4, 0.75),
name='conv2',
variables=variables)
conv3 = conv_layer(conv2,
depth=384,
window=3,
activation_fn=tf.nn.relu,
name='conv3',
variables=variables)
conv4 = conv_layer(conv3,
depth=384,
window=3,
activation_fn=tf.nn.relu,
name='conv4',
variables=variables)
conv5 = conv_layer(conv4,
depth=256,
window=3,
activation_fn=tf.nn.relu,
pool=(3, 2),
name='conv5',
variables=variables)
conv5r = conv_to_ff_layer(conv5)
# ff_layer(input_layer, depth, activation_fn=tf.nn.sigmoid, dropout=None, name=None, activation=True, variables=None):
fc6 = ff_layer(
conv5r,
depth=512, # TODO: return to 4096
activation_fn=tf.nn.relu,
dropout=self.keep_prob,
name='fc6',
variables=variables)
fc7 = ff_layer(
fc6,
depth=512, # TODO: return to 4096
activation_fn=tf.nn.relu,
dropout=self.keep_prob,
name='fc7',
variables=variables)
output = ff_layer(fc7,
depth=NUM_LABELS,
name='output',
activation=False,
variables=variables)
return output, variables
def model(self, data):
"""Construct a model.
:param data: the batched input images
"""
variables = defaultdict(list)
conv = conv_layer(data,
depth=64,
window=5,
pool=(2, 2),
name='conv1',
variables=variables)
conv = conv_layer(conv,
depth=32,
window=5,
pool=(2, 2),
name='conv2',
variables=variables)
reshape = conv_to_ff_layer(conv)
hidden = ff_layer(reshape, depth=512, name='ff1', variables=variables)
output = ff_layer(hidden,
depth=NUM_LABELS,
name='ff2',
activation=False,
variables=variables)
return output, variables
def model(self, data):
variables = defaultdict(list)
conv11 = conv_layer(data,
depth=64,
window=3,
name='conv11',
variables=variables)
conv12 = conv_layer(conv11,
depth=64,
window=3,
name='conv12',
variables=variables,
pool=(2, 2))
conv21 = conv_layer(conv12,
depth=128,
window=3,
name='conv21',
variables=variables)
conv22 = conv_layer(conv21,
depth=128,
window=3,
name='conv22',
variables=variables,
pool=(2, 2))
conv31 = conv_layer(conv22,
depth=256,
window=3,
name='conv31',
variables=variables)
conv32 = conv_layer(conv31,
depth=256,
window=3,
name='conv32',
variables=variables)
conv33 = conv_layer(conv32,
depth=256,
window=3,
name='conv33',
variables=variables,
pool=(2, 2))
conv41 = conv_layer(conv33,
depth=512,
window=3,
name='conv41',
variables=variables)
conv42 = conv_layer(conv41,
depth=512,
window=3,
name='conv42',
variables=variables)
conv43 = conv_layer(conv42,
depth=512,
window=3,
name='conv43',
variables=variables,
pool=(2, 2))
conv51 = conv_layer(conv43,
depth=512,
window=3,
name='conv51',
variables=variables)
conv52 = conv_layer(conv51,
depth=512,
window=3,
name='conv52',
variables=variables)
conv53 = conv_layer(conv52,
depth=512,
window=3,
name='conv53',
variables=variables,
pool=(2, 2))
conv5r = conv_to_ff_layer(conv53)
# ff_layer(input_layer, depth, activation_fn=tf.nn.sigmoid, dropout=None, name=None, activation=True, variables=None):
fc6 = ff_layer(conv5r,
depth=4096,
dropout=self.keep_prob,
name='fc6',
variables=variables)
fc7 = ff_layer(fc6,
depth=4096,
dropout=self.keep_prob,
name='fc7',
variables=variables)
output = ff_layer(fc7,
depth=NUM_LABELS,
name='output',
activation=False,
variables=variables)
return output, variables