def model_fn(net, X_len, max_reach, block_size, out_classes, batch_size, dtype,
**kwargs):
"""
Args:
net -> Input tensor shaped (batch_size, max_reach + block_size + max_reach, 3)
Returns:
logits -> Unscaled logits tensor in time_major form, (block_size, batch_size, out_classes)
"""
for block in range(1, 4):
with tf.variable_scope("block%d" % block):
for layer in range(kwargs['num_layers']):
with tf.variable_scope('layer_%d' % layer):
net = batch_normalization(net)
net = tf.nn.relu(net)
net = conv_1d(
net,
64,
3,
scope="conv_1d_%d" % layer,
weights_init=variance_scaling_initializer(dtype=dtype))
net = max_pool_1d(net, 2)
net = tf.nn.relu(net)
net = central_cut(net, block_size, 8)
net = tf.transpose(net, [1, 0, 2], name="Shift_to_time_major")
net = conv_1d(net, 9, 1, scope='logits')
return {
'logits': net,
'init_state': tf.constant(0),
'final_state': tf.constant(0),
}
def model_fn(net, X_len, max_reach, block_size, out_classes, batch_size, dtype,
**kwargs):
"""
Args:
net -> Input tensor shaped (batch_size, max_reach + block_size + max_reach, 3)
Returns:
logits -> Unscaled logits tensor in time_major form, (block_size, batch_size, out_classes)
"""
print("model in", net.get_shape())
for block in range(1, 4):
with tf.variable_scope("block%d" % block):
if block > 1:
net = tf.expand_dims(net, 3)
net = tf.layers.max_pooling2d(net, [1, 2], [1, 2])
net = tf.squeeze(net, axis=3)
for layer in range(kwargs['num_layers']):
with tf.variable_scope('layer_%d' % layer):
res = net
for sublayer in range(kwargs['num_sub_layers']):
res = batch_normalization(res,
scope='bn_%d' % sublayer)
res = tf.nn.relu(res)
res = conv_1d(
res,
32 * 2**(4 - block),
3,
scope="conv_1d_%d" % sublayer,
weights_init=variance_scaling_initializer(
dtype=dtype))
k = tf.get_variable(
"k",
initializer=tf.constant_initializer(1.0),
shape=[])
net = tf.nn.relu(k) * res + net
net = max_pool_1d(net, 2)
net = tf.nn.relu(net)
net = central_cut(net, block_size, 8)
print("after slice", net.get_shape())
net = tf.transpose(net, [1, 0, 2], name="Shift_to_time_major")
print("after transpose", net.get_shape())
net = conv_1d(net, 9, 1, scope='logits')
print("model out", net.get_shape())
return {
'logits': net,
'init_state': tf.constant(0),
'final_state': tf.constant(0),
}
def model_fn(net, X_len, max_reach, block_size, out_classes, batch_size, dtype, **kwargs):
"""
Args:
net -> Input tensor shaped (batch_size, max_reach + block_size + max_reach, 3)
Returns:
logits -> Unscaled logits tensor in time_major form, (block_size, batch_size, out_classes)
"""
for block in range(1, 3):
with tf.variable_scope("block%d" % block):
for layer in range(kwargs['num_layers']):
with tf.variable_scope('layer_%d' % layer):
res = net
for sublayer in range(kwargs['num_sub_layers']):
res = batch_normalization(
res, scope='bn_%d' % sublayer)
res = tf.nn.relu(res)
res = conv_1d(
res,
64,
3,
scope="conv_1d_%d" % sublayer,
weights_init=variance_scaling_initializer(
dtype=dtype)
)
k = tf.get_variable(
"k", initializer=tf.constant_initializer(1.0), shape=[])
net = tf.nn.relu(k) * res + net
net = max_pool_1d(net, 2)
net = tf.nn.relu(net)
net = central_cut(net, block_size, 4)
net = tf.transpose(net, [1, 0, 2], name="Shift_to_time_major")
# with tf.name_scope("RNN"):
# from tensorflow.contrib.cudnn_rnn import CudnnGRU, RNNParamsSaveable
# rnn_layer = CudnnGRU(
# num_layers=1,
# num_units=64,
# input_size=64,
# )
#
# print(rnn_layer.params_size())
# import sys
# sys.exit(0)
# rnn_params = tf.get_variable("rnn_params", shape=[rnn_layer.params_size()], validate_shape=False)
# params_saveable = RNNParamsSaveable(
# rnn_layer.params_to_canonical, rnn_layer.canonical_to_params, [rnn_params])
# tf.add_to_collection(tf.GraphKeys.SAVEABLE_OBJECTS, params_saveable)
#
with tf.name_scope("RNN"):
cell = tf.contrib.rnn.GRUCell(64)
init_state = cell.zero_state(batch_size, dtype=tf.float32)
outputs, final_state = tf.nn.dynamic_rnn(cell, net, initial_state=init_state, sequence_length=tf.div(X_len + 3, 4), time_major=True, parallel_iterations=128)
net = conv_1d(outputs, 9, 1, scope='logits')
return {
'logits': net,
'init_state': init_state,
'final_state': final_state
}
def model_fn(net, X_len, max_reach, block_size, out_classes, batch_size, dtype,
**kwargs):
"""
Args:
net -> Input tensor shaped (batch_size, max_reach + block_size + max_reach, 3)
Returns:
logits -> Unscaled logits tensor in time_major form, (block_size, batch_size, out_classes)
"""
with tf.name_scope("model"):
print("model in", net.get_shape())
for block in range(1, 4):
with tf.variable_scope("block%d" % block):
for layer in range(1, 20 + 1):
with tf.variable_scope('layer_%d' % layer):
res = net
for sublayer in [1, 2]:
res = batch_normalization(res,
scope='bn_%d' % sublayer)
res = tf.nn.relu(res)
res = conv_1d(
res,
64,
3,
scope="conv_1d_%d" % sublayer,
weights_init=variance_scaling_initializer(
dtype=dtype))
k = tf.get_variable(
"k",
initializer=tf.constant_initializer(1.0),
shape=[])
net = tf.nn.relu(k) * res + net
net = max_pool_1d(net, 2)
cut_size = tf.shape(net)[1] - tf.div(block_size, 8)
with tf.control_dependencies([
tf.cond(
tflearn.get_training_mode(), lambda: tf.assert_equal(
tf.mod(cut_size, 2), 0, name="cut_size_assert"),
lambda: tf.no_op())
]):
cut_size = tf.div(cut_size, 2)
net = tf.slice(net, [0, cut_size, 0],
[-1, tf.div(block_size, 8), -1],
name="Cutting")
print("after slice", net.get_shape())
net = tf.transpose(net, [1, 0, 2], name="Shift_to_time_major")
state_size = 64
outputs = net
print("outputs", outputs.get_shape())
with tf.variable_scope("Output"):
outputs = tf.reshape(outputs,
[block_size // 8 * batch_size, state_size],
name="flatten")
W = tf.get_variable("W", shape=[state_size, out_classes])
b = tf.get_variable("b", shape=[out_classes])
outputs = tf.matmul(outputs, W) + b
logits = tf.reshape(outputs,
[block_size // 8, batch_size, out_classes],
name="logits")
print("model out", logits.get_shape())
return {
'logits': logits,
'init_state': tf.constant(0),
'final_state': tf.constant(0),
}