def AddFCLayer(self, prev_layer, index):
"""Parse expression and add Fully Connected Layer.
Args:
prev_layer: Input tensor.
index: Position in model_str to start parsing
Returns:
Output tensor, end index in model_str.
"""
pattern = re.compile(R'(F)(s|t|r|l|m)({\w+})?(\d+)')
m = pattern.match(self.model_str, index)
if m is None:
return None, None
fn = self._NonLinearity(m.group(2))
name = self._GetLayerName(m.group(0), index, m.group(3))
depth = int(m.group(4))
input_depth = shapes.tensor_dim(prev_layer, 1) * shapes.tensor_dim(
prev_layer, 2) * shapes.tensor_dim(prev_layer, 3)
# The slim fully connected is actually a 1x1 conv, so we have to crush the
# dimensions on input.
# Everything except batch goes to depth, and therefore has to be known.
shaped = tf.reshape(
prev_layer, [-1, input_depth], name=name + '_reshape_in')
output = slim.fully_connected(shaped, depth, activation_fn=fn, scope=name)
# Width and height are collapsed to 1.
self.reduction_factors[1] = None
self.reduction_factors[2] = None
return tf.reshape(
output, [shapes.tensor_dim(prev_layer, 0), 1, 1, depth],
name=name + '_reshape_out'), m.end()
def _AddOutputs(self, prev_layer, out_dims, out_func, num_classes):
"""Adds the output layer and loss function.
Args:
prev_layer: Output of last layer of main network.
out_dims: Number of output dimensions, 0, 1 or 2.
out_func: Output non-linearity. 's' or 'c'=softmax, 'l'=logistic.
num_classes: Number of outputs/size of last output dimension.
"""
height_in = shapes.tensor_dim(prev_layer, dim=1)
logits, outputs = self._AddOutputLayer(prev_layer, out_dims, out_func,
num_classes)
if self.mode == 'train':
# Setup loss for training.
self.loss = self._AddLossFunction(logits, height_in, out_dims,
out_func)
tf.summary.scalar('loss', self.loss)
elif out_dims == 0:
# Be sure the labels match the output, even in eval mode.
self.labels = tf.slice(self.labels, [0, 0], [-1, 1])
self.labels = tf.reshape(self.labels, [-1])
logging.info('Final output=%s', outputs)
logging.info('Labels tensor=%s', self.labels)
self.output = outputs
def _AddOutputLayer(self, prev_layer, out_dims, out_func, num_classes):
"""Add the fully-connected logits and SoftMax/Logistic output Layer.
Args:
prev_layer: Output of last layer of main network.
out_dims: Number of output dimensions, 0, 1 or 2.
out_func: Output non-linearity. 's' or 'c'=softmax, 'l'=logistic.
num_classes: Number of outputs/size of last output dimension.
Returns:
logits: Pre-softmax/logistic fully-connected output shaped to out_dims.
outputs: Post-softmax/logistic shaped to out_dims.
Raises:
ValueError: if syntax is incorrect.
"""
# Reduce dimensionality appropriate to the output dimensions.
batch_in = shapes.tensor_dim(prev_layer, dim=0)
height_in = shapes.tensor_dim(prev_layer, dim=1)
width_in = shapes.tensor_dim(prev_layer, dim=2)
depth_in = shapes.tensor_dim(prev_layer, dim=3)
if out_dims:
# Combine any remaining height and width with batch and unpack after.
shaped = tf.reshape(prev_layer, [-1, depth_in])
else:
# Everything except batch goes to depth, and therefore has to be known.
shaped = tf.reshape(prev_layer,
[-1, height_in * width_in * depth_in])
logits = slim.fully_connected(shaped, num_classes, activation_fn=None)
if out_func == 'l':
raise ValueError('Logistic not yet supported!')
else:
output = tf.nn.softmax(logits)
# Reshape to the dessired output.
if out_dims == 2:
output_shape = [batch_in, height_in, width_in, num_classes]
elif out_dims == 1:
output_shape = [batch_in, height_in * width_in, num_classes]
else:
output_shape = [batch_in, num_classes]
output = tf.reshape(output, output_shape, name='Output')
logits = tf.reshape(logits, output_shape)
return logits, output
def _AddLossFunction(self, logits, height_in, out_dims, out_func):
"""Add the appropriate loss function.
Args:
logits: Pre-softmax/logistic fully-connected output shaped to out_dims.
height_in: Height of logits before going into the softmax layer.
out_dims: Number of output dimensions, 0, 1 or 2.
out_func: Output non-linearity. 's' or 'c'=softmax, 'l'=logistic.
Returns:
loss: That which is to be minimized.
Raises:
ValueError: if logistic is used.
"""
if out_func == 'c':
# Transpose batch to the middle.
ctc_input = tf.transpose(logits, [1, 0, 2])
# Compute the widths of each batch element from the input widths.
widths = self.layers.GetLengths(dim=2, factor=height_in)
cross_entropy = tf.nn.ctc_loss(ctc_input, self.sparse_labels,
widths)
elif out_func == 's':
if out_dims == 2:
self.labels = _PadLabels3d(logits, self.labels)
elif out_dims == 1:
self.labels = _PadLabels2d(shapes.tensor_dim(logits, dim=1),
self.labels)
else:
self.labels = tf.slice(self.labels, [0, 0], [-1, 1])
self.labels = tf.reshape(self.labels, [-1])
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=self.labels, name='xent')
else:
# TODO(rays) Labels need an extra dimension for logistic, so different
# padding functions are needed, as well as a different loss function.
raise ValueError('Logistic not yet supported!')
return tf.reduce_sum(cross_entropy)
def _LSTMLayer(self, prev_layer, direction, dim, summarize, depth, name):
"""Adds an LSTM layer with the given pre-parsed attributes.
Always maps 4-D to 4-D regardless of summarize.
Args:
prev_layer: Input tensor.
direction: 'forward' 'backward' or 'bidirectional'
dim: 'x' or 'y', dimension to consider as time.
summarize: True if we are to return only the last timestep.
depth: Output depth.
name: Some string naming the op.
Returns:
Output tensor.
"""
# If the target dimension is y, we need to transpose.
if dim == 'x':
lengths = self.GetLengths(2, 1)
inputs = prev_layer
else:
lengths = self.GetLengths(1, 1)
inputs = tf.transpose(prev_layer, [0, 2, 1, 3], name=name + '_ytrans_in')
input_batch = shapes.tensor_dim(inputs, 0)
num_slices = shapes.tensor_dim(inputs, 1)
num_steps = shapes.tensor_dim(inputs, 2)
input_depth = shapes.tensor_dim(inputs, 3)
# Reshape away the other dimension.
inputs = tf.reshape(
inputs, [-1, num_steps, input_depth], name=name + '_reshape_in')
# We need to replicate the lengths by the size of the other dimension, and
# any changes that have been made to the batch dimension.
tile_factor = tf.to_float(input_batch *
num_slices) / tf.to_float(tf.shape(lengths)[0])
lengths = tf.tile(lengths, [tf.cast(tile_factor, tf.int32)])
lengths = tf.cast(lengths, tf.int64)
outputs = nn_ops.rnn_helper(
inputs,
lengths,
cell_type='lstm',
num_nodes=depth,
direction=direction,
name=name,
stddev=0.1)
# Output depth is doubled if bi-directional.
if direction == 'bidirectional':
output_depth = depth * 2
else:
output_depth = depth
# Restore the other dimension.
if summarize:
outputs = tf.slice(
outputs, [0, num_steps - 1, 0], [-1, 1, -1], name=name + '_sum_slice')
outputs = tf.reshape(
outputs, [input_batch, num_slices, 1, output_depth],
name=name + '_reshape_out')
else:
outputs = tf.reshape(
outputs, [input_batch, num_slices, num_steps, output_depth],
name=name + '_reshape_out')
if dim == 'y':
outputs = tf.transpose(outputs, [0, 2, 1, 3], name=name + '_ytrans_out')
return outputs
def lstm_layer(inp,
length=None,
state=None,
memory=None,
num_nodes=None,
backward=False,
clip=50.0,
reg_func=tf.nn.l2_loss,
weight_reg=False,
weight_collection="LSTMWeights",
bias_reg=False,
stddev=None,
seed=None,
decode=False,
use_native_weights=False,
name=None):
"""Adds ops for an LSTM layer.
This adds ops for the following operations:
input => (forward-LSTM|backward-LSTM) => output
The direction of the LSTM is determined by `backward`. If it is false, the
forward LSTM is used, the backward one otherwise.
Args:
inp: A 3-D tensor of shape [`batch_size`, `max_length`, `feature_dim`].
length: A 1-D tensor of shape [`batch_size`] and type int64. Each element
represents the length of the corresponding sequence in `inp`.
state: If specified, uses it as the initial state.
memory: If specified, uses it as the initial memory.
num_nodes: The number of LSTM cells.
backward: If true, reverses the `inp` before adding the ops. The output is
also reversed so that the direction is the same as `inp`.
clip: Value used to clip the cell values.
reg_func: Function used for the weight regularization such as
`tf.nn.l2_loss`.
weight_reg: If true, regularize the filter weights with `reg_func`.
weight_collection: Collection to add the weights to for regularization.
bias_reg: If true, regularize the bias vector with `reg_func`.
stddev: Standard deviation used to initialize the variables.
seed: Seed used to initialize the variables.
decode: If true, does not add ops which are not used for inference.
use_native_weights: If true, uses weights in the same format as the native
implementations.
name: Name of the op.
Returns:
A 3-D tensor of shape [`batch_size`, `max_length`, `num_nodes`].
"""
with tf.variable_scope(name):
if backward:
if length is None:
inp = tf.reverse(inp, [1])
else:
inp = tf.reverse_sequence(inp, length, 1, 0)
num_prev = inp.get_shape()[2]
if stddev:
initializer = tf.truncated_normal_initializer(stddev=stddev,
seed=seed)
else:
initializer = tf.uniform_unit_scaling_initializer(seed=seed)
if use_native_weights:
with tf.variable_scope("LSTMCell"):
w = tf.get_variable(
"W_0",
shape=[num_prev + num_nodes, 4 * num_nodes],
initializer=initializer,
dtype=tf.float32)
w_i_m = tf.slice(w, [0, 0], [num_prev, 4 * num_nodes],
name="w_i_m")
w_m_m = tf.reshape(tf.slice(w, [num_prev, 0],
[num_nodes, 4 * num_nodes]),
[num_nodes, 4, num_nodes],
name="w_m_m")
else:
w_i_m = tf.get_variable("w_i_m", [num_prev, 4 * num_nodes],
initializer=initializer)
w_m_m = tf.get_variable("w_m_m", [num_nodes, 4, num_nodes],
initializer=initializer)
if not decode and weight_reg:
tf.add_to_collection(weight_collection,
reg_func(w_i_m, name="w_i_m_reg"))
tf.add_to_collection(weight_collection,
reg_func(w_m_m, name="w_m_m_reg"))
batch_size = shapes.tensor_dim(inp, dim=0)
num_frames = shapes.tensor_dim(inp, dim=1)
prev = tf.reshape(inp, tf.stack([batch_size * num_frames, num_prev]))
if use_native_weights:
with tf.variable_scope("LSTMCell"):
b = tf.get_variable("B",
shape=[4 * num_nodes],
initializer=tf.zeros_initializer(),
dtype=tf.float32)
biases = tf.identity(b, name="biases")
else:
biases = tf.get_variable("biases", [4 * num_nodes],
initializer=tf.constant_initializer(0.0))
if not decode and bias_reg:
tf.add_to_collection(weight_collection,
reg_func(biases, name="biases_reg"))
prev = tf.nn.xw_plus_b(prev, w_i_m, biases)
prev = tf.reshape(prev,
tf.stack([batch_size, num_frames, 4, num_nodes]))
if state is None:
state = tf.fill(tf.stack([batch_size, num_nodes]), 0.0)
if memory is None:
memory = tf.fill(tf.stack([batch_size, num_nodes]), 0.0)
out, _, mem = rnn.variable_lstm(prev, state, memory, w_m_m, clip=clip)
if backward:
if length is None:
out = tf.reverse(out, [1])
else:
out = tf.reverse_sequence(out, length, 1, 0)
return out, mem