def _compute_output_shape(self, input_shape):
input_shape = tuple(tensor_shape.TensorShape(input_shape).as_list())
if self._output_shape is None:
x = K.placeholder(shape=input_shape)
x = self.call(x)
if isinstance(x, list):
return [tensor_shape.TensorShape(K.int_shape(x_elem)) for x_elem in x]
else:
return tensor_shape.TensorShape(K.int_shape(x))
elif isinstance(self._output_shape, (tuple, list)):
if isinstance(input_shape, list):
num_samples = input_shape[0][0]
else:
num_samples = input_shape[0] if input_shape else None
return tensor_shape.TensorShape((num_samples,) +
tuple(self._output_shape))
else:
shape = self._output_shape(input_shape)
if not isinstance(shape, (list, tuple)):
raise ValueError(
'`output_shape` function must return a tuple or a list of tuples.')
if isinstance(shape, list):
if isinstance(shape[0], int) or shape[0] is None:
shape = tuple(shape)
return tensor_shape.TensorShape(shape)
def set_model(self, model):
self.model = model
self.sess = K.get_session()
if self.histogram_freq and self.merged is None:
for layer in self.model.layers:
for weight in layer.weights:
mapped_weight_name = weight.name.replace(':', '_')
tf_summary.histogram(mapped_weight_name, weight)
if self.write_grads:
grads = model.optimizer.get_gradients(model.total_loss, weight)
def is_indexed_slices(grad):
return type(grad).__name__ == 'IndexedSlices'
grads = [grad.values if is_indexed_slices(grad) else grad
for grad in grads]
tf_summary.histogram('{}_grad'.format(mapped_weight_name), grads)
if self.write_images:
w_img = array_ops.squeeze(weight)
shape = K.int_shape(w_img)
if len(shape) == 2: # dense layer kernel case
if shape[0] > shape[1]:
w_img = array_ops.transpose(w_img)
shape = K.int_shape(w_img)
w_img = array_ops.reshape(w_img, [1, shape[0], shape[1], 1])
elif len(shape) == 3: # convnet case
if K.image_data_format() == 'channels_last':
# switch to channels_first to display
# every kernel as a separate image
w_img = array_ops.transpose(w_img, perm=[2, 0, 1])
shape = K.int_shape(w_img)
w_img = array_ops.reshape(w_img,
[shape[0], shape[1], shape[2], 1])
elif len(shape) == 1: # bias case
w_img = array_ops.reshape(w_img, [1, shape[0], 1, 1])
else:
# not possible to handle 3D convnets etc.
continue
shape = K.int_shape(w_img)
assert len(shape) == 4 and shape[-1] in [1, 3, 4]
tf_summary.image(mapped_weight_name, w_img)
if hasattr(layer, 'output'):
tf_summary.histogram('{}_out'.format(layer.name), layer.output)
self.merged = tf_summary.merge_all()
if self.write_graph:
self.writer = tf_summary.FileWriter(self.log_dir, self.sess.graph)
else:
self.writer = tf_summary.FileWriter(self.log_dir)
def __call__(self, inputs, initial_state=None, constants=None, **kwargs):
inputs, initial_state, constants = self._standardize_args(
inputs, initial_state, constants)
if initial_state is None and constants is None:
return super(ConvRNN2D, self).__call__(inputs, **kwargs)
# If any of `initial_state` or `constants` are specified and are Keras
# tensors, then add them to the inputs and temporarily modify the
# input_spec to include them.
additional_inputs = []
additional_specs = []
if initial_state is not None:
kwargs['initial_state'] = initial_state
additional_inputs += initial_state
self.state_spec = []
for state in initial_state:
shape = K.int_shape(state)
self.state_spec.append(InputSpec(shape=shape))
additional_specs += self.state_spec
if constants is not None:
kwargs['constants'] = constants
additional_inputs += constants
self.constants_spec = [InputSpec(shape=K.int_shape(constant))
for constant in constants]
self._num_constants = len(constants)
additional_specs += self.constants_spec
# at this point additional_inputs cannot be empty
for tensor in additional_inputs:
if K.is_keras_tensor(tensor) != K.is_keras_tensor(additional_inputs[0]):
raise ValueError('The initial state or constants of an RNN'
' layer cannot be specified with a mix of'
' Keras tensors and non-Keras tensors')
if K.is_keras_tensor(additional_inputs[0]):
# Compute the full input spec, including state and constants
full_input = [inputs] + additional_inputs
full_input_spec = self.input_spec + additional_specs
# Perform the call with temporarily replaced input_spec
original_input_spec = self.input_spec
self.input_spec = full_input_spec
output = super(ConvRNN2D, self).__call__(full_input, **kwargs)
self.input_spec = original_input_spec
return output
else:
return super(ConvRNN2D, self).__call__(inputs, **kwargs)
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
shapes = [K.int_shape(p) for p in params]
accumulators = [K.zeros(shape) for shape in shapes]
delta_accumulators = [K.zeros(shape) for shape in shapes]
self.weights = accumulators + delta_accumulators
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / # pylint: disable=g-no-augmented-assignment
(1. + self.decay * K.cast(self.iterations,
K.dtype(self.decay))))
for p, g, a, d_a in zip(params, grads, accumulators, delta_accumulators):
# update accumulator
new_a = self.rho * a + (1. - self.rho) * K.square(g)
self.updates.append(K.update(a, new_a))
# use the new accumulator and the *old* delta_accumulator
update = g * K.sqrt(d_a + self.epsilon) / K.sqrt(new_a + self.epsilon)
new_p = p - lr * update
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
# update delta_accumulator
new_d_a = self.rho * d_a + (1 - self.rho) * K.square(update)
self.updates.append(K.update(d_a, new_d_a))
return self.updates
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr *= (1. /
(1. + self.decay * K.cast(self.iterations, K.dtype(self.decay))))
t = K.cast(self.iterations, K.floatx()) + 1
lr_t = lr / (1. - K.pow(self.beta_1, t))
shapes = [K.int_shape(p) for p in params]
# zero init of 1st moment
ms = [K.zeros(shape) for shape in shapes]
# zero init of exponentially weighted infinity norm
us = [K.zeros(shape) for shape in shapes]
self.weights = [self.iterations] + ms + us
for p, g, m, u in zip(params, grads, ms, us):
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
u_t = K.maximum(self.beta_2 * u, K.abs(g))
p_t = p - lr_t * m_t / (u_t + self.epsilon)
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(u, u_t))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / # pylint: disable=g-no-augmented-assignment
(1. + self.decay * K.cast(self.iterations,
K.dtype(self.decay))))
# momentum
shapes = [K.int_shape(p) for p in params]
moments = [K.zeros(shape) for shape in shapes]
self.weights = [self.iterations] + moments
for p, g, m in zip(params, grads, moments):
v = self.momentum * m - lr * g # velocity
self.updates.append(K.update(m, v))
if self.nesterov:
new_p = p + self.momentum * v - lr * g
else:
new_p = p + v
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def get_constants(self, inputs, training=None):
constants = []
if self.implementation != 0 and 0 < self.dropout < 1:
input_shape = K.int_shape(inputs)
input_dim = input_shape[-1]
ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
ones = K.tile(ones, (1, int(input_dim)))
def dropped_inputs():
return K.dropout(ones, self.dropout)
dp_mask = [
K.in_train_phase(dropped_inputs, ones, training=training)
for _ in range(3)
]
constants.append(dp_mask)
else:
constants.append([K.cast_to_floatx(1.) for _ in range(3)])
if 0 < self.recurrent_dropout < 1:
ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
ones = K.tile(ones, (1, self.units))
def dropped_inputs(): # pylint: disable=function-redefined
return K.dropout(ones, self.recurrent_dropout)
rec_dp_mask = [
K.in_train_phase(dropped_inputs, ones, training=training)
for _ in range(3)
]
constants.append(rec_dp_mask)
else:
constants.append([K.cast_to_floatx(1.) for _ in range(3)])
return constants
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [state_ops.assign_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * ( # pylint: disable=g-no-augmented-assignment
1. / (1. + self.decay * math_ops.cast(self.iterations,
K.dtype(self.decay))))
t = math_ops.cast(self.iterations, K.floatx()) + 1
lr_t = lr * (
K.sqrt(1. - math_ops.pow(self.beta_2, t)) /
(1. - math_ops.pow(self.beta_1, t)))
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
if self.amsgrad:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
else:
vhats = [K.zeros(1) for _ in params]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * math_ops.square(g)
if self.amsgrad:
vhat_t = math_ops.maximum(vhat, v_t)
p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(state_ops.assign(vhat, vhat_t))
else:
p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
self.updates.append(state_ops.assign(m, m_t))
self.updates.append(state_ops.assign(v, v_t))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(state_ops.assign(p, new_p))
return self.updates
def __call__(self, inputs, initial_state=None, **kwargs):
if isinstance(inputs, list):
if len(inputs) > 1:
initial_state = inputs[1:]
inputs = inputs[0]
if initial_state is None:
return super(Bidirectional, self).__call__(inputs, **kwargs)
# Standardize `initial_state` into list
if isinstance(initial_state, tuple):
initial_state = list(initial_state)
elif not isinstance(initial_state, list):
initial_state = [initial_state]
# Check if `initial_state` can be splitted into half
num_states = len(initial_state)
if num_states % 2 > 0:
raise ValueError(
'When passing `initial_state` to a Bidirectional RNN, the state '
'should be a list containing the states of the underlying RNNs. '
'Found: ' + str(initial_state))
# Applies the same workaround as in `RNN.__call__`, without handling
# constants
kwargs['initial_state'] = initial_state
additional_inputs = initial_state
additional_specs = [InputSpec(shape=K.int_shape(state))
for state in initial_state]
self.forward_layer.state_spec = additional_specs[:num_states // 2]
self.backward_layer.state_spec = additional_specs[num_states // 2:]
is_keras_tensor = K.is_keras_tensor(additional_inputs[0])
for tensor in additional_inputs:
if K.is_keras_tensor(tensor) != is_keras_tensor:
raise ValueError('The initial state of a Bidirectional'
' layer cannot be specified with a mix of'
' Keras tensors and non-Keras tensors'
' (a "Keras tensor" is a tensor that was'
' returned by a Keras layer, or by `Input`)')
if is_keras_tensor:
# Compute the full input spec, including state
full_input = [inputs] + additional_inputs
full_input_spec = self.input_spec + additional_specs
# Perform the call with temporarily replaced input_spec
original_input_spec = self.input_spec
self.input_spec = full_input_spec
output = super(Bidirectional, self).__call__(full_input, **kwargs)
self.input_spec = original_input_spec
return output
else:
return super(Bidirectional, self).__call__(inputs, **kwargs)
def call(self, inputs, training=None, mask=None):
kwargs = {}
if generic_utils.has_arg(self.layer.call, 'training'):
kwargs['training'] = training
uses_learning_phase = False # pylint: disable=redefined-outer-name
input_shape = K.int_shape(inputs)
if input_shape[0]:
# batch size matters, use rnn-based implementation
def step(x, _):
global uses_learning_phase # pylint: disable=global-variable-undefined
output = self.layer.call(x, **kwargs)
if hasattr(output, '_uses_learning_phase'):
uses_learning_phase = (output._uses_learning_phase or
uses_learning_phase)
return output, []
_, outputs, _ = K.rnn(
step,
inputs,
initial_states=[],
input_length=input_shape[0],
unroll=False)
y = outputs
else:
# No batch size specified, therefore the layer will be able
# to process batches of any size.
# We can go with reshape-based implementation for performance.
input_length = input_shape[1]
if not input_length:
input_length = array_ops.shape(inputs)[1]
# Shape: (num_samples * timesteps, ...). And track the
# transformation in self._input_map.
input_uid = generic_utils.object_list_uid(inputs)
inputs = array_ops.reshape(inputs, (-1,) + input_shape[2:])
self._input_map[input_uid] = inputs
# (num_samples * timesteps, ...)
y = self.layer.call(inputs, **kwargs)
if hasattr(y, '_uses_learning_phase'):
uses_learning_phase = y._uses_learning_phase
# Shape: (num_samples, timesteps, ...)
output_shape = self.compute_output_shape(input_shape).as_list()
y = array_ops.reshape(y, (-1, input_length) + tuple(output_shape[2:]))
# Apply activity regularizer if any:
if (hasattr(self.layer, 'activity_regularizer') and
self.layer.activity_regularizer is not None):
regularization_loss = self.layer.activity_regularizer(y)
self.add_loss(regularization_loss, inputs)
if uses_learning_phase:
y._uses_learning_phase = True
return y
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [state_ops.assign_add(self.iterations, 1)]
t = math_ops.cast(self.iterations, K.floatx()) + 1
# Due to the recommendations in [2], i.e. warming momentum schedule
momentum_cache_t = self.beta_1 * (
1. - 0.5 *
(math_ops.pow(K.cast_to_floatx(0.96), t * self.schedule_decay)))
momentum_cache_t_1 = self.beta_1 * (
1. - 0.5 *
(math_ops.pow(K.cast_to_floatx(0.96), (t + 1) * self.schedule_decay)))
m_schedule_new = self.m_schedule * momentum_cache_t
m_schedule_next = self.m_schedule * momentum_cache_t * momentum_cache_t_1
self.updates.append((self.m_schedule, m_schedule_new))
shapes = [K.int_shape(p) for p in params]
ms = [K.zeros(shape) for shape in shapes]
vs = [K.zeros(shape) for shape in shapes]
self.weights = [self.iterations] + ms + vs
for p, g, m, v in zip(params, grads, ms, vs):
# the following equations given in [1]
g_prime = g / (1. - m_schedule_new)
m_t = self.beta_1 * m + (1. - self.beta_1) * g
m_t_prime = m_t / (1. - m_schedule_next)
v_t = self.beta_2 * v + (1. - self.beta_2) * math_ops.square(g)
v_t_prime = v_t / (1. - math_ops.pow(self.beta_2, t))
m_t_bar = (
1. - momentum_cache_t) * g_prime + momentum_cache_t_1 * m_t_prime
self.updates.append(state_ops.assign(m, m_t))
self.updates.append(state_ops.assign(v, v_t))
p_t = p - self.lr * m_t_bar / (K.sqrt(v_t_prime) + self.epsilon)
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(state_ops.assign(p, new_p))
return self.updates
def pop(self):
"""Removes the last layer in the model.
Raises:
TypeError: if there are no layers in the model.
"""
if not self.layers:
raise TypeError('There are no layers in the model.')
self.layers.pop()
if not self.layers:
self.outputs = []
self._inbound_nodes = []
self._outbound_nodes = []
else:
self.layers[-1]._outbound_nodes = []
self.outputs = [self.layers[-1].output]
# update self._inbound_nodes
self._inbound_nodes[0].output_tensors = self.outputs
self._inbound_nodes[0].output_shapes = [K.int_shape(self.outputs[0])]
self.built = False
def transition_block(x, reduction, name):
"""A transition block.
Arguments:
x: input tensor.
reduction: float, compression rate at transition layers.
name: string, block label.
Returns:
output tensor for the block.
"""
bn_axis = 3 if K.image_data_format() == 'channels_last' else 1
x = BatchNormalization(axis=bn_axis, epsilon=1.001e-5, name=name + '_bn')(x)
x = Activation('relu', name=name + '_relu')(x)
x = Conv2D(
int(K.int_shape(x)[bn_axis] * reduction),
1,
use_bias=False,
name=name + '_conv')(
x)
x = AveragePooling2D(2, strides=2, name=name + '_pool')(x)
return x
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
shapes = [K.int_shape(p) for p in params]
accumulators = [K.zeros(shape) for shape in shapes]
self.weights = accumulators
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr *= (1. /
(1. + self.decay * K.cast(self.iterations, K.dtype(self.decay))))
for p, g, a in zip(params, grads, accumulators):
new_a = a + K.square(g) # update accumulator
self.updates.append(K.update(a, new_a))
new_p = p - lr * g / (K.sqrt(new_a) + self.epsilon)
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def _eager_metrics_fn(model, outputs, targets):
"""Calculates the metrics for each output of the given model.
Arguments:
model: The model on which metrics are being calculated.
outputs: The outputs of the given model.
targets: The predictions or targets of the given model.
Returns:
Returns the metric names and metric results for each output of the model.
"""
metric_names = []
metric_results = []
if not isinstance(outputs, list):
outputs = [outputs]
if not isinstance(targets, list):
targets = [targets]
for i in range(len(model.outputs)):
output_metrics = model.nested_metrics[i]
for nested_output_metric in output_metrics:
metric_name, metric_fn = _get_metrics_info(
nested_output_metric, backend.int_shape(model.outputs[i]),
model.loss_functions[i])
if len(model.output_names) > 1:
metric_name = model.output_names[i] + '_' + metric_name
if metric_name not in model.metrics_names:
model.metrics_names.append(metric_name)
with backend.name_scope(metric_name):
metric_result = metric_fn(outputs[i], targets[i])
metric_names.append(metric_name)
metric_results.append(backend.mean(metric_result))
return metric_results
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
accumulators = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
self.weights = accumulators
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / # pylint: disable=g-no-augmented-assignment
(1. + self.decay * K.cast(self.iterations,
K.dtype(self.decay))))
for p, g, a in zip(params, grads, accumulators):
# update accumulator
new_a = self.rho * a + (1. - self.rho) * K.square(g)
self.updates.append(K.update(a, new_a))
new_p = p - lr * g / (K.sqrt(new_a) + self.epsilon)
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def call(self, inputs, mask=None, training=None, initial_state=None):
# input shape: `(samples, time (padded with zeros), input_dim)`
# note that the .build() method of subclasses MUST define
# self.input_spec and self.state_spec with complete input shapes.
if isinstance(inputs, list):
initial_state = inputs[1:]
inputs = inputs[0]
elif initial_state is not None:
pass
elif self.stateful:
initial_state = self.states
else:
initial_state = self.get_initial_state(inputs)
if isinstance(mask, list):
mask = mask[0]
if len(initial_state) != len(self.states):
raise ValueError('Layer has ' + str(len(self.states)) +
' states but was passed ' + str(len(initial_state)) +
' initial states.')
input_shape = K.int_shape(inputs)
if self.unroll and input_shape[1] is None:
raise ValueError('Cannot unroll a RNN if the '
'time dimension is undefined. \n'
'- If using a Sequential model, '
'specify the time dimension by passing '
'an `input_shape` or `batch_input_shape` '
'argument to your first layer. If your '
'first layer is an Embedding, you can '
'also use the `input_length` argument.\n'
'- If using the functional API, specify '
'the time dimension by passing a `shape` '
'or `batch_shape` argument to your Input layer.')
constants = self.get_constants(inputs, training=None)
preprocessed_input = self.preprocess_input(inputs, training=None)
last_output, outputs, states = K.rnn(
self.step,
preprocessed_input,
initial_state,
go_backwards=self.go_backwards,
mask=mask,
constants=constants,
unroll=self.unroll)
if self.stateful:
updates = []
for i in range(len(states)):
updates.append((self.states[i], states[i]))
self.add_update(updates, inputs)
# Properly set learning phase
if 0 < self.dropout + self.recurrent_dropout:
last_output._uses_learning_phase = True
outputs._uses_learning_phase = True
if not self.return_sequences:
outputs = last_output
if self.return_state:
if not isinstance(states, (list, tuple)):
states = [states]
else:
states = list(states)
return [outputs] + states
return outputs
def preprocess_weights_for_loading(layer,
weights,
original_keras_version=None,
original_backend=None):
"""Converts layers weights from Keras 1 format to Keras 2.
Arguments:
layer: Layer instance.
weights: List of weights values (Numpy arrays).
original_keras_version: Keras version for the weights, as a string.
original_backend: Keras backend the weights were trained with,
as a string.
Returns:
A list of weights values (Numpy arrays).
"""
if layer.__class__.__name__ == 'Bidirectional':
num_weights_per_layer = len(weights) // 2
forward_weights = preprocess_weights_for_loading(
layer.forward_layer, weights[:num_weights_per_layer],
original_keras_version, original_backend)
backward_weights = preprocess_weights_for_loading(
layer.backward_layer, weights[num_weights_per_layer:],
original_keras_version, original_backend)
weights = forward_weights + backward_weights
if original_keras_version == '1':
if layer.__class__.__name__ == 'TimeDistributed':
weights = preprocess_weights_for_loading(
layer.layer, weights, original_keras_version, original_backend)
if layer.__class__.__name__ == 'Conv1D':
shape = weights[0].shape
# Handle Keras 1.1 format
if shape[:2] != (layer.kernel_size[0], 1) or shape[3] != layer.filters:
# Legacy shape:
# (filters, input_dim, filter_length, 1)
assert shape[0] == layer.filters and shape[2:] == (layer.kernel_size[0],
1)
weights[0] = np.transpose(weights[0], (2, 3, 1, 0))
weights[0] = weights[0][:, 0, :, :]
if layer.__class__.__name__ == 'Conv2D':
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, stack_size, filters)
weights[0] = np.transpose(weights[0], (2, 3, 1, 0))
if layer.__class__.__name__ == 'Conv2DTranspose':
if layer.data_format == 'channels_last':
# old: (kernel_rows, kernel_cols, stack_size, filters)
# new: (kernel_rows, kernel_cols, filters, stack_size)
weights[0] = np.transpose(weights[0], (0, 1, 3, 2))
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, filters, stack_size)
weights[0] = np.transpose(weights[0], (2, 3, 0, 1))
if layer.__class__.__name__ == 'Conv3D':
if layer.data_format == 'channels_first':
# old: (filters, stack_size, ...)
# new: (..., stack_size, filters)
weights[0] = np.transpose(weights[0], (2, 3, 4, 1, 0))
if layer.__class__.__name__ == 'GRU':
if len(weights) == 9:
kernel = np.concatenate([weights[0], weights[3], weights[6]], axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[4], weights[7]], axis=-1)
bias = np.concatenate([weights[2], weights[5], weights[8]], axis=-1)
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ == 'LSTM':
if len(weights) == 12:
# old: i, c, f, o
# new: i, f, c, o
kernel = np.concatenate(
[weights[0], weights[6], weights[3], weights[9]], axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[7], weights[4], weights[10]], axis=-1)
bias = np.concatenate(
[weights[2], weights[8], weights[5], weights[11]], axis=-1)
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ == 'ConvLSTM2D':
if len(weights) == 12:
kernel = np.concatenate(
[weights[0], weights[6], weights[3], weights[9]], axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[7], weights[4], weights[10]], axis=-1)
bias = np.concatenate(
[weights[2], weights[8], weights[5], weights[11]], axis=-1)
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, stack_size, filters)
kernel = np.transpose(kernel, (2, 3, 1, 0))
recurrent_kernel = np.transpose(recurrent_kernel, (2, 3, 1, 0))
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ in ['Model', 'Sequential']:
new_weights = []
# trainable weights
for sublayer in layer.layers:
num_weights = len(sublayer.trainable_weights)
if num_weights > 0:
new_weights.extend(
preprocess_weights_for_loading(
layer=sublayer,
weights=weights[:num_weights],
original_keras_version=original_keras_version,
original_backend=original_backend))
weights = weights[num_weights:]
# non-trainable weights
for sublayer in layer.layers:
num_weights = len([
l for l in sublayer.weights if l not in sublayer.trainable_weights
])
if num_weights > 0:
new_weights.extend(
preprocess_weights_for_loading(
layer=sublayer,
weights=weights[:num_weights],
original_keras_version=original_keras_version,
original_backend=original_backend))
weights = weights[num_weights:]
weights = new_weights
conv_layers = ['Conv1D', 'Conv2D', 'Conv3D', 'Conv2DTranspose', 'ConvLSTM2D']
if layer.__class__.__name__ in conv_layers:
if original_backend == 'theano':
weights[0] = conv_utils.convert_kernel(weights[0])
if layer.__class__.__name__ == 'ConvLSTM2D':
weights[1] = conv_utils.convert_kernel(weights[1])
if K.int_shape(layer.weights[0]) != weights[0].shape:
weights[0] = np.transpose(weights[0], (3, 2, 0, 1))
if layer.__class__.__name__ == 'ConvLSTM2D':
weights[1] = np.transpose(weights[1], (3, 2, 0, 1))
# Convert the weights of CuDNNLSTM so that they could be loaded into LSTM
if layer.__class__.__name__ == 'LSTM' and len(weights) == 3:
# Determine if loading a CuDNNLSTM layer from the number of bias weights:
# CuDNNLSTM has (units * 8) weights; while LSTM has (units * 4)
# if there's no bias weight in the file, skip this conversion
units = weights[1].shape[0]
bias = weights[2]
if len(bias) == units * 8:
# reshape the kernels
kernels = np.split(weights[0], 4, axis=1)
kernels = [
kernel.reshape(-1).reshape(kernel.shape, order='F')
for kernel in kernels
]
weights[0] = np.concatenate(kernels, axis=1)
# transpose the recurrent kernels
recurrent_kernels = np.split(weights[1], 4, axis=1)
recurrent_kernels = [kernel.T for kernel in recurrent_kernels]
weights[1] = np.concatenate(recurrent_kernels, axis=1)
# split the bias into half and merge
weights[2] = bias[:units * 4] + bias[units * 4:]
return weights
def call(self,
inputs,
mask=None,
training=None,
initial_state=None,
constants=None):
# note that the .build() method of subclasses MUST define
# self.input_spec and self.state_spec with complete input shapes.
if isinstance(inputs, list):
inputs = inputs[0]
if initial_state is not None:
pass
elif self.stateful:
initial_state = self.states
else:
initial_state = self.get_initial_state(inputs)
if isinstance(mask, list):
mask = mask[0]
if len(initial_state) != len(self.states):
raise ValueError('Layer has ' + str(len(self.states)) +
' states but was passed ' +
str(len(initial_state)) +
' initial states.')
timesteps = K.int_shape(inputs)[1]
kwargs = {}
if generic_utils.has_arg(self.cell.call, 'training'):
kwargs['training'] = training
if constants:
if not generic_utils.has_arg(self.cell.call, 'constants'):
raise ValueError('RNN cell does not support constants')
def step(inputs, states):
constants = states[-self._num_constants:]
states = states[:-self._num_constants]
return self.cell.call(inputs, states, constants=constants,
**kwargs)
else:
def step(inputs, states):
return self.cell.call(inputs, states, **kwargs)
last_output, outputs, states = K.rnn(step,
inputs,
initial_state,
constants=constants,
go_backwards=self.go_backwards,
mask=mask,
input_length=timesteps)
if self.stateful:
updates = []
for i in range(len(states)):
updates.append(K.update(self.states[i], states[i]))
self.add_update(updates, inputs=True)
if self.return_sequences:
output = outputs
else:
output = last_output
# Properly set learning phase
if getattr(last_output, '_uses_learning_phase', False):
output._uses_learning_phase = True
if self.return_state:
if not isinstance(states, (list, tuple)):
states = [states]
else:
states = list(states)
return [output] + states
else:
return output
def __call__(self, inputs, initial_state=None, constants=None, **kwargs):
"""`Bidirectional.__call__` implements the same API as the wrapped `RNN`."""
inputs, initial_state, constants = _standardize_args(
inputs, initial_state, constants, self._num_constants)
if isinstance(inputs, list):
if len(inputs) > 1:
initial_state = inputs[1:]
inputs = inputs[0]
if initial_state is None and constants is None:
return super(Bidirectional, self).__call__(inputs, **kwargs)
# Applies the same workaround as in `RNN.__call__`
additional_inputs = []
additional_specs = []
if initial_state is not None:
# Check if `initial_state` can be splitted into half
num_states = len(initial_state)
if num_states % 2 > 0:
raise ValueError(
'When passing `initial_state` to a Bidirectional RNN, '
'the state should be a list containing the states of '
'the underlying RNNs. '
'Found: ' + str(initial_state))
kwargs['initial_state'] = initial_state
additional_inputs += initial_state
state_specs = [InputSpec(shape=K.int_shape(state))
for state in initial_state]
self.forward_layer.state_spec = state_specs[:num_states // 2]
self.backward_layer.state_spec = state_specs[num_states // 2:]
additional_specs += state_specs
if constants is not None:
kwargs['constants'] = constants
additional_inputs += constants
constants_spec = [InputSpec(shape=K.int_shape(constant))
for constant in constants]
self.forward_layer.constants_spec = constants_spec
self.backward_layer.constants_spec = constants_spec
additional_specs += constants_spec
self._num_constants = len(constants)
self.forward_layer._num_constants = self._num_constants
self.backward_layer._num_constants = self._num_constants
is_keras_tensor = K.is_keras_tensor(additional_inputs[0])
for tensor in additional_inputs:
if K.is_keras_tensor(tensor) != is_keras_tensor:
raise ValueError('The initial state of a Bidirectional'
' layer cannot be specified with a mix of'
' Keras tensors and non-Keras tensors'
' (a "Keras tensor" is a tensor that was'
' returned by a Keras layer, or by `Input`)')
if is_keras_tensor:
# Compute the full input spec, including state
full_input = [inputs] + additional_inputs
full_input_spec = self.input_spec + additional_specs
# Perform the call with temporarily replaced input_spec
original_input_spec = self.input_spec
self.input_spec = full_input_spec
output = super(Bidirectional, self).__call__(full_input, **kwargs)
self.input_spec = original_input_spec
return output
else:
return super(Bidirectional, self).__call__(inputs, **kwargs)
def inception_resnet_block(x, scale, block_type, block_idx, activation='relu'):
"""Adds a Inception-ResNet block.
This function builds 3 types of Inception-ResNet blocks mentioned
in the paper, controlled by the `block_type` argument (which is the
block name used in the official TF-slim implementation):
- Inception-ResNet-A: `block_type='block35'`
- Inception-ResNet-B: `block_type='block17'`
- Inception-ResNet-C: `block_type='block8'`
Arguments:
x: input tensor.
scale: scaling factor to scale the residuals (i.e., the output of
passing `x` through an inception module) before adding them
to the shortcut branch. Let `r` be the output from the residual
branch, the output of this block will be `x + scale * r`.
block_type: `'block35'`, `'block17'` or `'block8'`, determines
the network structure in the residual branch.
block_idx: an `int` used for generating layer names. The Inception-ResNet
blocks
are repeated many times in this network. We use `block_idx` to
identify
each of the repetitions. For example, the first Inception-ResNet-A
block
will have `block_type='block35', block_idx=0`, ane the layer names
will have
a common prefix `'block35_0'`.
activation: activation function to use at the end of the block
(see [activations](../activations.md)).
When `activation=None`, no activation is applied
(i.e., "linear" activation: `a(x) = x`).
Returns:
Output tensor for the block.
Raises:
ValueError: if `block_type` is not one of `'block35'`,
`'block17'` or `'block8'`.
"""
if block_type == 'block35':
branch_0 = conv2d_bn(x, 32, 1)
branch_1 = conv2d_bn(x, 32, 1)
branch_1 = conv2d_bn(branch_1, 32, 3)
branch_2 = conv2d_bn(x, 32, 1)
branch_2 = conv2d_bn(branch_2, 48, 3)
branch_2 = conv2d_bn(branch_2, 64, 3)
branches = [branch_0, branch_1, branch_2]
elif block_type == 'block17':
branch_0 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(x, 128, 1)
branch_1 = conv2d_bn(branch_1, 160, [1, 7])
branch_1 = conv2d_bn(branch_1, 192, [7, 1])
branches = [branch_0, branch_1]
elif block_type == 'block8':
branch_0 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(branch_1, 224, [1, 3])
branch_1 = conv2d_bn(branch_1, 256, [3, 1])
branches = [branch_0, branch_1]
else:
raise ValueError('Unknown Inception-ResNet block type. '
'Expects "block35", "block17" or "block8", '
'but got: ' + str(block_type))
block_name = block_type + '_' + str(block_idx)
channel_axis = 1 if K.image_data_format() == 'channels_first' else 3
mixed = Concatenate(axis=channel_axis,
name=block_name + '_mixed')(branches)
up = conv2d_bn(mixed,
K.int_shape(x)[channel_axis],
1,
activation=None,
use_bias=True,
name=block_name + '_conv')
x = Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
arguments={'scale': scale},
name=block_name)([x, up])
if activation is not None:
x = Activation(activation, name=block_name + '_ac')(x)
return x
def call(self,
inputs,
mask=None,
training=None,
initial_state=None,
constants=None):
# note that the .build() method of subclasses MUST define
# self.input_spec and self.state_spec with complete input shapes.
if isinstance(inputs, list):
inputs = inputs[0]
if initial_state is not None:
pass
elif self.stateful:
initial_state = self.states
else:
initial_state = self.get_initial_state(inputs)
if isinstance(mask, list):
mask = mask[0]
if len(initial_state) != len(self.states):
raise ValueError('Layer has ' + str(len(self.states)) +
' states but was passed ' +
str(len(initial_state)) + ' initial states.')
timesteps = K.int_shape(inputs)[1]
kwargs = {}
if generic_utils.has_arg(self.cell.call, 'training'):
kwargs['training'] = training
if constants:
if not generic_utils.has_arg(self.cell.call, 'constants'):
raise ValueError('RNN cell does not support constants')
def step(inputs, states):
constants = states[-self._num_constants:]
states = states[:-self._num_constants]
return self.cell.call(inputs,
states,
constants=constants,
**kwargs)
else:
def step(inputs, states):
return self.cell.call(inputs, states, **kwargs)
last_output, outputs, states = K.rnn(step,
inputs,
initial_state,
constants=constants,
go_backwards=self.go_backwards,
mask=mask,
input_length=timesteps)
if self.stateful:
updates = []
for i in range(len(states)):
updates.append(K.update(self.states[i], states[i]))
self.add_update(updates, inputs=True)
if self.return_sequences:
output = outputs
else:
output = last_output
# Properly set learning phase
if getattr(last_output, '_uses_learning_phase', False):
output._uses_learning_phase = True
if self.return_state:
if not isinstance(states, (list, tuple)):
states = [states]
else:
states = list(states)
return [output] + states
else:
return output
def add(self, layer):
"""Adds a layer instance on top of the layer stack.
Arguments:
layer: layer instance.
Raises:
TypeError: If `layer` is not a layer instance.
ValueError: In case the `layer` argument does not
know its input shape.
ValueError: In case the `layer` argument has
multiple output tensors, or is already connected
somewhere else (forbidden in `Sequential` models).
"""
if not isinstance(layer, (base_layer.Layer, base_layer.TFBaseLayer)):
raise TypeError('The added layer must be '
'an instance of class Layer. '
'Found: ' + str(layer))
if not self.outputs:
# First layer in model: check that it is an input layer.
if not isinstance(layer, InputLayer):
# Create an input layer.
# First, we need to infer its expected input shape and dtype.
if isinstance(layer, (Model, Sequential)):
# We were passed a model as first layer.
# This requires a specific way to figure out the
# input shape and dtype.
if not layer.layers:
raise ValueError('Cannot add an empty model '
'to a `Sequential` model.')
# In case of nested models: recover the first layer
# of the deepest model to infer input shape and dtype.
first_layer = layer.layers[0]
while isinstance(first_layer, (Model, Sequential)):
first_layer = first_layer.layers[0]
batch_shape = first_layer._batch_input_shape
dtype = first_layer.dtype
else:
# We were passed a regular layer, and it should
# know about its input shape. Otherwise, that's an error.
if not hasattr(layer, '_batch_input_shape'):
raise ValueError('The first layer in a '
'Sequential model must '
'get an `input_shape` argument.')
batch_shape = layer._batch_input_shape
dtype = layer.dtype
# Instantiate the input layer.
x = Input(batch_shape=batch_shape,
dtype=dtype,
name=layer.name + '_input')
# This will build the current layer
# and create the node connecting the current layer
# to the input layer we just created.
layer(x)
if len(layer._inbound_nodes[-1].output_tensors) != 1:
raise ValueError('All layers in a Sequential model '
'should have a single output tensor. '
'For multi-output layers, '
'use the functional API.')
self.outputs = [layer._inbound_nodes[-1].output_tensors[0]]
self.inputs = network.get_source_inputs(self.outputs[0])
# We create an input node, which we will keep updated
# as we add more layers
base_layer.Node(outbound_layer=self,
inbound_layers=[],
node_indices=[],
tensor_indices=[],
input_tensors=self.inputs,
output_tensors=self.outputs)
else:
output_tensor = layer(self.outputs[0])
if isinstance(output_tensor, list):
raise TypeError('All layers in a Sequential model '
'should have a single output tensor. '
'For multi-output layers, '
'use the functional API.')
self.outputs = [output_tensor]
# update self._inbound_nodes
self._inbound_nodes[0].output_tensors = self.outputs
self._inbound_nodes[0].output_shapes = [
K.int_shape(self.outputs[0])
]
self._layers.append(layer)
self.built = False
def preprocess_weights_for_loading(layer,
weights,
original_keras_version=None,
original_backend=None):
"""Converts layers weights from Keras 1 format to Keras 2.
Arguments:
layer: Layer instance.
weights: List of weights values (Numpy arrays).
original_keras_version: Keras version for the weights, as a string.
original_backend: Keras backend the weights were trained with,
as a string.
Returns:
A list of weights values (Numpy arrays).
"""
if layer.__class__.__name__ == 'Bidirectional':
num_weights_per_layer = len(weights) // 2
forward_weights = preprocess_weights_for_loading(
layer.forward_layer, weights[:num_weights_per_layer],
original_keras_version, original_backend)
backward_weights = preprocess_weights_for_loading(
layer.backward_layer, weights[num_weights_per_layer:],
original_keras_version, original_backend)
weights = forward_weights + backward_weights
if original_keras_version == '1':
if layer.__class__.__name__ == 'TimeDistributed':
weights = preprocess_weights_for_loading(
layer.layer, weights, original_keras_version, original_backend)
if layer.__class__.__name__ == 'Conv1D':
shape = weights[0].shape
# Handle Keras 1.1 format
if shape[:2] != (layer.kernel_size[0], 1) or shape[3] != layer.filters:
# Legacy shape:
# (filters, input_dim, filter_length, 1)
assert shape[0] == layer.filters and shape[2:] == (layer.kernel_size[0],
1)
weights[0] = np.transpose(weights[0], (2, 3, 1, 0))
weights[0] = weights[0][:, 0, :, :]
if layer.__class__.__name__ == 'Conv2D':
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, stack_size, filters)
weights[0] = np.transpose(weights[0], (2, 3, 1, 0))
if layer.__class__.__name__ == 'Conv2DTranspose':
if layer.data_format == 'channels_last':
# old: (kernel_rows, kernel_cols, stack_size, filters)
# new: (kernel_rows, kernel_cols, filters, stack_size)
weights[0] = np.transpose(weights[0], (0, 1, 3, 2))
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, filters, stack_size)
weights[0] = np.transpose(weights[0], (2, 3, 0, 1))
if layer.__class__.__name__ == 'Conv3D':
if layer.data_format == 'channels_first':
# old: (filters, stack_size, ...)
# new: (..., stack_size, filters)
weights[0] = np.transpose(weights[0], (2, 3, 4, 1, 0))
if layer.__class__.__name__ == 'GRU':
if len(weights) == 9:
kernel = np.concatenate([weights[0], weights[3], weights[6]], axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[4], weights[7]], axis=-1)
bias = np.concatenate([weights[2], weights[5], weights[8]], axis=-1)
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ == 'LSTM':
if len(weights) == 12:
# old: i, c, f, o
# new: i, f, c, o
kernel = np.concatenate(
[weights[0], weights[6], weights[3], weights[9]], axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[7], weights[4], weights[10]], axis=-1)
bias = np.concatenate(
[weights[2], weights[8], weights[5], weights[11]], axis=-1)
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ == 'ConvLSTM2D':
if len(weights) == 12:
kernel = np.concatenate(
[weights[0], weights[6], weights[3], weights[9]], axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[7], weights[4], weights[10]], axis=-1)
bias = np.concatenate(
[weights[2], weights[8], weights[5], weights[11]], axis=-1)
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, stack_size, filters)
kernel = np.transpose(kernel, (2, 3, 1, 0))
recurrent_kernel = np.transpose(recurrent_kernel, (2, 3, 1, 0))
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ in ['Model', 'Sequential']:
new_weights = []
# trainable weights
for sublayer in layer.layers:
num_weights = len(sublayer.trainable_weights)
if num_weights > 0:
new_weights.extend(
preprocess_weights_for_loading(
layer=sublayer,
weights=weights[:num_weights],
original_keras_version=original_keras_version,
original_backend=original_backend))
weights = weights[num_weights:]
# non-trainable weights
for sublayer in layer.layers:
num_weights = len([
l for l in sublayer.weights if l not in sublayer.trainable_weights
])
if num_weights > 0:
new_weights.extend(
preprocess_weights_for_loading(
layer=sublayer,
weights=weights[:num_weights],
original_keras_version=original_keras_version,
original_backend=original_backend))
weights = weights[num_weights:]
weights = new_weights
conv_layers = ['Conv1D', 'Conv2D', 'Conv3D', 'Conv2DTranspose', 'ConvLSTM2D']
if layer.__class__.__name__ in conv_layers:
if original_backend == 'theano':
weights[0] = conv_utils.convert_kernel(weights[0])
if layer.__class__.__name__ == 'ConvLSTM2D':
weights[1] = conv_utils.convert_kernel(weights[1])
if K.int_shape(layer.weights[0]) != weights[0].shape:
weights[0] = np.transpose(weights[0], (3, 2, 0, 1))
if layer.__class__.__name__ == 'ConvLSTM2D':
weights[1] = np.transpose(weights[1], (3, 2, 0, 1))
return _convert_rnn_weights(layer, weights)
def add(self, layer):
"""Adds a layer instance on top of the layer stack.
Arguments:
layer: layer instance.
Raises:
TypeError: If `layer` is not a layer instance.
ValueError: In case the `layer` argument does not
know its input shape.
ValueError: In case the `layer` argument has
multiple output tensors, or is already connected
somewhere else (forbidden in `Sequential` models).
"""
if not isinstance(layer, (base_layer.Layer, base_layer.TFBaseLayer)):
raise TypeError('The added layer must be '
'an instance of class Layer. '
'Found: ' + str(layer))
if not self.outputs:
# First layer in model: check that it is an input layer.
if not isinstance(layer, InputLayer):
# Create an input layer.
# First, we need to infer its expected input shape and dtype.
if isinstance(layer, (Model, Sequential)):
# We were passed a model as first layer.
# This requires a specific way to figure out the
# input shape and dtype.
if not layer.layers:
raise ValueError('Cannot add an empty model '
'to a `Sequential` model.')
# In case of nested models: recover the first layer
# of the deepest model to infer input shape and dtype.
first_layer = layer.layers[0]
while isinstance(first_layer, (Model, Sequential)):
first_layer = first_layer.layers[0]
batch_shape = first_layer._batch_input_shape
dtype = first_layer.dtype
else:
# We were passed a regular layer, and it should
# know about its input shape. Otherwise, that's an error.
if not hasattr(layer, '_batch_input_shape'):
raise ValueError('The first layer in a '
'Sequential model must '
'get an `input_shape` argument.')
batch_shape = layer._batch_input_shape
dtype = layer.dtype
# Instantiate the input layer.
x = Input(
batch_shape=batch_shape, dtype=dtype, name=layer.name + '_input')
# This will build the current layer
# and create the node connecting the current layer
# to the input layer we just created.
layer(x)
if len(layer._inbound_nodes[-1].output_tensors) != 1:
raise ValueError('All layers in a Sequential model '
'should have a single output tensor. '
'For multi-output layers, '
'use the functional API.')
self.outputs = [layer._inbound_nodes[-1].output_tensors[0]]
self.inputs = network.get_source_inputs(self.outputs[0])
# We create an input node, which we will keep updated
# as we add more layers
base_layer.Node(
outbound_layer=self,
inbound_layers=[],
node_indices=[],
tensor_indices=[],
input_tensors=self.inputs,
output_tensors=self.outputs)
else:
output_tensor = layer(self.outputs[0])
if isinstance(output_tensor, list):
raise TypeError('All layers in a Sequential model '
'should have a single output tensor. '
'For multi-output layers, '
'use the functional API.')
self.outputs = [output_tensor]
# update self._inbound_nodes
self._inbound_nodes[0].output_tensors = self.outputs
self._inbound_nodes[0].output_shapes = [K.int_shape(self.outputs[0])]
self._layers.append(layer)
self.built = False
def inception_resnet_block(x, scale, block_type, block_idx, activation='relu'):
"""Adds a Inception-ResNet block.
This function builds 3 types of Inception-ResNet blocks mentioned
in the paper, controlled by the `block_type` argument (which is the
block name used in the official TF-slim implementation):
- Inception-ResNet-A: `block_type='block35'`
- Inception-ResNet-B: `block_type='block17'`
- Inception-ResNet-C: `block_type='block8'`
Arguments:
x: input tensor.
scale: scaling factor to scale the residuals (i.e., the output of
passing `x` through an inception module) before adding them
to the shortcut branch. Let `r` be the output from the residual
branch,
the output of this block will be `x + scale * r`.
block_type: `'block35'`, `'block17'` or `'block8'`, determines
the network structure in the residual branch.
block_idx: an `int` used for generating layer names. The Inception-ResNet
blocks
are repeated many times in this network. We use `block_idx` to
identify
each of the repetitions. For example, the first Inception-ResNet-A
block
will have `block_type='block35', block_idx=0`, ane the layer names
will have
a common prefix `'block35_0'`.
activation: activation function to use at the end of the block.
When `activation=None`, no activation is applied
(i.e., "linear" activation: `a(x) = x`).
Returns:
Output tensor for the block.
Raises:
ValueError: if `block_type` is not one of `'block35'`,
`'block17'` or `'block8'`.
"""
if block_type == 'block35':
branch_0 = conv2d_bn(x, 32, 1)
branch_1 = conv2d_bn(x, 32, 1)
branch_1 = conv2d_bn(branch_1, 32, 3)
branch_2 = conv2d_bn(x, 32, 1)
branch_2 = conv2d_bn(branch_2, 48, 3)
branch_2 = conv2d_bn(branch_2, 64, 3)
branches = [branch_0, branch_1, branch_2]
elif block_type == 'block17':
branch_0 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(x, 128, 1)
branch_1 = conv2d_bn(branch_1, 160, [1, 7])
branch_1 = conv2d_bn(branch_1, 192, [7, 1])
branches = [branch_0, branch_1]
elif block_type == 'block8':
branch_0 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(branch_1, 224, [1, 3])
branch_1 = conv2d_bn(branch_1, 256, [3, 1])
branches = [branch_0, branch_1]
else:
raise ValueError('Unknown Inception-ResNet block type. '
'Expects "block35", "block17" or "block8", '
'but got: ' + str(block_type))
block_name = block_type + '_' + str(block_idx)
channel_axis = 1 if K.image_data_format() == 'channels_first' else 3
mixed = Concatenate(axis=channel_axis, name=block_name + '_mixed')(branches)
up = conv2d_bn(
mixed,
K.int_shape(x)[channel_axis],
1,
activation=None,
use_bias=True,
name=block_name + '_conv')
x = Lambda(
lambda inputs, scale: inputs[0] + inputs[1] * scale,
output_shape=K.int_shape(x)[1:],
arguments={'scale': scale},
name=block_name)([x, up])
if activation is not None:
x = Activation(activation, name=block_name + '_ac')(x)
return x
def _adjust_block(p, ip, filters, block_id=None):
"""Adjusts the input `previous path` to match the shape of the `input`.
Used in situations where the output number of filters needs to be changed.
Arguments:
p: Input tensor which needs to be modified
ip: Input tensor whose shape needs to be matched
filters: Number of output filters to be matched
block_id: String block_id
Returns:
Adjusted Keras tensor
"""
channel_dim = 1 if K.image_data_format() == 'channels_first' else -1
img_dim = 2 if K.image_data_format() == 'channels_first' else -2
ip_shape = K.int_shape(ip)
if p is not None:
p_shape = K.int_shape(p)
with K.name_scope('adjust_block'):
if p is None:
p = ip
elif p_shape[img_dim] != ip_shape[img_dim]:
with K.name_scope('adjust_reduction_block_%s' % block_id):
p = Activation('relu', name='adjust_relu_1_%s' % block_id)(p)
p1 = AveragePooling2D(
(1, 1),
strides=(2, 2),
padding='valid',
name='adjust_avg_pool_1_%s' % block_id)(p)
p1 = Conv2D(filters // 2, (1, 1),
padding='same',
use_bias=False,
name='adjust_conv_1_%s' % block_id,
kernel_initializer='he_normal')(p1)
p2 = ZeroPadding2D(padding=((0, 1), (0, 1)))(p)
p2 = Cropping2D(cropping=((1, 0), (1, 0)))(p2)
p2 = AveragePooling2D(
(1, 1),
strides=(2, 2),
padding='valid',
name='adjust_avg_pool_2_%s' % block_id)(p2)
p2 = Conv2D(filters // 2, (1, 1),
padding='same',
use_bias=False,
name='adjust_conv_2_%s' % block_id,
kernel_initializer='he_normal')(p2)
p = concatenate([p1, p2], axis=channel_dim)
p = BatchNormalization(axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='adjust_bn_%s' % block_id)(p)
elif p_shape[channel_dim] != filters:
with K.name_scope('adjust_projection_block_%s' % block_id):
p = Activation('relu')(p)
p = Conv2D(filters, (1, 1),
strides=(1, 1),
padding='same',
name='adjust_conv_projection_%s' % block_id,
use_bias=False,
kernel_initializer='he_normal')(p)
p = BatchNormalization(axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='adjust_bn_%s' % block_id)(p)
return p
def preprocess_weights_for_loading(layer,
weights,
original_keras_version=None,
original_backend=None):
"""Converts layers weights from Keras 1 format to Keras 2.
Arguments:
layer: Layer instance.
weights: List of weights values (Numpy arrays).
original_keras_version: Keras version for the weights, as a string.
original_backend: Keras backend the weights were trained with,
as a string.
Returns:
A list of weights values (Numpy arrays).
"""
if layer.__class__.__name__ == 'Bidirectional':
num_weights_per_layer = len(weights) // 2
forward_weights = preprocess_weights_for_loading(
layer.forward_layer, weights[:num_weights_per_layer],
original_keras_version, original_backend)
backward_weights = preprocess_weights_for_loading(
layer.backward_layer, weights[num_weights_per_layer:],
original_keras_version, original_backend)
weights = forward_weights + backward_weights
if original_keras_version == '1':
if layer.__class__.__name__ == 'TimeDistributed':
weights = preprocess_weights_for_loading(layer.layer, weights,
original_keras_version,
original_backend)
if layer.__class__.__name__ == 'Conv1D':
shape = weights[0].shape
# Handle Keras 1.1 format
if shape[:2] != (layer.kernel_size[0],
1) or shape[3] != layer.filters:
# Legacy shape:
# (filters, input_dim, filter_length, 1)
assert shape[0] == layer.filters and shape[2:] == (
layer.kernel_size[0], 1)
weights[0] = np.transpose(weights[0], (2, 3, 1, 0))
weights[0] = weights[0][:, 0, :, :]
if layer.__class__.__name__ == 'Conv2D':
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, stack_size, filters)
weights[0] = np.transpose(weights[0], (2, 3, 1, 0))
if layer.__class__.__name__ == 'Conv2DTranspose':
if layer.data_format == 'channels_last':
# old: (kernel_rows, kernel_cols, stack_size, filters)
# new: (kernel_rows, kernel_cols, filters, stack_size)
weights[0] = np.transpose(weights[0], (0, 1, 3, 2))
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, filters, stack_size)
weights[0] = np.transpose(weights[0], (2, 3, 0, 1))
if layer.__class__.__name__ == 'Conv3D':
if layer.data_format == 'channels_first':
# old: (filters, stack_size, ...)
# new: (..., stack_size, filters)
weights[0] = np.transpose(weights[0], (2, 3, 4, 1, 0))
if layer.__class__.__name__ == 'GRU':
if len(weights) == 9:
kernel = np.concatenate([weights[0], weights[3], weights[6]],
axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[4], weights[7]], axis=-1)
bias = np.concatenate([weights[2], weights[5], weights[8]],
axis=-1)
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ == 'LSTM':
if len(weights) == 12:
# old: i, c, f, o
# new: i, f, c, o
kernel = np.concatenate(
[weights[0], weights[6], weights[3], weights[9]], axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[7], weights[4], weights[10]], axis=-1)
bias = np.concatenate(
[weights[2], weights[8], weights[5], weights[11]], axis=-1)
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ == 'ConvLSTM2D':
if len(weights) == 12:
kernel = np.concatenate(
[weights[0], weights[6], weights[3], weights[9]], axis=-1)
recurrent_kernel = np.concatenate(
[weights[1], weights[7], weights[4], weights[10]], axis=-1)
bias = np.concatenate(
[weights[2], weights[8], weights[5], weights[11]], axis=-1)
if layer.data_format == 'channels_first':
# old: (filters, stack_size, kernel_rows, kernel_cols)
# new: (kernel_rows, kernel_cols, stack_size, filters)
kernel = np.transpose(kernel, (2, 3, 1, 0))
recurrent_kernel = np.transpose(recurrent_kernel,
(2, 3, 1, 0))
weights = [kernel, recurrent_kernel, bias]
if layer.__class__.__name__ in ['Model', 'Sequential']:
new_weights = []
# trainable weights
for sublayer in layer.layers:
num_weights = len(sublayer.trainable_weights)
if num_weights > 0:
new_weights.extend(
preprocess_weights_for_loading(
layer=sublayer,
weights=weights[:num_weights],
original_keras_version=original_keras_version,
original_backend=original_backend))
weights = weights[num_weights:]
# non-trainable weights
for sublayer in layer.layers:
num_weights = len([
l for l in sublayer.weights
if l not in sublayer.trainable_weights
])
if num_weights > 0:
new_weights.extend(
preprocess_weights_for_loading(
layer=sublayer,
weights=weights[:num_weights],
original_keras_version=original_keras_version,
original_backend=original_backend))
weights = weights[num_weights:]
weights = new_weights
conv_layers = [
'Conv1D', 'Conv2D', 'Conv3D', 'Conv2DTranspose', 'ConvLSTM2D'
]
if layer.__class__.__name__ in conv_layers:
if original_backend == 'theano':
weights[0] = conv_utils.convert_kernel(weights[0])
if layer.__class__.__name__ == 'ConvLSTM2D':
weights[1] = conv_utils.convert_kernel(weights[1])
if K.int_shape(layer.weights[0]) != weights[0].shape:
weights[0] = np.transpose(weights[0], (3, 2, 0, 1))
if layer.__class__.__name__ == 'ConvLSTM2D':
weights[1] = np.transpose(weights[1], (3, 2, 0, 1))
return _convert_rnn_weights(layer, weights)
def _adjust_block(p, ip, filters, block_id=None):
"""Adjusts the input `previous path` to match the shape of the `input`.
Used in situations where the output number of filters needs to be changed.
Arguments:
p: Input tensor which needs to be modified
ip: Input tensor whose shape needs to be matched
filters: Number of output filters to be matched
block_id: String block_id
Returns:
Adjusted Keras tensor
"""
channel_dim = 1 if K.image_data_format() == 'channels_first' else -1
img_dim = 2 if K.image_data_format() == 'channels_first' else -2
ip_shape = K.int_shape(ip)
if p is not None:
p_shape = K.int_shape(p)
with K.name_scope('adjust_block'):
if p is None:
p = ip
elif p_shape[img_dim] != ip_shape[img_dim]:
with K.name_scope('adjust_reduction_block_%s' % block_id):
p = Activation('relu', name='adjust_relu_1_%s' % block_id)(p)
p1 = AveragePooling2D(
(1, 1),
strides=(2, 2),
padding='valid',
name='adjust_avg_pool_1_%s' % block_id)(
p)
p1 = Conv2D(
filters // 2, (1, 1),
padding='same',
use_bias=False,
name='adjust_conv_1_%s' % block_id,
kernel_initializer='he_normal')(
p1)
p2 = ZeroPadding2D(padding=((0, 1), (0, 1)))(p)
p2 = Cropping2D(cropping=((1, 0), (1, 0)))(p2)
p2 = AveragePooling2D(
(1, 1),
strides=(2, 2),
padding='valid',
name='adjust_avg_pool_2_%s' % block_id)(
p2)
p2 = Conv2D(
filters // 2, (1, 1),
padding='same',
use_bias=False,
name='adjust_conv_2_%s' % block_id,
kernel_initializer='he_normal')(
p2)
p = concatenate([p1, p2], axis=channel_dim)
p = BatchNormalization(
axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='adjust_bn_%s' % block_id)(
p)
elif p_shape[channel_dim] != filters:
with K.name_scope('adjust_projection_block_%s' % block_id):
p = Activation('relu')(p)
p = Conv2D(
filters, (1, 1),
strides=(1, 1),
padding='same',
name='adjust_conv_projection_%s' % block_id,
use_bias=False,
kernel_initializer='he_normal')(
p)
p = BatchNormalization(
axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='adjust_bn_%s' % block_id)(
p)
return p