def convert_all_kernels_in_model(model):
"""Converts all convolution kernels in a model from Theano to TensorFlow.
Also works from TensorFlow to Theano.
This is used for converting legacy Theano-saved model files.
Arguments:
model: target model for the conversion.
"""
# Note: SeparableConvolution not included
# since only supported by TF.
conv_classes = {
'Conv1D',
'Conv2D',
'Conv3D',
'Conv2DTranspose',
}
to_assign = []
for layer in model.layers:
if layer.__class__.__name__ in conv_classes:
original_kernel = K.get_value(layer.kernel)
converted_kernel = convert_kernel(original_kernel)
to_assign.append((layer.kernel, converted_kernel))
K.batch_set_value(to_assign)
def _flip_conv_weights(self):
# Theano > Tensorflow, just flips the weight arrays in the first 2 dims. Doesn't change shape.
for layer in self.net.layers:
if layer.__class__.__name__ == 'Conv2D':
from tensorflow.keras import backend
original_w = backend.eval(layer.kernel)
from tensorflow.python.keras.utils.conv_utils import convert_kernel
converted_w = convert_kernel(original_w)
layer.kernel.assign(converted_w)
def convert_all_kernels_in_model(model):
"""Converts all convolution kernels in a model from Theano to TensorFlow.
Also works from TensorFlow to Theano.
Arguments:
model: target model for the conversion.
"""
# Note: SeparableConvolution not included
# since only supported by TF.
conv_classes = {
'Conv1D',
'Conv2D',
'Conv3D',
'Conv2DTranspose',
}
to_assign = []
for layer in model.layers:
if layer.__class__.__name__ in conv_classes:
original_kernel = K.get_value(layer.kernel)
converted_kernel = convert_kernel(original_kernel)
to_assign.append((layer.kernel, converted_kernel))
K.batch_set_value(to_assign)
def preprocess_weights_for_loading(layer,
weights,
original_keras_version=None,
original_backend=None):
"""Preprocess layer weights between different Keras formats.
Converts layers weights from Keras 1 format to Keras 2 and also weights of
CuDNN layers in 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).
"""
def convert_nested_bidirectional(weights):
"""Converts layers nested in `Bidirectional` wrapper.
This function uses `preprocess_weights_for_loading()` for converting
layers.
Arguments:
weights: List of weights values (Numpy arrays).
Returns:
A list of weights values (Numpy arrays).
"""
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)
return forward_weights + backward_weights
def convert_nested_time_distributed(weights):
"""Converts layers nested in `TimeDistributed` wrapper.
This function uses `preprocess_weights_for_loading()` for converting nested
layers.
Arguments:
weights: List of weights values (Numpy arrays).
Returns:
A list of weights values (Numpy arrays).
"""
return preprocess_weights_for_loading(
layer.layer, weights, original_keras_version, original_backend)
def convert_nested_model(weights):
"""Converts layers nested in `Model` or `Sequential`.
This function uses `preprocess_weights_for_loading()` for converting nested
layers.
Arguments:
weights: List of weights values (Numpy arrays).
Returns:
A list of weights values (Numpy arrays).
"""
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:]
return new_weights
# Convert layers nested in Bidirectional/Model/Sequential.
# Both transformation should be ran for both Keras 1->2 conversion
# and for conversion of CuDNN layers.
if layer.__class__.__name__ == 'Bidirectional':
weights = convert_nested_bidirectional(weights)
if layer.__class__.__name__ == 'TimeDistributed':
weights = convert_nested_time_distributed(weights)
elif layer.__class__.__name__ in ['Model', 'Sequential']:
weights = convert_nested_model(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]
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 CuDNN layers
return _convert_rnn_weights(layer, weights)
def preprocess_weights_for_loading(layer,
weights,
original_keras_version=None,
original_backend=None):
"""Preprocess layer weights between different Keras formats.
Converts layers weights from Keras 1 format to Keras 2 and also weights of
CuDNN layers in 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).
"""
def convert_nested_bidirectional(weights):
"""Converts layers nested in `Bidirectional` wrapper.
This function uses `preprocess_weights_for_loading()` for converting
layers.
Arguments:
weights: List of weights values (Numpy arrays).
Returns:
A list of weights values (Numpy arrays).
"""
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)
return forward_weights + backward_weights
def convert_nested_time_distributed(weights):
"""Converts layers nested in `TimeDistributed` wrapper.
This function uses `preprocess_weights_for_loading()` for converting nested
layers.
Arguments:
weights: List of weights values (Numpy arrays).
Returns:
A list of weights values (Numpy arrays).
"""
return preprocess_weights_for_loading(layer.layer, weights,
original_keras_version,
original_backend)
def convert_nested_model(weights):
"""Converts layers nested in `Model` or `Sequential`.
This function uses `preprocess_weights_for_loading()` for converting nested
layers.
Arguments:
weights: List of weights values (Numpy arrays).
Returns:
A list of weights values (Numpy arrays).
"""
trainable_weights = weights[:len(layer.trainable_weights)]
non_trainable_weights = weights[len(layer.trainable_weights):]
new_trainable_weights = []
new_non_trainable_weights = []
for sublayer in layer.layers:
num_trainable_weights = len(sublayer.trainable_weights)
num_non_trainable_weights = len(sublayer.non_trainable_weights)
if sublayer.weights:
preprocessed = preprocess_weights_for_loading(
layer=sublayer,
weights=(
trainable_weights[:num_trainable_weights] +
non_trainable_weights[:num_non_trainable_weights]),
original_keras_version=original_keras_version,
original_backend=original_backend)
new_trainable_weights.extend(
preprocessed[:num_trainable_weights])
new_non_trainable_weights.extend(
preprocessed[num_trainable_weights:])
trainable_weights = trainable_weights[num_trainable_weights:]
non_trainable_weights = non_trainable_weights[
num_non_trainable_weights:]
return new_trainable_weights + new_non_trainable_weights
# Convert layers nested in Bidirectional/Model/Sequential.
# Both transformation should be ran for both Keras 1->2 conversion
# and for conversion of CuDNN layers.
if layer.__class__.__name__ == 'Bidirectional':
weights = convert_nested_bidirectional(weights)
if layer.__class__.__name__ == 'TimeDistributed':
weights = convert_nested_time_distributed(weights)
elif layer.__class__.__name__ in ['Model', 'Sequential']:
weights = convert_nested_model(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]
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 CuDNN layers
return _convert_rnn_weights(layer, weights)
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 create_cnn_model(weights_path=None):
# creates our cnn model
#filters which total weights is “n*m*k*l” (Here the input has l=32 feature maps as inputs, k=64 feature maps as outputs)
#Then there is a term called bias for each feature map. So, the total number of parameters are “(n*m*l+1)*k”.
'''
PARAMETERS
https://towardsdatascience.com/understanding-and-calculating-the-number-of-parameters-in-convolution-neural-networks-cnns-fc88790d530d
https://medium.com/@shashikachamod4u/calculate-output-size-and-number-of-trainable-parameters-in-a-convolution-layer-1d64cae6c009
https://medium.com/@iamvarman/how-to-calculate-the-number-of-parameters-in-the-cnn-5bd55364d7ca
https://cs231n.github.io/convolutional-networks/
'''
input = Input(shape=(1, IMG_WIDTH, IMG_HEIGHT))
input_pad = ZeroPadding2D(padding=(3, 3))(input)
conv1_1_3x3_s1 = Conv2D(32, (3, 3),
strides=(1, 1),
padding='same',
activation='relu',
name='conv1_1/3x3_s1',
kernel_regularizer=l2(l2_regulizer))(input_pad)
conv1_2_3x3_s1 = Conv2D(
32, (3, 3),
strides=(1, 1),
padding='same',
activation='relu',
name='conv1_2/3x3_s1',
kernel_regularizer=l2(l2_regulizer))(conv1_1_3x3_s1)
conv1_zero_pad = ZeroPadding2D(padding=(1, 1))(conv1_2_3x3_s1)
pool1_helper = PoolHelper()(conv1_zero_pad)
pool1_2_2x2_s1 = MaxPooling2D(pool_size=(2, 2),
strides=(1, 1),
padding='same',
name='pool1/2x2_s1')(pool1_helper)
pool1_norm1 = LRN(name='pool1/norm1')(pool1_2_2x2_s1)
conv2_1_3x3_reduce = Conv2D(
64, (1, 1),
padding='same',
activation='relu',
name='conv2_1/3x3_reduce',
kernel_regularizer=l2(l2_regulizer))(pool1_norm1)
conv2_2_3x3 = Conv2D(
64, (3, 3),
padding='same',
activation='relu',
name='conv2_2/3x3',
kernel_regularizer=l2(l2_regulizer))(conv2_1_3x3_reduce)
conv2_norm2 = LRN(name='conv2/norm2')(conv2_2_3x3)
conv2_zero_pad = ZeroPadding2D(padding=(1, 1))(conv2_norm2)
pool2_helper = PoolHelper()(conv2_zero_pad)
pool2_3x3_s2 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
padding='same',
name='pool2/3x3_s2')(pool2_helper)
conv3_1_3x3_s1 = Conv2D(128, (3, 3),
strides=(1, 1),
padding='same',
activation='relu',
name='conv3_1/3x3_s1',
kernel_regularizer=l2(l2_regulizer))(pool2_3x3_s2)
conv3_2_3x3_s1 = Conv2D(
128, (3, 3),
strides=(1, 1),
padding='same',
activation='relu',
name='conv3_2/3x3_s1',
kernel_regularizer=l2(l2_regulizer))(conv3_1_3x3_s1)
conv3_zero_pad = ZeroPadding2D(padding=(1, 1))(conv3_2_3x3_s1)
pool3_helper = PoolHelper()(conv3_zero_pad)
pool3_2_2x2_s1 = MaxPooling2D(pool_size=(2, 2),
strides=(1, 1),
padding='same',
name='pool3/2x2_s1')(pool3_helper)
pool3_norm1 = LRN(name='pool3/norm1')(pool3_2_2x2_s1)
conv4_1_3x3_reduce = Conv2D(
256, (1, 1),
padding='same',
activation='relu',
name='conv4_1/3x3_reduce',
kernel_regularizer=l2(l2_regulizer))(pool3_norm1)
conv4_2_3x3 = Conv2D(
256, (3, 3),
padding='same',
activation='relu',
name='conv4_2/3x3',
kernel_regularizer=l2(l2_regulizer))(conv4_1_3x3_reduce)
conv4_norm2 = LRN(name='conv4/norm2')(conv4_2_3x3)
conv4_zero_pad = ZeroPadding2D(padding=(1, 1))(conv4_norm2)
pool4_helper = PoolHelper()(conv4_zero_pad)
pool4_3x3_s2 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
padding='same',
name='pool4/3x3_s2')(pool4_helper)
conv5_1_3x3_s1 = Conv2D(512, (3, 3),
strides=(1, 1),
padding='same',
activation='relu',
name='conv5_1/3x3_s1',
kernel_regularizer=l2(l2_regulizer))(pool4_3x3_s2)
conv5_2_3x3_s1 = Conv2D(
512, (3, 3),
strides=(1, 1),
padding='same',
activation='relu',
name='conv5_2/3x3_s1',
kernel_regularizer=l2(l2_regulizer))(conv5_1_3x3_s1)
conv5_zero_pad = ZeroPadding2D(padding=(1, 1))(conv5_2_3x3_s1)
pool5_helper = PoolHelper()(conv5_zero_pad)
pool5_2_2x2_s1 = MaxPooling2D(pool_size=(2, 2),
strides=(1, 1),
padding='same',
name='pool5/2x2_s1')(pool5_helper)
pool5_norm1 = LRN(name='pool5/norm1')(pool5_2_2x2_s1)
conv6_1_3x3_reduce = Conv2D(
1024, (1, 1),
padding='same',
activation='relu',
name='conv6_1/3x3_reduce',
kernel_regularizer=l2(l2_regulizer))(pool5_norm1)
conv6_2_3x3 = Conv2D(
1024, (3, 3),
padding='same',
activation='relu',
name='conv6_2/3x3',
kernel_regularizer=l2(l2_regulizer))(conv6_1_3x3_reduce)
conv6_norm2 = LRN(name='conv6/norm2')(conv6_2_3x3)
conv6_zero_pad = ZeroPadding2D(padding=(1, 1))(conv6_norm2)
pool6_helper = PoolHelper()(conv6_zero_pad)
pool6_3x3_s2 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
padding='same',
name='pool6/3x3_s2')(pool6_helper)
pool7_2x2_s1 = AveragePooling2D(pool_size=(2, 2),
strides=(1, 1),
name='pool7/2x2_s1')(pool6_3x3_s2)
loss_flat = Flatten()(pool7_2x2_s1)
pool7_drop_2x2_s1 = Dropout(rate=0.5)(loss_flat)
loss_classifier = Dense(
num_classes,
name='loss3/classifier',
kernel_regularizer=l2(l2_regulizer))(pool7_drop_2x2_s1)
loss_classifier_act = Activation('softmax', name='prob')(loss_classifier)
mynet = Model(inputs=input, outputs=[loss_classifier_act])
if weights_path:
mynet.load_weights(weights_path)
if keras.backend.backend() == 'tensorflow':
# convert the convolutional kernels for tensorflow
ops = []
for layer in mynet.layers:
if layer.__class__.__name__ == 'Conv2D':
original_w = K.get_value(layer.kernel)
converted_w = convert_kernel(original_w)
ops.append(tf.assign(layer.kernel, converted_w).op)
K.get_session().run(ops)
return mynet
