Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Adding optional parameter activation_dtype to models #327

Open
wants to merge 7 commits into
base: master
Choose a base branch
from
Open

Adding optional parameter activation_dtype to models #327

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
ba9b183
update on requirements to get noisy students weights
phborba Apr 18, 2020
766be51
activation_dtype on inception_resnet and inception_v3
phborba Apr 18, 2020
a43d081
fpn and linknet with activation_dtype
phborba Apr 18, 2020
e52c86d
unet and pspnet
phborba Apr 18, 2020
cda0ec1
setting restrictions only to softmax activations
phborba Apr 18, 2020
f342915
removing parameter where it does not make sense
phborba Apr 19, 2020
f0b0727
bug fix
phborba Apr 19, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion requirements.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,3 @@
keras_applications>=1.0.7,<=1.0.8
image-classifiers==1.0.0
efficientnet==1.0.0
efficientnet==1.1.0
119 changes: 69 additions & 50 deletions segmentation_models/backbones/inception_resnet_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,7 @@ def conv2d_bn(x,
strides=1,
padding='same',
activation='relu',
activation_dtype=None,
use_bias=False,
name=None):
"""Utility function to apply conv + BN.
Expand All @@ -54,6 +55,8 @@ def conv2d_bn(x,
strides: strides in `Conv2D`.
padding: padding mode in `Conv2D`.
activation: activation in `Conv2D`.
activation_dtype: Optional type parameter to force activations
to be treated in certain type. Used when mixed_precision is enabled.
use_bias: whether to use a bias in `Conv2D`.
name: name of the ops; will become `name + '_ac'` for the activation
and `name + '_bn'` for the batch norm layer.
Expand All @@ -74,11 +77,14 @@ def conv2d_bn(x,
name=bn_name)(x)
if activation is not None:
ac_name = None if name is None else name + '_ac'
x = layers.Activation(activation, name=ac_name)(x)
if activation_dtype is None:
x = layers.Activation(activation, name=ac_name)(x)
else:
x = layers.Activation(activation, name=ac_name, dtype=activation_dtype)(x)
return x


def inception_resnet_block(x, scale, block_type, block_idx, activation='relu'):
def inception_resnet_block(x, scale, block_type, block_idx, activation='relu', activation_dtype=None):
"""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
Expand Down Expand Up @@ -108,31 +114,33 @@ def inception_resnet_block(x, scale, block_type, block_idx, activation='relu'):
(see [activations](../activations.md)).
When `activation=None`, no activation is applied
(i.e., "linear" activation: `a(x) = x`).
activation_dtype: Optional type parameter to force activations
to be treated in certain type. Used when mixed_precision is enabled.
# 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)
branch_0 = conv2d_bn(x, 32, 1, activation_dtype=activation_dtype)
branch_1 = conv2d_bn(x, 32, 1, activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 32, 3, activation_dtype=activation_dtype)
branch_2 = conv2d_bn(x, 32, 1, activation_dtype=activation_dtype)
branch_2 = conv2d_bn(branch_2, 48, 3, activation_dtype=activation_dtype)
branch_2 = conv2d_bn(branch_2, 64, 3, activation_dtype=activation_dtype)
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])
branch_0 = conv2d_bn(x, 192, 1, activation_dtype=activation_dtype)
branch_1 = conv2d_bn(x, 128, 1, activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 160, [1, 7], activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 192, [7, 1], activation_dtype=activation_dtype)
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])
branch_0 = conv2d_bn(x, 192, 1, activation_dtype=activation_dtype)
branch_1 = conv2d_bn(x, 192, 1, activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 224, [1, 3], activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 256, [3, 1], activation_dtype=activation_dtype)
branches = [branch_0, branch_1]
else:
raise ValueError('Unknown Inception-ResNet block type. '
Expand All @@ -148,14 +156,18 @@ def inception_resnet_block(x, scale, block_type, block_idx, activation='relu'):
1,
activation=None,
use_bias=True,
name=block_name + '_conv')
name=block_name + '_conv',
activation_dtype=activation_dtype)

x = layers.Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
output_shape=backend.int_shape(x)[1:],
arguments={'scale': scale},
name=block_name)([x, up])
if activation is not None:
x = layers.Activation(activation, name=block_name + '_ac')(x)
if activation_dtype is None:
x = layers.Activation(activation, name=block_name + '_ac')(x)
else:
x = layers.Activation(activation, name=block_name + '_ac', dtype=activation_dtype)(x)
return x


Expand All @@ -165,6 +177,7 @@ def InceptionResNetV2(include_top=True,
input_shape=None,
pooling=None,
classes=1000,
activation_dtype=None,
**kwargs):
"""Instantiates the Inception-ResNet v2 architecture.
Optionally loads weights pre-trained on ImageNet.
Expand Down Expand Up @@ -234,23 +247,23 @@ def InceptionResNetV2(include_top=True,
img_input = input_tensor

# Stem block: 35 x 35 x 192
x = conv2d_bn(img_input, 32, 3, strides=2, padding='same')
x = conv2d_bn(x, 32, 3, padding='same')
x = conv2d_bn(x, 64, 3, padding='same')
x = layers.MaxPooling2D(3, strides=2, padding='same')(x)
x = conv2d_bn(x, 80, 1, padding='same')
x = conv2d_bn(x, 192, 3, padding='same')
x = layers.MaxPooling2D(3, strides=2, padding='same')(x)
x = conv2d_bn(img_input, 32, 3, strides=2, padding='same', activation_dtype=activation_dtype)
x = conv2d_bn(x, 32, 3, padding='same', activation_dtype=activation_dtype)
x = conv2d_bn(x, 64, 3, padding='same', activation_dtype=activation_dtype)
x = layers.MaxPooling2D(3, strides=2, padding='same', activation_dtype=activation_dtype)(x)
x = conv2d_bn(x, 80, 1, padding='same', activation_dtype=activation_dtype)
x = conv2d_bn(x, 192, 3, padding='same', activation_dtype=activation_dtype)
x = layers.MaxPooling2D(3, strides=2, padding='same', activation_dtype=activation_dtype)(x)

# Mixed 5b (Inception-A block): 35 x 35 x 320
branch_0 = conv2d_bn(x, 96, 1, padding='same')
branch_1 = conv2d_bn(x, 48, 1, padding='same')
branch_1 = conv2d_bn(branch_1, 64, 5, padding='same')
branch_2 = conv2d_bn(x, 64, 1, padding='same')
branch_2 = conv2d_bn(branch_2, 96, 3, padding='same')
branch_2 = conv2d_bn(branch_2, 96, 3, padding='same')
branch_pool = layers.AveragePooling2D(3, strides=1, padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 64, 1, padding='same')
branch_0 = conv2d_bn(x, 96, 1, padding='same', activation_dtype=activation_dtype)
branch_1 = conv2d_bn(x, 48, 1, padding='same', activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 64, 5, padding='same', activation_dtype=activation_dtype)
branch_2 = conv2d_bn(x, 64, 1, padding='same', activation_dtype=activation_dtype)
branch_2 = conv2d_bn(branch_2, 96, 3, padding='same', activation_dtype=activation_dtype)
branch_2 = conv2d_bn(branch_2, 96, 3, padding='same', activation_dtype=activation_dtype)
branch_pool = layers.AveragePooling2D(3, strides=1, padding='same', activation_dtype=activation_dtype)(x)
branch_pool = conv2d_bn(branch_pool, 64, 1, padding='same', activation_dtype=activation_dtype)
branches = [branch_0, branch_1, branch_2, branch_pool]
channel_axis = 1 if backend.image_data_format() == 'channels_first' else 3
x = layers.Concatenate(axis=channel_axis, name='mixed_5b')(branches)
Expand All @@ -263,11 +276,11 @@ def InceptionResNetV2(include_top=True,
block_idx=block_idx)

# Mixed 6a (Reduction-A block): 17 x 17 x 1088
branch_0 = conv2d_bn(x, 384, 3, strides=2, padding='same')
branch_1 = conv2d_bn(x, 256, 1, padding='same')
branch_1 = conv2d_bn(branch_1, 256, 3, padding='same')
branch_1 = conv2d_bn(branch_1, 384, 3, strides=2, padding='same')
branch_pool = layers.MaxPooling2D(3, strides=2, padding='same')(x)
branch_0 = conv2d_bn(x, 384, 3, strides=2, padding='same', activation_dtype=activation_dtype)
branch_1 = conv2d_bn(x, 256, 1, padding='same', activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 256, 3, padding='same', activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 384, 3, strides=2, padding='same', activation_dtype=activation_dtype)
branch_pool = layers.MaxPooling2D(3, strides=2, padding='same', activation_dtype=activation_dtype)(x)
branches = [branch_0, branch_1, branch_pool]
x = layers.Concatenate(axis=channel_axis, name='mixed_6a')(branches)

Expand All @@ -279,14 +292,14 @@ def InceptionResNetV2(include_top=True,
block_idx=block_idx)

# Mixed 7a (Reduction-B block): 8 x 8 x 2080
branch_0 = conv2d_bn(x, 256, 1, padding='same')
branch_0 = conv2d_bn(branch_0, 384, 3, strides=2, padding='same')
branch_1 = conv2d_bn(x, 256, 1, padding='same')
branch_1 = conv2d_bn(branch_1, 288, 3, strides=2, padding='same')
branch_2 = conv2d_bn(x, 256, 1, padding='same')
branch_2 = conv2d_bn(branch_2, 288, 3, padding='same')
branch_2 = conv2d_bn(branch_2, 320, 3, strides=2, padding='same')
branch_pool = layers.MaxPooling2D(3, strides=2, padding='same')(x)
branch_0 = conv2d_bn(x, 256, 1, padding='same', activation_dtype=activation_dtype)
branch_0 = conv2d_bn(branch_0, 384, 3, strides=2, padding='same', activation_dtype=activation_dtype)
branch_1 = conv2d_bn(x, 256, 1, padding='same', activation_dtype=activation_dtype)
branch_1 = conv2d_bn(branch_1, 288, 3, strides=2, padding='same', activation_dtype=activation_dtype)
branch_2 = conv2d_bn(x, 256, 1, padding='same', activation_dtype=activation_dtype)
branch_2 = conv2d_bn(branch_2, 288, 3, padding='same', activation_dtype=activation_dtype)
branch_2 = conv2d_bn(branch_2, 320, 3, strides=2, padding='same', activation_dtype=activation_dtype)
branch_pool = layers.MaxPooling2D(3, strides=2, padding='same', activation_dtype=activation_dtype)(x)
branches = [branch_0, branch_1, branch_2, branch_pool]
x = layers.Concatenate(axis=channel_axis, name='mixed_7a')(branches)

Expand All @@ -295,20 +308,26 @@ def InceptionResNetV2(include_top=True,
x = inception_resnet_block(x,
scale=0.2,
block_type='block8',
block_idx=block_idx)
block_idx=block_idx,
activation_dtype=activation_dtype)
x = inception_resnet_block(x,
scale=1.,
activation=None,
block_type='block8',
block_idx=10)
block_idx=10,
activation_dtype=activation_dtype)

# Final convolution block: 8 x 8 x 1536
x = conv2d_bn(x, 1536, 1, name='conv_7b')
x = conv2d_bn(x, 1536, 1, name='conv_7b', activation_dtype=activation_dtype)

if include_top:
# Classification block
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
x = layers.Dense(classes, activation='softmax', name='predictions')(x)
if activation_dtype is None:
x = layers.Dense(classes, activation='softmax', name='predictions')(x)
else:
x = layers.Dense(classes, name='dense_logists')(x)
x = layers.Activation('softmax', dtype=activation_dtype, name='predictions')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D()(x)
Expand Down
11 changes: 10 additions & 1 deletion segmentation_models/backbones/inception_v3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,6 +77,7 @@ def InceptionV3(include_top=True,
input_tensor=None,
input_shape=None,
pooling=None,
activation_dtype=None,
classes=1000,
**kwargs):
"""Instantiates the Inception v3 architecture.
Expand Down Expand Up @@ -109,6 +110,8 @@ def InceptionV3(include_top=True,
the output of the model will be a 2D tensor.
- `max` means that global max pooling will
be applied.
activation_dtype: Optional type parameter to force activations
to be treated in certain type. Used when mixed_precision is enabled.
classes: optional number of classes to classify images
into, only to be specified if `include_top` is True, and
if no `weights` argument is specified.
Expand Down Expand Up @@ -345,7 +348,13 @@ def InceptionV3(include_top=True,
if include_top:
# Classification block
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
x = layers.Dense(classes, activation='softmax', name='predictions')(x)
if activation_dtype is None:
x = layers.Dense(classes, activation='softmax', name='predictions')(x)
else:
#only softmax activation must be cast when using mixed precision
x = layers.Dense(classes, name='dense_logits')(x)
x = layers.Activation('softmax', dtype=activation_dtype, name='predictions')(x)

else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D()(x)
Expand Down
6 changes: 5 additions & 1 deletion segmentation_models/models/_common_blocks.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@ def Conv2dBn(
data_format=None,
dilation_rate=(1, 1),
activation=None,
activation_dtype=None,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
Expand Down Expand Up @@ -62,7 +63,10 @@ def wrapper(input_tensor):
x = layers.BatchNormalization(axis=bn_axis, name=bn_name)(x)

if activation:
x = layers.Activation(activation, name=act_name)(x)
if activation_dtype is None:
x = layers.Activation(activation, name=act_name)(x)
else:
x = layers.Activation(activation, name=act_name, dtype=activation_dtype)(x)

return x

Expand Down
Loading