vato007/fast-depth-tf
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Build Functional Model, remove subclassed model

Browse Source 
Functional models are way easier to work with,
and I don't need any advanced features that would
require model subclassing
This commit is contained in:
Piv
2021-03-17 18:24:46 +10:30
parent b25b9be4eb
commit 00762f3e86
2 changed files with 42 additions and 112 deletions
Download Patch File Download Diff File Expand all files Collapse all files

90
fast_depth_functional.py
View File

@@ -5,59 +5,53 @@ Functional version of fastdepth model. Note that this doesn't work at the moment
'''
def _depthwise_conv_block(inputs,
pointwise_conv_filters,
depth_multiplier=1,
strides=(1, 1),
block_id=1):
channel_axis = 1 if keras.backend.image_data_format() == 'channels_first' else -1
pointwise_conv_filters = int(pointwise_conv_filters)
if strides == (1, 1):
x = inputs
else:
x = keras.layers.ZeroPadding2D(((0, 1), (0, 1)), name='conv_pad_%d' % block_id)(
inputs)
x = keras.layers.DepthwiseConv2D((3, 3),
padding='same' if strides == (1, 1) else 'valid',
depth_multiplier=depth_multiplier,
strides=strides,
use_bias=False,
name='conv_dw_%d' % block_id)(
x)
x = keras.layers.BatchNormalization(
axis=channel_axis, name='conv_dw_%d_bn' % block_id)(
x)
x = keras.layers.ReLU(6., name='conv_dw_%d_relu' % block_id)(x)
x = keras.layers.Conv2D(
pointwise_conv_filters, (1, 1),
padding='same',
use_bias=False,
strides=(1, 1),
name='conv_pw_%d' % block_id)(
x)
x = keras.layers.BatchNormalization(
axis=channel_axis, name='conv_pw_%d_bn' % block_id)(
x)
def FDDepthwiseBlock(inputs,
out_channels,
block_id=1):
x = keras.layers.DepthwiseConv2D(5, padding='same')(inputs)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.ReLU(6.)(x)
x = keras.layers.Conv2D(out_channels, 1, padding='same')(x)
x = keras.layers.BatchNormalization()(x)
return keras.layers.ReLU(6., name='conv_pw_%d_relu' % block_id)(x)
def make_fastdepth_functional():
# This doesn't work, at least right now...
mobilenet = keras.applications.MobileNet(include_top=False)
input = keras.layers.Input(shape=(224, 224, 3))
x = mobilenet(input)
x = _depthwise_conv_block(x, 512, block_id=14)
x = _depthwise_conv_block(x, 256, block_id=15)
x = keras.layers.Add()([x, mobilenet.get_layer('conv_pw_5_relu').output])
x = _depthwise_conv_block(x, 128, block_id=16)
x = keras.layers.Add()([x, mobilenet.get_layer('conv_pw_3_relu').output])
x = _depthwise_conv_block(x, 64, block_id=17)
x = keras.layers.Add()([x, mobilenet.get_layer('conv_pw_1_relu').output])
x = _depthwise_conv_block(x, 32, block_id=18)
x = input
mobilenet = keras.applications.MobileNet(include_top=False, weights=None)
for layer in mobilenet.layers:
x = layer(x)
if layer.name == 'conv_pw_5_relu':
conv5 = x
elif layer.name == 'conv_pw_3_relu':
conv3 = x
elif layer.name == 'conv_pw_1_relu':
conv1 = x
x = keras.layers.Conv2D(1, 1)(x)
x = FDDepthwiseBlock(x, 512, block_id=14)
# TODO: Bilinear interpolation
# x = keras.layers.experimental.preprocessing.Resizing(14, 14)
# Nearest neighbour interpolation, used by fast depth paper
x = keras.layers.experimental.preprocessing.Resizing(14, 14, interpolation='nearest')(x)
x = FDDepthwiseBlock(x, 256, block_id=15)
x = keras.layers.experimental.preprocessing.Resizing(28, 28, interpolation='nearest')(x)
x = keras.layers.Add()([x, conv5])
x = FDDepthwiseBlock(x, 128, block_id=16)
x = keras.layers.experimental.preprocessing.Resizing(56, 56, interpolation='nearest')(x)
x = keras.layers.Add()([x, conv3])
x = FDDepthwiseBlock(x, 64, block_id=17)
x = keras.layers.experimental.preprocessing.Resizing(112, 112, interpolation='nearest')(x)
x = keras.layers.Add()([x, conv1])
x = FDDepthwiseBlock(x, 32, block_id=18)
x = keras.layers.experimental.preprocessing.Resizing(224, 224, interpolation='nearest')(x)
x = keras.layers.Conv2D(1, 1, padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.ReLU()(x)
return keras.Model(input, x, name="fast_depth")
x = keras.layers.ReLU(6.)(x)
return keras.Model(inputs=input, outputs=x, name="fast_depth")
if __name__ == '__main__':
make_fastdepth_functional().summary()