def read_model(model, input_shape, f_activation='relu'):
"""Reads the operations on a single EfficientNet block.
Args:
model: efficientnet_model.Model,
input_shape: int, square image assumed.
f_activation: str or None, one of 'relu', 'swish', None.
Returns:
list, of operations.
"""
_ = model(torch.ones(input_shape))
input_size = input_shape[2] # Assuming square
ops = []
# 1
l_name = 'conv1'
layer = getattr(model, l_name)
# aa = list(layer.weight[0].shape)
layer_temp = counting.Conv2D(
input_size, list(layer.weight.shape), layer.stride,
layer.padding, True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
# Input size might have changed.
input_size = counting.get_conv_output_size(
image_size=input_size, filter_size=layer_temp.kernel_shape[2],
padding=layer_temp.padding, stride=layer_temp.strides[0])
for idx, block in enumerate(model.layers):
# print(idx)
if idx == 0 or idx == 1 or idx == 10 or idx == 16:
block_ops, input_size = read_block(block, input_size, f_activation=f_activation)
ops.append(('block_%d' % idx, block_ops))
else:
block_ops, input_size = read_block_without_shortcut(block, input_size, f_activation=f_activation)
ops.append(('block_%d' % idx, block_ops))
l_name = 'conv2'
layer = getattr(model, l_name)
# aa = list(layer.weight[0].shape)
layer_temp = counting.Conv2D(
input_size, list(layer.weight.shape), layer.stride,
layer.padding, True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
# Input size might have changed.
input_size = counting.get_conv_output_size(
image_size=input_size, filter_size=layer_temp.kernel_shape[2],
padding=layer_temp.padding, stride=layer_temp.strides[0])
# Blocks
# Head
l_name = 'linear'
layer = getattr(model, l_name)
shapetemp = [layer.in_features, layer.out_features]
ops.append(('_fc', counting.FullyConnected(
list(shapetemp), True, None)))
return ops
def read_block(block, input_size, f_activation='swish'):
"""Reads the operations on a single EfficientNet block.
Args:
block: efficientnet_model.MBConvBlock,
input_shape: int, square image assumed.
f_activation: str or None, one of 'relu', 'swish', None.
Returns:
list, of operations.
"""
ops = []
#shortcut does not need to change the input size
l_name = 'shortcut'
layer = getattr(block, l_name)[0]
layer_temp = counting.Conv2D(
input_size, list(layer.weight.shape), layer.stride,
layer.padding, True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
l_name = 'conv1'
layer = getattr(block, l_name)
layer_temp = counting.Conv2D(
input_size, list(layer.weight.shape), layer.stride,
layer.padding, True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
# Input size might have changed.
input_size = counting.get_conv_output_size(
image_size=input_size, filter_size=layer_temp.kernel_shape[2],
padding=layer_temp.padding, stride=layer_temp.strides[0])
l_name = 'conv2'
layer = getattr(block, l_name)
layer_temp = counting.Conv2D(
input_size, list(layer.weight.shape), layer.stride,
layer.padding, True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
# Input size might have changed.
input_size = counting.get_conv_output_size(
image_size=input_size, filter_size=layer_temp.kernel_shape[2],
padding=layer_temp.padding, stride=layer_temp.strides[0])
l_name = 'conv3'
layer = getattr(block, l_name)
layer_temp = counting.Conv2D(
input_size, list(layer.weight.shape), layer.stride,
layer.padding, True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
# Input size might have changed.
input_size = counting.get_conv_output_size(
image_size=input_size, filter_size=layer_temp.kernel_shape[2],
padding=layer_temp.padding, stride=layer_temp.strides[0])
return ops, input_size
def read_model(model, input_shape, f_activation='swish'):
"""Reads the operations on a single EfficientNet block.
Args:
model: efficientnet_model.Model,
input_shape: int, square image assumed.
f_activation: str or None, one of 'relu', 'swish', None.
Returns:
list, of operations.
"""
# Ensure that the input run through model
_ = model(tf.ones(input_shape))
input_size = input_shape[1] # Assuming square
ops = []
# 1
l_name = '_conv_stem'
layer = getattr(model, l_name)
layer_temp = counting.Conv2D(
input_size, layer.weights[0].shape.as_list(), layer.strides,
layer.padding, True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
# Input size might have changed.
input_size = counting.get_conv_output_size(
image_size=input_size, filter_size=layer_temp.kernel_shape[0],
padding=layer_temp.padding, stride=layer_temp.strides[0])
# Blocks
for idx, block in enumerate(model._blocks):
block_ops, input_size = read_block(block, input_size,
f_activation=f_activation)
ops.append(('block_%d' % idx, block_ops))
# Head
l_name = '_conv_head'
layer = getattr(model, l_name)
layer_temp = counting.Conv2D(
input_size, layer.weights[0].shape.as_list(), layer.strides,
layer.padding, True, f_activation) # Use bias true since batch_norm
n_channels_out = layer.weights[0].shape.as_list()[-1]
ops.append((l_name, layer_temp))
ops.append(('_avg_pooling', counting.GlobalAvg(input_size, n_channels_out)))
return ops
def read_block(block, input_size, f_activation='swish'):
"""Reads the operations on a single EfficientNet block.
Args:
block: efficientnet_model.MBConvBlock,
input_shape: int, square image assumed.
f_activation: str or None, one of 'relu', 'swish', None.
Returns:
list, of operations.
"""
conv_counter_class = {
efficientnet_model.MBConvBlock: counting.Conv2D,
efficientnet_model.MBConvBlockBinary: counting.Conv2DBinary,
}[type(block)]
ops = []
# 1
l_name = '_expand_conv'
if hasattr(block, l_name):
layer = getattr(block, l_name)
layer_temp = conv_counter_class(
input_size, layer.kernel.shape.as_list(), layer.strides, layer.padding,
True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
# 2
l_name = '_depthwise_conv'
layer = getattr(block, l_name)
layer_temp = counting.DepthWiseConv2D(
input_size, layer.weights[0].shape.as_list(), layer.strides,
layer.padding, True, f_activation) # Use bias true since batch_norm
ops.append((l_name, layer_temp))
# Input size might have changed.
input_size = counting.get_conv_output_size(
image_size=input_size, filter_size=layer_temp.kernel_shape[0],
padding=layer_temp.padding, stride=layer_temp.strides[0])
# 3
if block._has_se:
se_reduce = getattr(block, '_se_reduce')
se_expand = getattr(block, '_se_expand')
# Kernel has the input features in its second dimension.
n_channels = se_reduce.kernel.shape.as_list()[2]
ops.append(('_se_reduce_mean', counting.GlobalAvg(input_size, n_channels)))
# input size is 1
layer_temp = conv_counter_class(
1, se_reduce.kernel.shape.as_list(), se_reduce.strides,
se_reduce.padding, True, f_activation)
ops.append(('_se_reduce', layer_temp))
layer_temp = conv_counter_class(
1, se_expand.kernel.shape.as_list(), se_expand.strides,
se_expand.padding, True, 'sigmoid')
ops.append(('_se_expand', layer_temp))
ops.append(('_se_scale', counting.Scale(input_size, n_channels)))
# 4
l_name = '_project_conv'
layer = getattr(block, l_name)
layer_temp = conv_counter_class(
input_size, layer.kernel.shape.as_list(), layer.strides, layer.padding,
True, None) # Use bias true since batch_norm, no activation
ops.append((l_name, layer_temp))
if (block._block_args.id_skip
and all(s == 1 for s in block._block_args.strides)
and block._block_args.input_filters == block._block_args.output_filters):
ops.append(('_skip_add', counting.Add(input_size, n_channels)))
return ops, input_size