def model_core_squeezenet17(x, mode, params, info):
y, info = layer.conv2d_batch_norm_relu_dropout_l(x, mode, params, info, filters=128, kernel_size=3, strides=1)
y, info = layer.sqnet_expand(y, mode, params, info, 2*64)
y, info = layer.maxpool_l(y, info) # output 128x128
#y, info = layer.sqnet_squeeze_pool(y, mode, params, info, 80)
y, info = layer.sqnet_squeeze(y, mode, params, info, 80)
y, info = layer.sqnet_expand(y, mode, params, info, 2*96)
y, info = layer.maxpool_l(y, info) # output 64x64
#y, info = layer.sqnet_squeeze_pool(y, mode, params, info, 104)
y, info = layer.sqnet_squeeze(y, mode, params, info, 104)
y, info = layer.sqnet_expand(y, mode, params, info, 2*112)
y, info = layer.sqnet_squeeze(y, mode, params, info, 120)
y, info = layer.sqnet_expand(y, mode, params, info, 2*128)
y, info = layer.maxpool_l(y, info) # output 32x32
#y, info = layer.sqnet_squeeze_pool(y, mode, params, info, 120)
y, info = layer.sqnet_squeeze(y, mode, params, info, 120)
y, info = layer.sqnet_expand(y, mode, params, info, 2*112)
y, info = layer.sqnet_squeeze(y, mode, params, info, 104)
y, info = layer.sqnet_expand(y, mode, params, info, 2*96)
y, info = layer.maxpool_l(y, info) # output 16x16
#y, info = layer.sqnet_squeeze_pool(y, mode, params, info, 88)
y, info = layer.sqnet_squeeze(y, mode, params, info, 88)
y, info = layer.sqnet_expand(y, mode, params, info, 2*80)
y, info = layer.sqnet_squeeze(y, mode, params, info, 72)
y, info = layer.sqnet_expand(y, mode, params, info, 2*65, last=True)
return y, info
def model_core_squeezenet12(x, mode, params, info):
y, info = layer.conv2d_batch_norm_relu_dropout_l(x, mode, params, info, filters=32, kernel_size=6, strides=2) # output 128x128
y, info = layer.maxpool_l(y, info) # output 64x64
y, info = layer.sqnet_squeeze(y, mode, params, info, 21)
y, info = layer.sqnet_expand(y, mode, params, info, 2*26)
y, info = layer.sqnet_squeeze(y, mode, params, info, 36)
y, info = layer.sqnet_expand(y, mode, params, info, 2*46)
y, info = layer.maxpool_l(y, info) # output 32x32
y, info = layer.sqnet_squeeze(y, mode, params, info, 41)
y, info = layer.sqnet_expand(y, mode, params, info, 2*36)
y, info = layer.sqnet_squeeze(y, mode, params, info, 31)
y, info = layer.sqnet_expand(y, mode, params, info, 2*26)
y, info = layer.maxpool_l(y, info) # output 16x16
y, info = layer.sqnet_squeeze(y, mode, params, info, 21)
y, info = layer.sqnet_expand(y, mode, params, info, 2*16, last=True)
return y, info
def model_core_darknet(x, mode, params, info):
y, info = layer.conv2d_batch_norm_relu_dropout_l(x, mode, params, info, filters=64, kernel_size=3, strides=1)
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=50, kernel_size=1, strides=1)
y, info = layer.maxpool_l(y, info) # output 128x128
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=52, kernel_size=3, strides=1)
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=54, kernel_size=1, strides=1)
y, info = layer.maxpool_l(y, info) # output 64x64
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=56, kernel_size=3, strides=1)
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=58, kernel_size=1, strides=1)
y, info = layer.maxpool_l(y, info) # output 32x32
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=60, kernel_size=3, strides=1)
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=62, kernel_size=1, strides=1)
y, info = layer.maxpool_l(y, info) # output 16x16
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=64, kernel_size=3, strides=1)
y, info = layer.conv2d_batch_norm_relu_dropout_l(y, mode, params, info, filters=65, kernel_size=1, strides=1)
# the last number of filters must be multiple of 5 so that the YOLO head can be added afterwards
return y, info
def model_core_configurable_squeezenet(x, mode, params, info):
"""This configurable model tries to spread the layers evenly between the maxpool layers
and also spread depth increases and decreases in a progressive way. Depth at the end is
the same as it was initially."""
nlayers = params["layers"]
depth_increment = params["depth_increment"]
first_layer_fdepth = params["first_layer_filter_depth"]
first_layer_fstride = params["first_layer_filter_stride"]
first_layer_fsize = params["first_layer_filter_size"]
layers_n, first_expand_layer, depth_incr_doubler = layer.sqnet_spread_layers(nlayers)
def sqnet_squeeze_expand(x, mode, params, info, depth_increment, last=False):
"""Squeezenet "fire" module, i.e. a "squeeze" module followed by an "expand" module."""
depth = int(x.get_shape()[3])//2
depth += depth_increment
x, info = layer.sqnet_squeeze(x, mode, params, info, depth)
depth += depth_increment
x, info = layer.sqnet_expand(x, mode, params, info, 2*depth, last=last)
return x, info
y, info = layer.conv2d_batch_norm_relu_dropout_l(x, mode, params, info,
filters=first_layer_fdepth,
kernel_size=first_layer_fsize,
strides=first_layer_fstride)
if first_expand_layer == 1:
y, info = layer.sqnet_expand(y, mode, params, info, first_layer_fdepth)
y, info = layer.maxpool_l(y, info)
for _ in range(layers_n[1]//2):
y, info = sqnet_squeeze_expand(y, mode, params, info, depth_increment*depth_incr_doubler[0])
if first_layer_fstride == 1:
y, info = layer.maxpool_l(y, info)
for _ in range(layers_n[2]//2):
y, info = sqnet_squeeze_expand(y, mode, params, info, depth_increment*depth_incr_doubler[1])
y, info = layer.maxpool_l(y, info)
for _ in range(layers_n[3]//2):
y, info = sqnet_squeeze_expand(y, mode, params, info, -depth_increment*depth_incr_doubler[2])
y, info = layer.maxpool_l(y, info)
for i in range(layers_n[4]//2, 0, -1):
y, info = sqnet_squeeze_expand(y, mode, params, info, -depth_increment*depth_incr_doubler[3], last=(i==1))
return y, info