def conv_feat_layers(inputs, width, detect, training):
#
# convolutional features maps for detection and recognition
#
# detect = True for detection
# detect = False for recognition
#
# width reduce: per 8, for detection
# height reduce: per 8, for detection
#
# width reduce: per 8, for recognition
# height reduce: per 64, for recognition
#
#
# recog-inputs should have shape [ batch, 64, width, channel]
#
# height_norm = 64
#
# 64, 32, 16, 8
# 6, 4, 2, 1
#
#
# detection
#
# [3,3; 1,1],
# [6,6; 2,2], [8,8; 2,2], [10,10; 2,2]
# [20,20; 4,4], [22,22; 4,4], [24,24; 4,4]
# [48,48; 8,8], [50,50; 8,8], [52,52; 8,8],
#
# anchor width: 8,
# anchor height: 8, 16, 24, 32, 48,
#
# feature_layer --> receptive_field
# [0,0] --> [0:52, 0:52]
# [0,1] --> [0:52, 8:52+8]
# [i,j] --> [8*i:52+8*i, 8*j:52+8*j]
#
# feature_layer --> anchor_center
# [0,0] --> [26, 26]
# [0,1] --> [26, 26+8]
# [i,j] --> [26+8*i, 26+8*j]
#
#
layer_params = [
[64, (3, 3), (1, 1), 'same', True, True, 'conv1'],
[64, (3, 3), (1, 1), 'same', True, True, 'conv2'],
[64, (2, 2), (2, 2), 'valid', True, True, 'pool1'], # for pool
[128, (3, 3), (1, 1), 'same', True, True, 'conv3'],
[128, (3, 3), (1, 1), 'same', True, True, 'conv4'],
[128, (2, 2), (2, 2), 'valid', True, True, 'pool2'], # for pool
[256, (3, 3), (1, 1), 'same', True, True, 'conv5'],
[256, (3, 3), (1, 1), 'same', True, True, 'conv6'],
[256, (2, 2), (2, 2), 'valid', True, True, 'pool3']
] # for pool
layers_detect = [[512, (3, 3), (1, 1), 'same', True, True,
'feat_d']] # for feat
layers_recog = [[256, (3, 1), (1, 1), 'valid', True, True, 'conv_r1'],
[256, (3, 1), (1, 1), 'valid', True, True, 'conv_r2'],
[256, (3, 1), (1, 1), 'valid', True, True, 'conv_r3'],
[512, (2, 1), (1, 1), 'valid', True, True,
'feat_r']] # for feat
#
with tf.variable_scope("conv_comm"):
#
inputs = layers.conv_layer(inputs, layer_params[0], training)
inputs = layers.conv_layer(inputs, layer_params[1], training)
inputs = layers.padd_layer(inputs, [[0, 0], [0, 0], [0, 1], [0, 0]],
name='padd1')
inputs = layers.conv_layer(inputs, layer_params[2], training)
#inputs = layers.maxpool_layer(inputs, (2,2), (2,2), 'valid', 'pool1')
#
params = [[64, 3, (1, 1), 'same', True, True, 'conv1'],
[64, 3, (1, 1), 'same', True, False, 'conv2']]
inputs = layers.block_resnet(inputs, params, training, 'res1')
#
inputs = layers.conv_layer(inputs, layer_params[3], training)
inputs = layers.conv_layer(inputs, layer_params[4], training)
inputs = layers.padd_layer(inputs, [[0, 0], [0, 0], [0, 1], [0, 0]],
name='padd2')
inputs = layers.conv_layer(inputs, layer_params[5], training)
#inputs = layers.maxpool_layer(inputs, (2,2), (2,2), 'valid', 'pool2')
#
params = [[128, 3, (1, 1), 'same', True, True, 'conv1'],
[128, 3, (1, 1), 'same', True, False, 'conv2']]
inputs = layers.block_resnet(inputs, params, training, 'res2')
#
inputs = layers.conv_layer(inputs, layer_params[6], training)
inputs = layers.conv_layer(inputs, layer_params[7], training)
inputs = layers.padd_layer(inputs, [[0, 0], [0, 0], [0, 1], [0, 0]],
name='padd3')
inputs = layers.conv_layer(inputs, layer_params[8], training)
#inputs = layers.maxpool_layer(inputs, (2,2), (2,2), 'valid', 'pool3')
#
params = [[256, 3, (1, 1), 'same', True, True, 'conv1'],
[256, 3, (1, 1), 'same', True, False, 'conv2']]
inputs = layers.block_resnet(inputs, params, training, 'res3')
#
if detect:
#
inputs = layers.conv_layer(inputs, layers_detect[0], training)
#
conv_feat = tf.squeeze(inputs, axis=0, name='detect_feat') # squeeze
#
else: # recog
#
inputs = layers.conv_layer(inputs, layers_recog[0], training)
inputs = layers.conv_layer(inputs, layers_recog[1], training)
inputs = layers.conv_layer(inputs, layers_recog[2], training)
#
params = [[256, 2, (1, 1), 'same', True, True, 'conv1'],
[256, 2, (1, 1), 'same', True, False, 'conv2']]
inputs = layers.block_resnet(inputs, params, training, 'res_r1')
#
inputs = layers.conv_layer(inputs, layers_recog[3], training)
#
conv_feat = tf.squeeze(inputs, axis=0, name='recog_feat') # squeeze
#
# Calculate resulting sequence length from original image widths
#
two = tf.constant(2, dtype=tf.float32, name='two')
#
w = tf.cast(width, tf.float32)
#
w = tf.div(w, two)
w = tf.ceil(w)
#
w = tf.div(w, two)
w = tf.ceil(w)
#
w = tf.div(w, two)
w = tf.ceil(w)
#
w = tf.cast(w, tf.int32)
#
# Vectorize
sequence_length = tf.reshape(w, [-1], name='seq_len')
#
#
return conv_feat, sequence_length
def conv_feat_layers(inputs, width, training):
#
# convolutional features maps for detection and recognition
#
#
# recog-inputs should have shape [ batch, 64, width, channel]
#
# height_norm = 36
#
#
# detection
#
# [3,1; 1,1],
# [9,2; 3,2], [9,2; 3,2], [9,2; 3,2]
# [18,4; 6,4], [18,4; 6,4], [18,4; 6,4]
# [36,8; 12,8], [36,8; 12,8], [36,8; 12,8],
#
# anchor width: 8,
# anchor height: 12, 24, 36, 48,
#
# feature_layer --> receptive_field
# [0,0] --> [0:36, 0:8]
# [0,1] --> [0:36, 8:8+8]
# [i,j] --> [12*i:36+12*i, 8*j:8+8*j]
#
# feature_layer --> anchor_center
# [0,0] --> [18, 4]
# [0,1] --> [18, 4+8]
# [i,j] --> [18+12*i, 4+8*j]
#
#
layer_params = [
[64, (3, 3), (1, 1), 'same', True, True, 'conv1'],
[64, (3, 3), (1, 1), 'same', True, True, 'conv2'],
[64, (2, 2), (2, 2), 'valid', True, True, 'pool1'], # for pool
[128, (3, 3), (1, 1), 'same', True, True, 'conv3'],
[128, (3, 3), (1, 1), 'same', True, True, 'conv4'],
[128, (2, 2), (2, 2), 'valid', True, True, 'pool2'], # for pool
[256, (3, 3), (1, 1), 'same', True, True, 'conv5'],
[256, (3, 3), (1, 1), 'same', True, True, 'conv6'],
[256, (3, 2), (3, 2), 'valid', True, True, 'pool3'], # for pool
[512, (3, 1), (1, 1), 'valid', True, True, 'feat_c']
] # for feat
#
with tf.variable_scope("conv_comm"):
#
inputs = layers.conv_layer(inputs, layer_params[0], training)
inputs = layers.conv_layer(inputs, layer_params[1], training)
inputs = layers.padd_layer(inputs, [[0, 0], [0, 0], [0, 1], [0, 0]],
name='padd1')
inputs = layers.conv_layer(inputs, layer_params[2], training)
#inputs = layers.maxpool_layer(inputs, (2,2), (2,2), 'valid', 'pool1')
#
params = [[64, 3, (1, 1), 'same', True, True, 'conv1'],
[64, 3, (1, 1), 'same', True, False, 'conv2']]
inputs = layers.block_resnet_others(inputs, params, True, training,
'res1')
#
inputs = layers.conv_layer(inputs, layer_params[3], training)
inputs = layers.conv_layer(inputs, layer_params[4], training)
inputs = layers.padd_layer(inputs, [[0, 0], [0, 0], [0, 1], [0, 0]],
name='padd2')
inputs = layers.conv_layer(inputs, layer_params[5], training)
#inputs = layers.maxpool_layer(inputs, (2,2), (2,2), 'valid', 'pool2')
#
params = [[128, 3, (1, 1), 'same', True, True, 'conv1'],
[128, 3, (1, 1), 'same', True, False, 'conv2']]
inputs = layers.block_resnet_others(inputs, params, True, training,
'res2')
#
inputs = layers.conv_layer(inputs, layer_params[6], training)
inputs = layers.conv_layer(inputs, layer_params[7], training)
inputs = layers.padd_layer(inputs, [[0, 0], [0, 0], [0, 1], [0, 0]],
name='padd3')
inputs = layers.conv_layer(inputs, layer_params[8], training)
#inputs = layers.maxpool_layer(inputs, (3,2), (3,2), 'valid', 'pool3')
#
params = [[256, 3, (1, 1), 'same', True, True, 'conv1'],
[256, 3, (1, 1), 'same', True, False, 'conv2']]
inputs = layers.block_resnet_others(inputs, params, True, training,
'res3')
#
inputs = layers.conv_layer(inputs, layer_params[9], training)
#
conv_feat = tf.squeeze(inputs, axis=0, name='conv_feat') # squeeze
#
#
# Calculate resulting sequence length from original image widths
#
two = tf.constant(2, dtype=tf.float32, name='two')
#
w = tf.cast(width, tf.float32)
#
w = tf.div(w, two)
w = tf.ceil(w)
#
w = tf.div(w, two)
w = tf.ceil(w)
#
w = tf.div(w, two)
w = tf.ceil(w)
#
w = tf.cast(w, tf.int32)
#
# Vectorize
sequence_length = tf.reshape(w, [-1], name='seq_len')
#
#
return conv_feat, sequence_length
def conv_feat_layers(inputs, training):
#
# convolutional features maps for detection
#
# [3,1; 1,1],
# [9,2; 3,2], [9,2; 3,2], [9,2; 3,2]
# [18,4; 6,4], [18,4; 6,4], [18,4; 6,4]
# [36,8; 12,8], [36,8; 12,8], [36,8; 12,8],
#
# anchor width: 8,
# anchor height: 6, 12, 24, 36,
#
# feature_layer --> receptive_field
# [0,0] --> [0:36, 0:8]
# [0,1] --> [0:36, 8:8+8]
# [i,j] --> [12*i:36+12*i, 8*j:8+8*j]
#
# feature_layer --> anchor_center
# [0,0] --> [18, 4]
# [0,1] --> [18, 4+8]
# [i,j] --> [18+12*i, 4+8*j]
#
#
with tf.variable_scope("conv_comm"):
#
inputs = layers.conv_layer(
inputs, [64, (7, 7),
(2, 2), 'same', True, True, 'conv1'], training)
inputs = layers.maxpool_layer(inputs, (3, 3), (2, 2), 'same', 'pool1')
inputs = layers.block_resnet(inputs,
64,
relu=True,
training=training,
name='res1_1')
inputs = layers.block_resnet(inputs,
64,
relu=True,
training=training,
name='res1_2')
inputs = layers.block_resnet_half(inputs,
128,
relu=True,
training=training,
name='res2_1')
inputs = layers.block_resnet(inputs,
128,
relu=True,
training=training,
name='res2_2')
inputs = layers.block_resnet_half(inputs,
256,
relu=True,
training=training,
name='res3_1')
inputs = layers.block_resnet(inputs,
256,
relu=True,
training=training,
name='res3_2')
inputs = layers.block_resnet_half(inputs,
512,
relu=True,
training=training,
name='res4_1')
inputs = layers.block_resnet(inputs,
512,
relu=True,
training=training,
name='res4_2')
#
# conv_feat = layers.conv_layer(inputs, [512, (3, 1), (1, 1), 'valid', True, True, 'conv_feat'], training)
#
feat_size = tf.shape(inputs)
#
#
# Calculate resulting sequence length from original image widths
#
# two = tf.constant(2, dtype=tf.float32, name='two')
# #
# w = tf.cast(width, tf.float32)
# for _ in range(5):
# w = tf.div(w, two)
# w = tf.ceil(w)
# #
# w = tf.cast(w, tf.int32)
# #
# print("w before tile:", w)
# w = tf.tile(w, [feat_size[1]])
# print("w before flatten shape:", w.shape)
# #
# # Flatten -> Vectorize
# sequence_length = tf.reshape(w, [-1], name='seq_len')
# print("w after flatten shape:", sequence_length.shape)
#
return inputs #, sequence_length