def net():
n = caffe.NetSpec()
n.data = L.Input(input_param=dict(shape=dict(dim=data_shape)))
n.dataout = L.Log(n.data, base=_base, scale=_scale, shift=_shift)
return n.to_proto()
def fcn(split):
n = caffe.NetSpec()
n.data, n.sem, n.geo = L.Python(
module='siftflow_layers',
layer='SIFTFlowSegDataLayer',
ntop=3,
param_str=str(
dict(siftflow_dir='../data/sift-flow/data/',
split=split,
seed=1337)))
# the base net
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=100)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
n.pool3 = max_pool(n.relu3_3)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 512)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 512)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 512)
n.pool4 = max_pool(n.relu4_3)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 512)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 512)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 512)
n.pool5 = max_pool(n.relu5_3)
# fully conv
n.fc6, n.relu6 = conv_relu(n.pool5, 4096, ks=7, pad=0)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7, n.relu7 = conv_relu(n.drop6, 4096, ks=1, pad=0)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.score_fr_sem = L.Convolution(
n.drop7,
num_output=33,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.upscore2_sem = L.Deconvolution(n.score_fr_sem,
convolution_param=dict(num_output=33,
kernel_size=4,
stride=2,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_pool4_sem = L.Convolution(
n.pool4,
num_output=33,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.score_pool4_semc = crop(n.score_pool4_sem, n.upscore2_sem)
n.fuse_pool4_sem = L.Eltwise(n.upscore2_sem,
n.score_pool4_semc,
operation=P.Eltwise.SUM)
n.upscore_pool4_sem = L.Deconvolution(n.fuse_pool4_sem,
convolution_param=dict(
num_output=33,
kernel_size=4,
stride=2,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_pool3_sem = L.Convolution(
n.pool3,
num_output=33,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.score_pool3_semc = crop(n.score_pool3_sem, n.upscore_pool4_sem)
n.fuse_pool3_sem = L.Eltwise(n.upscore_pool4_sem,
n.score_pool3_semc,
operation=P.Eltwise.SUM)
n.upscore8_sem = L.Deconvolution(n.fuse_pool3_sem,
convolution_param=dict(num_output=33,
kernel_size=16,
stride=8,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_sem = crop(n.upscore8_sem, n.data)
# loss to make score happy (o.w. loss_sem)
n.loss = L.SoftmaxWithLoss(n.score_sem,
n.sem,
loss_param=dict(normalize=False,
ignore_label=255))
n.score_fr_geo = L.Convolution(
n.drop7,
num_output=3,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.upscore2_geo = L.Deconvolution(n.score_fr_geo,
convolution_param=dict(num_output=3,
kernel_size=4,
stride=2,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_pool4_geo = L.Convolution(
n.pool4,
num_output=3,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.score_pool4_geoc = crop(n.score_pool4_geo, n.upscore2_geo)
n.fuse_pool4_geo = L.Eltwise(n.upscore2_geo,
n.score_pool4_geoc,
operation=P.Eltwise.SUM)
n.upscore_pool4_geo = L.Deconvolution(n.fuse_pool4_geo,
convolution_param=dict(
num_output=3,
kernel_size=4,
stride=2,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_pool3_geo = L.Convolution(
n.pool3,
num_output=3,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.score_pool3_geoc = crop(n.score_pool3_geo, n.upscore_pool4_geo)
n.fuse_pool3_geo = L.Eltwise(n.upscore_pool4_geo,
n.score_pool3_geoc,
operation=P.Eltwise.SUM)
n.upscore8_geo = L.Deconvolution(n.fuse_pool3_geo,
convolution_param=dict(num_output=3,
kernel_size=16,
stride=8,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_geo = crop(n.upscore8_geo, n.data)
n.loss_geo = L.SoftmaxWithLoss(n.score_geo,
n.geo,
loss_param=dict(normalize=False,
ignore_label=255))
return n.to_proto()
'evaluate_difficult_gt': False,
'name_size_file': name_size_file,
}
### Hopefully you don't need to change the following ###
# Check file.
check_if_exist(train_data)
check_if_exist(test_data)
check_if_exist(label_map_file)
check_if_exist(pretrain_model)
make_if_not_exist(save_dir)
make_if_not_exist(job_dir)
make_if_not_exist(snapshot_dir)
# Create train net.
net = caffe.NetSpec()
net.data, net.label = CreateAnnotatedDataLayer(
train_data,
batch_size=batch_size_per_device,
train=True,
output_label=True,
label_map_file=label_map_file,
transform_param=train_transform_param,
batch_sampler=batch_sampler)
VGGNetBody(net,
from_layer='data',
fully_conv=True,
reduced=True,
dilated=True,
dropout=False)
def net(split):
n = caffe.NetSpec()
loss_param = dict(normalize=False)
if split == 'train':
data_params = dict(mean=(104.00699, 116.66877, 122.67892))
# 图像与标签
data_params['root'] = './datasets/ICDAR2013_TCB'
data_params['source'] = "ICDAR2013_TCB.lst"
data_params['shuffle'] = True
data_params['ignore_label'] = -1
n.data, n.label = L.Python(module='pylayer_old', layer='ImageLabelmapDataLayer', ntop=2, \
param_str=str(data_params))
if data_params.has_key('ignore_label'):
loss_param['ignore_label'] = int(data_params['ignore_label'])
elif split == 'test':
n.data = L.Input(name='data',
input_param=dict(shape=dict(dim=[1, 3, 500, 500])))
else:
raise Exception("Invalid phase")
#第一个卷积阶段
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=1)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
#第二个卷积阶段
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
#第三个卷积阶段
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
# 第三个卷积阶段最后一个卷积层，接一个MCFE模块, Channel: 64, kernel: 3*3
n.conv3_dilation1 = conv_dilation01(n.conv3_3, mult=[100, 1, 200, 0])
n.conv3_dilation2 = conv_dilation03(n.conv3_3, mult=[100, 1, 200, 0])
n.conv3_dilation3 = conv_dilation05(n.conv3_3, mult=[100, 1, 200, 0])
n.conv3_dilation4 = conv_dilation07(n.conv3_3, mult=[100, 1, 200, 0])
# 在Channel维度上进行拼接
n.concat_conv33 = L.Concat(n.conv3_dilation1,
n.conv3_dilation2,
n.conv3_dilation3,
n.conv3_dilation4,
concat_param=dict({'concat_dim': 1}))
# MCFE模块后接BLSTM module
# # ===================== prepare lstm inputs =====================
n.im2col_conv33 = L.Im2col(n.concat_conv33,
convolution_param=dict(kernel_size=3, pad=1))
n.im2col_transpose_conv33 = L.Transpose(
n.im2col_conv33, transpose_param=dict(dim=[3, 2, 0, 1]))
n.lstm_input_conv33 = L.Reshape(n.im2col_transpose_conv33,
reshape_param=dict(shape=dict(dim=-1),
axis=1,
num_axes=2))
# 前向LSTM
n.lstm_conv33 = L.Lstm(n.lstm_input_conv33,
lstm_param=dict(num_output=128,
weight_filler=dict(type='gaussian',
std=0.01),
bias_filler=dict(type='constant'),
clipping_threshold=1))
#后向LSTM
n.rlstm_input_conv33 = L.Reverse(n.lstm_input_conv33,
name='lstm_reverse1_conv33',
reverse_param=dict(axis=0))
n.rlstm_output_conv33 = L.Lstm(n.rlstm_input_conv33,
name='rlstm_conv33',
lstm_param=dict(num_output=128))
n.rlstm_conv33 = L.Reverse(n.rlstm_output_conv33,
name='lstm_reverse2_conv33',
reverse_param=dict(axis=0))
# lstm_conv33 和 rlstm_conv33经过Concat拼接，n*c*(h1+h2+...+hk)*w
n.merge_lstm_rlstm_conv33 = L.Concat(n.lstm_conv33,
n.rlstm_conv33,
concat_param=dict(axis=2))
n.lstm_output_reshape_conv33 = L.Reshape(n.merge_lstm_rlstm_conv33,
reshape_param=dict(
shape=dict(dim=[-1, 1]),
axis=1,
num_axes=1))
# transpose size of output as (N, C, H, W)
n.lstm_output_conv33 = L.Transpose(n.lstm_output_reshape_conv33,
transpose_param=dict(dim=[2, 3, 1, 0]))
n.pool3 = max_pool(n.relu3_3)
# 第四个卷积阶段
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 512)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 512)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 512)
# 第三个卷积阶段最后一个卷积层，接一个MCFE模块, Channel: 128, kernel: 3*3
n.conv4_dilation1 = conv_dilation1(n.conv4_3, mult=[100, 1, 200, 0])
n.conv4_dilation2 = conv_dilation3(n.conv4_3, mult=[100, 1, 200, 0])
n.conv4_dilation3 = conv_dilation5(n.conv4_3, mult=[100, 1, 200, 0])
n.conv4_dilation4 = conv_dilation7(n.conv4_3, mult=[100, 1, 200, 0])
# 在Channel维度上进行拼接， n*(c1+c2+...+ck)*h*w
n.concat_conv43 = L.Concat(n.conv4_dilation1,
n.conv4_dilation2,
n.conv4_dilation3,
n.conv4_dilation4,
concat_param=dict({'concat_dim': 1}))
# BLSTM module
# # ===================== prepare lstm inputs =====================
n.im2col_conv43 = L.Im2col(n.concat_conv43,
convolution_param=dict(kernel_size=3, pad=1))
n.im2col_transpose_conv43 = L.Transpose(
n.im2col_conv43, transpose_param=dict(dim=[3, 2, 0, 1]))
n.lstm_input_conv43 = L.Reshape(n.im2col_transpose_conv43,
reshape_param=dict(shape=dict(dim=-1),
axis=1,
num_axes=2))
# 前向LSTM
n.lstm_conv43 = L.Lstm(n.lstm_input_conv43,
lstm_param=dict(num_output=256,
weight_filler=dict(type='gaussian',
std=0.01),
bias_filler=dict(type='constant'),
clipping_threshold=1))
# 后向LSTM
n.rlstm_input_conv43 = L.Reverse(n.lstm_input_conv43,
name='lstm_reverse1_conv43',
reverse_param=dict(axis=0))
n.rlstm_output_conv43 = L.Lstm(n.rlstm_input_conv43,
name='rlstm_conv43',
lstm_param=dict(num_output=256))
n.rlstm_conv43 = L.Reverse(n.rlstm_output_conv43,
name='lstm_reverse2_conv43',
reverse_param=dict(axis=0))
#lstm_conv43 和 rlstm_conv43经Concat拼接，n*c*(h1+h2+...+hk)*w
n.merge_lstm_rlstm_conv43 = L.Concat(n.lstm_conv43,
n.rlstm_conv43,
concat_param=dict(axis=2))
n.lstm_output_reshape_conv43 = L.Reshape(n.merge_lstm_rlstm_conv43,
reshape_param=dict(
shape=dict(dim=[-1, 1]),
axis=1,
num_axes=1))
# transpose size of output as (N, C, H, W)
n.lstm_output_conv43 = L.Transpose(n.lstm_output_reshape_conv43,
transpose_param=dict(dim=[2, 3, 1, 0]))
n.pool4 = max_pool(n.relu4_3)
# The fiveth conv stage
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 512)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 512)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 512)
# MCFE inception module, Channel: 128, kernel: 3*3
n.conv5_dilation1 = conv_dilation1(n.conv5_3, mult=[100, 1, 200, 0])
n.conv5_dilation2 = conv_dilation3(n.conv5_3, mult=[100, 1, 200, 0])
n.conv5_dilation3 = conv_dilation5(n.conv5_3, mult=[100, 1, 200, 0])
n.conv5_dilation4 = conv_dilation7(n.conv5_3, mult=[100, 1, 200, 0])
n.concat_conv53 = L.Concat(n.conv5_dilation1,
n.conv5_dilation2,
n.conv5_dilation3,
n.conv5_dilation4,
concat_param=dict({'concat_dim': 1}))
# BLSTM module
# ===================== prepare lstm inputs =====================
n.im2col_conv53 = L.Im2col(n.concat_conv53,
convolution_param=dict(kernel_size=3, pad=1))
n.im2col_transpose_conv53 = L.Transpose(
n.im2col_conv53, transpose_param=dict(dim=[3, 2, 0, 1]))
n.lstm_input_conv53 = L.Reshape(n.im2col_transpose_conv53,
reshape_param=dict(shape=dict(dim=-1),
axis=1,
num_axes=2))
# 前向LSTM
n.lstm_conv53 = L.Lstm(n.lstm_input_conv53,
lstm_param=dict(num_output=256,
weight_filler=dict(type='gaussian',
std=0.01),
bias_filler=dict(type='constant'),
clipping_threshold=1))
#后向LSTM
n.rlstm_input_conv53 = L.Reverse(n.lstm_input_conv53,
name='lstm_reverse1_conv53',
reverse_param=dict(axis=0))
n.rlstm_output_conv53 = L.Lstm(n.rlstm_input_conv53,
name='rlstm_conv53',
lstm_param=dict(num_output=256))
n.rlstm_conv53 = L.Reverse(n.rlstm_output_conv53,
name='lstm_reverse2_conv53',
reverse_param=dict(axis=0))
# lstm_conv53和rlstm_conv53经过Concat拼接，n*c*(h1+h2+...+hk)*w
n.merge_lstm_rlstm_conv53 = L.Concat(n.lstm_conv53,
n.rlstm_conv53,
concat_param=dict(axis=2))
n.lstm_output_reshape_conv53 = L.Reshape(n.merge_lstm_rlstm_conv53,
reshape_param=dict(
shape=dict(dim=[-1, 1]),
axis=1,
num_axes=1))
# transpose size of output as (N, C, H, W)
n.lstm_output_conv53 = L.Transpose(n.lstm_output_reshape_conv53,
transpose_param=dict(dim=[2, 3, 1, 0]))
# 第三个阶段，BLSTM的输出，经过1x1的卷积降维，4x上采样，裁剪成与原图像大小相同
n.score_dsn3 = conv1x1(n.lstm_output_conv33,
lr=[0.01, 1, 0.02, 0],
wf=dict(type='gaussian', std=0.01))
n.score_dsn3_up = upsample(n.score_dsn3, stride=4)
n.upscore_dsn3 = L.Crop(n.score_dsn3_up, n.data)
# BalanceCrossEntropyLoss
if split == 'train':
n.loss3 = L.BalanceCrossEntropyLoss(n.upscore_dsn3,
n.label,
loss_param=loss_param)
if split == 'test':
n.sigmoid_dsn3 = L.Sigmoid(n.upscore_dsn3)
#第四个阶段，BLSTM的输出，经过1x1的卷积降维，8x上采样，裁剪成与原图像大小相同
n.score_dsn4 = conv1x1(n.lstm_output_conv43,
lr=[0.01, 1, 0.02, 0],
wf=dict(type='gaussian', std=0.01))
n.score_dsn4_up = upsample(n.score_dsn4, stride=8)
n.upscore_dsn4 = L.Crop(n.score_dsn4_up, n.data)
# BalanceCrossEntropyLoss
if split == 'train':
n.loss4 = L.BalanceCrossEntropyLoss(n.upscore_dsn4,
n.label,
loss_param=loss_param)
if split == 'test':
n.sigmoid_dsn4 = L.Sigmoid(n.upscore_dsn4)
# 第五个阶段，BLSTM的输出，经过1x1的卷积降维，16x上采样，裁剪成与原图像大小相同
n.score_dsn5 = conv1x1(n.lstm_output_conv53,
lr=[0.01, 1, 0.02, 0],
wf=dict(type='gaussian', std=0.01))
n.score_dsn5_up = upsample(n.score_dsn5, stride=16)
n.upscore_dsn5 = L.Crop(n.score_dsn5_up, n.data)
# BalanceCrossEntropyLoss
if split == 'train':
n.loss5 = L.BalanceCrossEntropyLoss(n.upscore_dsn5,
n.label,
loss_param=loss_param)
if split == 'test':
n.sigmoid_dsn5 = L.Sigmoid(n.upscore_dsn5)
# 将三个阶段的输出，在Channel维度上进行拼接，作为Attention模块的输入
n.concat_upscore = L.Concat(n.upscore_dsn3,
n.upscore_dsn4,
n.upscore_dsn5,
name='concat',
concat_param=dict({'concat_dim': 1}))
# upscore_dsn3，upscore_dsn4，upscore_dsn5经3X3的卷积, 降维
n.output_mask_product03 = L.Convolution(
n.upscore_dsn3,
num_output=1,
kernel_size=3,
pad=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant'),
engine=1)
n.output_mask_product04 = L.Convolution(
n.upscore_dsn4,
num_output=1,
kernel_size=3,
pad=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant'),
engine=1)
n.output_mask_product05 = L.Convolution(
n.upscore_dsn5,
num_output=1,
kernel_size=3,
pad=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant'),
engine=1)
### Attention 模块
# 第一个卷积层num_output=512, kernel_size：3x3
n.att_conv1_mask_512 = L.Convolution(
n.concat_upscore,
num_output=512,
kernel_size=3,
pad=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
engine=1)
n.relu_att_conv1 = L.ReLU(n.att_conv1_mask_512, in_place=True)
n.drop_att_conv1_mask = L.Dropout(n.relu_att_conv1,
dropout_ratio=0.5,
in_place=True)
# 第二个卷积层num_output=3, kernel_size：1x1
n.att_fc_mask_512 = L.Convolution(
n.drop_att_conv1_mask,
num_output=3,
kernel_size=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
engine=1)
n.attention = L.Softmax(n.att_fc_mask_512)
# 生成三个注意力权重
n.attention3, n.attention4, n.attention5 = L.Slice(n.attention,
name='slice_attention',
slice_param=dict(
axis=1,
slice_point=[1, 2]),
ntop=3)
# 注意力权重与feature map相乘，进行融合
n.output_mask3 = L.Eltwise(n.attention3,
n.output_mask_product03,
operation=P.Eltwise.PROD)
n.output_mask4 = L.Eltwise(n.attention4,
n.output_mask_product04,
operation=P.Eltwise.PROD)
n.output_mask5 = L.Eltwise(n.attention5,
n.output_mask_product05,
operation=P.Eltwise.PROD)
n.output_fusion = L.Eltwise(n.output_mask3,
n.output_mask4,
n.output_mask5,
operation=P.Eltwise.SUM)
#作为对比，不经过Attention模块， 将三个阶段的输出，在Channel维度上进行拼接，经1X1的卷积，输出
n.upscore_fuse = L.Convolution(n.concat_upscore,
name='new-score-weighting',
num_output=1,
kernel_size=1,
param=[
dict(lr_mult=0.001, decay_mult=1),
dict(lr_mult=0.002, decay_mult=0)
],
weight_filler=dict(type='constant',
value=0.2),
engine=1)
if split == 'train':
n.loss_fuse = L.BalanceCrossEntropyLoss(n.upscore_fuse,
n.label,
loss_param=loss_param)
n.loss_output_fusion = L.BalanceCrossEntropyLoss(n.output_fusion,
n.label,
loss_param=loss_param)
if split == 'test':
n.sigmoid_fuse = L.Sigmoid(n.upscore_fuse)
n.sigmoid_output_fusion = L.Sigmoid(n.output_fusion)
return n.to_proto()
def write_prototxt(is_train, output_folder, \
filename, main_branch, \
num_output_stage1, \
blocks, sync_bn, uni_bn):
netspec = caffe.NetSpec()
#### Input Setting ####
crop_size = 112
width = 170
height = 128
length = 16
step = 1
num_segments = 1
if is_train:
use_global_stats = False
else:
use_global_stats = True
#### Data layer ####
if is_train:
data_train_params = dict(name='data', \
ntop=2, \
video4d_data_param=dict( \
source="../kinetics_train_list.txt", \
batch_size=24, \
new_width=width, \
new_height=height, \
new_length=length, \
num_segments=num_segments, \
modality=0, \
step=step, \
rand_step=True, \
name_pattern='image_%06d.jpg', \
shuffle=True), \
transform_param=dict(
crop_size=crop_size, \
mirror=True, \
multi_scale=True, \
max_distort=1, \
scale_ratios=[1, 0.875, 0.75, 0.66], \
mean_value=[104]*length+[117]*length+[123]*length), \
include=dict(phase=0))
data_val_params = dict(name='vdata', \
ntop=2, \
video4d_data_param=dict(
source="../kinetics_val_list.txt", \
batch_size=1, \
new_width=width, \
new_height=height, \
new_length=length, \
num_segments=num_segments, \
modality=0, \
step=step, \
name_pattern='image_%06d.jpg'), \
transform_param=dict(
crop_size=crop_size, \
mirror=False, \
mean_value=[104]*length+[117]*length+[123]*length), \
include=dict(phase=1))
# pdb.set_trace()
netspec.data, netspec.label = BaseModule('Video4dData',
data_train_params).attach(
netspec, [])
netspec.vdata, netspec.vlabel = BaseModule('Video4dData',
data_val_params).attach(
netspec, [])
else:
data_params = dict(name='data', \
dummy_data_param=dict( \
shape=dict(\
dim=[10, 3, length, crop_size, crop_size])))
netspec.data = BaseModule('DummyData', data_params).attach(netspec, [])
#### (Optional) Reshape Layer ####
if is_train:
reshape_params = dict(name='data_reshape', \
reshape_param=dict( \
shape=dict(dim=[-1, 3, length, crop_size, crop_size])))
netspec.data_reshape = BaseModule('Reshape', reshape_params).attach(
netspec, [netspec.data])
#### Stage 1 ####
channels = 3 * 7 * 7 * 3 * 64 / (7 * 7 * 3 + 3 * 64)
name = '1_s'
conv1xdxd_params = dict(name='conv'+name, \
num_output=channels, \
kernel_size=[1, 7, 7], \
pad=[0, 3, 3], \
stride=[1, 2, 2], \
engine=2)
conv1xdxd = BaseModule('Convolution', conv1xdxd_params).attach(
netspec, [netspec.data_reshape if is_train else netspec.data])
name = '1_t'
stage1 = SgpAttenModule(name_template=name, \
bn_params=dict(frozen=False), \
stride=2, \
num_output=64, \
t_conv=False, \
sync_bn=sync_bn, \
uni_bn=uni_bn).attach(netspec, [conv1xdxd])
num_output = num_output_stage1
#### Stages 2 - 5 ####
last = stage1
for stage in range(4):
for block in range(blocks[stage]):
# First block usually projection
if block == 0:
shortcut = 'projection'
stride = 2
if stage == 0:
shortcut = 'identity'
stride = 1
else:
shortcut = 'identity'
stride = 1
name = str(stage + 2) + num2letter[int(block)]
curr_num_output = num_output * (2**(stage))
if uni_bn:
params = dict(name=name,
num_output=curr_num_output,
shortcut=shortcut,
main_branch=main_branch,
stride=stride,
frozen=False)
else:
params = dict(name=name,
num_output=curr_num_output,
shortcut=shortcut,
main_branch=main_branch,
stride=stride,
use_global_stats=use_global_stats)
last = PreActWiderDecoupSgpAttenMixBlock(name_template=name, \
shortcut=shortcut, \
num_output=curr_num_output, \
stride=stride, \
t_conv=False, \
sync_bn=sync_bn, \
uni_bn=uni_bn).attach(netspec, [last])
# else:
# last = PreActWiderDecoupSgpAttenStrongestBlock(name_template=name, \
# shortcut=shortcut, \
# num_output=curr_num_output, \
# stride=stride, \
# t_conv=True, \
# sync_bn=sync_bn, \
# uni_bn=uni_bn).attach(netspec, [last])
#### Last Norm & ReLU ####
if uni_bn:
bn_params = dict(frozen=False)
else:
bn_params = dict(use_global_stats=use_global_stats)
last = BNReLUModule(name_template='5b', \
bn_params=bn_params, \
sync_bn=sync_bn, \
uni_bn=uni_bn).attach(netspec, [last])
#### pool5 ####
pool_params = dict(global_pooling=True, pool=P.Pooling.AVE, name='pool5')
pool = BaseModule('Pooling', pool_params).attach(netspec, [last])
#### pool5_reshape ####
reshape_params = dict(shape=dict(dim=[-1, num_output_stage1 * 8]),
name='pool5_reshape')
reshape = BaseModule('Reshape', reshape_params).attach(netspec, [pool])
#### dropout ####
dropout_params = dict(dropout_ratio=0.2, name='dropout')
dropout = BaseModule('Dropout', dropout_params).attach(netspec, [reshape])
#### ip ####
ip_params = dict(name='fc400', num_output=400)
ip = BaseModule('InnerProduct', ip_params).attach(netspec, [dropout])
if is_train:
#### Softmax Loss ####
smax_params = dict(name='loss')
smax_loss = BaseModule('SoftmaxWithLoss',
smax_params).attach(netspec,
[ip, netspec.label])
#### Top1 Accuracy ####
top1_params = dict(name='top1',
accuracy_param=dict(top_k=1),
include=dict(phase=1))
top1 = BaseModule('Accuracy',
top1_params).attach(netspec, [ip, netspec.label])
#### Top5 Accuracy ####
top5_params = dict(name='top5',
accuracy_param=dict(top_k=5),
include=dict(phase=1))
top5 = BaseModule('Accuracy',
top5_params).attach(netspec, [ip, netspec.label])
filepath = os.path.join(output_folder, filename)
fp = open(filepath, 'w')
print >> fp, netspec.to_proto()
fp.close()
def build_VGG16Net(split,
num_classes,
batch_size,
resize_w,
resize_h,
crop_w=0,
crop_h=0,
crop_margin=0,
mirror=0,
rotate=0,
HSV_prob=0,
HSV_jitter=0,
train=True,
deploy=False):
weight_param = dict(lr_mult=1, decay_mult=1)
bias_param = dict(lr_mult=2, decay_mult=0)
learned_param = [weight_param, bias_param]
frozen_param = [dict(lr_mult=0)] * 2
n = caffe.NetSpec()
pydata_params = dict(split=split,
mean=(103.939, 116.779,
123.68)) #For VGG16 different mean than AlexNet
pydata_params['dir'] = '../../../datasets/SocialMedia'
pydata_params['train'] = train
pydata_params['batch_size'] = batch_size
pydata_params['resize_w'] = resize_w
pydata_params['resize_h'] = resize_h
pydata_params['crop_w'] = crop_w
pydata_params['crop_h'] = crop_h
pydata_params['crop_margin'] = crop_margin
pydata_params['mirror'] = mirror
pydata_params['rotate'] = rotate
pydata_params['HSV_prob'] = HSV_prob
pydata_params['HSV_jitter'] = HSV_jitter
pydata_params['num_classes'] = num_classes
pylayer = 'customDataLayer'
n.data, n.label = L.Python(module='layers',
layer=pylayer,
ntop=2,
param_str=str(pydata_params))
# conv
n.conv1_1, n.relu1_1 = conv_relu(n.data, 3, 64, pad=1, param=froozen_param)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1,
3,
64,
pad=1,
param=froozen_param)
n.pool1 = max_pool(n.relu1_2, 2, stride=2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1,
3,
128,
pad=1,
param=froozen_param)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1,
3,
128,
pad=1,
param=froozen_param)
n.pool2 = max_pool(n.relu2_2, 2, stride=2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2,
3,
256,
pad=1,
param=froozen_param)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1,
3,
256,
pad=1,
param=froozen_param)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2,
3,
256,
pad=1,
param=froozen_param)
n.pool3 = max_pool(n.relu3_3, 2, stride=2)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3,
3,
512,
pad=1,
param=froozen_param)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1,
3,
512,
pad=1,
param=froozen_param)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2,
3,
512,
pad=1,
param=froozen_param)
n.pool4 = max_pool(n.relu4_3, 2, stride=2)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4,
3,
512,
pad=1,
param=learned_param)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1,
3,
512,
pad=1,
param=learned_param)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2,
3,
512,
pad=1,
param=learned_param)
n.pool5 = max_pool(n.relu5_3, 2, stride=2)
# fully conn
n.fc6, n.relu6 = fc_relu(n.pool5, 4096, param=boosted_param)
if train:
n.drop6 = fc7input = L.Dropout(n.relu6,
in_place=True,
dropout_ratio=0.5) #0.5
else:
fc7input = n.relu6
n.fc7, n.relu7 = fc_relu(fc7input, 4096, param=boosted_param)
if train:
n.drop7 = fc8input = L.Dropout(n.relu7,
in_place=True,
dropout_ratio=0.5) #0.5
else:
fc8input = n.relu7
n.fc8C = L.InnerProduct(fc8input,
num_output=num_classes,
param=boosted_param)
if not deploy:
n.loss = L.SigmoidCrossEntropyLoss(n.fc8C, n.label)
if deploy:
n.probs = L.Sigmoid(n.fc8C)
with open('deploy.prototxt', 'w') as f:
f.write(str(n.to_proto()))
return f.name
else:
if train:
with open('train.prototxt', 'w') as f:
f.write(str(n.to_proto()))
return f.name
else:
with open('val.prototxt', 'w') as f:
f.write(str(n.to_proto()))
return f.name
if train:
with open('train.prototxt', 'w') as f:
f.write(str(n.to_proto()))
return f.name
else:
with open('val.prototxt', 'w') as f:
f.write(str(n.to_proto()))
return f.name
def dump_model(operation='create', redo=False):
# Creates graph from saved GraphDef.
create_graph()
sess = tf.InteractiveSession()
# Creates caffe model.
deploy_net_file = 'models/inception_v3/inception_v3_deploy.prototxt'
model_file = 'models/inception_v3/inception_v3.caffemodel'
net = []
if operation == 'create' and (not os.path.exists(deploy_net_file) or redo):
net = caffe.NetSpec()
elif operation == 'save' and (not os.path.exists(model_file) or redo):
caffe.set_device(1)
caffe.set_mode_gpu()
net = caffe.Net(deploy_net_file, caffe.TEST)
else:
return
# dump the preprocessing parameters
dump_inputlayer(sess, net, operation)
# dump the filters
dump_convbn(sess, net, 'data', 'conv', operation)
dump_convbn(sess, net, 'conv', 'conv_1', operation)
dump_convbn(sess, net, 'conv_1', 'conv_2', operation)
dump_pool(sess, net, 'conv_2', 'pool', operation)
dump_convbn(sess, net, 'pool', 'conv_3', operation)
dump_convbn(sess, net, 'conv_3', 'conv_4', operation)
dump_pool(sess, net, 'conv_4', 'pool_1', operation)
# inceptions with 1x1, 3x3, 5x5 convolutions
from_layer = 'pool_1'
for inception_id in range(0, 3):
if inception_id == 0:
out_layer = 'mixed'
else:
out_layer = 'mixed_{}'.format(inception_id)
dump_tower(sess, net, from_layer, out_layer, ['conv'], operation)
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv', 'conv_1'], operation)
dump_tower(sess, net, from_layer, '{}/tower_1'.format(out_layer),
['conv', 'conv_1', 'conv_2'], operation)
dump_tower(sess, net, from_layer, '{}/tower_2'.format(out_layer),
['pool', 'conv'], operation)
dump_inception(
sess, net, out_layer,
['conv', 'tower/conv_1', 'tower_1/conv_2', 'tower_2/conv'],
operation)
from_layer = '{}/join'.format(out_layer)
# inceptions with 1x1, 3x3(in sequence) convolutions
out_layer = 'mixed_3'
dump_tower(sess, net, from_layer, out_layer, ['conv'], operation)
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv', 'conv_1', 'conv_2'], operation)
dump_tower(sess, net, from_layer, out_layer, ['pool'], operation)
dump_inception(sess, net, out_layer, ['conv', 'tower/conv_2', 'pool'],
operation)
from_layer = '{}/join'.format(out_layer)
# inceptions with 1x1, 7x1, 1x7 convolutions
for inception_id in range(4, 8):
out_layer = 'mixed_{}'.format(inception_id)
dump_tower(sess, net, from_layer, out_layer, ['conv'], operation)
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv', 'conv_1', 'conv_2'], operation)
dump_tower(sess, net, from_layer, '{}/tower_1'.format(out_layer),
['conv', 'conv_1', 'conv_2', 'conv_3', 'conv_4'], operation)
dump_tower(sess, net, from_layer, '{}/tower_2'.format(out_layer),
['pool', 'conv'], operation)
dump_inception(
sess, net, out_layer,
['conv', 'tower/conv_2', 'tower_1/conv_4', 'tower_2/conv'],
operation)
from_layer = '{}/join'.format(out_layer)
# inceptions with 1x1, 3x3, 1x7, 7x1 filters
out_layer = 'mixed_8'
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv', 'conv_1'], operation)
dump_tower(sess, net, from_layer, '{}/tower_1'.format(out_layer),
['conv', 'conv_1', 'conv_2', 'conv_3'], operation)
dump_tower(sess, net, from_layer, out_layer, ['pool'], operation)
dump_inception(sess, net, out_layer,
['tower/conv_1', 'tower_1/conv_3', 'pool'], operation)
from_layer = '{}/join'.format(out_layer)
for inception_id in range(9, 11):
out_layer = 'mixed_{}'.format(inception_id)
dump_tower(sess, net, from_layer, out_layer, ['conv'], operation)
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv'], operation)
dump_tower(sess, net, '{}/tower/conv'.format(out_layer),
'{}/tower/mixed'.format(out_layer), ['conv'], operation)
dump_tower(sess, net, '{}/tower/conv'.format(out_layer),
'{}/tower/mixed'.format(out_layer), ['conv_1'], operation)
dump_inception(sess, net, '{}/tower/mixed'.format(out_layer),
['conv', 'conv_1'], operation, False)
dump_tower(sess, net, from_layer, '{}/tower_1'.format(out_layer),
['conv', 'conv_1'], operation)
dump_tower(sess, net, '{}/tower_1/conv_1'.format(out_layer),
'{}/tower_1/mixed'.format(out_layer), ['conv'], operation)
dump_tower(sess, net, '{}/tower_1/conv_1'.format(out_layer),
'{}/tower_1/mixed'.format(out_layer), ['conv_1'], operation)
dump_inception(sess, net, '{}/tower_1/mixed'.format(out_layer),
['conv', 'conv_1'], operation, False)
dump_tower(sess, net, from_layer, '{}/tower_2'.format(out_layer),
['pool', 'conv'], operation)
dump_inception(
sess, net, out_layer,
['conv', 'tower/mixed', 'tower_1/mixed', 'tower_2/conv'],
operation)
from_layer = '{}/join'.format(out_layer)
dump_pool(sess, net, from_layer, 'pool_3', operation)
dump_softmax(sess, net, 'pool_3', 'softmax', operation)
if operation == 'create' and (not os.path.exists(deploy_net_file) or redo):
model_dir = os.path.dirname(deploy_net_file)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
with open(deploy_net_file, 'w') as f:
print('name: "inception_v3_deploy"', file=f)
print(net.to_proto(), file=f)
elif operation == 'save' and (not os.path.exists(model_file) or redo):
net.save(model_file)
sess.close()
def fcn(split, tops):
n = caffe.NetSpec()
n.data, n.label = L.Python(module='nyud_layers',
layer='NYUDSegDataLayer',
ntop=2,
param_str=str(
dict(nyud_dir='../data',
split=split,
tops=tops,
seed=1337)))
# the base net
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=100)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
n.pool3 = max_pool(n.relu3_3)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 512)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 512)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 512)
n.pool4 = max_pool(n.relu4_3)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 512)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 512)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 512)
n.pool5 = max_pool(n.relu5_3)
# fully conv
n.fc6_new, n.relu6 = conv_relu(n.pool5, 4096, ks=7, pad=0)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7_new, n.relu7 = conv_relu(n.drop6, 4096, ks=1, pad=0)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.score_fr = L.Convolution(
n.drop7,
num_output=40,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.upscore = L.Deconvolution(n.score_fr,
convolution_param=dict(num_output=40,
kernel_size=64,
stride=32,
bias_term=False),
param=[dict(lr_mult=0)])
n.score = crop(n.upscore, n.data)
n.loss = L.SoftmaxWithLoss(n.score,
n.label,
loss_param=dict(normalize=False,
ignore_label=255))
return n.to_proto()
def alexnet(train_lmdb, test_lmdb, mean_file, train_batch_size, test_batch_size, output_path):
# train
ntrain = caffe.NetSpec()
# val
nval = caffe.NetSpec()
# deploy
ndeploy = caffe.NetSpec()
#--------------------------------------------------
# train + val
ntrain.data, ntrain.label = L.Data(name='data', batch_size=train_batch_size, backend=P.Data.LMDB,
source=train_lmdb, transform_param=dict(mirror=True, crop_size=227, mean_file=mean_file), include=dict(phase=caffe.TRAIN), ntop=2)
nval.data, nval.label = L.Data(name='data', batch_size=test_batch_size, backend=P.Data.LMDB,
source=test_lmdb, transform_param=dict(mirror=False, crop_size=227, mean_file=mean_file), include=dict(phase=caffe.TEST), ntop=2)
ntrain.conv1 = L.Convolution(ntrain.data, name='conv1', kernel_size=11, num_output=96, stride=4, weight_filler=dict(
type='gaussian', std=0.01), bias_filler=dict(type='constant', value=0), param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
ntrain.relu1 = L.ReLU(ntrain.conv1, name='relu1', in_place=True)
ntrain.norm1 = L.LRN(ntrain.relu1, name='norm1',
local_size=5, alpha=1e-4, beta=0.75)
ntrain.pool1 = L.Pooling(ntrain.norm1, name='pool1', kernel_size=3,
stride=2, pool=P.Pooling.MAX)
ntrain.conv2 = L.Convolution(ntrain.pool1, name='conv2', kernel_size=5, num_output=256, pad=2, group=2, weight_filler=dict(
type='gaussian', std=0.01), bias_filler=dict(type='constant', value=0.1), param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
ntrain.relu2 = L.ReLU(ntrain.conv2, name='relu2', in_place=True)
ntrain.norm2 = L.LRN(ntrain.relu2, name='norm2',
local_size=5, alpha=1e-4, beta=0.75)
ntrain.pool2 = L.Pooling(ntrain.norm2, name='pool2', kernel_size=3,
stride=2, pool=P.Pooling.MAX)
ntrain.conv3 = L.Convolution(ntrain.pool2, name='conv3', kernel_size=3, num_output=384, pad=1, weight_filler=dict(
type='gaussian', std=0.01), bias_filler=dict(type='constant', value=0), param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
ntrain.relu3 = L.ReLU(ntrain.conv3, name='relu3', in_place=True)
ntrain.conv4 = L.Convolution(ntrain.relu3, name='conv4', kernel_size=3, num_output=384, pad=1, group=2, weight_filler=dict(
type='gaussian', std=0.01), bias_filler=dict(type='constant', value=0.1), param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
ntrain.relu4 = L.ReLU(ntrain.conv4, name='relu4', in_place=True)
ntrain.conv5 = L.Convolution(ntrain.relu4, name='conv5', kernel_size=3, num_output=256, pad=1, group=2, weight_filler=dict(
type='gaussian', std=0.01), bias_filler=dict(type='constant', value=0.1), param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
ntrain.relu5 = L.ReLU(ntrain.conv5, name='relu5', in_place=True)
ntrain.pool5 = L.Pooling(ntrain.relu5, name='pool5', kernel_size=3,
stride=2, pool=P.Pooling.MAX)
ntrain.fc6 = L.InnerProduct(ntrain.pool5, name='fc6', num_output=4096,
weight_filler=dict(type='gaussian', std=0.005), bias_filler=dict(type='constant', value=1e-1), param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
ntrain.relu6 = L.ReLU(ntrain.fc6, name='relu6', in_place=True)
ntrain.drop6 = L.Dropout(ntrain.relu6, name='drop6',
dropout_ratio=0.5, in_place=True)
ntrain.fc7 = L.InnerProduct(ntrain.drop6, name='fc7', num_output=4096,
weight_filler=dict(type='gaussian', std=0.005), bias_filler=dict(type='constant', value=1e-1), param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
ntrain.relu7 = L.ReLU(ntrain.fc7, name='relu7', in_place=True)
ntrain.drop7 = L.Dropout(ntrain.relu7, name='drop7',
dropout_ratio=0.5, in_place=True)
ntrain.fc8 = L.InnerProduct(ntrain.drop7, name='fc8', num_output=1000,
weight_filler=dict(type='gaussian', std=0.01), bias_filler=dict(type='constant', value=0), param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
ntrain.accuracy = L.Accuracy(ntrain.fc8, ntrain.label, name='accuracy',
include=dict(phase=caffe.TEST))
ntrain.loss = L.SoftmaxWithLoss(ntrain.fc8, ntrain.label, name='loss')
#--------------------------------------------------
# deploy，删去lr_mult、decay_mult、weight_filler、bias_filler
# ({'shape': {'dim': [batch_size, channels, n_rows, n_cols]}})
ndeploy.data = L.Input(input_param={'shape': {'dim': [10, 13, 227, 227]}})
ndeploy.conv1 = L.Convolution(
ndeploy.data, name='conv1', kernel_size=11, num_output=96, stride=4)
ndeploy.relu1 = L.ReLU(ndeploy.conv1, name='relu1', in_place=True)
ndeploy.norm1 = L.LRN(ndeploy.relu1, name='norm1',
local_size=5, alpha=1e-4, beta=0.75)
ndeploy.pool1 = L.Pooling(ndeploy.norm1, name='pool1', kernel_size=3,
stride=2, pool=P.Pooling.MAX)
ndeploy.conv2 = L.Convolution(
ndeploy.pool1, name='conv2', kernel_size=5, num_output=256, pad=2, group=2)
ndeploy.relu2 = L.ReLU(ndeploy.conv2, name='relu2', in_place=True)
ndeploy.norm2 = L.LRN(ndeploy.relu2, name='norm2',
local_size=5, alpha=1e-4, beta=0.75)
ndeploy.pool2 = L.Pooling(ndeploy.norm2, name='pool2', kernel_size=3,
stride=2, pool=P.Pooling.MAX)
ndeploy.conv3 = L.Convolution(
ndeploy.pool2, name='conv3', kernel_size=3, num_output=384, pad=1)
ndeploy.relu3 = L.ReLU(ndeploy.conv3, name='relu3', in_place=True)
ndeploy.conv4 = L.Convolution(
ndeploy.relu3, name='conv4', kernel_size=3, num_output=384, pad=1, group=2)
ndeploy.relu4 = L.ReLU(ndeploy.conv4, name='relu4', in_place=True)
ndeploy.conv5 = L.Convolution(
ndeploy.relu4, name='conv5', kernel_size=3, num_output=256, pad=1, group=2)
ndeploy.relu5 = L.ReLU(ndeploy.conv5, name='relu5', in_place=True)
ndeploy.pool5 = L.Pooling(ndeploy.relu5, name='pool5', kernel_size=3,
stride=2, pool=P.Pooling.MAX)
ndeploy.fc6 = L.InnerProduct(ndeploy.pool5, name='fc6', num_output=4096)
ndeploy.relu6 = L.ReLU(ndeploy.fc6, name='relu6', in_place=True)
ndeploy.drop6 = L.Dropout(ndeploy.relu6, name='drop6',
dropout_ratio=5e-1, in_place=True)
ndeploy.fc7 = L.InnerProduct(ndeploy.drop6, name='fc7', num_output=4096)
ndeploy.relu7 = L.ReLU(ndeploy.fc7, name='relu7', in_place=True)
ndeploy.drop7 = L.Dropout(ndeploy.relu7, name='drop7',
dropout_ratio=5e-1, in_place=True)
ndeploy.fc8 = L.InnerProduct(ndeploy.drop7, name='fc8', num_output=1000)
ndeploy.prob = L.Softmax(ndeploy.fc8, name='prob')
out_train_val = str('name: "AlexNet"\n') + \
str(nval.to_proto()) + str(ntrain.to_proto())
with open(output_path + '/alexnet_train_val.prototxt', 'w') as f:
f.write(out_train_val)
out_deploy = str('name: "AlexNet"\n') + str(ndeploy.to_proto())
with open(output_path + '/alexnet_deploy.prototxt', 'w') as f:
f.write(out_deploy)
def net():
n = caffe.NetSpec()
n.data = L.Input(input_param=dict(shape=dict(dim=data_shape)))
n.dataout = L.Reduction(n.data, axis=0, coeff=1, operation=_operation)
return n.to_proto()
def net():
n = caffe.NetSpec()
n.data = L.Input(input_param=dict(shape=dict(dim=data_shape)))
n.dataout = L.Tile(n.data, axis=3, tiles=3)
return n.to_proto()
def net():
n = caffe.NetSpec()
n.data = L.Input(input_param=dict(shape=dict(dim=data_shape)))
n.dataout = L.Flatten(n.data, axis=_axis, end_axis=_end_axis)
return n.to_proto()
def net():
n = caffe.NetSpec()
n.data1 = L.Input(input_param=dict(shape=dict(dim=data_shape)))
n.data2 = L.Input(input_param=dict(shape=dict(dim=data_shape)))
n.dataout = L.Crop(n.data1, n.data2, axis=_axis, offset=_offset)
return n.to_proto()
def net():
n = caffe.NetSpec()
n.data = L.Input(input_param=dict(shape=dict(dim=data_shape)))
n.dataout = L.AbsVal(n.data)
return n.to_proto()
def gen_prototxt(nangles,
max_order,
scales,
filter_size_factor=wavelet.DEFAULT_SIZE,
nchannels_input=3,
intput_shape=[256, 256],
data=None,
verbose=False,
output_path=None):
n = caffe.NetSpec()
if data is None:
data = L.Input(shape=dict(dim=[1, nchannels_input] + intput_shape))
n.data = data
scat_count = -1
dim_total = nchannels_input
layers = [[(data, [None], 0)]]
for o in range(max_order):
layer = []
for s in scales:
kernel_size = s * filter_size_factor * 2
delta_offset = kernel_size // 2
for c0, s0, offset in layers[-1]:
if s0[-1] is not None and s <= s0[-1]:
continue
scat_count += 1
dim_in = nchannels_input * nangles**o
dim_out = nchannels_input * nangles**(o + 1)
dim_total += dim_out
name = 'scat%i_%i_%ito%i' % (s, scat_count, dim_in, dim_out)
c = scat_layer(c0,
dim=dim_out,
kernel_size=kernel_size,
name=name,
group=dim_in)
layer.append((c, s0 + [s], offset + delta_offset))
if verbose:
print "%s:" % name
print " kernel size: %i" % kernel_size
print " %s (%i)" % ("->".join(map(str,
(s0 + [s]))), dim_out)
layers.append(layer)
if verbose:
print "Total output dimensionality: %i" % dim_total
# Crop the coefficients before concatenation
# The last coefficient is the smallest because it's having the highest order.
last_coefficient = layers[-1][-1][0]
max_offset = layers[-1][-1][2]
coefficients = []
for layer in layers:
for c, _, offset in layer:
coefficients.append(
L.Crop(c, last_coefficient, offset=max_offset - offset))
concat = L.Concat(*coefficients)
# Do the final gaussian blur and resampling
kernel_size = scales[-1] * filter_size_factor * 2
stride = scales[-1]
c = conv_layer(concat,
dim=dim_total,
group=dim_total,
kernel_size=kernel_size,
name='psi',
stride=stride)
n.output = c
proto_str = str(n.to_proto())
if output_path:
with open(output_path, 'w+') as f:
f.write(proto_str)
return f.name
else:
with tempfile.NamedTemporaryFile(delete=False) as f:
f.write(proto_str)
return f.name
def inception_resnet_v2_proto(self, batch_size, phase='TRAIN'):
n = caffe.NetSpec()
if phase == 'TRAIN':
source_data = self.train_data
mirror = True
else:
source_data = self.test_data
mirror = False
n.data, n.label = L.Data(source=source_data, backend=P.Data.LMDB, batch_size=batch_size, ntop=2,
transform_param=dict(crop_size=299, mean_value=[104, 117, 123], mirror=mirror))
# stem
n.conv1_3x3_s2, n.conv1_3x3_s2_bn, n.conv1_3x3_s2_scale, n.conv1_3x3_s2_relu, n.conv2_3x3_s1, n.conv2_3x3_s1_bn, \
n.conv2_3x3_s1_scale, n.conv2_3x3_s1_relu, n.conv3_3x3_s1, n.conv3_3x3_s1_bn, n.conv3_3x3_s1_scale, n.conv3_3x3_s1_relu, \
n.inception_stem1_3x3_s2, n.inception_stem1_3x3_s2_bn, n.inception_stem1_3x3_s2_scale, n.inception_stem1_3x3_s2_relu, \
n.inception_stem1_pool, n.inception_stem1, n.inception_stem2_3x3_reduce, n.inception_stem2_3x3_reduce_bn, \
n.inception_stem2_3x3_reduce_scale, n.inception_stem2_3x3_reduce_relu, n.inception_stem2_3x3, \
n.inception_stem2_3x3_bn, n.inception_stem2_3x3_scale, n.inception_stem2_3x3_relu, n.inception_stem2_7x1_reduce, \
n.inception_stem2_7x1_reduce_bn, n.inception_stem2_7x1_reduce_scale, n.inception_stem2_7x1_reduce_relu, \
n.inception_stem2_7x1, n.inception_stem2_7x1_bn, n.inception_stem2_7x1_scale, n.inception_stem2_7x1_relu, \
n.inception_stem2_1x7, n.inception_stem2_1x7_bn, n.inception_stem2_1x7_scale, n.inception_stem2_1x7_relu, \
n.inception_stem2_3x3_2, n.inception_stem2_3x3_2_bn, n.inception_stem2_3x3_2_scale, n.inception_stem2_3x3_2_relu, \
n.inception_stem2, n.inception_stem3_3x3_s2, n.inception_stem3_3x3_s2_bn, n.inception_stem3_3x3_s2_scale, \
n.inception_stem3_3x3_s2_relu, n.inception_stem3_pool, n.inception_stem3 = \
stem_299x299(n.data) # 384x35x35
# 5 x inception_resnet_a
for i in xrange(5):
if i == 0:
bottom = 'n.inception_stem3'
else:
bottom = 'n.inception_resnet_a(order)_residual_eltwise'.replace('(order)', str(i))
exec (inception_resnet_a.replace('(order)', str(i + 1)).replace('bottom', bottom)) # 384x35x35
# reduction_a
n.reduction_a_pool, n.reduction_a_3x3, n.reduction_a_3x3_bn, n.reduction_a_3x3_scale, n.reduction_a_3x3_relu, \
n.reduction_a_3x3_2_reduce, n.reduction_a_3x3_2_reduce_bn, n.reduction_a_3x3_2_reduce_scale, \
n.reduction_a_3x3_2_reduce_relu, n.reduction_a_3x3_2, n.reduction_a_3x3_2_bn, n.reduction_a_3x3_2_scale, \
n.reduction_a_3x3_2_relu, n.reduction_a_3x3_3, n.reduction_a_3x3_3_bn, n.reduction_a_3x3_3_scale, \
n.reduction_a_3x3_3_relu, n.reduction_a_concat = \
reduction_a(n.inception_resnet_a5_residual_eltwise) # 1152x17x17
# 10 x inception_resnet_b
for i in xrange(10):
if i == 0:
bottom = 'n.reduction_a_concat'
else:
bottom = 'n.inception_resnet_b(order)_residual_eltwise'.replace('(order)', str(i))
exec (inception_resnet_b.replace('(order)', str(i + 1)).replace('bottom', bottom)) # 1152x17x17
# reduction_b
n.reduction_b_pool, n.reduction_b_3x3_reduce, n.reduction_b_3x3_reduce_bn, n.reduction_b_3x3_reduce_scale, \
n.reduction_b_3x3_reduce_relu, n.reduction_b_3x3, n.reduction_b_3x3_bn, n.reduction_b_3x3_scale, \
n.reduction_b_3x3_relu, n.reduction_b_3x3_2_reduce, n.reduction_b_3x3_2_reduce_bn, n.reduction_b_3x3_2_reduce_scale, \
n.reduction_b_3x3_2_reduce_relu, n.reduction_b_3x3_2, n.reduction_b_3x3_2_bn, n.reduction_b_3x3_2_scale, \
n.reduction_b_3x3_2_relu, n.reduction_b_3x3_3_reduce, n.reduction_b_3x3_3_reduce_bn, n.reduction_b_3x3_3_reduce_scale, \
n.reduction_b_3x3_3_reduce_relu, n.reduction_b_3x3_3, n.reduction_b_3x3_3_bn, n.reduction_b_3x3_3_scale, \
n.reduction_b_3x3_3_relu, n.reduction_b_3x3_4, n.reduction_b_3x3_4_bn, n.reduction_b_3x3_4_scale, \
n.reduction_b_3x3_4_relu, n.reduction_b_concat = \
reduction_b(n.inception_resnet_b10_residual_eltwise) # 2048x8x8
# 5 x inception_resnet_c
for i in xrange(5):
if i == 0:
bottom = 'n.reduction_b_concat'
else:
bottom = 'n.inception_resnet_c(order)_residual_eltwise'.replace('(order)', str(i))
exec (inception_resnet_c.replace('(order)', str(i + 1)).replace('bottom', bottom)) # 2048x8x8
n.pool_8x8_s1 = L.Pooling(n.inception_resnet_c5_residual_eltwise,
pool=P.Pooling.AVE,
global_pooling=True) # 2048x1x1
n.pool_8x8_s1_drop = L.Dropout(n.pool_8x8_s1, dropout_param=dict(dropout_ratio=0.2))
n.classifier = L.InnerProduct(n.pool_8x8_s1_drop, num_output=self.classifier_num,
param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type='xavier'),
bias_filler=dict(type='constant', value=0))
n.loss = L.SoftmaxWithLoss(n.classifier, n.label)
if phase == 'TRAIN':
pass
else:
n.accuracy_top1 = L.Accuracy(n.classifier, n.label, include=dict(phase=1))
n.accuracy_top5 = L.Accuracy(n.classifier, n.label, include=dict(phase=1),
accuracy_param=dict(top_k=5))
return n.to_proto()
def DAPNet_Train():
time_postfix = time.strftime("%m-%d_%H-%M-%S", time.localtime())
################################################################################
os.chdir(caffe_root)
################################################################################
# work dir
ProjectName = "{}_{}_{}".format(BaseNet, Models, Ver)
work_dir = "{}/{}/{}".format(Results_dir, Project, ProjectName)
make_if_not_exist(work_dir)
################################################################################
# work and model dirs
proto_dir = "{}/Proto".format(work_dir)
log_dir = "{}/Logs".format(work_dir)
model_dir = "{}/Models".format(work_dir)
pic_dir = "{}/Pics".format(work_dir)
job_dir = "{}/Job".format(work_dir)
make_if_not_exist(proto_dir)
make_if_not_exist(log_dir)
make_if_not_exist(model_dir)
make_if_not_exist(pic_dir)
make_if_not_exist(job_dir)
################################################################################
# work file
log_file = "{}/{}.log".format(log_dir, time_postfix)
train_net_file = "{}/train.prototxt".format(proto_dir)
test_net_file = "{}/test.prototxt".format(proto_dir)
solver_file = "{}/solver.prototxt".format(proto_dir)
snapshot_prefix = "{}/{}".format(model_dir, ProjectName)
job_file = "{}/train.sh".format(job_dir)
################################################################################
# TRAIN
net = caffe.NetSpec()
net = get_DAPDataLayer(net, train=True, batchsize=batchsize_per_device)
net = FaceBoxFPNNet(net, train=True, data_layer="data", gt_label="label",\
net_width=resized_width, net_height=resized_height)
# net = DAPNet_hand_pool1(net, train=True, data_layer="data", gt_label="label", \
# net_width=resized_width, net_height=resized_height)
with open(train_net_file, 'w') as f:
print('name: "{}_train"'.format(ProjectName), file=f)
print(net.to_proto(), file=f)
################################################################################
# TEST
if isinstance(val_list, list):
test_net_files = []
for id_val in xrange(len(val_list)):
net = caffe.NetSpec()
net = get_DAPDataLayer(net,
train=False,
batchsize=1,
id_val=id_val)
test_net_file = "{}/test{}.prototxt".format(proto_dir, id_val)
net = FaceBoxFPNNet(net, train=False, data_layer="data", gt_label="label", \
net_width=resized_width, net_height=resized_height)
with open(test_net_file, 'w') as f:
print('name: "{}_test{}"'.format(ProjectName, id_val), file=f)
print(net.to_proto(), file=f)
test_net_files.append(test_net_file)
test_net_file = test_net_files
else:
net = caffe.NetSpec()
net = get_DAPDataLayer(net, train=False, batchsize=1)
net = FaceBoxFPNNet(net, train=False, data_layer="data", gt_label="label", \
net_width=resized_width, net_height=resized_height)
# net = DAPNet_hand_pool1(net, train=False, data_layer="data", gt_label="label", \
# net_width=resized_width, net_height=resized_height)
with open(test_net_file, 'w') as f:
print('name: "{}_test"'.format(ProjectName), file=f)
print(net.to_proto(), file=f)
test_net_file = [
test_net_file,
]
################################################################################
# Solver
solver_param = get_solver_param()
solver = caffe_pb2.SolverParameter(train_net=train_net_file, \
test_net=test_net_file,snapshot_prefix=snapshot_prefix,**solver_param)
with open(solver_file, 'w') as f:
print(solver, file=f)
################################################################################
# CaffeModel & Snapshot
max_iter = 0
for file in os.listdir(model_dir):
if file.endswith(".solverstate"):
basename = os.path.splitext(file)[0]
iter = int(basename.split("{}_iter_".format(ProjectName))[1])
if iter > max_iter:
max_iter = iter
if fine_tuning:
train_param = '--weights="{}" \\\n'.format(get_pretained_model())
else:
train_param = ''
if resume_training:
if max_iter > 0:
train_param = '--snapshot="{}_iter_{}.solverstate" \\\n'.format(
snapshot_prefix, max_iter)
################################################################################
# job scripts
with open(job_file, 'w') as f:
f.write('cd {}\n'.format(caffe_root))
f.write('./build/tools/caffe train \\\n')
f.write('--solver="{}" \\\n'.format(solver_file))
f.write(train_param)
if solver_param['solver_mode'] == P.Solver.GPU:
f.write('--gpu {} 2>&1 | tee {}\n'.format(get_gpus(), log_file))
else:
f.write('2>&1 | tee {}.log\n'.format(log_file))
os.chmod(job_file, stat.S_IRWXU)
# ==========================================================================
# Training
subprocess.call(job_file, shell=True)
def fcn(split):
n = caffe.NetSpec()
pydata_params = dict(split=split,
mean=(104.00699, 116.66877, 122.67892),
seed=1337)
pydata_params['cocostuff_dir'] = 'cocostuff'
pylayer = 'COCOSTUFFSegDataLayer'
n.data, n.label = L.Python(module='cocostuff_layers',
layer=pylayer,
ntop=2,
param_str=str(pydata_params))
# the base net
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=100)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
n.pool3 = max_pool(n.relu3_3)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 512)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 512)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 512)
n.pool4 = max_pool(n.relu4_3)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 512)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 512)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 512)
n.pool5 = max_pool(n.relu5_3)
# fully conv
n.fc6, n.relu6 = conv_relu(n.pool5, 4096, ks=7, pad=0)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7, n.relu7 = conv_relu(n.drop6, 4096, ks=1, pad=0)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.score_fr = L.Convolution(
n.drop7,
num_output=182,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.upscore = L.Deconvolution(n.score_fr,
convolution_param=dict(num_output=182,
kernel_size=64,
stride=32,
bias_term=False),
param=[dict(lr_mult=0)])
n.score = crop(n.upscore, n.data)
n.loss = L.SoftmaxWithLoss(n.score,
n.label,
loss_param=dict(normalize=False,
ignore_label=255))
return n.to_proto()
def deploy_pose_ren_net(dataset):
n = caffe.NetSpec()
point_num_ = util.get_joint_num(dataset)
# input layers
# n.data = 'input: \"data\"\
# input_shape \{\
# dim: 1\
# dim: 1\
# dim: 96\
# dim: 96\
# \}'
# n.prev_pose = 'input: \"prev_pose\"\
# input_shape \{\
# dim: 1\
# dim: {}\
# \}'.format(point_num_*3)
n.data = L.Input(name="data", shape=dict(dim=[1, 1, 96, 96]))
n.prev_pose = L.Input(name="prev_pose",
shape=dict(dim=[1, point_num_ * 3]))
print str(n.to_proto())
# the base net
n.conv0, n.relu0 = conv_relu(n.data, 16)
n.conv1 = conv(n.relu0, 16)
n.pool1 = max_pool(n.conv1)
n.relu1 = L.ReLU(n.pool1, in_place=True)
n.conv2_0, n.relu2_0 = conv_relu(n.pool1, 32, ks=1, pad=0)
n.conv2, n.relu2 = conv_relu(n.relu2_0, 32)
n.conv3 = conv(n.relu2, 32)
n.res1 = L.Eltwise(n.conv2_0, n.conv3)
n.pool2 = max_pool(n.res1)
n.relu3 = L.ReLU(n.pool2, in_place=True)
n.conv3_0, n.relu3_0 = conv_relu(n.relu3, 64, ks=1, pad=0)
n.conv4, n.relu4 = conv_relu(n.relu3_0, 64)
n.conv5 = conv(n.relu4, 64)
n.res2 = L.Eltwise(n.conv3_0, n.conv5)
n.pool3 = max_pool(n.res2)
n.relu5 = L.ReLU(n.pool3, in_place=True)
# pose guided region ensemble
for idx in xrange(point_num_):
if idx not in get_guided_joints(dataset):
continue
rois = 'rois_{}'.format(idx)
n[rois] = L.Python(n.prev_pose,
module='python_layers.py_generate_roi_layer',
layer='PyGenerateROILayer',
ntop=1,
param_str=str(
dict(joint_idx=idx,
roi_h=6,
roi_w=6,
img_h=96,
img_w=96,
spatial_mul=8)))
roipool = 'roi_pool_{}'.format(idx)
n[roipool] = L.ROIPooling(n.pool3,
n[rois],
roi_pooling_param=dict(pooled_w=7,
pooled_h=7,
spatial_scale=0.125))
# fc
fc1 = 'fc1_{}'.format(idx)
relu6 = 'relu6_{}'.format(idx)
drop1 = 'drop1_{}'.format(idx)
n[fc1], n[relu6], n[drop1] = fc_relu_dropout(n[roipool], 2048, 0.5)
# structure connection
# connect_structure_1 = [[0,1,3], [0,4,6], [0,7,9], [0,10,12], [0,13,15]]
connect_structure_1 = get_connect_structure(dataset)
concate_bottom_final = []
for idx in xrange(len(connect_structure_1)):
concate_bottom = []
for jdx in xrange(len(connect_structure_1[idx])):
drop1 = 'drop1_{}'.format(connect_structure_1[idx][jdx])
concate_bottom.append(n[drop1])
concate_1 = 'concate_1_{}'.format(idx)
n[concate_1] = L.Concat(*concate_bottom)
fc2 = 'fc2_{}'.format(idx)
relu7 = 'relu7_{}'.format(idx)
drop2 = 'drop2_{}'.format(idx)
n[fc2], n[relu7], n[drop2] = fc_relu_dropout(n[concate_1], 2048, 0.5)
concate_bottom_final.append(n[drop2])
n.fc_concat = L.Concat(*concate_bottom_final)
n.fc3_0 = fc(n.fc_concat, point_num_ * 3)
return str(n.to_proto())
def silent_net():
n = caffe.NetSpec()
n.data, n.data2 = L.DummyData(shape=dict(dim=3), ntop=2)
n.silence_data = L.Silence(n.data, ntop=0)
n.silence_data2 = L.Silence(n.data2, ntop=0)
return n.to_proto()
def net(split):
n = caffe.NetSpec()
loss_param = dict(normalize=False)
if split == 'train':
data_params = dict(mean=(104.00699, 116.66877, 122.67892))
# 图像与标签
data_params['root'] = './datasets/CTW1500_Total_TCB'
data_params['source'] = "CTW1500_Total_TCB.lst"
data_params['shuffle'] = True
data_params['ignore_label'] = -1
n.data, n.label = L.Python(module='pylayer_old', layer='ImageLabelmapDataLayer', ntop=2, \
param_str=str(data_params))
if data_params.has_key('ignore_label'):
loss_param['ignore_label'] = int(data_params['ignore_label'])
elif split == 'test':
n.data = L.Input(name='data',
input_param=dict(shape=dict(dim=[1, 3, 500, 500])))
else:
raise Exception("Invalid phase")
# The first conv stage
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=1)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
# The second conv stage
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
# The third conv stage
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
n.pool3 = max_pool(n.relu3_3)
n.conv3_dilation1 = conv_dilation01(n.conv3_3, mult=[100, 1, 200, 0])
n.conv3_dilation2 = conv_dilation03(n.conv3_3, mult=[100, 1, 200, 0])
n.conv3_dilation3 = conv_dilation05(n.conv3_3, mult=[100, 1, 200, 0])
n.conv3_dilation4 = conv_dilation07(n.conv3_3, mult=[100, 1, 200, 0])
n.concat_conv33 = L.Concat(n.conv3_dilation1,
n.conv3_dilation2,
n.conv3_dilation3,
n.conv3_dilation4,
concat_param=dict({'concat_dim': 1}))
# The fourth conv stage
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 512)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 512)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 512)
n.pool4 = max_pool(n.relu4_3)
n.conv4_dilation1 = conv_dilation1(n.conv4_3, mult=[100, 1, 200, 0])
n.conv4_dilation2 = conv_dilation3(n.conv4_3, mult=[100, 1, 200, 0])
n.conv4_dilation3 = conv_dilation5(n.conv4_3, mult=[100, 1, 200, 0])
n.conv4_dilation4 = conv_dilation7(n.conv4_3, mult=[100, 1, 200, 0])
n.concat_conv43 = L.Concat(n.conv4_dilation1,
n.conv4_dilation2,
n.conv4_dilation3,
n.conv4_dilation4,
concat_param=dict({'concat_dim': 1}))
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 512)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 512)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 512)
n.conv5_dilation1 = conv_dilation1(n.conv5_3, mult=[100, 1, 200, 0])
n.conv5_dilation2 = conv_dilation3(n.conv5_3, mult=[100, 1, 200, 0])
n.conv5_dilation3 = conv_dilation5(n.conv5_3, mult=[100, 1, 200, 0])
n.conv5_dilation4 = conv_dilation7(n.conv5_3, mult=[100, 1, 200, 0])
n.concat_conv53 = L.Concat(n.conv5_dilation1,
n.conv5_dilation2,
n.conv5_dilation3,
n.conv5_dilation4,
concat_param=dict({'concat_dim': 1}))
# # DSN3
n.score_dsn3 = conv1x1(n.concat_conv33,
lr=[0.01, 1, 0.02, 0],
wf=dict(type='gaussian', std=0.01))
n.score_dsn3_up = upsample(n.score_dsn3, stride=4)
n.upscore_dsn3 = L.Crop(n.score_dsn3_up, n.data)
if split == 'train':
n.loss3 = L.BalanceCrossEntropyLoss(n.upscore_dsn3,
n.label,
loss_param=loss_param)
if split == 'test':
n.sigmoid_dsn3 = L.Sigmoid(n.upscore_dsn3)
# # DSN4
n.score_dsn4 = conv1x1(n.concat_conv43,
lr=[0.01, 1, 0.02, 0],
wf=dict(type='gaussian', std=0.01))
n.score_dsn4_up = upsample(n.score_dsn4, stride=8)
n.upscore_dsn4 = L.Crop(n.score_dsn4_up, n.data)
if split == 'train':
n.loss4 = L.BalanceCrossEntropyLoss(n.upscore_dsn4,
n.label,
loss_param=loss_param)
if split == 'test':
n.sigmoid_dsn4 = L.Sigmoid(n.upscore_dsn4)
# DSN5
n.score_dsn5 = conv1x1(n.concat_conv53,
lr=[0.01, 1, 0.02, 0],
wf=dict(type='gaussian', std=0.01))
n.score_dsn5_up = upsample(n.score_dsn5, stride=16)
n.upscore_dsn5 = L.Crop(n.score_dsn5_up, n.data)
if split == 'train':
n.loss5 = L.BalanceCrossEntropyLoss(n.upscore_dsn5,
n.label,
loss_param=loss_param)
if split == 'test':
n.sigmoid_dsn5 = L.Sigmoid(n.upscore_dsn5)
# ############### concatenation and pass through attention model #########
n.concat_upscore = L.Concat(n.upscore_dsn3,
n.upscore_dsn4,
n.upscore_dsn5,
name='concat',
concat_param=dict({'concat_dim': 1}))
n.output_mask_product03 = L.Convolution(
n.upscore_dsn3,
num_output=1,
kernel_size=3,
pad=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant'),
engine=1)
n.output_mask_product04 = L.Convolution(
n.upscore_dsn4,
num_output=1,
kernel_size=3,
pad=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant'),
engine=1)
n.output_mask_product05 = L.Convolution(
n.upscore_dsn5,
num_output=1,
kernel_size=3,
pad=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant'),
engine=1)
### attention model
n.att_conv1_mask_512 = L.Convolution(
n.concat_upscore,
num_output=512,
kernel_size=3,
pad=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
engine=1)
n.relu_att_conv1 = L.ReLU(n.att_conv1_mask_512, in_place=True)
n.drop_att_conv1_mask = L.Dropout(n.relu_att_conv1,
dropout_ratio=0.5,
in_place=True)
n.att_fc_mask_512 = L.Convolution(
n.drop_att_conv1_mask,
num_output=3,
kernel_size=1,
param=[dict(lr_mult=10, decay_mult=1),
dict(lr_mult=20, decay_mult=0)],
engine=1)
n.attention = L.Softmax(n.att_fc_mask_512)
n.attention3, n.attention4, n.attention5 = L.Slice(n.attention,
name='slice_attention',
slice_param=dict(
axis=1,
slice_point=[1, 2]),
ntop=3)
# # ---- multiply attention weights ----
n.output_mask3 = L.Eltwise(n.attention3,
n.output_mask_product03,
operation=P.Eltwise.PROD)
n.output_mask4 = L.Eltwise(n.attention4,
n.output_mask_product04,
operation=P.Eltwise.PROD)
n.output_mask5 = L.Eltwise(n.attention5,
n.output_mask_product05,
operation=P.Eltwise.PROD)
n.output_fusion = L.Eltwise(n.output_mask3,
n.output_mask4,
n.output_mask5,
operation=P.Eltwise.SUM)
n.upscore_fuse = L.Convolution(n.concat_upscore,
name='new-score-weighting',
num_output=1,
kernel_size=1,
param=[
dict(lr_mult=0.001, decay_mult=1),
dict(lr_mult=0.002, decay_mult=0)
],
weight_filler=dict(type='constant',
value=0.2),
engine=1)
if split == 'test':
n.sigmoid_fuse = L.Sigmoid(n.upscore_fuse)
n.sigmoid_output_fusion = L.Sigmoid(n.output_fusion)
if split == 'train':
n.loss_fuse = L.BalanceCrossEntropyLoss(n.upscore_fuse,
n.label,
loss_param=loss_param)
n.loss_output_fusion = L.BalanceCrossEntropyLoss(n.output_fusion,
n.label,
loss_param=loss_param)
# n.loss_fuse = L.BalanceCrossEntropyLoss(n.upscore_fuse, n.label, loss_param=loss_param)
return n.to_proto()
def fcn(split):
n = caffe.NetSpec()
pydata_params = dict(split=split,
mean=(104.00699, 116.66877, 122.67892),
seed=1337)
if split == 'train':
pydata_params[
'sbdd_dir'] = '/home/wen/caffe-master/semantic/fcn/data/sbdd/benchmark/benchmark_RELEASE/dataset'
pylayer = 'SBDDSegDataLayer'
else:
pydata_params[
'voc_dir'] = '/home/wen/caffe-master/semantic/fcn/data/pascal/VOC2012'
pylayer = 'VOCSegDataLayer'
n.data, n.label = L.Python(module='voc_layers',
layer=pylayer,
ntop=2,
param_str=str(pydata_params))
n.conv1_1, n.relu1_1 = conv_relu(n.data, 16 * 4, pad=10)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 16 * 4)
n.pool1 = max_pool(n.relu1_2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 32 * 4)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 32 * 4)
n.pool2 = max_pool(n.relu2_2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 64 * 4)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 64 * 4)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 64 * 4)
n.pool3 = max_pool(n.relu3_3)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 128 * 4)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 128 * 4)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 128 * 4)
n.pool4 = max_pool(n.relu4_3)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 128 * 4)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 128 * 4)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 128 * 4)
n.pool5 = max_pool(n.relu5_3)
n.fc6_new, n.relu6 = conv_relu(n.pool5, 1024, ks=3, pad=0)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7_new, n.relu7 = conv_relu(n.drop6, 1024, ks=1, pad=0)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.score_fr_new = L.Convolution(
n.drop7,
num_output=21,
kernel_size=1,
pad=0,
weight_filler=dict(type='xavier'),
bias_filler=dict(type='constant'),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.upscore_new = L.Deconvolution(n.score_fr_new,
convolution_param=dict(
num_output=21,
kernel_size=128,
stride=8,
bias_term=False,
),
param=[dict(lr_mult=0)])
n.score = L.Crop(n.upscore_new, n.data)
n.loss = L.SoftmaxWithLoss(n.score,
n.label,
loss_param=dict(normalize=False,
ignore_label=255))
return n.to_proto()
def cnn(split):
n = caffe.NetSpec()
if split == 'train':
pydata_params = dict(
dataset_dir='/home/kevin/dataset/ws_exp/gp_labelled',
split=split,
mean=(104.00698793, 116.66876762, 122.67891434),
seed=1337,
img_size=(224, 224),
crop_size=(224, 224, 224, 224))
pylayer = 'WashingtonDataLayerSS'
pydata_params['randomize'] = True
pydata_params['batch_size'] = 64
elif split == 'test':
pydata_params = dict(
dataset_dir='/home/kevin/dataset/washington_rgbd_dataset',
split=split,
mean=(104.00698793, 116.66876762, 122.67891434),
seed=1337,
img_size=(224, 224),
crop_size=(224, 224, 224, 224))
pylayer = 'WashingtonDataLayer'
pydata_params['randomize'] = False
pydata_params['batch_size'] = 1
else:
n.img = L.Input(
name='input',
ntop=2,
shape=[dict(dim=1),
dict(dim=1),
dict(dim=224),
dict(dim=224)])
#---------------------------------Data Layer---------------------------------------#
n.rgb, n.depth, n.label = L.Python(
name="data",
module='data_layers.washington_data_layer',
layer=pylayer,
ntop=3,
param_str=str(pydata_params))
#n.rgb_crop = L.Python(n.rgb, name="crop_rgb", module='data_layers.random_crop_layer', layer='RandomCropLayer', ntop=1, param_str=str(dict(h=224,w=224)))
#n.depth_crop = L.Python(n.depth, name="crop_depth", module='data_layers.random_crop_layer', layer='RandomCropLayer', ntop=1, param_str=str(dict(h=227,w=227)))
#---------------------------------RGB-Net---------------------------------------#
# the vgg 16 base net
n.conv1_1, n.relu1_1 = conv_relu("conv1_1", n.rgb, 64, pad=1, lr1=0, lr2=0)
n.conv1_2, n.relu1_2 = conv_relu("conv1_2", n.relu1_1, 64, lr1=0, lr2=0)
n.rgb_pool1 = max_pool(n.relu1_2)
n.conv2_1, n.relu2_1 = conv_relu("conv2_1", n.rgb_pool1, 128, lr1=0, lr2=0)
n.conv2_2, n.relu2_2 = conv_relu("conv2_2", n.relu2_1, 128, lr1=0, lr2=0)
n.rgb_pool2 = max_pool(n.relu2_2)
n.conv3_1, n.relu3_1 = conv_relu("conv3_1", n.rgb_pool2, 256, lr1=0, lr2=0)
n.conv3_2, n.relu3_2 = conv_relu("conv3_2", n.relu3_1, 256, lr1=0, lr2=0)
n.conv3_3, n.relu3_3 = conv_relu("conv3_3", n.relu3_2, 256, lr1=0, lr2=0)
n.rgb_pool3 = max_pool(n.relu3_3)
n.conv4_1, n.relu4_1 = conv_relu("conv4_1", n.rgb_pool3, 512, lr1=0, lr2=0)
n.conv4_2, n.relu4_2 = conv_relu("conv4_2", n.relu4_1, 512, lr1=0, lr2=0)
n.conv4_3, n.relu4_3 = conv_relu("conv4_3", n.relu4_2, 512, lr1=0, lr2=0)
n.rgb_pool4 = max_pool(n.relu4_3)
n.conv5_1, n.relu5_1 = conv_relu("conv5_1", n.rgb_pool4, 512, lr1=0, lr2=0)
n.conv5_2, n.relu5_2 = conv_relu("conv5_2", n.relu5_1, 512, lr1=0, lr2=0)
n.conv5_3, n.relu5_3 = conv_relu("conv5_3", n.relu5_2, 512, lr1=0, lr2=0)
n.rgb_pool5 = max_pool(n.relu5_3)
# fully conv
n.rgb_fc6, n.rgb_relu6 = fc_relu(n.rgb_pool5, 4096, lr1=0, lr2=0)
n.rgb_drop6 = L.Dropout(n.rgb_relu6, dropout_ratio=0.5, in_place=True)
n.rgb_fc7, n.rgb_relu7 = fc_relu(n.rgb_drop6, 4096, lr1=0, lr2=0)
n.rgb_drop7 = L.Dropout(n.rgb_relu7, dropout_ratio=0.5, in_place=True)
n.rgb_fc8 = fc(n.rgb_drop7, 51, lr1=0, lr2=0)
#---------------------------------Depth-Net---------------------------------------#
# the base net
n.conv1, n.relu1 = conv_relu("conv1",
n.depth,
128,
ks=5,
stride=2,
pad=2,
lr1=1,
lr2=2)
n.depth_pool1 = max_pool(n.relu1)
n.norm1 = L.LRN(n.depth_pool1,
lrn_param=dict(local_size=5, alpha=0.0005, beta=0.75, k=2))
n.conv2, n.relu2 = conv_relu("conv2",
n.norm1,
256,
ks=5,
stride=1,
pad=2,
lr1=1,
lr2=2)
n.depth_pool2 = max_pool(n.relu2)
n.norm2 = L.LRN(n.depth_pool2,
lrn_param=dict(local_size=5, alpha=0.0005, beta=0.75, k=2))
n.conv3, n.relu3 = conv_relu("conv3",
n.norm2,
384,
ks=3,
pad=1,
group=2,
lr1=1,
lr2=2)
n.depth_pool3 = max_pool(n.relu3)
n.conv4, n.relu4 = conv_relu("conv4",
n.depth_pool3,
512,
ks=3,
pad=1,
group=1,
lr1=1,
lr2=2)
n.conv5, n.relu5 = conv_relu("conv5",
n.relu4,
512,
ks=3,
pad=1,
group=1,
lr1=1,
lr2=2)
n.depth_pool5 = max_pool(n.relu5)
n.depth_fc6, n.depth_relu6 = fc_relu(n.depth_pool5, 4096, lr1=1, lr2=2)
n.depth_drop6 = L.Dropout(n.depth_relu6, dropout_ratio=0.5, in_place=True)
n.depth_fc7, n.depth_relu7 = fc_relu(n.depth_drop6, 4096, lr1=1, lr2=2)
n.depth_drop7 = L.Dropout(n.depth_relu7, dropout_ratio=0.5, in_place=True)
n.depth_fc8 = fc(n.depth_drop7, 51, lr1=1, lr2=2)
#-----------------------------------final output---------------------------------#
# Concatenation
#n.concat = L.Concat(n.rgb_drop7, n.depth_drop7, axis=1)
#n.rgbd_fc8 = fc(n.concat, 6, lr1=1, lr2=2)
if split != 'deploy':
n.rgb_accuracy = L.Accuracy(n.rgb_fc8, n.label)
n.rgb_loss = L.SoftmaxWithLoss(n.rgb_fc8, n.label)
n.depth_accuracy = L.Accuracy(n.depth_fc8, n.label)
n.depth_loss = L.SoftmaxWithLoss(n.depth_fc8, n.label)
#n.accuracy = L.Accuracy(n.rgbd_fc8, n.label)
#n.loss = L.SoftmaxWithLoss(n.rgbd_fc8, n.label)
return n.to_proto()
def generate_net(lmdb, label_file, PHASE, batch_size):
net = caffe.NetSpec()
if(PHASE=="TRAIN"):
# data layer
net.data, net.label = caffe.layers.AnnotatedData(ntop=2, include={'phase':caffe.TRAIN},
transform_param=dict(mirror=True, mean_value=[104, 117, 123],
resize_param=dict(prob=1.0, resize_mode=caffe.params.Resize.WARP, height=300, width=300,
interp_mode=[caffe.params.Resize.LINEAR,caffe.params.Resize.AREA,caffe.params.Resize.NEAREST,caffe.params.Resize.CUBIC,caffe.params.Resize.LANCZOS4]),
emit_constraint=dict(emit_type=0),
distort_param=dict(brightness_prob=0.5, brightness_delta=32.0,
contrast_prob=0.5, contrast_lower=0.5, contrast_upper=1.5, hue_prob=0.5, hue_delta=18.0,
saturation_prob=0.5, saturation_lower=0.5, saturation_upper=1.5, random_order_prob=0.0),
expand_param=dict(prob=0.5, max_expand_ratio=4.0)),
data_param=dict(source=lmdb, batch_size=batch_size, backend=caffe.params.Data.LMDB),
annotated_data_param=dict(
batch_sampler=[dict(max_sample=1, max_trials=1),
dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0),
sample_constraint=dict(min_jaccard_overlap=0.1), max_sample=1, max_trials=50),
dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0),
sample_constraint=dict(min_jaccard_overlap=0.3), max_sample=1, max_trials=50),
dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0),
sample_constraint=dict(min_jaccard_overlap=0.5), max_sample=1, max_trials=50),
dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0),
sample_constraint=dict(min_jaccard_overlap=0.7), max_sample=1, max_trials=50),
dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0),
sample_constraint=dict(min_jaccard_overlap=0.9), max_sample=1, max_trials=50),
dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0),
sample_constraint=dict(min_jaccard_overlap=1.0), max_sample=1, max_trials=50)],
label_map_file=label_file))
elif(PHASE=="DEPLOY"):
net.data = caffe.layers.Input(shape={'dim':[1,3,300,300]})
# bone
net.conv1 = caffe.layers.Convolution(net.data, num_output=57, kernel_size=3, stride=2, weight_filler={"type":"xavier"},
param=[dict(lr_mult=1.0,decay_mult=0.0),dict(lr_mult=1.0,decay_mult=0.0)])
net.relu_conv1 = caffe.layers.ReLU(net.conv1,in_place=True)
net.pool1 = caffe.layers.Pooling(net.relu_conv1, pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
# fire1
net.tops['fire2/concat'] = fire(net, net.pool1, 'fire2', 15,49, 53)
# fire2
net.tops['fire3/concat'] = fire(net, net.tops['fire2/concat'], 'fire3', 15, 54, 52)
net.pool3 = caffe.layers.Pooling(net.tops['fire3/concat'], pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
# fire3
net.tops['fire4/concat'] = fire(net, net.pool3, 'fire4', 29, 92, 94)
# fire4
net.tops['fire5/concat'] = fire(net, net.tops['fire4/concat'], 'fire5', 29, 90, 83)
net.pool5 = caffe.layers.Pooling(net.tops['fire5/concat'], pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
# fire5
net.tops['fire6/concat'] = fire(net, net.pool5, 'fire6', 44, 166, 161)
# fire6
net.tops['fire7/concat'] = fire(net, net.tops['fire6/concat'], 'fire7', 45, 155, 146)
# fire7
net.tops['fire8/concat'] = fire(net, net.tops['fire7/concat'], 'fire8', 49, 163, 171)
# fire8
net.tops['fire9/concat'] = fire(net, net.tops['fire8/concat'], 'fire9', 25, 29, 54)
net.pool9 = caffe.layers.Pooling(net.tops['fire9/concat'], pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
# fire9
net.tops['fire10/concat'] = fire(net, net.pool9, 'fire10', 37, 45, 56)
net.pool10 = caffe.layers.Pooling(net.tops['fire10/concat'], pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
# fire10
net.tops['fire11/concat'] = fire(net, net.pool10, 'fire11', 38, 41, 44)
# conv12
net.conv12_1 = caffe.layers.Convolution(net.tops['fire11/concat'], param=[dict(lr_mult=1.0, decay_mult=1.0)],
convolution_param={'num_output':51, 'bias_term':False, 'kernel_size':1, 'weight_filler':{'type':'msra'}})
net.tops['conv12_1/bn'] = caffe.layers.BatchNorm(net.conv12_1, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
net.tops['conv12_1/scale'] = caffe.layers.Scale(net.tops['conv12_1/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)],
scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
net.tops['conv12_1/relu'] = caffe.layers.ReLU(net.tops['conv12_1/scale'], in_place=True)
net.conv12_2 = caffe.layers.Convolution(net.tops['conv12_1/relu'], param=[dict(lr_mult=1.0, decay_mult=1.0)],
convolution_param={'num_output':46, 'bias_term':False, 'pad':1, 'kernel_size':3, 'stride':2, 'weight_filler':{'type':'msra'}})
net.tops['conv12_2/bn'] = caffe.layers.BatchNorm(net.conv12_2, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
net.tops['conv12_2/scale'] = caffe.layers.Scale(net.tops['conv12_2/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)],
scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
net.tops['conv12_2/relu'] = caffe.layers.ReLU(net.tops['conv12_2/scale'], in_place=True)
# conv13
net.conv13_1 = caffe.layers.Convolution(net.tops['conv12_2/relu'], param=[dict(lr_mult=1.0, decay_mult=1.0)],
convolution_param={'num_output':55, 'bias_term':False, 'kernel_size':1, 'weight_filler':{'type':'msra'}})
net.tops['conv13_1/bn'] = caffe.layers.BatchNorm(net.conv13_1, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
net.tops['conv13_1/scale'] = caffe.layers.Scale(net.tops['conv13_1/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)],
scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
net.tops['conv13_1/relu'] = caffe.layers.ReLU(net.conv13_1, in_place=True)
net.conv13_2 = caffe.layers.Convolution(net.tops['conv13_1/relu'], param=[dict(lr_mult=1.0, decay_mult=1.0)],
convolution_param={'num_output':85, 'bias_term':False, 'pad':1, 'kernel_size':3, 'stride':2, 'weight_filler':{'type':'msra'}})
net.tops['conv13_2/bn'] = caffe.layers.BatchNorm(net.conv13_2, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
net.tops['conv13_2/scale'] = caffe.layers.Scale(net.tops['conv13_2/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)],
scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
net.tops['conv13_2/relu'] = caffe.layers.ReLU(net.tops['conv13_2/scale'], in_place=True)
# fire 5 prior box
prior_box(net, net.tops['fire5/concat'], 'fire5', 16, 84, 21.0, 45.0, [2.0], 8)
# fire 9 prior box
prior_box(net, net.tops['fire9/concat'], 'fire9', 24, 126, 45.0, 99.0, [2.0, 3.0], 16)
# fire 10 prior box
prior_box(net, net.tops['fire10/concat'], 'fire10', 24, 126, 99.0, 153.0, [2.0, 3.0], 32)
# fire 11 prior box
prior_box(net, net.tops['fire11/concat'], 'fire11', 24, 126, 153.0, 207.0, [2.0, 3.0], 64)
# conv12_2 prior box
prior_box(net, net.tops['conv12_2'], 'conv12_2', 24, 126, 207.0, 261.0, [2.0, 3.0], 100)
# conv13_2 prior box
prior_box(net, net.tops['conv13_2'], 'conv13_2', 16, 84, 261.0, 315.0, [2.0], 300)
# last process
net.tops['mbox_loc'] = caffe.layers.Concat(net.tops['fire5_mbox_loc_flat'], net.tops['fire9_mbox_loc_flat'], net.tops['fire10_mbox_loc_flat'], net.tops['fire11_mbox_loc_flat'],
net.tops['conv12_2_mbox_loc_flat'], net.tops['conv13_2_mbox_loc_flat'], concat_param={'axis':1})
net.tops['mbox_conf'] = caffe.layers.Concat(net.tops['fire5_mbox_conf_flat'], net.tops['fire9_mbox_conf_flat'], net.tops['fire10_mbox_conf_flat'], net.tops['fire11_mbox_conf_flat'],
net.tops['conv12_2_mbox_conf_flat'], net.tops['conv13_2_mbox_conf_flat'], concat_param={'axis':1})
net.tops['mbox_priorbox'] = caffe.layers.Concat(net.tops['fire5_mbox_priorbox'], net.tops['fire9_mbox_priorbox'], net.tops['fire10_mbox_priorbox'], net.tops['fire11_mbox_priorbox'],
net.tops['conv12_2_mbox_priorbox'], net.tops['conv13_2_mbox_priorbox'], concat_param={'axis':2})
if(PHASE=='TRAIN'):
net.tops['mbox_loss'] = caffe.layers.MultiBoxLoss(net.tops['mbox_loc'], net.tops['mbox_conf'], net.tops['mbox_priorbox'], net.label, include={'phase':caffe.TRAIN},
propagate_down=[True, True, False, False], loss_param={'normalization':caffe.params.Loss.VALID}, multibox_loss_param={'loc_loss_type':caffe.params.MultiBoxLoss.SMOOTH_L1,
'conf_loss_type':caffe.params.MultiBoxLoss.SOFTMAX, 'loc_weight':1.0,
'num_classes':21, 'share_location':True, 'match_type':caffe.params.MultiBoxLoss.PER_PREDICTION, 'overlap_threshold':0.5, 'use_prior_for_matching':True,
'background_label_id':0, 'use_difficult_gt':True, 'neg_pos_ratio':3.0, 'neg_overlap':0.5,
'code_type':caffe.params.PriorBox.CENTER_SIZE, 'ignore_cross_boundary_bbox':False, 'mining_type':caffe.params.MultiBoxLoss.MAX_NEGATIVE})
elif(PHASE=='DEPLOY'):
net.tops['mbox_conf_reshape'] = caffe.layers.Reshape(net.tops['mbox_conf'], reshape_param={'shape':{'dim':[0,-1,21]}})
net.tops['mbox_conf_softmax'] = caffe.layers.Softmax(net.tops['mbox_conf_reshape'], softmax_param={'axis':2})
net.tops['mbox_conf_flatten'] = caffe.layers.Flatten(net.tops['mbox_conf_softmax'], flatten_param={'axis':1})
net.tops['detection_out'] = caffe.layers.DetectionOutput(net.tops['mbox_loc'], net.tops['mbox_conf_flatten'], net.tops['mbox_priorbox'], include={
'phase':caffe.TEST}, detection_output_param={'num_classes':21, 'share_location':True, 'background_label_id':0,
'nms_param':{'nms_threshold':0.45, 'top_k':100}, 'code_type':caffe.params.PriorBox.CENTER_SIZE, 'keep_top_k':100, 'confidence_threshold':0.25})
return str(net.to_proto())
def fcn(obj_cls, part, split):
n = caffe.NetSpec()
n.data, n.label = L.Python(
module='pascalpart_layers',
layer='PASCALPartSegDataLayer',
ntop=2,
param_str=str(
dict(voc_dir='/home/cv/hdl/caffe/data/pascal/VOC',
part_dir='/home/cv/hdl/caffe/data/pascal/pascal-part',
obj_cls=obj_cls,
part=part,
split=split,
seed=1337)))
# the base net
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=100)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
n.pool3 = max_pool(n.relu3_3)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 512)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 512)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 512)
n.pool4 = max_pool(n.relu4_3)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 512)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 512)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 512)
n.pool5 = max_pool(n.relu5_3)
# fully conv
n.fc6, n.relu6 = conv_relu(n.pool5, 4096, ks=7, pad=0)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7, n.relu7 = conv_relu(n.drop6, 4096, ks=1, pad=0)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.score_fr = L.Convolution(
n.drop7,
num_output=11,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.upscore2 = L.Deconvolution(n.score_fr,
convolution_param=dict(num_output=11,
kernel_size=4,
stride=2,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_pool4 = L.Convolution(
n.pool4,
num_output=11,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.score_pool4c = crop(n.score_pool4, n.upscore2)
n.fuse_pool4 = L.Eltwise(n.upscore2,
n.score_pool4c,
operation=P.Eltwise.SUM)
n.upscore16 = L.Deconvolution(n.fuse_pool4,
convolution_param=dict(num_output=11,
kernel_size=32,
stride=16,
bias_term=False),
param=[dict(lr_mult=0)])
n.score = crop(n.upscore16, n.data)
n.loss = L.SoftmaxWithLoss(n.score,
n.label,
loss_param=dict(normalize=False,
ignore_label=255))
return n.to_proto()
def convert_symbol2proto(symbol):
def looks_like_weight(name):
"""Internal helper to figure out if node should be hidden with `hide_weights`.
"""
if name.endswith("_weight"):
return True
if name.endswith("_bias"):
return True
if name.endswith("_beta") or name.endswith("_gamma") or name.endswith("_moving_var") or name.endswith(
"_moving_mean"):
return True
return False
json_symbol = json.loads(symbol.tojson())
all_nodes = json_symbol['nodes']
no_weight_nodes = []
for node in all_nodes:
op = node['op']
name = node['name']
if op == 'null':
if looks_like_weight(name):
continue
no_weight_nodes.append(node)
# build next node dict
next_node = dict()
for node in no_weight_nodes:
node_name = node['name']
for input in node['inputs']:
last_node_name = all_nodes[input[0]]['name']
if last_node_name in next_node:
next_node[last_node_name].append(node_name)
else:
next_node[last_node_name] = [node_name]
supported_op_type = ['null', 'BatchNorm', 'Convolution', 'Activation', 'Pooling', 'elemwise_add', 'SliceChannel',
'FullyConnected', 'SoftmaxOutput', '_maximum', 'add_n', 'Concat', '_mul_scalar', 'Deconvolution', 'UpSampling']
top_dict = dict()
caffe_net = caffe.NetSpec()
for node in no_weight_nodes:
if node['op'] == 'null':
input_param = dict()
if node['name'] == 'data':
input_param['shape'] = dict(dim=[1, 3, 160, 160])
else:
input_param['shape'] = dict(dim=[1])
top_data = CL.Input(ntop=1, input_param=input_param)
top_dict[node['name']] = [top_data]
setattr(caffe_net, node['name'], top_data)
elif node['op'].endswith('_copy'):
pass
elif node['op'] == 'BatchNorm':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
attr = node['attrs']
in_place = False
if len(next_node[bottom_node_name]) == 1:
in_place = True
if 'momentum' in attr:
momentum = float(attr['momentum'])
else:
momentum = 0.9
if 'eps' in attr:
eps = float(attr['eps'])
else:
eps = 0.001
if NO_INPLACE:
in_place = False
bn_top = CL.BatchNorm(top_dict[bottom_node_name][input[1]], ntop=1,
batch_norm_param=dict(use_global_stats=True,
moving_average_fraction=momentum,
eps=eps), in_place=in_place)
setattr(caffe_net, node['name'], bn_top)
scale_top = CL.Scale(bn_top, ntop=1, scale_param=dict(bias_term=True), in_place=not NO_INPLACE)
top_dict[node['name']] = [scale_top]
setattr(caffe_net, node['name'] + '_scale', scale_top)
elif node['op'] == 'Convolution':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
attr = node['attrs']
convolution_param = dict()
if 'kernel' in attr:
kernel_size = eval(attr['kernel'])
assert kernel_size[0] == kernel_size[1]
convolution_param['kernel_size'] = kernel_size[0]
else:
convolution_param['kernel_size'] = 1
if 'no_bias' in attr:
convolution_param['bias_term'] = not eval(attr['no_bias'])
if 'num_group' in attr:
convolution_param['group'] = int(attr['num_group'])
convolution_param['num_output'] = int(attr['num_filter'])
if 'pad' in attr:
pad_size = eval(attr['pad'])
assert pad_size[0] == pad_size[1]
convolution_param['pad'] = pad_size[0]
if 'stride' in attr:
stride_size = eval(attr['stride'])
assert stride_size[0] == stride_size[1]
convolution_param['stride'] = stride_size[0]
conv_top = CL.Convolution(top_dict[bottom_node_name][input[1]], ntop=1, convolution_param=convolution_param)
top_dict[node['name']] = [conv_top]
setattr(caffe_net, node['name'], conv_top)
elif node['op'] == 'Deconvolution':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
attr = node['attrs']
convolution_param = dict()
if 'kernel' in attr:
kernel_size = eval(attr['kernel'])
assert kernel_size[0] == kernel_size[1]
convolution_param['kernel_size'] = kernel_size[0]
else:
convolution_param['kernel_size'] = 1
if 'no_bias' in attr:
convolution_param['bias_term'] = not eval(attr['no_bias'])
else:
convolution_param['bias_term'] = False
if 'num_group' in attr:
convolution_param['group'] = int(attr['num_group'])
convolution_param['num_output'] = int(attr['num_filter'])
if 'pad' in attr:
pad_size = eval(attr['pad'])
assert pad_size[0] == pad_size[1]
convolution_param['pad'] = pad_size[0]
if 'stride' in attr:
stride_size = eval(attr['stride'])
assert stride_size[0] == stride_size[1]
convolution_param['stride'] = stride_size[0]
conv_top = CL.Deconvolution(top_dict[bottom_node_name][input[1]], ntop=1, convolution_param=convolution_param)
top_dict[node['name']] = [conv_top]
setattr(caffe_net, node['name'], conv_top)
elif node['op'] == 'UpSampling':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
attr = node['attrs']
convolution_param = dict()
if 'scale' in attr:
kernel_size = 2 * eval(attr['scale']) - eval(attr['scale']) % 2
convolution_param['kernel_size'] = kernel_size
else:
convolution_param['kernel_size'] = 1
convolution_param['bias_term'] = False
convolution_param['num_output'] = int(attr['num_filter'])
convolution_param['group'] = int(attr['num_filter'])
convolution_param['pad'] = int(math.ceil((eval(attr['scale']) - 1) / 2.))
convolution_param['stride'] = eval(attr['scale'])
conv_top = CL.Deconvolution(top_dict[bottom_node_name][input[1]], ntop=1,
convolution_param=convolution_param)
top_dict[node['name']] = [conv_top]
setattr(caffe_net, node['name'], conv_top)
elif node['op'] == 'Activation':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
attr = node['attrs']
in_place = False
if len(next_node[bottom_node_name]) == 1:
in_place = True
if NO_INPLACE:
in_place = False
if attr['act_type'] == 'relu':
ac_top = CL.ReLU(top_dict[bottom_node_name][input[1]], ntop=1, in_place=in_place)
elif attr['act_type'] == 'sigmoid':
ac_top = CL.Sigmoid(top_dict[bottom_node_name][input[1]], ntop=1, in_place=in_place)
elif attr['act_type'] == 'tanh':
ac_top = CL.TanH(top_dict[bottom_node_name][input[1]], ntop=1, in_place=in_place)
top_dict[node['name']] = [ac_top]
setattr(caffe_net, node['name'], ac_top)
elif node['op'] == 'Pooling':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
attr = node['attrs']
pooling_param = dict()
if attr['pool_type'] == 'avg':
pooling_param['pool'] = 1
elif attr['pool_type'] == 'max':
pooling_param['pool'] = 0
else:
assert False, attr['pool_type']
if 'global_pool' in attr and eval(attr['global_pool']) is True:
pooling_param['global_pooling'] = True
else:
if 'kernel' in attr:
kernel_size = eval(attr['kernel'])
assert kernel_size[0] == kernel_size[1]
pooling_param['kernel_size'] = kernel_size[0]
if 'pad' in attr:
pad_size = eval(attr['pad'])
assert pad_size[0] == pad_size[1]
pooling_param['pad'] = pad_size[0]
if 'stride' in attr:
stride_size = eval(attr['stride'])
assert stride_size[0] == stride_size[1]
pooling_param['stride'] = stride_size[0]
pool_top = CL.Pooling(top_dict[bottom_node_name][input[1]], ntop=1, pooling_param=pooling_param)
top_dict[node['name']] = [pool_top]
setattr(caffe_net, node['name'], pool_top)
elif node['op'] == 'elemwise_add' or node['op'] == 'add_n':
input_a = node['inputs'][0]
while True:
if all_nodes[input_a[0]]['op'] not in supported_op_type:
input_a = all_nodes[input_a[0]]['inputs'][0]
else:
break
input_b = node['inputs'][1]
while True:
if all_nodes[input_b[0]]['op'] not in supported_op_type:
input_b = all_nodes[input_b[0]]['inputs'][0]
else:
break
bottom_node_name_a = all_nodes[input_a[0]]['name']
bottom_node_name_b = all_nodes[input_b[0]]['name']
eltwise_param = dict()
eltwise_param['operation'] = 1
ele_add_top = CL.Eltwise(top_dict[bottom_node_name_a][input_a[1]], top_dict[bottom_node_name_b][input_b[1]],
ntop=1, eltwise_param=eltwise_param)
top_dict[node['name']] = [ele_add_top]
setattr(caffe_net, node['name'], ele_add_top)
elif node['op'] == '_maximum':
input_a = node['inputs'][0]
while True:
if all_nodes[input_a[0]]['op'] not in supported_op_type:
input_a = all_nodes[input_a[0]]['inputs'][0]
else:
break
input_b = node['inputs'][1]
while True:
if all_nodes[input_b[0]]['op'] not in supported_op_type:
input_b = all_nodes[input_b[0]]['inputs'][0]
else:
break
bottom_node_name_a = all_nodes[input_a[0]]['name']
bottom_node_name_b = all_nodes[input_b[0]]['name']
eltwise_param = dict()
eltwise_param['operation'] = 2
ele_add_top = CL.Eltwise(top_dict[bottom_node_name_a][input_a[1]], top_dict[bottom_node_name_b][input_b[1]],
ntop=1, eltwise_param=eltwise_param)
top_dict[node['name']] = [ele_add_top]
setattr(caffe_net, node['name'], ele_add_top)
elif node['op'] == '_mul_scalar':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
attr = node['attrs']
in_place = False
if len(next_node[bottom_node_name]) == 1:
in_place = True
if NO_INPLACE:
in_place = False
scale_top = CL.Scale(top_dict[bottom_node_name][input[1]], ntop=1, scale_param=dict(bias_term=False, filler=dict(value=-1)), in_place=in_place)
# scale_top = CL.Power(top_dict[bottom_node_name][input[1]], power=1.0, scale=float(attr['scalar']), shift=0, in_place=in_place)
top_dict[node['name']] = [scale_top]
setattr(caffe_net, node['name'], scale_top)
elif node['op'] == 'SliceChannel':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
slice_param = dict()
slice_param['slice_dim'] = 1
slice_num = 2
slice_outputs = CL.Slice(top_dict[bottom_node_name][input[1]], ntop=slice_num, slice_param=slice_param)
top_dict[node['name']] = slice_outputs
for idx, output in enumerate(slice_outputs):
setattr(caffe_net, node['name'] + '_' + str(idx), output)
elif node['op'] == 'FullyConnected':
input = node['inputs'][0]
while True:
if all_nodes[input[0]]['op'] not in supported_op_type:
input = all_nodes[input[0]]['inputs'][0]
else:
break
bottom_node_name = all_nodes[input[0]]['name']
attr = node['attrs']
inner_product_param = dict()
inner_product_param['num_output'] = int(attr['num_hidden'])
fc_top = CL.InnerProduct(top_dict[bottom_node_name][input[1]], ntop=1,
inner_product_param=inner_product_param)
top_dict[node['name']] = [fc_top]
setattr(caffe_net, node['name'], fc_top)
elif node['op'] == 'SoftmaxOutput':
input_a = node['inputs'][0]
while True:
if all_nodes[input_a[0]]['op'] not in supported_op_type:
input_a = all_nodes[input_a[0]]['inputs'][0]
else:
break
input_b = node['inputs'][1]
while True:
if all_nodes[input_b[0]]['op'] not in supported_op_type:
input_b = all_nodes[input_b[0]]['inputs'][0]
else:
break
bottom_node_name_a = all_nodes[input_a[0]]['name']
bottom_node_name_b = all_nodes[input_b[0]]['name']
softmax_loss = CL.SoftmaxWithLoss(top_dict[bottom_node_name_a][input_a[1]],
top_dict[bottom_node_name_b][input_b[1]], ntop=1)
top_dict[node['name']] = [softmax_loss]
setattr(caffe_net, node['name'], softmax_loss)
elif node['op'] == 'Concat':
if len(node['inputs']) == 2:
input_a = node['inputs'][0]
while True:
if all_nodes[input_a[0]]['op'] not in supported_op_type:
input_a = all_nodes[input_a[0]]['inputs'][0]
else:
break
input_b = node['inputs'][1]
while True:
if all_nodes[input_b[0]]['op'] not in supported_op_type:
input_b = all_nodes[input_b[0]]['inputs'][0]
else:
break
bottom_node_name_a = all_nodes[input_a[0]]['name']
bottom_node_name_b = all_nodes[input_b[0]]['name']
concat_top = CL.Concat(top_dict[bottom_node_name_a][input_a[1]], top_dict[bottom_node_name_b][input_b[1]], ntop=1)
top_dict[node['name']] = [concat_top]
setattr(caffe_net, node['name'], concat_top)
elif len(node['inputs']) == 3:
input_a = node['inputs'][0]
while True:
if all_nodes[input_a[0]]['op'] not in supported_op_type:
input_a = all_nodes[input_a[0]]['inputs'][0]
else:
break
input_b = node['inputs'][1]
while True:
if all_nodes[input_b[0]]['op'] not in supported_op_type:
input_b = all_nodes[input_b[0]]['inputs'][0]
else:
break
input_c = node['inputs'][2]
while True:
if all_nodes[input_c[0]]['op'] not in supported_op_type:
input_c = all_nodes[input_c[0]]['inputs'][0]
else:
break
bottom_node_name_a = all_nodes[input_a[0]]['name']
bottom_node_name_b = all_nodes[input_b[0]]['name']
bottom_node_name_c = all_nodes[input_c[0]]['name']
concat_top = CL.Concat(top_dict[bottom_node_name_a][input_a[1]],
top_dict[bottom_node_name_b][input_b[1]],
top_dict[bottom_node_name_c][input_c[1]], ntop=1)
top_dict[node['name']] = [concat_top]
setattr(caffe_net, node['name'], concat_top)
else:
logging.warn('unknown op type = %s' % node['op'])
return caffe_net.to_proto()
def create_neural_net(input_file, batch_size=50):
net = caffe.NetSpec()
net.data, net.label = L.Data(batch_size=batch_size,
source=input_file,
backend=caffe.params.Data.LMDB,
ntop=2,
include=dict(phase=caffe.TEST),
name='juniward04')
## pre-process
net.conv1 = L.Convolution(net.data,
num_output=16,
kernel_size=4,
stride=1,
pad=1,
weight_filler=dict(type='dct4'),
param=[{
'lr_mult': 0,
'decay_mult': 0
}],
bias_term=False)
TRUNCABS = caffe_pb2.QuantTruncAbsParameter.TRUNCABS
net.quanttruncabs = L.QuantTruncAbs(net.conv1,
process=TRUNCABS,
threshold=8,
in_place=True)
## block 1
[
net.conv1_proj, net.bn2, net.scale2, net.conv512_1, net.bn2_1,
net.scale2_1, net.relu512_1, net.conv512_to_256, net.bn2_2,
net.scale2_2, net.res512_to_256, net.relu512_to_256
] = add_downsampling_block(net.quanttruncabs, 12)
## block 2
[
net.conv256_1, net.bn2_3, net.scale2_3, net.relu256_1, net.conv256_2,
net.bn2_4, net.scale2_4, net.res256_2, net.relu256_2
] = add_skip_block(net.res512_to_256, 24)
## block 2_1
[
net.conv256_4, net.bn3_1, net.scale3_1, net.relu256_4, net.conv256_5,
net.bn3_2, net.scale3_2, net.res256_5, net.relu256_5
] = add_skip_block(net.res256_2, 24)
## block 2_2
[
net.conv256_6, net.bn4_1, net.scale4_1, net.relu256_6, net.conv256_7,
net.bn4_2, net.scale4_2, net.res256_7, net.relu256_7
] = add_skip_block(net.res256_5, 24)
## block 3
[
net.res256_2_proj, net.bn2_5, net.scale2_5, net.conv256_3, net.bn2_6,
net.scale2_6, net.relu256_3, net.conv256_to_128, net.bn2_7,
net.scale2_7, net.res256_to_128, net.relu256_to_128
] = add_downsampling_block(net.res256_7, 24)
## block 4 cur
[
net.conv128_1, net.bn2_8, net.scale2_8, net.relu128_1, net.conv128_2,
net.bn2_9, net.scale2_9, net.res128_2, net.relu128_2
] = add_skip_block(net.res256_to_128, 48)
## block 4_1
[
net.conv128_4, net.bn3_3, net.scale3_3, net.relu128_4, net.conv128_5,
net.bn3_4, net.scale3_4, net.res128_5, net.relu128_5
] = add_skip_block(net.res128_2, 48)
## block 4_2
[
net.conv128_6, net.bn4_3, net.scale4_3, net.relu128_6, net.conv128_7,
net.bn4_4, net.scale4_4, net.res128_7, net.relu128_7
] = add_skip_block(net.res128_5, 48)
## block 5
[
net.res128_2_proj, net.bn2_10, net.scale2_10, net.conv128_3,
net.bn2_11, net.scale2_11, net.relu128_3, net.conv128_to_64,
net.bn2_12, net.scale2_12, net.res128_to_64, net.relu128_to_64
] = add_downsampling_block(net.res128_7, 48)
## block 6
[
net.conv64_1, net.bn2_13, net.scale2_13, net.relu64_1, net.conv64_2,
net.bn2_14, net.scale2_14, net.res64_2, net.relu64_2
] = add_skip_block(net.res128_to_64, 96)
## block 6_1
[
net.conv64_4, net.bn3_5, net.scale3_5, net.relu64_4, net.conv64_5,
net.bn3_6, net.scale3_6, net.res64_5, net.relu64_5
] = add_skip_block(net.res64_2, 96)
## block 6_2
[
net.conv64_6, net.bn4_5, net.scale4_5, net.relu64_6, net.conv64_7,
net.bn4_6, net.scale4_6, net.res64_7, net.relu64_7
] = add_skip_block(net.res64_5, 96)
## block 7
[
net.res64_2_proj, net.bn2_15, net.scale2_15, net.conv64_3, net.bn2_16,
net.scale2_16, net.relu64_3, net.conv64_to_32, net.bn2_17,
net.scale2_17, net.res64_to_32, net.relu64_to_32
] = add_downsampling_block(net.res64_7, 96)
## block 8
[
net.conv32_1, net.bn2_18, net.scale2_18, net.relu32_1, net.conv32_2,
net.bn2_19, net.scale2_19, net.res32_2, net.relu32_2
] = add_skip_block(net.res64_to_32, 192)
## block 8_1
[
net.conv32_4, net.bn3_7, net.scale3_7, net.relu32_4, net.conv32_5,
net.bn3_8, net.scale3_8, net.res32_5, net.relu32_5
] = add_skip_block(net.res32_2, 192)
## block 8_2
[
net.conv32_6, net.bn4_7, net.scale4_7, net.relu32_6, net.conv32_7,
net.bn4_8, net.scale4_8, net.res32_7, net.relu32_7
] = add_skip_block(net.res32_5, 192)
## block 9
[
net.res32_2_proj, net.bn2_20, net.scale2_20, net.conv32_3, net.bn2_21,
net.scale2_21, net.relu32_3, net.conv32_to_16, net.bn2_22,
net.scale2_22, net.res32_to_16, net.relu32_to_16
] = add_downsampling_block(net.res32_7, 192)
## block 10
[
net.conv16_1, net.bn2_23, net.scale2_23, net.relu16_1, net.conv16_2,
net.bn2_24, net.scale2_24, net.res16_2, net.relu16_2
] = add_skip_block(net.res32_to_16, 384)
## block 10_1
[
net.conv16_3, net.bn3_9, net.scale3_9, net.relu16_3, net.conv16_4,
net.bn3_10, net.scale3_10, net.res16_4, net.relu16_4
] = add_skip_block(net.res16_2, 384)
## ## block 10_2
## [net.conv16_5, net.bn4_9, net.scale4_9, net.relu16_5, net.conv16_6, net.bn4_10,
## net.scale4_10, net.res16_6, net.relu16_6] = add_skip_block(net.res16_4, 384)
## global pool
AVE = caffe_pb2.PoolingParameter.AVE
net.global_pool = L.Pooling(net.res16_4, pool=AVE, kernel_size=8, stride=1)
## full connecting
net.fc = L.InnerProduct(net.global_pool,
param=[{
'lr_mult': 1
}, {
'lr_mult': 2
}],
num_output=2,
weight_filler=dict(type='xavier'),
bias_filler=dict(type='constant'))
## accuracy
net.accuracy = L.Accuracy(net.fc,
net.label,
include=dict(phase=caffe.TEST))
## loss
net.loss = L.SoftmaxWithLoss(net.fc, net.label)
return net.to_proto()
def net(split):
n=caffe.NetSpec()
loss_param=dict(normalize=False)
if split == 'train':
data_params=dict(mean=(104.00699, 116.66877, 122.67892))
data_params['root']='/home/liang/rcf/data/HED-BSDS'
data_params['source']='bsds_pascal_train_pair.lst'
data_params['shuffle']=True
n.data,n.label=L.Python(module='ImageLabelmapData',layer='ImageLabelmapDataLayer',ntop=2,param_str=str(data_params))
elif split == 'test':
n.data=L.Input(name='data',input_param=dict(shape=dict(dim=[1,3,500,500])))
else:
raise Exception('Invalid split')
#vgg architecture
n.conv1_1,n.relu1_1=conv_relu(n.data,num_out=64)
n.conv1_2,n.relu1_2=conv_relu(n.relu1_1,num_out=64)
n.pool1=max_pool(n.relu1_2)
n.conv2_1,n.relu2_1=conv_relu(n.pool1,num_out=128)
n.conv2_2,n.relu2_2=conv_relu(n.relu2_1,num_out=128)
n.pool2=max_pool(n.relu2_2)
n.conv3_1,n.relu3_1=conv_relu(n.pool2,num_out=256)
n.conv3_2,n.relu3_2=conv_relu(n.relu3_1,num_out=256)
n.conv3_3,n.relu3_3=conv_relu(n.relu3_2,num_out=256)
n.pool3=max_pool(n.relu3_3)
n.conv4_1,n.relu4_1=conv_relu(n.pool3,num_out=512)
n.conv4_2,n.relu4_2=conv_relu(n.relu4_1,num_out=512)
n.conv4_3,n.relu4_3=conv_relu(n.relu4_2,num_out=512)
n.pool4=max_pool(n.relu4_3)
n.conv5_1,n.relu5_1=conv_relu(n.pool4,num_out=512,lr=[100,1,200,0])
n.conv5_2,n.relu5_2=conv_relu(n.relu5_1,num_out=512,lr=[100,1,200,0])
n.conv5_3,n.relu5_3=conv_relu(n.relu5_2,num_out=512,lr=[100,1,200,0])
#conv1
n.dsn1=conv1x1(n.conv1_2)
n.dsn1_crop=crop(n.dsn1,n.data)
if split=='train':
n.dsn1_loss=L.SigmoidCrossEntropyLoss(n.dsn1_crop,n.label)
else:
n.sigmoid_dsn1=L.Sigmoid(n.dsn1_crop)
#conv2
n.dsn2=conv1x1(n.conv2_2)
n.dsn2_up=upsample(n.dsn2,stride=2)
n.dsn2_crop=crop(n.dsn2_up,n.data)
if split=='train':
n.dsn2_loss=L.SigmoidCrossEntropyLoss(n.dsn2_crop,n.label)
else:
n.sigmoid_dsn2=L.Sigmoid(n.dsn2_crop)
#conv3
n.dsn3=conv1x1(n.conv3_3)
n.dsn3_up=upsample(n.dsn3,stride=4)
n.dsn3_crop=crop(n.dsn3_up,n.data)
if split=='train':
n.dsn3_loss=L.SigmoidCrossEntropyLoss(n.dsn3_crop,n.label)
else:
n.sigmoid_dsn3=L.Sigmoid(n.dsn3_crop)
#conv4
n.dsn4=conv1x1(n.conv4_3)
n.dsn4_up=upsample(n.dsn4,stride=8)
n.dsn4_crop=crop(n.dsn4_up,n.data)
if split=='train':
n.dsn4_loss=L.SigmoidCrossEntropyLoss(n.dsn4_crop,n.label)
else:
n.sigmoid_dsn4=L.Sigmoid(n.dsn4_crop)
#conv5
n.dsn5=conv1x1(n.conv5_3)
n.dsn5_up=upsample(n.dsn5,stride=16)
n.dsn5_crop=crop(n.dsn5_up,n.data)
if split=='train':
n.dsn5_loss=L.SigmoidCrossEntropyLoss(n.dsn5_crop,n.label)
else:
n.sigmoid_dsn5=L.Sigmoid(n.dsn5_crop)
#concat
n.concat_5=L.Concat(n.dsn1_crop,n.dsn2_crop,n.dsn3_crop,n.dsn4_crop,n.dsn5_crop,name='concat',concat_param=dict(concat_dim=1))
n.dsn=L.Convolution(n.concat_5,name='dsn',num_output=1,kernel_size=1,param=[dict(lr_mult=0.001, decay_mult=1), dict(lr_mult=0.002, decay_mult=0)],weight_filler=dict(type='constant', value=0.2))
if split=='train':
n.fuse_loss=L.SigmoidCrossEntropyLoss(n.dsn,n.label)
else:
n.sigmoid_fuss=L.Sigmoid(n.dsn)
return n.to_proto()
def vgg_lowmem(data,
labels=None,
train=False,
param=learned_param,
num_classes=100,
with_labels=True):
"""
Returns a protobuf text file specifying a variant of VGG
- The Fully Connected (FC) layers (fc6 and fc7) have smaller dimensions
due to the lower resolution of mini-places images (128x128) compared
with ImageNet images (usually resized to 256x256)
"""
n = caffe.NetSpec()
n.data = data
conv_kwargs = dict(param=param, train=train)
n.conv1_1, n.relu1_1 = conv_relu(n.data,
3,
64,
stride=1,
pad=1,
**conv_kwargs)
n.conv1_1, n.relu1_1 = conv_relu(n.data,
3,
64,
stride=1,
pad=1,
**conv_kwargs)
n.pool1 = max_pool(n.relu1_1, 3, stride=2, train=train)
n.conv2, n.relu2 = conv_relu(n.pool1,
5,
256,
pad=2,
group=2,
**conv_kwargs)
n.pool2 = max_pool(n.relu2, 3, stride=2, train=train)
n.conv3, n.relu3 = conv_relu(n.pool2, 3, 384, pad=1, **conv_kwargs)
n.conv4, n.relu4 = conv_relu(n.relu3,
3,
384,
pad=1,
group=2,
**conv_kwargs)
n.conv5, n.relu5 = conv_relu(n.relu4,
3,
256,
pad=1,
group=2,
**conv_kwargs)
n.pool5 = max_pool(n.relu5, 3, stride=2, train=train)
n.fc6, n.relu6 = fc_relu(n.pool5, 1024, param=param)
n.drop6 = L.Dropout(n.relu6, in_place=True)
n.fc7, n.relu7 = fc_relu(n.drop6, 1024, param=param)
n.drop7 = L.Dropout(n.relu7, in_place=True)
n.fc8 = L.InnerProduct(n.drop7, num_output=num_classes, param=param)
preds = n.fc8
if not train:
# Compute the per-label probabilities at test/inference time.
preds = n.probs = L.Softmax(n.fc8)
if with_labels:
n.label = labels
n.loss = L.SoftmaxWithLoss(n.fc8, n.label)
n.accuracy_at_1 = L.Accuracy(preds, n.label)
n.accuracy_at_5 = L.Accuracy(preds,
n.label,
accuracy_param=dict(top_k=5))
else:
n.ignored_label = labels
n.silence_label = L.Silence(n.ignored_label, ntop=0)
return to_tempfile(str(n.to_proto()))
def net():
n = caffe.NetSpec()
n.data = L.Input(input_param=dict(shape=dict(dim=data_shape)))
n.dataout = L.Power(n.data, power=_power, scale=_scale, shift=_shift)
return n.to_proto()
