def alexnet(train_data_source,
test_data_source,
train_batch_size=128,
test_batch_size=50):
n = caffe.NetSpec()
n.data, n.label = L.ImageData(include=dict(phase=caffe.TRAIN),
transform_param=dict(mirror=False,
crop_size=227,
scale=1. / 255),
batch_size=train_batch_size,
source=train_data_source,
new_height=256,
new_width=256,
is_color=False,
ntop=2)
n.test_data, n.test_label = L.ImageData(include=dict(phase=caffe.TEST),
transform_param=dict(mirror=False,
crop_size=227,
scale=1. /
255),
batch_size=test_batch_size,
source=test_data_source,
new_height=256,
new_width=256,
is_color=False,
name='data',
top=['data', 'label'],
in_place=True,
ntop=2)
n.conv1 = L.Convolution(
n.data,
name='conv1',
num_output=96,
kernel_size=7,
stride=2,
group=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)])
n.conv1_bn = L.BatchNorm(n.conv1, eps=0)
n.relu1 = L.ReLU(n.conv1_bn, in_place=True)
n.pool1 = L.Pooling(n.relu1, kernel_size=3, stride=2, pool=P.Pooling.MAX)
n.conv2 = L.Convolution(
n.pool1,
name='conv2',
num_output=192,
pad=1,
kernel_size=5,
stride=2,
group=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)])
n.conv2_bn = L.BatchNorm(n.conv2, eps=0)
n.relu2 = L.ReLU(n.conv2_bn, in_place=True)
n.pool2 = L.Pooling(n.relu2, kernel_size=3, stride=2, pool=P.Pooling.MAX)
n.conv3 = L.Convolution(
n.pool2,
name='conv3',
num_output=384,
pad=1,
kernel_size=3,
stride=1,
group=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)])
n.conv3_bn = L.BatchNorm(n.conv3, eps=0)
n.relu3 = L.ReLU(n.conv3_bn, in_place=True)
n.conv4 = L.Convolution(
n.relu3,
name='conv4',
num_output=384,
pad=1,
kernel_size=3,
stride=1,
group=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)])
n.conv4_bn = L.BatchNorm(n.conv4, eps=0)
n.relu4 = L.ReLU(n.conv4_bn, in_place=True)
n.conv5 = L.Convolution(
n.relu4,
name='conv5',
num_output=192,
pad=1,
kernel_size=3,
stride=1,
group=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)])
n.conv5_bn = L.BatchNorm(n.conv5, eps=0)
n.relu5 = L.ReLU(n.conv5_bn, in_place=True)
n.pool5 = L.Pooling(n.relu5, kernel_size=3, stride=2, pool=P.Pooling.MAX)
n.fc6 = L.InnerProduct(
n.pool5,
name='fc6',
num_output=2048,
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)])
n.fc6_bn = L.BatchNorm(n.fc6, eps=0)
n.relu6 = L.ReLU(n.fc6_bn, in_place=True)
n.drop6 = L.Dropout(n.relu6, in_place=True)
n.fc7 = L.InnerProduct(
n.drop6,
name='fc7',
num_output=1024,
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)])
n.fc7_bn = L.BatchNorm(n.fc7, eps=0)
n.relu7 = L.ReLU(n.fc7_bn, in_place=True)
n.drop7 = L.Dropout(n.relu7, in_place=True)
n.fc8 = L.InnerProduct(n.drop7, name='fc8', num_output=1221)
n.accuracy = L.Accuracy(n.fc8, n.label, include=dict(phase=caffe.TEST))
n.loss = L.SoftmaxWithLoss(n.fc8, n.label)
return 'name: "AlexNet"\n' + str(n.to_proto())
def resnet_layers_proto(self,
batch_size,
phase='TRAIN',
stages=(3, 4, 6, 3)):
"""
(3, 4, 6, 3) for 50 layers; (3, 4, 23, 3) for 101 layers; (3, 8, 36, 3) for 152 layers
"""
global use_global_stats
n = caffe.NetSpec()
if phase == 'TRAIN':
n.data, n.label = L.Data(source=self.train_data,
backend=P.Data.LMDB,
batch_size=batch_size,
ntop=2,
include=dict(phase=0),
transform_param=dict(
crop_size=224,
mean_value=[104, 117, 123],
mirror=True))
use_global_stats = False
else:
n.data, n.label = L.Data(source=self.test_data,
backend=P.Data.LMDB,
batch_size=batch_size,
ntop=2,
include=dict(phase=1),
transform_param=dict(
crop_size=224,
mean_value=[104, 117, 123],
mirror=False))
use_global_stats = True
n.conv1, n.conv1_bn, n.conv1_scale, n.conv1_relu = conv_bn_scale_relu(
n.data,
num_output=64,
kernel_size=7,
stride=2,
pad=3,
bias_term=True)
n.pool1 = L.Pooling(n.conv1,
kernel_size=3,
stride=2,
pool=P.Pooling.MAX) # 64x56x56
for num in xrange(len(stages)): # num = 0, 1, 2, 3
for i in xrange(stages[num]):
if i == 0:
stage_string = branch_shortcut_string
bottom_string = [
'n.pool1',
'n.res2b%s' % str(stages[0] - 1),
'n.res3b%s' % str(stages[1] - 1),
'n.res4b%s' % str(stages[2] - 1)
][num]
else:
stage_string = branch_string
if i == 1:
bottom_string = 'n.res%sa' % str(num + 2)
else:
bottom_string = 'n.res%sb%s' % (str(num + 2),
str(i - 1))
exec(
stage_string.replace('(stage)', str(num + 2)).replace(
'(bottom)', bottom_string).replace(
'(num)', str(2**num * 64)).replace(
'(order)',
str(i)).replace('(stride)',
str(int(num > 0) + 1)))
exec 'n.pool5 = L.Pooling((bottom), pool=P.Pooling.AVE, global_pooling=True)'.replace(
'(bottom)', 'n.res5b%s' % str(stages[3] - 1))
n.classifier = L.InnerProduct(n.pool5, num_output=self.classifier_num)
n.loss = L.SoftmaxWithLoss(n.classifier, n.label)
if phase == 'TEST':
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()
train_net.data, train_net.label = L.Data(source=train_data,backend=P.Data.LMDB, batch_size=64,ntop=2,transform_param=dict(scale=0.00390625))
# 生成LeNet5的主体结构
lenet5_body(train_net, 'data')
# 生成误差损失层
train_net.loss = L.SoftmaxWithLoss(train_net.ip2, train_net.label)
# 测试网络
test_net = caffe.NetSpec() # 基础网络
# 带标签的数据输入层
test_net.data, test_net.label = L.Data(source=test_data, batch_size=100, backend=P.Data.LMDB, ntop=2,transform_param=dict(scale=0.00390625))
# 生成LeNet5的主体结构
lenet5_body(test_net, 'data')
# 生成误差损失层
test_net.loss = L.SoftmaxWithLoss(test_net.ip2, test_net.label)
# 添加一个精确层
test_net.accuracy = L.Accuracy(test_net.ip2, test_net.label)
# 实施网络
deploy_net = caffe.NetSpec() # 基础网络
# 带标签的数据输入层
deploy_net.data = L.Input(input_param=dict(shape=dict(dim=[64,1,28,28])))
# 生成LeNet5的主体结构
lenet5_body(deploy_net, 'data')
deploy_net.prob = L.Softmax(deploy_net.ip2)
# 保存训练文件
with open(model_root+'train.prototxt', 'w') as f:
print('name: "LenNet5_train"', file=f)
print(train_net.to_proto(), file=f)
with open(model_root+'test.prototxt', 'w') as f:
def vgg_face(split, mean, opt):
n = caffe.NetSpec()
# config python data layer
if split == 'train':
batch_size = opt.train_batch_size
if split == 'val':
batch_size = opt.val_batch_size
if split == 'test':
batch_size = opt.test_batch_size
if split == 'train' or split == 'val':
dataset_name = opt.train_dataset_name
else:
dataset_name = opt.test_dataset_name
pydata_params = dict(split=split,
data_dir=opt.data_dir,
batch_size=batch_size,
mean=mean,
dataset=dataset_name,
load_size=opt.load_size,
crop_size=opt.crop_size)
n.data, n.label = L.Python(module='faceData_layers',
layer='FaceDataLayer',
ntop=2,
param_str=str(pydata_params))
# vgg-face net
# conv layers
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64)
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)
# drop out and fc layers
n.fc6, n.relu6, n.drop6 = fc_relu_dropout(n.pool5, 4096, 0.5)
n.fc7, n.relu7, n.drop7 = fc_relu_dropout(n.fc6, 4096, 0.5)
lr_ratio = 100 # lr multiplier for truncated layers
n.fc8_face = L.InnerProduct(n.fc7,
num_output=1024,
param=[
dict(lr_mult=1 * lr_ratio, decay_mult=1),
dict(lr_mult=2 * lr_ratio, decay_mult=0)
],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', value=0))
n.fc9_face = L.InnerProduct(n.fc8_face,
num_output=2,
param=[
dict(lr_mult=1 * lr_ratio, decay_mult=1),
dict(lr_mult=2 * lr_ratio, decay_mult=0)
],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', value=0))
# loss layer
n.loss = L.SoftmaxWithLoss(n.fc9_face, n.label)
# loss and accuracy layer
n.acc = L.Accuracy(n.fc9_face, n.label)
return n.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 16
[
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_1(net.quanttruncabs, 12)
# [net.conv1_proj, net.bn2, net.scale2, net.conv512_1, net.bn2_1, net.scale2_1,
# net.relu512_1, net.conv512_2, net.bn2_2, net.scale2_2, net.relu512_2, net.conv512_to_256,
# net.bn2_3, net.scale2_3, net.res512_to_256,
# net.relu512_to_256] = add_downsampling_block(net.quanttruncabs, 12)
## block 2 13
[
net.conv256_1, net.bn2_4, net.scale2_4, net.relu256_1, net.conv256_2,
net.bn2_5, net.scale2_5, net.relu256_2, net.conv256_3, net.bn2_6,
net.scale2_6, net.res256_3, net.relu256_3
] = add_skip_block(net.res512_to_256, 24)
## block 3 16
[
net.res256_3_proj, net.bn2_7, net.scale2_7, net.conv256_4, net.bn2_8,
net.scale2_8, net.relu256_4, net.conv256_5, net.bn2_9, net.scale2_9,
net.relu256_5, net.conv256_to_128, net.bn2_10, net.scale2_10,
net.res256_to_128, net.relu256_to_128
] = add_downsampling_block(net.res256_3, 24)
## block 4 13
[
net.conv128_1, net.bn2_11, net.scale2_11, net.relu128_1, net.conv128_2,
net.bn2_12, net.scale2_12, net.relu128_2, net.conv128_3, net.bn2_13,
net.scale2_13, net.res128_3, net.relu128_3
] = add_skip_block(net.res256_to_128, 48)
## block 5 16
[
net.res128_3_proj, net.bn2_14, net.scale2_14, net.conv128_4,
net.bn2_15, net.scale2_15, net.relu128_4, net.conv128_5, net.bn2_16,
net.scale2_16, net.relu128_5, net.conv128_to_64, net.bn2_17,
net.scale2_17, net.res128_to_64, net.relu128_to_64
] = add_downsampling_block(net.res128_3, 48)
## block 6 13
[
net.conv64_1, net.bn2_18, net.scale2_18, net.relu64_1, net.conv64_2,
net.bn2_19, net.scale2_19, net.relu64_2, net.conv64_3, net.bn2_20,
net.scale2_20, net.res64_3, net.relu64_3
] = add_skip_block(net.res128_to_64, 96)
## block 7 16
[
net.res64_3_proj, net.bn2_21, net.scale2_21, net.conv64_4, net.bn2_22,
net.scale2_22, net.relu64_4, net.con64_5, net.bn2_23, net.scale2_23,
net.relu64_5, net.conv64_to_32, net.bn2_24, net.scale2_24,
net.res64_to_32, net.relu64_to_32
] = add_downsampling_block(net.res64_3, 96)
## block 8 13
[
net.conv32_1, net.bn2_25, net.scale2_25, net.relu32_1, net.conv32_2,
net.bn2_26, net.scale2_26, net.relu32_2, net.conv32_3, net.bn2_27,
net.scale2_27, net.res32_3, net.relu32_3
] = add_skip_block(net.res64_to_32, 192)
## ## block 9 16
## [net.res32_3_proj, net.bn2_28, net.scale2_28, net.conv32_4, net.bn2_29, net.scale2_29,
## net.relu32_4, net.con32_5, net.bn2_30, net.scale2_30, net.relu32_5, net.conv32_to_16,
## net.bn2_31, net.scale2_31, net.res32_to_16,
## net.relu32_to_16] = add_downsampling_block(net.res32_3, 192)
[net.res32_3_proj, net.bn2_28,
net.scale2_28] = add_module(net.res32_3, 2 * 192, 1, 3, 2)
## block 10 13
[
net.conv16_1, net.bn2_32, net.scale2_32, net.relu16_1, net.conv16_2,
net.bn2_33, net.scale2_33, net.relu16_2, net.conv16_3, net.bn2_34,
net.scale2_34, net.res16_3, net.relu16_3
] = add_skip_block(net.res32_3_proj, 384)
## global pool
AVE = caffe_pb2.PoolingParameter.AVE
net.global_pool = L.Pooling(net.res16_3, 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 buildnet( inputdb, mean_file, batch_size, height, width, nchannels, net_type="train"):
net = caffe.NetSpec()
crop_size = -1
if augment_data:
crop_size = width
train = False
if net_type=="train":
train = True
data_layers,label = lt.data_layer_trimese( net, inputdb, mean_file, batch_size, net_type, height, width, nchannels, [1,2], crop_size=768 )
# First conv layer
branch_ends = []
for n,layer in enumerate(data_layers):
conv1 = lt.convolution_layer( net, layer, "plane%d_conv1"%(n), "tri_conv1_plane%d"%(n), 64, 2, 5, 3, 0.05, addbatchnorm=True, train=train )
pool1 = lt.pool_layer( net, conv1, "plane%d_pool1"%(n), 3, 1 )
conv2 = lt.convolution_layer( net, pool1, "plane%d_conv2"%(n), "tri_conv2_plane%d"%(n), 64, 2, 3, 3, 0.05, addbatchnorm=True, train=train )
conv3 = lt.convolution_layer( net, conv2, "plane%d_conv3"%(n), "tri_conv3_plane%d"%(n), 64, 2, 3, 3, 0.05, addbatchnorm=True, train=train )
pool3 = lt.pool_layer( net, conv3, "plane%d_pool3"%(n), 3, 1 )
branch_ends.append( pool3 )
concat = lt.concat_layer( net, "mergeplanes", *branch_ends )
resnet1 = lt.resnet_module( net, concat, "resnet1", 64*3, 3, 1, 1,8,32, use_batch_norm, train)
resnet2 = lt.resnet_module( net, resnet1, "resnet2", 32, 3, 1, 1,8,32, use_batch_norm, train)
resnet3 = lt.resnet_module( net, resnet2, "resnet3", 32, 3, 1, 1,16,64, use_batch_norm, train)
resnet4 = lt.resnet_module( net, resnet3, "resnet4", 64, 3, 1, 1,16,64, use_batch_norm, train)
resnet5 = lt.resnet_module( net, resnet4, "resnet5", 64, 3, 1, 1,16,64, use_batch_norm, train)
resnet6 = lt.resnet_module( net, resnet5, "resnet6", 64, 3, 1, 1,32,128, use_batch_norm, train)
resnet7 = lt.resnet_module( net, resnet6, "resnet7", 128, 3, 1, 1, 32,128, use_batch_norm, train)
resnet8 = lt.resnet_module( net, resnet7, "resnet8", 128, 3, 1, 1, 32,128, use_batch_norm, train)
resnet9 = lt.resnet_module( net, resnet8, "resnet9", 128, 3, 1, 1, 64,256, use_batch_norm, train)
net.lastpool = lt.pool_layer( net, resnet9, "lastpool", 5, 1, P.Pooling.AVE )
lastpool_layer = net.lastpool
if use_dropout:
net.lastpool_dropout = L.Dropout(net.lastpool,
in_place=True,
dropout_param=dict(dropout_ratio=0.5))
lastpool_layer = net.lastpool_dropout
fc1 = lt.final_fully_connect( net, lastpool_layer, nclasses=512 )
fc2 = lt.final_fully_connect( net, fc1, nclasses=4096 )
fc3 = lt.final_fully_connect( net, fc2, nclasses=2 )
if train:
net.loss = L.SoftmaxWithLoss(fc3, net.label )
net.acc = L.Accuracy(fc3,net.label)
else:
net.probt = L.Softmax( fc3 )
net.acc = L.Accuracy(fc3,net.label)
return net
def generate_mobile(method, CHOICE_LIST):
# new net and new data
global net
net = caffe.NetSpec()
train_part = generate_data(method)
input_blob_name = "data"
input_size = 224
input_channel = 3
# conv0
input_blob_name, input_size, input_channel = generate_conv(
input_blob_name,
"conv0",
input_size,
input_channel,
3,
2,
1,
16,
(method == "train_test"),
True,
)
# assert
assert len(CHOICE_LIST) == len(TYPICAL_CHOICE)
assert sum([int(v1 != v2)
for v1, v2 in zip(CHOICE_LIST, TYPICAL_CHOICE)]) in [0, 1]
for i in range(12):
layer_name = (["cL%d", "L%d"
][int(CHOICE_LIST[i] == TYPICAL_CHOICE[i])]) % (i + 1)
input_blob_name, input_size, input_channel = generate_resnet_block(
input_blob_name,
layer_name,
input_size,
input_channel,
CHOICE_LIST[i],
i,
method,
)
# conv_tile
input_blob_name, input_size, input_channel = generate_conv(
input_blob_name,
"conv_tile",
input_size,
input_channel,
1,
1,
1,
1504,
(method == "train_test"),
True,
)
# avg_pool
input_blob_name, input_size, input_channel = generate_pooling(
input_blob_name,
"conv_tile",
input_size,
input_channel,
Params.Pooling.AVE,
7,
7,
)
# fc loss and accuracy layer
input_blob_name = generate_fc(input_blob_name, "FC", 1000, False, False)
if method == "train_test":
## loss and accuracy
net.tops["loss"] = Layers.SoftmaxWithLoss(
net.tops[input_blob_name],
net.label,
name="loss",
ntop=1,
include=dict(phase=caffe.TRAIN),
)
net.tops["accuracy"] = Layers.Accuracy(
net.tops[input_blob_name],
net.label,
name="accuracy",
ntop=1,
include=dict(phase=caffe.TEST),
)
return train_part + str(net.to_proto())
else:
return str(net.to_proto())
def InceptionResNetV2(train_lmdb,
test_lmdb,
input_size=299,
batch_size=256,
stages=[0, 5, 10, 5],
first_output=32,
include_acc=False):
# now, this code can't recognize include phase, so there will only be a TEST phase data layer
data, label = L.Data(source=train_lmdb,
backend=P.Data.LMDB,
batch_size=batch_size,
ntop=2,
transform_param=dict(crop_size=input_size,
mean_value=[104, 117, 123],
mirror=True),
include=dict(phase=getattr(caffe_pb2, 'TRAIN')))
data, label = L.Data(source=test_lmdb,
backend=P.Data.LMDB,
batch_size=batch_size,
ntop=2,
transform_param=dict(crop_size=input_size,
mean_value=[104, 117, 123],
mirror=True),
include=dict(phase=getattr(caffe_pb2, 'TEST')))
data, label = L.MemoryData(batch_size=batch_size,
height=input_size,
width=input_size,
channels=3,
ntop=2,
transform_param=dict(mean_value=[104, 117, 123],
mirror=True),
include=dict(phase=getattr(caffe_pb2, 'TEST')))
Inception_ResNet_A_input = stem(bottom=data,
conv1_num=32,
conv2_num=32,
conv3_num=64,
conv4_num=96,
conv5_num=64,
conv6_num=96,
conv7_num=64,
conv8_num=64,
conv9_num=64,
conv10_num=96,
conv11_num=192)
for i in xrange(stages[1]):
Inception_ResNet_A_input = Inception_ResNet_A(
bottom=Inception_ResNet_A_input,
bottom_size=384,
num1x1=32,
num3x3=48,
num3x3double=64)
Inception_ResNet_B_input = ReductionA(bottom=Inception_ResNet_A_input,
num1x1_k=256,
num3x3_l=256,
num3x3_n=384,
num3x3_m=384)
for i in xrange(stages[2]):
Inception_ResNet_B_input = Inception_ResNet_B(
bottom=Inception_ResNet_B_input,
bottom_size=1152,
num1x1=192,
num1x1double=128,
num7x1=160,
num1x7=192)
Inception_ResNet_C_input = ReductionB(bottom=Inception_ResNet_B_input,
num1x1=256,
num3x3=384,
num3x3double=288,
num3x3three=320)
for i in xrange(stages[3]):
Inception_ResNet_C_input = Inception_ResNet_C(
bottom=Inception_ResNet_C_input,
bottom_size=2144,
num1x1=192,
num1x3=224,
num3x1=256)
glb_pool = L.Pooling(Inception_ResNet_C_input,
pool=P.Pooling.AVE,
global_pooling=True)
dropout = L.Dropout(glb_pool, dropout_ratio=0.2)
fc = L.InnerProduct(dropout, num_output=1000)
loss = L.SoftmaxWithLoss(fc, label)
acc = L.Accuracy(fc, label, include=dict(phase=getattr(caffe_pb2, 'TEST')))
return to_proto(loss, acc)
def accuracy(bottom, label, test_only=False):
if test_only:
return L.Accuracy(bottom, label, include=dict(phase=caffe.TEST))
return L.Accuracy(bottom, label)
def cnn(split):
n = caffe.NetSpec()
pydata_params = dict(dataset_dir='/home/kevin/dataset/washington_rgbd_dataset', split=split, mean=(104.00698793, 116.66876762, 122.67891434),
seed=1337, batch_size=128, img_size=(227,227))
if split == 'deploy':
n.img = L.Input(name='input', ntop=2, shape=[dict(dim=1),dict(dim=1),dict(dim=224),dict(dim=224)])
else:
pylayer = 'WashingtonDataLayer'
#---------------------------------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))
#---------------------------------RGB-Net---------------------------------------#
# the caffe-net (alex-net)
n.rgb_conv1, n.rgb_relu1 = conv_relu(n.rgb, 96, ks=11, stride=4, pad=0)
n.rgb_pool1 = max_pool(n.rgb_relu1, ks=3)
n.rgb_norm1 = L.LRN(n.rgb_pool1, lrn_param=dict(local_size=5, alpha=0.0005, beta=0.75, k=2))
n.rgb_conv2, n.rgb_relu2 = conv_relu(n.rgb_norm1, 256, ks=5, pad=2, group=2)
n.rgb_pool2 = max_pool(n.rgb_relu2, ks=3)
n.rgb_norm2 = L.LRN(n.rgb_pool2, lrn_param=dict(local_size=5, alpha=0.0005, beta=0.75, k=2))
n.rgb_conv3, n.rgb_relu3 = conv_relu(n.rgb_norm2, 384, ks=3, pad=1, lr1=1, lr2=2)
n.rgb_conv4, n.rgb_relu4 = conv_relu(n.rgb_relu3, 384, ks=3, pad=1, group=2, lr1=1, lr2=2)
n.rgb_conv5, n.rgb_relu5 = conv_relu(n.rgb_relu4, 256, ks=3, pad=1, group=2, lr1=1, lr2=2)
n.rgb_pool5 = max_pool(n.rgb_relu5, ks=3)
# fully conv
n.rgb_fc6, n.rgb_relu6 = fc_relu(n.rgb_pool5, 4096, lr1=1, lr2=2)
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=1, lr2=2)
n.rgb_drop7 = L.Dropout(n.rgb_relu7, dropout_ratio=0.5, in_place=True)
n.rgb_fc8 = fc(n.rgb_drop7, 51, lr1=1, lr2=2)
#---------------------------------Depth-Net---------------------------------------#
# the caffe-net (alex-net)
n.depth_conv1, n.depth_relu1 = conv_relu(n.depth, 96, ks=11, stride=4, pad=0)
n.depth_pool1 = max_pool(n.depth_relu1, ks=3)
n.depth_norm1 = L.LRN(n.depth_pool1, lrn_param=dict(local_size=5, alpha=0.0005, beta=0.75, k=2))
n.depth_conv2, n.depth_relu2 = conv_relu(n.depth_norm1, 256, ks=5, pad=2, group=2)
n.depth_pool2 = max_pool(n.depth_relu2, ks=3)
n.depth_norm2 = L.LRN(n.depth_pool2, lrn_param=dict(local_size=5, alpha=0.0005, beta=0.75, k=2))
n.depth_conv3, n.depth_relu3 = conv_relu(n.depth_norm2, 384, ks=3, pad=1, lr1=1, lr2=2)
n.depth_conv4, n.depth_relu4 = conv_relu(n.depth_relu3, 384, ks=3, pad=1, group=2, lr1=1, lr2=2)
n.depth_conv5, n.depth_relu5 = conv_relu(n.depth_relu4, 256, ks=3, pad=1, group=2, lr1=1, lr2=2)
n.depth_pool5 = max_pool(n.depth_relu5, ks=3)
# fully conv
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, 51, 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.overall_accuracy = L.Accuracy(n.rgbd_fc8, n.label)
#n.overall_loss = L.SoftmaxWithLoss(n.rgbd_fc8, n.label)
return n.to_proto()
def alexnet_bn_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=227,
mean_value=[104, 117, 123],
mirror=mirror))
n.conv1, n.conv1_bn, n.conv1_scale, n.conv1_relu = \
factorization_conv_bn_scale_relu(n.data, num_output=96, kernel_size=11, stride=4,) # 96x55x55
n.pool1 = L.Pooling(n.conv1,
kernel_size=3,
stride=2,
pool=P.Pooling.MAX) # 96x27x27
n.conv2, n.conv2_bn, n.conv2_scale, n.conv2_relu = \
factorization_conv_bn_scale_relu(n.pool1, num_output=256, kernel_size=5, pad=2) # 256x27x27
n.pool2 = L.Pooling(n.conv2,
kernel_size=3,
stride=2,
pool=P.Pooling.MAX) # 256x13x13
n.conv3, n.conv3_bn, n.conv3_scale, n.conv3_relu = \
factorization_conv_bn_scale_relu(n.pool2, num_output=384, kernel_size=3, pad=1) # 384x13x13
n.conv4, n.conv4_bn, n.conv4_scale, n.conv4_relu = \
factorization_conv_bn_scale_relu(n.conv3, num_output=384, kernel_size=3, pad=1) # 384x13x13
n.conv5, n.conv5_bn, n.conv5_scale, n.conv5_relu = \
factorization_conv_bn_scale_relu(n.conv4, num_output=256, kernel_size=3, pad=1) # 256x13x13
n.pool5 = L.Pooling(n.conv5,
kernel_size=3,
stride=2,
pool=P.Pooling.MAX) # 256x6x16
n.fc6, n.relu6, n.drop6 = fc_relu_drop(n.pool5,
num_output=2048) # 1024x1x1
n.fc7, n.relu7, n.drop7 = fc_relu_drop(n.fc6,
num_output=2048) # 1024x1x1
n.fc8 = L.InnerProduct(n.fc7,
num_output=self.classifier_num,
param=[
dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)
],
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', value=0))
if phase == 'TRAIN':
pass
else:
n.accuracy_top1 = L.Accuracy(n.fc8, n.label, include=dict(phase=1))
n.accuracy_top5 = L.Accuracy(n.fc8,
n.label,
include=dict(phase=1),
accuracy_param=dict(top_k=5))
n.loss = L.SoftmaxWithLoss(n.fc8, n.label)
return n.to_proto()
def buildnet(inputdb,
mean_file,
batch_size,
height,
width,
nchannels,
net_type="train"):
net = caffe.NetSpec()
crop_size = -1
if augment_data:
crop_size = width
train = False
if net_type == "train":
train = True
data_layers, label = lt.data_layer_stacked(net,
inputdb,
mean_file,
batch_size,
net_type,
height,
width,
nchannels,
crop_size=crop_size)
# First conv layer
conv1 = lt.convolution_layer(net,
data_layers[0],
"conv1",
"conv1",
16,
2,
7,
3,
0.05,
addbatchnorm=True,
train=train)
pool1 = lt.pool_layer(net, conv1, "pool1", 5, 3)
resnet2 = lt.resnet_module(net, pool1, "resnet2", 16, 3, 1, 1, 8, 16,
use_batch_norm, train)
resnet3 = lt.resnet_module(net, resnet2, "resnet3", 16, 3, 1, 1, 8, 16,
use_batch_norm, train)
resnet4 = lt.resnet_module(net, resnet3, "resnet4", 16, 3, 1, 1, 8, 32,
use_batch_norm, train)
resnet5 = lt.resnet_module(net, resnet4, "resnet5", 32, 3, 1, 1, 8, 32,
use_batch_norm, train)
resnet6 = lt.resnet_module(net, resnet5, "resnet6", 32, 3, 1, 1, 8, 32,
use_batch_norm, train)
resnet7 = lt.resnet_module(net, resnet6, "resnet7", 32, 3, 1, 1, 16, 64,
use_batch_norm, train)
resnet8 = lt.resnet_module(net, resnet7, "resnet8", 64, 3, 1, 1, 16, 64,
use_batch_norm, train)
resnet9 = lt.resnet_module(net, resnet8, "resnet9", 64, 3, 1, 1, 16, 64,
use_batch_norm, train)
resnet10 = lt.resnet_module(net, resnet9, "resnet10", 64, 3, 1, 1, 32, 128,
use_batch_norm, train)
net.lastpool = lt.pool_layer(net, resnet10, "lastpool", 7, 1,
P.Pooling.AVE)
lastpool_layer = net.lastpool
if use_dropout:
net.lastpool_dropout = L.Dropout(net.lastpool,
in_place=True,
dropout_param=dict(dropout_ratio=0.5))
lastpool_layer = net.lastpool_dropout
fc2 = lt.final_fully_connect(net, lastpool_layer)
if train:
net.loss = L.SoftmaxWithLoss(fc2, net.label)
net.acc = L.Accuracy(fc2, net.label)
else:
net.probt = L.Softmax(fc2)
net.acc = L.Accuracy(fc2, net.label)
return net
def create_net(lmdb, mean_file, batch_size, include_acc=False):
# 网络规范
net = caffe.NetSpec()
net.data, net.label = L.Data(source=lmdb,
backend=P.Data.LMDB,
batch_size=batch_size,
ntop=2,
transform_param=dict(crop_size=227,
mean_file=mean_file,
mirror=True))
net.conv1 = L.Convolution(
net.data,
num_output=96,
kernel_size=11,
stride=4,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type="gaussian", std=0.01),
bias_filler=dict(type="constant", value=0))
net.relu1 = L.ReLU(net.conv1, in_place=True)
net.norm1 = L.LRN(net.conv1, local_size=5, alpha=0.0001, beta=0.75)
net.pool1 = L.Pooling(net.norm1,
pool=P.Pooling.MAX,
kernel_size=3,
stride=2)
net.conv2 = L.Convolution(
net.pool1,
num_output=256,
pad=2,
kernel_size=5,
group=2,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type="gaussian", std=0.01),
bias_filler=dict(type="constant", value=0.1))
net.relu2 = L.ReLU(net.conv2, in_place=True)
net.norm2 = L.LRN(net.conv2, local_size=5, alpha=0.0001, beta=0.75)
net.pool2 = L.Pooling(net.norm2,
pool=P.Pooling.MAX,
kernel_size=3,
stride=2)
net.conv3 = L.Convolution(
net.pool2,
num_output=384,
pad=1,
kernel_size=3,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type="gaussian", std=0.01),
bias_filler=dict(type="constant", value=0))
net.relu3 = L.ReLU(net.conv3, in_place=True)
net.conv4 = L.Convolution(
net.conv3,
num_output=384,
pad=1,
kernel_size=3,
group=2,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type="gaussian", std=0.01),
bias_filler=dict(type="constant", value=0.1))
net.relu4 = L.ReLU(net.conv4, in_place=True)
net.conv5 = L.Convolution(
net.conv4,
num_output=256,
pad=1,
kernel_size=3,
group=2,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type="gaussian", std=0.01),
bias_filler=dict(type="constant", value=0.1))
net.relu5 = L.ReLU(net.conv5, in_place=True)
net.pool5 = L.Pooling(net.conv5,
pool=P.Pooling.MAX,
kernel_size=3,
stride=2)
net.fc6 = L.InnerProduct(
net.pool5,
num_output=4096,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type="gaussian", std=0.005),
bias_filler=dict(type="constant", value=0.1))
net.relu6 = L.ReLU(net.fc6, in_place=True)
net.drop6 = L.Dropout(net.fc6, dropout_ratio=0.5, in_place=True)
net.fc7 = L.InnerProduct(
net.fc6,
num_output=4096,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type="gaussian", std=0.005),
bias_filler=dict(type="constant", value=0.1))
net.relu7 = L.ReLU(net.fc7, in_place=True)
net.drop7 = L.Dropout(net.fc7, dropout_ratio=0.5, in_place=True)
net.fc8 = L.InnerProduct(
net.fc7,
num_output=1000,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type="gaussian", std=0.01),
bias_filler=dict(type="constant", value=0.1))
net.loss = L.SoftmaxWithLoss(net.fc8, net.label)
if include_acc:
net.acc = L.Accuracy(net.fc8, net.label)
return net.to_proto()
return net.to_proto()
def mobilenet(data, label=None, num_classes=2):
"""Returns a NetSpec specifying MobileNet."""
n = caffe.NetSpec()
n["data"] = data
# conv1
n["conv1"], n["conv1/bn"], n["conv1/scale"], n["relu1"] = conv(
n["data"], 32)
# depthwise conv2_1
n["conv2_1/dw"], n["conv2_1/dw/bn"], n["conv2_1/dw/scale"], n[
"relu2_1/dw"] = conv_dw(n["relu1"], 32)
# pointwise conv2_1
n["conv2_1/sep"], n["conv2_1/sep/bn"], n["conv2_1/sep/scale"], n[
"relu2_1/sep"] = conv_pw(n["relu2_1/dw"], 64)
# depthwise conv2_2
n["conv2_2/dw"], n["conv2_2/dw/bn"], n["conv2_2/dw/scale"], n[
"relu2_2/dw"] = conv_dw(n["relu2_1/sep"], 64, 2)
# pointwise conv2_2
n["conv2_2/sep"], n["conv2_2/sep/bn"], n["conv2_2/sep/scale"], n[
"relu2_2/sep"] = conv_pw(n["relu2_2/dw"], 128)
# depthwise conv3_1
n["conv3_1/dw"], n["conv3_1/dw/bn"], n["conv3_1/dw/scale"], n[
"relu3_1/dw"] = conv_dw(n["relu2_2/sep"], 128)
# pointwise conv3_1
n["conv3_1/sep"], n["conv3_1/sep/bn"], n["conv3_1/sep/scale"], n[
"relu3_1/sep"] = conv_pw(n["relu3_1/dw"], 128)
# depthwise conv3_2
n["conv3_2/dw"], n["conv3_2/dw/bn"], n["conv3_2/dw/scale"], n[
"relu3_2/dw"] = conv_dw(n["relu3_1/sep"], 128, 2)
# pointwise conv3_2
n["conv3_2/sep"], n["conv3_2/sep/bn"], n["conv3_2/sep/scale"], n[
"relu3_2/sep"] = conv_pw(n["relu3_2/dw"], 256)
# depthwise conv4_1
n["conv4_1/dw"], n["conv4_1/dw/bn"], n["conv4_1/dw/scale"], n[
"relu4_1/dw"] = conv_dw(n["relu3_2/sep"], 256)
# pointwise conv4_1
n["conv4_1/sep"], n["conv4_1/sep/bn"], n["conv4_1/sep/scale"], n[
"relu4_1/sep"] = conv_pw(n["relu4_1/dw"], 256)
# depthwise conv4_2
n["conv4_2/dw"], n["conv4_2/dw/bn"], n["conv4_2/dw/scale"], n[
"relu4_2/dw"] = conv_dw(n["relu4_1/sep"], 256, 2)
# pointwise conv4_2
n["conv4_2/sep"], n["conv4_2/sep/bn"], n["conv4_2/sep/scale"], n[
"relu4_2/sep"] = conv_pw(n["relu4_2/dw"], 512)
# depthwise conv5_1
n["conv5_1/dw"], n["conv5_1/dw/bn"], n["conv5_1/dw/scale"], n[
"relu5_1/dw"] = conv_dw(n["relu4_2/sep"], 512)
# pointwise conv5_1
n["conv5_1/sep"], n["conv5_1/sep/bn"], n["conv5_1/sep/scale"], n[
"relu5_1/sep"] = conv_pw(n["relu5_1/dw"], 512)
# depthwise conv5_2
n["conv5_2/dw"], n["conv5_2/dw/bn"], n["conv5_2/dw/scale"], n[
"relu5_2/dw"] = conv_dw(n["relu5_1/sep"], 512)
# pointwise conv5_2
n["conv5_2/sep"], n["conv5_2/sep/bn"], n["conv5_2/sep/scale"], n[
"relu5_2/sep"] = conv_pw(n["relu5_2/dw"], 512)
# depthwise conv5_3
n["conv5_3/dw"], n["conv5_3/dw/bn"], n["conv5_3/dw/scale"], n[
"relu5_3/dw"] = conv_dw(n["relu5_2/sep"], 512)
# pointwise conv5_3
n["conv5_3/sep"], n["conv5_3/sep/bn"], n["conv5_3/sep/scale"], n[
"relu5_3/sep"] = conv_pw(n["relu5_3/dw"], 512)
# depthwise conv5_4
n["conv5_4/dw"], n["conv5_4/dw/bn"], n["conv5_4/dw/scale"], n[
"relu5_4/dw"] = conv_dw(n["relu5_3/sep"], 512)
# pointwise conv5_4
n["conv5_4/sep"], n["conv5_4/sep/bn"], n["conv5_4/sep/scale"], n[
"relu5_4/sep"] = conv_pw(n["relu5_4/dw"], 512)
# depthwise conv5_5
n["conv5_5/dw"], n["conv5_5/dw/bn"], n["conv5_5/dw/scale"], n[
"relu5_5/dw"] = conv_dw(n["relu5_4/sep"], 512)
# pointwise conv5_5
n["conv5_5/sep"], n["conv5_5/sep/bn"], n["conv5_5/sep/scale"], n[
"relu5_5/sep"] = conv_pw(n["relu5_5/dw"], 512)
# depthwise conv5_6
n["conv5_6/dw"], n["conv5_6/dw/bn"], n["conv5_6/dw/scale"], n[
"relu5_6/dw"] = conv_dw(n["relu5_5/sep"], 512, 2)
# pointwise conv5_6
n["conv5_6/sep"], n["conv5_6/sep/bn"], n["conv5_6/sep/scale"], n[
"relu5_6/sep"] = conv_pw(n["relu5_6/dw"], 1024)
# depthwise conv6
n["conv6/dw"], n["conv6/dw/bn"], n["conv6/dw/scale"], n[
"relu6/dw"] = conv_dw(n["relu5_6/sep"], 1024)
# pointwise conv6
n["conv6/sep"], n["conv6/sep/bn"], n["conv6/sep/scale"], n[
"relu6/sep"] = conv_pw(n["relu6/dw"], 1024)
# pool6
n["pool6"] = L.Pooling(n["relu6/sep"],
global_pooling=True,
pool=P.Pooling.AVE)
# fc7_ft
n["fc7_ft"] = L.Convolution(
n["pool6"],
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
kernel_size=1,
num_output=num_classes,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant', value=0))
if label is not None:
n["label"] = label
n["loss"] = L.SoftmaxWithLoss(n["fc7_ft"], n["label"])
n["acc"] = L.Accuracy(n["fc7_ft"], n["label"])
else:
n["prob"] = L.Softmax(n["fc7_ft"])
return n.to_proto()
def accuracy(self, bottom_data, bottom_label, axis=1, ignore_label=-1):
return L.Accuracy(bottom_data,
bottom_label,
axis=axis,
ignore_label=ignore_label)
def resnet_layers_proto(self,
batch_size,
phase='TRAIN',
stages=(3, 4, 6, 3)):
n = caffe.NetSpec()
if phase == 'TRAIN':
source_data = self.train_data
need_mirror = True
else:
source_data = self.test_data
need_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=224,
mean_value=[128, 128, 128],
mirror=need_mirror))
n.conv1, n.conv1_bn, n.conv1_scale, n.conv1_relu = \
block_conv_bn_scale_relu( n.data, num_output = 64, kernel_size = 7, stride = 2, pad = 3 ) # 64x112x112
n.pool1 = L.Pooling(n.conv1,
kernel_size=3,
stride=2,
pool=P.Pooling.MAX)
residual_num = 0
for num in xrange(len(stages)):
for i in xrange(stages[num]):
residual_num = residual_num + 1
if num == 0 and i == 0:
stage_string = skip_connect_with_dimen_match_no_patch_reduce
if residual_num == 1:
bottom_string = 'n.pool1'
else:
bottom_string = 'n.res%s_eletwise' % (
str(residual_num - 1))
elif i == 0 and num > 0:
stage_string = skip_connect_with_dimen_match
if residual_num == 1:
bottom_string = 'n.pool1'
else:
bottom_string = 'n.res%s_eletwise' % (
str(residual_num - 1))
else:
stage_string = skip_connect_no_dimen_match
bottom_string = 'n.res%s_eletwise' % (str(residual_num -
1))
exec(
stage_string.replace('(stage)', str(residual_num)).replace(
'(bottom)',
bottom_string).replace('(num)', str(2**num * 64)))
exec 'n.pool5 = L.Pooling( bottom_string, kernel_size=7, stride=1, pool=P.Pooling.AVE)'.replace(
'bottom_string', 'n.res%s_eletwise' % str(residual_num))
n.classifier = L.InnerProduct(n.pool5,
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 cnn(split):
n = caffe.NetSpec()
pydata_params = dict(dataset_dir='/home/kevin/dataset/processed_data3',
variable='depth_map',
split=split,
mean=(2),
seed=1337,
batch_size=256,
frame_num=30,
img_size=(227, 227))
if split == 'deploy':
n.img = L.Input(
name='input',
ntop=2,
shape=[dict(dim=1),
dict(dim=1),
dict(dim=227),
dict(dim=227)])
else:
if split is 'train':
pydata_params['dtype'] = 'frame'
pylayer = 'ModelNetDataLayer'
else:
pydata_params['dtype'] = 'object'
pylayer = 'ModelNetDataLayer'
n.img, n.label = L.Python(module='data_layers.model_net_layer',
layer=pylayer,
ntop=2,
param_str=str(pydata_params))
# the base net
n.conv1, n.relu1 = conv_relu("conv1", n.img, 96, ks=11, stride=4, pad=0)
n.pool1 = max_pool(n.relu1, ks=3)
n.norm1 = L.LRN(n.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, pad=2, group=2)
n.pool2 = max_pool(n.relu2, ks=3)
n.norm2 = L.LRN(n.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)
n.conv4, n.relu4 = conv_relu("conv4", n.relu3, 384, ks=3, pad=1, group=2)
n.conv5, n.relu5 = conv_relu("conv5", n.relu4, 256, ks=3, pad=1, group=2)
n.pool5 = max_pool(n.relu5, ks=3)
n.fc6, n.relu6 = fc_relu(n.pool5, 4096, lr1=1, lr2=2)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7, n.relu7 = fc_relu(n.drop6, 4096, lr1=1, lr2=2)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.fc8 = fc(n.drop7, 40, lr1=1, lr2=2)
if split != 'deploy':
n.accuracy = L.Accuracy(n.fc8, n.label)
n.loss = L.SoftmaxWithLoss(n.fc8, n.label)
#n.loss = L.Python(n.fc8, n.label, loss_weight=1, module='nn_layers.max_softmax_loss_layer', layer='MaxSoftmaxLossLayer')
# n.display = L.Scale(n.corr, param=[dict(lr_mult=0)], filler=dict(type='constant',value=1.0))
# n.fc9_bn = L.BatchNorm(n.relu9, param=[dict(lr_mult=0),dict(lr_mult=0),dict(lr_mult=0)], batch_norm_param=dict(use_global_stats=True))
return n.to_proto()
def cnn(split):
n = caffe.NetSpec()
pydata_params = dict(dataset_dir='/home/kevin/dataset/rgbd',
split=split,
mean=(104.00698793, 116.66876762, 122.67891434),
seed=1337,
img_size=(224, 224),
crop_size=(224, 224, 224, 224))
if split == 'train':
pylayer = 'RGBDDataLayer'
pydata_params['randomize'] = True
pydata_params['batch_size'] = 64
elif split == 'test':
pylayer = 'RGBDDataLayer'
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.rgbd_data_layer',
layer=pylayer,
ntop=3,
param_str=str(pydata_params))
#---------------------------------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=1, lr2=2)
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=1, lr2=2)
n.rgb_drop7 = L.Dropout(n.rgb_relu7, dropout_ratio=0.5, in_place=True)
# n.rgb_fc8 = fc(n.rgb_drop7, 11, lr1=1, lr2=2)
#---------------------------------Depth-Net---------------------------------------#
# the base net
n.conv1, n.relu1 = conv_relu("conv1",
n.depth,
128,
ks=5,
stride=2,
pad=2,
lr1=0,
lr2=0)
n.depth_pool1 = max_pool(n.relu1, ks=3)
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=0,
lr2=0)
n.depth_pool2 = max_pool(n.relu2, ks=3)
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=0,
lr2=0)
n.depth_pool3 = max_pool(n.relu3, ks=3)
n.conv4, n.relu4 = conv_relu("conv4",
n.depth_pool3,
512,
ks=3,
pad=1,
group=1,
lr1=0,
lr2=0)
n.conv5, n.relu5 = conv_relu("conv5",
n.relu4,
512,
ks=3,
pad=1,
group=1,
lr1=0,
lr2=0)
n.depth_pool5 = max_pool(n.relu5, ks=3)
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, 11, lr1=1, lr2=2)
#-----------------------------------final output---------------------------------#
# Concatenation
n.concat = L.Concat(n.rgb_drop7, n.depth_drop7, axis=1)
n.fuse_fc1, n.fuse_relu1 = fc_relu(n.concat, 1024, lr1=1, lr2=2)
n.fuse_drop1 = L.Dropout(n.fuse_fc1, dropout_ratio=0.5, in_place=True)
n.fuse_fc2, n.fuse_relu2 = fc_relu(n.fuse_drop1, 1024, lr1=1, lr2=2)
n.fuse_drop2 = L.Dropout(n.fuse_fc2, dropout_ratio=0.5, in_place=True)
n.rgbd_fc8 = fc(n.fuse_drop2, 11, 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.rgbd_accuracy = L.Accuracy(n.rgbd_fc8, n.label)
n.rgbd_loss = L.SoftmaxWithLoss(n.rgbd_fc8, n.label)
return n.to_proto()
def segnet_network(data_source, label_source, mode='train'):
""" Builds a Caffe Network Definition object for SegNet
Args:
data_source (str): path to the data LMDB
label_source (str): path to the label LMDB
mode (str, optional): 'train', 'test' or 'deploy' (defaults to 'train')
Returns:
obj: SegNet (Caffe Network Definition object)
"""
n = caffe.NetSpec()
if MEAN_PIXEL is None:
transform_param = {}
else:
transform_param = {'mean_value': MEAN_PIXEL}
if mode == 'deploy':
n.data = L.Input(input_param={ 'shape':\
{ 'dim': [BATCH_SIZE, 3, test_patch_size[0], test_patch_size[1]] }
})
else:
n.data = L.Data(batch_size=BATCH_SIZE, backend=P.Data.LMDB,\
transform_param=transform_param, source=data_source)
n.label = L.Data(batch_size=BATCH_SIZE, backend=P.Data.LMDB, source=label_source)
convolution_block(n, n.data, "conv1_{}", 2, planes=(64,64,64), lr_mult=0.5)
n.pool1, n.pool1_mask = L.Pooling(n.conv1_2, pool=P.Pooling.MAX, kernel_size=2, stride=2, ntop=2)
convolution_block(n, n.pool1, "conv2_{}", 2, planes=(128,128,128), lr_mult=0.5)
n.pool2, n.pool2_mask = L.Pooling(n.conv2_2, pool=P.Pooling.MAX, kernel_size=2, stride=2, ntop=2)
convolution_block(n, n.pool2, "conv3_{}", 3, planes=(256,256,256), lr_mult=0.5)
n.pool3, n.pool3_mask = L.Pooling(n.conv3_3, pool=P.Pooling.MAX, kernel_size=2, stride=2, ntop=2)
convolution_block(n, n.pool3, "conv4_{}", 3, planes=(512,512,512), lr_mult=0.5)
n.pool4, n.pool4_mask = L.Pooling(n.conv4_3, pool=P.Pooling.MAX, kernel_size=2, stride=2, ntop=2)
convolution_block(n, n.pool4, "conv5_{}", 3, planes=(512,512,512), lr_mult=0.5)
n.pool5, n.pool5_mask = L.Pooling(n.conv5_3, pool=P.Pooling.MAX, kernel_size=2, stride=2, ntop=2)
n.upsample5 = L.Upsample(n.pool5, n.pool5_mask, scale=2)
convolution_block(n, n.upsample5, "conv5_{}_D", 3, planes=(512,512,512), lr_mult=1, reverse=True)
n.upsample4 = L.Upsample(n.conv5_1_D, n.pool4_mask, scale=2)
convolution_block(n, n.upsample4, "conv4_{}_D", 3, planes=(512,512,256), lr_mult=1, reverse=True)
n.upsample3 = L.Upsample(n.conv4_1_D, n.pool3_mask, scale=2)
convolution_block(n, n.upsample3, "conv3_{}_D", 3, planes=(256,256,128), lr_mult=1, reverse=True)
n.upsample2 = L.Upsample(n.conv3_1_D, n.pool2_mask, scale=2)
convolution_block(n, n.upsample2, "conv2_{}_D", 2, planes=(128,128,64), lr_mult=1, reverse=True)
n.upsample1 = L.Upsample(n.conv2_1_D, n.pool1_mask, scale=2)
n.conv1_2_D, n.conv1_2_D_bn, n.conv1_2_D_scale, n.conv1_2_D_relu =\
convolution_unit(n.upsample1, 3, 1, 64, lr_mult=1)
n.conv1_1_D, _, _, _ = convolution_unit(n.conv1_2_D, 3, 1, 6, lr_mult=1)
if mode == 'train' or mode == 'test':
n.loss = L.SoftmaxWithLoss(n.conv1_1_D, n.label, loss_param={'ignore_label': IGNORE_LABEL})
n.accuracy = L.Accuracy(n.conv1_1_D, n.label)
return n
def ZFNetBody(net, from_layer, for_training=True):
net.conv1 = L.Convolution(
net[from_layer],
kernel_size=k_conv1,
stride=s_conv1,
num_output=d_conv1,
pad=p_conv1,
bias_term=True,
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', std=0),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
net.relu1 = L.ReLU(net.conv1, in_place=True)
net.pool1 = L.Pooling(net.relu1,
pool=P.Pooling.MAX,
kernel_size=k_pool1,
stride=s_pool1)
net.norm1 = L.LRN(net.pool1,
lrn_param=dict(local_size=local_size_norm1,
alpha=alpha_norm1,
beta=beta_norm1))
net.conv2 = L.Convolution(
net.norm1,
kernel_size=k_conv2,
stride=s_conv2,
num_output=d_conv2, #pad=p_conv2,
bias_term=True,
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', std=0),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
net.relu2 = L.ReLU(net.conv2, in_place=True)
net.pool2 = L.Pooling(net.relu2,
pool=P.Pooling.MAX,
kernel_size=k_pool2,
stride=s_pool2)
net.norm2 = L.LRN(net.pool2,
lrn_param=dict(local_size=local_size_norm2,
alpha=alpha_norm2,
beta=beta_norm2))
net.conv3 = L.Convolution(
net.norm2,
kernel_size=k_conv3,
stride=s_conv3,
num_output=d_conv3,
pad=p_conv3,
bias_term=True,
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', std=0),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
net.relu3 = L.ReLU(net.conv3, in_place=True)
net.conv4 = L.Convolution(
net.relu3,
kernel_size=k_conv4,
stride=s_conv4,
num_output=d_conv4,
pad=p_conv4,
bias_term=True,
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', std=0),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
net.relu4 = L.ReLU(net.conv4, in_place=True)
net.conv5 = L.Convolution(
net.relu4,
kernel_size=k_conv5,
stride=s_conv5,
num_output=d_conv5,
pad=p_conv5,
bias_term=True,
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', std=0),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
net.relu5 = L.ReLU(net.conv5, in_place=True)
net.pool5 = L.Pooling(net.relu5,
pool=P.Pooling.MAX,
kernel_size=k_pool5,
stride=s_pool5)
net.fc6 = L.InnerProduct(
net.pool5,
num_output=k_ip6,
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', std=0),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
net.relu6 = L.ReLU(net.fc6, in_place=True)
net.drop6 = L.Dropout(net.relu6,
dropout_param=dict(dropout_ratio=r_drop6),
in_place=True)
net.fc7 = L.InnerProduct(
net.fc6,
num_output=k_ip7,
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', std=0),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
net.relu7 = L.ReLU(net.fc7, in_place=True)
net.drop7 = L.Dropout(net.relu7,
dropout_param=dict(dropout_ratio=r_drop7),
in_place=True)
net.fc8 = L.InnerProduct(
net.fc7,
num_output=k_ip8,
weight_filler=dict(type='gaussian', std=0.01),
bias_filler=dict(type='constant', std=0),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
if not for_training:
net.acc = L.Accuracy(net.fc8,
net.label,
include=dict(phase=caffe_pb2.Phase.Value('TEST')))
net.loss = L.SoftmaxWithLoss(net.fc8, net.label)
return net
def shuffle_net(group,
scale_f,
input_size,
se=False,
num_classes=1000,
asoft=True):
# figure out network structure
group_defs = {
1: [36, 72, 144],
2: [50, 100, 200],
3: [60, 120, 240],
4: [68, 136, 272],
8: [96, 192, 384],
}
nouts_list = [int(v * scale_f) for v in group_defs[group]]
nunits_list = [3, 7, 3]
f_size = 24
# setup the first couple of layers
n = caffe.NetSpec()
net = n.__dict__['tops']
n.data, n.label = L.ImageData(batch_size=128,
source="../data/train.list",
root_folder="/",
ntop=2,
include=dict(phase=0),
transform_param=dict(crop_size=input_size,
mirror=True,
scale=1 / 128.))
# The data mean
n.conv1 = L.Convolution(n.data,
kernel_size=3,
stride=2,
num_output=f_size,
pad=1,
bias_term=False,
param=[dict(lr_mult=1, decay_mult=1)],
weight_filler=dict(type="msra"))
n.conv1_bn = L.BatchNorm(
n.conv1,
param=[dict(lr_mult=0),
dict(lr_mult=0),
dict(lr_mult=0)],
in_place=False)
n.conv1_scale = L.Scale(
n.conv1_bn,
scale_param=dict(bias_term=True),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=1)],
in_place=True)
n.conv1_relu = L.ReLU(n.conv1_scale, in_place=True)
n.conv1_pool = L.Pooling(n.conv1_relu, stride=2, kernel_size=3)
# make the convolutional body
last_size = f_size / 4
for i, (nout, nunit) in enumerate(zip(nouts_list, nunits_list)):
s = 'Step' + str(i + 1) + '_reduction_'
if i == 0:
standard_unit(n,
nout - last_size,
s,
group,
se=se,
newdepth=True,
is_first=True)
else:
standard_unit(n, nout - last_size, s, group, se=se, newdepth=True)
last_size = nout
for unit in range(nunit):
s = 'Step' + str(i + 1) + '_' + str(unit + 1) + '_'
standard_unit(n, nout, s, group, se=se)
# add the end layers
net = n.__dict__['tops']
bottom = net[list(net.keys())[-1]] #find the last layer in netspec
n.global_pool = L.Pooling(bottom,
pooling_param=dict(pool=1, global_pooling=True))
n.score = L.InnerProduct(n.global_pool,
num_output=num_classes,
bias_term=False,
param=[dict(lr_mult=1, decay_mult=1)],
weight_filler=dict(type="msra"))
n.loss = L.SoftmaxWithLoss(n.score, n.label)
n.accuracy = L.Accuracy(n.score, n.label)
return n
def vgg_16(lmdb, bs_train=32, bs_val=10, lmdb_flag=True, not_deploy=True):
n = caffe.NetSpec()
if not_deploy:
if lmdb_flag:
n.data, n.label = L.Data(source=lmdb + 'cub200_2011_train_lmdb',
backend=P.Data.LMDB,
include=dict(phase=caffe_pb2.TRAIN),
batch_size=bs_train,
ntop=2,
transform_param=dict(
crop_size=224,
mean_value=[110, 127, 123],
mirror=True))
data_str = n.to_proto()
n.data, n.label = L.Data(source=lmdb + 'cub200_2011_val_lmdb',
backend=P.Data.LMDB,
include=dict(phase=caffe_pb2.TEST),
batch_size=bs_val,
ntop=2,
transform_param=dict(
crop_size=224,
mean_value=[110, 127, 123],
mirror=False))
else:
n.data, n.label = L.Data(source=lmdb + 'cub200_2011_train_leveldb',
backend=P.Data.LEVELDB,
include=dict(phase=caffe_pb2.TRAIN),
batch_size=bs_train,
ntop=2,
transform_param=dict(
crop_size=224,
mean_value=[110, 127, 123],
mirror=True))
data_str = n.to_proto()
n.data, n.label = L.Data(source=lmdb + 'cub200_2011_val_leveldb',
backend=P.Data.LEVELDB,
include=dict(phase=caffe_pb2.TEST),
batch_size=bs_val,
ntop=2,
transform_param=dict(
crop_size=224,
mean_value=[110, 127, 123],
mirror=False))
else:
data_str = 'input: "data"\ninput_dim: 1\ninput_dim: 3\ninput_dim: 224\ninput_dim: 224'
n.data = L.Data()
# the net itself
n.conv1_1, n.relu1_1 = conv_relu(n.data, nout=64, pad=1, ks=3)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, nout=64, pad=1, ks=3)
n.pool1 = max_pool(n.relu1_2, ks=2, stride=2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, nout=128, pad=1, ks=3)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, nout=128, pad=1, ks=3)
n.pool2 = max_pool(n.relu2_2, ks=2, stride=2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, nout=256, pad=1, ks=3)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, nout=256, pad=1, ks=3)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, nout=256, pad=1, ks=3)
n.pool3 = max_pool(n.relu3_3, ks=2, stride=2)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, nout=512, pad=1, ks=3)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, nout=512, pad=1, ks=3)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, nout=512, pad=1, ks=3)
n.pool4 = max_pool(n.relu4_3, ks=2, stride=2)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, nout=512, pad=1, ks=3)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, nout=512, pad=1, ks=3)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, nout=512, pad=1, ks=3)
n.pool5 = ave_pool(n.relu5_3, ks=14, stride=1)
n.softmax = L.Convolution(n.pool5,
kernel_size=1,
num_output=200,
param=[
dict(lr_mult=10, decay_mult=10),
dict(lr_mult=20, decay_mult=0)
])
if not_deploy:
n.loss = L.SoftmaxWithLoss(n.softmax, n.label)
n.acc_top_1 = L.Accuracy(n.softmax, n.label, top_k=1)
else:
n.prob = L.Softmax(n.softmax)
model_str = str(n.to_proto())
if not not_deploy:
model_str = model_str[54:-1]
return str(data_str) + '\n' + model_str
def pj_x(mode, batchsize, T, exp_T, question_vocab_size, exp_vocab_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode':mode, 'batchsize':batchsize})
n.data, n.cont, n.img_feature, n.label, n.exp, n.exp_out, n.exp_cont_1, n.exp_cont_2 = \
L.Python(module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=8)
n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08), param=fixed_weights)
n.embed = L.TanH(n.embed_ba)
n.exp_embed_ba = L.Embed(n.exp, input_dim=exp_vocab_size, num_output=300, \
weight_filler=dict(type='uniform', min=-0.08, max=0.08))
n.exp_embed = L.TanH(n.exp_embed_ba)
# LSTM1
n.lstm1 = L.LSTM(\
n.embed, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)),
param=fixed_weights_lstm)
tops1 = L.Slice(n.lstm1, ntop=T, slice_param={'axis':0})
for i in range(T-1):
n.__setattr__('slice_first'+str(i), tops1[int(i)])
n.__setattr__('silence_data_first'+str(i), L.Silence(tops1[int(i)],ntop=0))
n.lstm1_out = tops1[T-1]
n.lstm1_reshaped = L.Reshape(n.lstm1_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm1_reshaped_droped = L.Dropout(n.lstm1_reshaped,dropout_param={'dropout_ratio':0.3})
n.lstm1_droped = L.Dropout(n.lstm1,dropout_param={'dropout_ratio':0.3})
# LSTM2
n.lstm2 = L.LSTM(\
n.lstm1_droped, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)),
param=fixed_weights_lstm)
tops2 = L.Slice(n.lstm2, ntop=T, slice_param={'axis':0})
for i in range(T-1):
n.__setattr__('slice_second'+str(i), tops2[int(i)])
n.__setattr__('silence_data_second'+str(i), L.Silence(tops2[int(i)],ntop=0))
n.lstm2_out = tops2[T-1]
n.lstm2_reshaped = L.Reshape(n.lstm2_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm2_reshaped_droped = L.Dropout(n.lstm2_reshaped,dropout_param={'dropout_ratio':0.3})
concat_botom = [n.lstm1_reshaped_droped, n.lstm2_reshaped_droped]
n.lstm_12 = L.Concat(*concat_botom)
# Tile question feature
n.q_emb_resh = L.Reshape(n.lstm_12, reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.q_emb_tiled_1 = L.Tile(n.q_emb_resh, axis=2, tiles=14)
n.q_emb_resh_tiled = L.Tile(n.q_emb_tiled_1, axis=3, tiles=14)
# Embed image feature
n.i_emb = L.Convolution(n.img_feature, kernel_size=1, stride=1,
num_output=2048, pad=0, weight_filler=dict(type='xavier'),
param=fixed_weights)
# Eltwise product and normalization
n.eltwise = L.Eltwise(n.q_emb_resh_tiled, n.i_emb, eltwise_param={'operation': P.Eltwise.PROD})
n.eltwise_sqrt = L.SignedSqrt(n.eltwise)
n.eltwise_l2 = L.L2Normalize(n.eltwise_sqrt)
n.eltwise_drop = L.Dropout(n.eltwise_l2, dropout_param={'dropout_ratio': 0.3})
# Attention for VQA
n.att_conv1 = L.Convolution(n.eltwise_drop, kernel_size=1, stride=1, num_output=512, pad=0, weight_filler=dict(type='xavier'), param=fixed_weights)
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu, kernel_size=1, stride=1, num_output=1, pad=0, weight_filler=dict(type='xavier'), param=fixed_weights)
n.att_reshaped = L.Reshape(n.att_conv2,reshape_param=dict(shape=dict(dim=[-1,1,14*14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att_map = L.Reshape(n.att_softmax,reshape_param=dict(shape=dict(dim=[-1,1,14,14])))
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]), data_filler=dict(type='constant', value=1), ntop=1)
n.att_feature = L.SoftAttention(n.img_feature, n.att_map, dummy)
n.att_feature_resh = L.Reshape(n.att_feature, reshape_param=dict(shape=dict(dim=[-1,2048])))
# eltwise product + normalization again for VQA
n.i_emb2 = L.InnerProduct(n.att_feature_resh, num_output=2048, weight_filler=dict(type='xavier'), param=fixed_weights)
n.eltwise2 = L.Eltwise(n.lstm_12, n.i_emb2, eltwise_param={'operation': P.Eltwise.PROD})
n.eltwise2_sqrt = L.SignedSqrt(n.eltwise2)
n.eltwise2_l2 = L.L2Normalize(n.eltwise2_sqrt)
n.eltwise2_drop = L.Dropout(n.eltwise2_l2, dropout_param={'dropout_ratio': 0.3})
n.prediction = L.InnerProduct(n.eltwise2_drop, num_output=3000, weight_filler=dict(type='xavier'), param=fixed_weights)
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
# Embed VQA GT answer during training
n.exp_emb_ans = L.Embed(n.label, input_dim=3000, num_output=300, \
weight_filler=dict(type='uniform', min=-0.08, max=0.08))
n.exp_emb_ans_tanh = L.TanH(n.exp_emb_ans)
n.exp_emb_ans2 = L.InnerProduct(n.exp_emb_ans_tanh, num_output=2048, weight_filler=dict(type='xavier'))
# Merge VQA answer and visual+textual feature
n.exp_emb_resh = L.Reshape(n.exp_emb_ans2, reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.exp_emb_tiled_1 = L.Tile(n.exp_emb_resh, axis=2, tiles=14)
n.exp_emb_tiled = L.Tile(n.exp_emb_tiled_1, axis=3, tiles=14)
n.eltwise_emb = L.Convolution(n.eltwise, kernel_size=1, stride=1, num_output=2048, pad=0, weight_filler=dict(type='xavier'))
n.exp_eltwise = L.Eltwise(n.eltwise_emb, n.exp_emb_tiled, eltwise_param={'operation': P.Eltwise.PROD})
n.exp_eltwise_sqrt = L.SignedSqrt(n.exp_eltwise)
n.exp_eltwise_l2 = L.L2Normalize(n.exp_eltwise_sqrt)
n.exp_eltwise_drop = L.Dropout(n.exp_eltwise_l2, dropout_param={'dropout_ratio': 0.3})
# Attention for Explanation
n.exp_att_conv1 = L.Convolution(n.exp_eltwise_drop, kernel_size=1, stride=1, num_output=512, pad=0, weight_filler=dict(type='xavier'))
n.exp_att_conv1_relu = L.ReLU(n.exp_att_conv1)
n.exp_att_conv2 = L.Convolution(n.exp_att_conv1_relu, kernel_size=1, stride=1, num_output=1, pad=0, weight_filler=dict(type='xavier'))
n.exp_att_reshaped = L.Reshape(n.exp_att_conv2,reshape_param=dict(shape=dict(dim=[-1,1,14*14])))
n.exp_att_softmax = L.Softmax(n.exp_att_reshaped, axis=2)
n.exp_att_map = L.Reshape(n.exp_att_softmax,reshape_param=dict(shape=dict(dim=[-1,1,14,14])))
exp_dummy = L.DummyData(shape=dict(dim=[batchsize, 1]), data_filler=dict(type='constant', value=1), ntop=1)
n.exp_att_feature_prev = L.SoftAttention(n.img_feature, n.exp_att_map, exp_dummy)
n.exp_att_feature_resh = L.Reshape(n.exp_att_feature_prev, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.exp_att_feature_embed = L.InnerProduct(n.exp_att_feature_resh, num_output=2048, weight_filler=dict(type='xavier'))
n.exp_lstm12_embed = L.InnerProduct(n.lstm_12, num_output=2048, weight_filler=dict(type='xavier'))
n.exp_eltwise2 = L.Eltwise(n.exp_lstm12_embed, n.exp_att_feature_embed, eltwise_param={'operation': P.Eltwise.PROD})
n.exp_att_feature = L.Eltwise(n.exp_emb_ans2, n.exp_eltwise2, eltwise_param={'operation': P.Eltwise.PROD})
# LSTM1 for Explanation
n.exp_lstm1 = L.LSTM(\
n.exp_embed, n.exp_cont_1,\
recurrent_param=dict(\
num_output=2048,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
n.exp_lstm1_dropped = L.Dropout(n.exp_lstm1,dropout_param={'dropout_ratio':0.3})
# merge with LSTM1 for explanation
n.exp_att_resh = L.Reshape(n.exp_att_feature, reshape_param=dict(shape=dict(dim=[1, -1, 2048])))
n.exp_att_tiled = L.Tile(n.exp_att_resh, axis=0, tiles=exp_T)
n.exp_eltwise_all = L.Eltwise(n.exp_lstm1_dropped, n.exp_att_tiled, eltwise_param={'operation': P.Eltwise.PROD})
n.exp_eltwise_all_sqrt = L.SignedSqrt(n.exp_eltwise_all)
n.exp_eltwise_all_l2 = L.L2Normalize(n.exp_eltwise_all_sqrt)
n.exp_eltwise_all_drop = L.Dropout(n.exp_eltwise_all_l2, dropout_param={'dropout_ratio': 0.3})
# LSTM2 for Explanation
n.exp_lstm2 = L.LSTM(\
n.exp_eltwise_all_drop, n.exp_cont_2,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
n.exp_lstm2_dropped = L.Dropout(n.exp_lstm2,dropout_param={'dropout_ratio':0.3})
n.exp_prediction = L.InnerProduct(n.exp_lstm2_dropped, num_output=exp_vocab_size, weight_filler=dict(type='xavier'), axis=2)
n.exp_loss = L.SoftmaxWithLoss(n.exp_prediction, n.exp_out,
loss_param=dict(ignore_label=-1),
softmax_param=dict(axis=2))
n.exp_accuracy = L.Accuracy(n.exp_prediction, n.exp_out, axis=2, ignore_label=-1)
return n.to_proto()
def accuracy(input_blob, label_blob):
return L.Accuracy(input_blob, label_blob, include=dict(phase=caffe.TEST))
def add_accuracy_layer(net, bottom, name):
""" Add accuracy layer """
net[name] = L.Accuracy(bottom[0], bottom[1])
def construct_fcn(image_lmdb, contour_lmdb, batch_size=1, include_acc=False):
net = caffe.NetSpec()
# args for convlution layers
weight_filler = dict(type='gaussian', mean=0.0, 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)]
net.data = L.Data(source=image_lmdb,
backend=P.Data.LMDB,
batch_size=batch_size,
ntop=1,
transform_param=dict(crop_size=0,
mean_value=[77],
mirror=False))
net.label = L.Data(source=contour_lmdb,
backend=P.Data.LMDB,
batch_size=batch_size,
ntop=1)
# conv-relu-pool 1
net.conv1 = L.Convolution(net.data,
kernel_size=5,
stride=2,
num_output=100,
pad=50,
group=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
param=param)
net.relu1 = L.ReLU(net.conv1, in_place=True)
net.pool1 = L.Pooling(net.relu1,
pool=P.Pooling.MAX,
kernel_size=2,
stride=2)
# conv-relu-pool 2
net.conv2 = L.Convolution(net.pool1,
kernel_size=5,
stride=2,
num_output=200,
pad=0,
group=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
param=param)
net.relu2 = L.ReLU(net.conv2, in_place=True)
net.pool2 = L.Pooling(net.relu2,
pool=P.Pooling.MAX,
kernel_size=2,
stride=2)
net.conv3 = L.Convolution(net.pool2,
kernel_size=3,
stride=1,
num_output=300,
pad=0,
group=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
param=param)
net.relu3 = L.ReLU(net.conv3, in_place=True)
net.conv4 = L.Convolution(net.relu3,
kernel_size=3,
stride=1,
num_output=300,
pad=0,
group=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
param=param)
net.relu4 = L.ReLU(net.conv4, in_place=True)
net.drop = L.Dropout(net.relu4, dropout_ratio=0.1, in_place=True)
net.score_classes = L.Convolution(net.drop,
kernel_size=1,
stride=1,
num_output=2,
pad=0,
group=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
param=param)
net.upscore = L.Deconvolution(net.score_classes)
net.score = L.Crop(net.upscore, net.data)
net.loss = L.SoftmaxWithLoss(net.score,
net.label,
loss_param=dict(normalize=True))
if include_acc:
net.accuracy = L.Accuracy(net.score, net.label)
return net.to_proto()
def create_net(phase):
global train_transform_param
global test_transform_param
train_transform_param = {'mirror': False, 'mean_file': Params['mean_file']}
test_transform_param = {'mean_file': Params['mean_file']}
if phase == 'train':
lmdb_file = Params['train_lmdb']
transform_param = train_transform_param
batch_size = Params['batch_size_per_device']
else:
lmdb_file = Params['test_lmdb']
transform_param = test_transform_param
batch_size = Params['test_batch_size']
net = caffe.NetSpec()
if phase == 'test':
net.data, net.label = L.Data(batch_size=batch_size,
backend=P.Data.LMDB,
source=lmdb_file,
transform_param=transform_param,
ntop=2)
elif phase == 'train':
net.data = L.Input(shape=dict(dim=[128, 3, 4, 12]))
net.label = L.Input(shape=dict(dim=[128]))
elif phase == 'deploy':
net.data = L.Input(shape=dict(dim=[1, 3, 4, 12]))
kwargs = {
'param':
[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
'weight_filler': dict(type='gaussian', std=0.0001),
'bias_filler': dict(type='constant')
}
net.conv1 = L.Convolution(net.data, num_output=16, kernel_size=3, **kwargs)
net.pool1 = L.Pooling(net.conv1,
pool=P.Pooling.MAX,
kernel_size=3,
stride=2)
net.relu1 = L.ReLU(net.pool1, in_place=True)
kwargs = {
'param':
[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
'weight_filler': dict(type='gaussian', std=0.005),
'bias_filler': dict(type='constant')
}
net.fc2 = L.InnerProduct(net.pool1, num_output=16, **kwargs)
net.relu2 = L.ReLU(net.fc2, in_place=True)
net.drop2 = L.Dropout(net.fc2,
in_place=True,
dropout_param=dict(dropout_ratio=0.5))
kwargs = {
'param':
[dict(lr_mult=1, decay_mult=100),
dict(lr_mult=2, decay_mult=0)],
'weight_filler': dict(type='gaussian', std=0.01),
'bias_filler': dict(type='constant', value=0)
}
net.fc3 = L.InnerProduct(net.fc2, num_output=2, **kwargs)
if phase == 'train':
net.loss = L.SoftmaxWithLoss(net.fc3, net.label)
elif phase == 'test':
net.accuracy = L.Accuracy(net.fc3, net.label)
else:
net.prob = L.Softmax(net.fc3)
net_proto = net.to_proto()
net_proto.name = '{}_{}'.format(Params['model_name'], phase)
return net_proto
def densenet(mode, data_file, bs, nlayer, nclass, first_nout=16, growth_rate=16, dropout=0.2):
net = caffe.NetSpec()
# data layer ---------------------------------------------------------------
mirror = True
shuffle = True
if mode == 1: # TEST phase
mirror = False
shuffle = False
transform = dict(scale = 0.0078125,
mirror = mirror,
#crop_size = 224,
mean_value = [127.5, 127.5, 127.5])
net.data, net.label = L.Data(#include = dict(phase = mode),
transform_param = transform,
source = data_file,
batch_size = bs,
backend = P.Data.LMDB,
ntop = 2)
# net.data, net.label = L.ImageData(#include = dict(phase = mode),
# transform_param = transform,
# source = data_file,
# batch_size = bs,
# shuffle = shuffle,
# #new_height = 256,
# #new_width = 256,
# #is_color = True,
# ntop = 2)
pre_fmap = 0 # total number of previous feature maps
# first convolution --------------------------------------------------------
net.conv_1 = L.Convolution(net.data, num_output=first_nout,
kernel_size=7, stride=2, pad=3,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'),
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
net.relu_1 = L.PReLU(net.conv_1, in_place=True)
net.pool_1 = L.Pooling(net.relu_1, pool=P.Pooling.MAX,
kernel_size=3, stride=2)
pre_layer = net.pool_1
pre_fmap += first_nout
# DB + TD ------------------------------------------------------------------
# +1 in order to make the index values from 1
for major in xrange(len(nlayer)-1):
# DB
for minor in xrange(nlayer[major]):
pre_layer = cat_layer(net, mode, major+1, minor+1, pre_layer, growth_rate, dropout)
pre_fmap += growth_rate
# TD
pre_layer = transition_down(net, mode, major+1, pre_layer, pre_fmap, dropout)
pre_fmap = pre_fmap // 2
# last DB, without TD
major = len(nlayer)
for minor in xrange(nlayer[-1]):
pre_layer = cat_layer(net, mode, major, minor+1, pre_layer, growth_rate, dropout)
pre_fmap += growth_rate
# final layers -------------------------------------------------------------
use_global_stats = False
if mode == 1: # TEST phase
use_global_stats = True
net.bn_final = L.BatchNorm(pre_layer, in_place=False,
batch_norm_param = dict(use_global_stats=use_global_stats),
param=[dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)])
net.scale_finel = L.Scale(net.bn_final, bias_term=True, in_place=True,
filler=dict(value=1), bias_filler=dict(value=0))
net.relu_final = L.PReLU(net.scale_finel, in_place=True)
net.pool_final = L.Pooling(net.relu_final, pool=P.Pooling.AVE, global_pooling=True)
net.fc_class = L.InnerProduct(net.pool_final, num_output=nclass,
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)])
net.loss = L.SoftmaxWithLoss(net.fc_class, net.label)
if mode == 1:
net.accuracy = L.Accuracy(net.fc_class, net.label)
return str(net.to_proto())
def resnet_layers_proto(self,
batch_size,
phase='TRAIN',
stages=(3, 4, 6, 3)):
"""
:param batch_size: the batch_size of train and test phase
:param phase: TRAIN or TEST
:param stages: the num of layers = 2 + 3*sum(stages), layers would better be chosen from [50, 101, 152]
{every stage is composed of 1 residual_branch_shortcut module and stage[i]-1 residual_branch
modules, each module consists of 3 conv layers}
(3, 4, 6, 3) for 50 layers; (3, 4, 23, 3) for 101 layers; (3, 8, 36, 3) for 152 layers
"""
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=224,
mean_value=[104, 117, 123],
mirror=mirror))
n.conv1, n.conv1_bn, n.conv1_scale, n.conv1_relu = \
conv_bn_scale_relu(n.data, num_output=64, kernel_size=7, stride=2, pad=3) # 64x112x112
n.pool1 = L.Pooling(n.conv1,
kernel_size=3,
stride=2,
pool=P.Pooling.MAX) # 64x56x56
for num in xrange(len(stages)): # num = 0, 1, 2, 3
for i in xrange(stages[num]):
if i == 0:
stage_string = branch_shortcut_string
bottom_string = [
'n.pool1',
'n.res2b%s' % str(stages[0] - 1),
'n.res3b%s' % str(stages[1] - 1),
'n.res4b%s' % str(stages[2] - 1)
][num]
else:
stage_string = branch_string
if i == 1:
bottom_string = 'n.res%sa' % str(num + 2)
else:
bottom_string = 'n.res%sb%s' % (str(num + 2),
str(i - 1))
exec(
stage_string.replace('(stage)', str(num + 2)).replace(
'(bottom)', bottom_string).replace(
'(num)', str(2**num * 64)).replace(
'(order)',
str(i)).replace('(stride)',
str(int(num > 0) + 1)))
exec 'n.pool5 = L.Pooling((bottom), pool=P.Pooling.AVE, global_pooling=True)'.\
replace('(bottom)', 'n.res5b%s' % str(stages[3] - 1))
n.classifier = L.InnerProduct(n.pool5,
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 create_net(train_list,batch_size,include_acc=False):
spec = caffe.NetSpec()
'''NetSpec可以用作命名,下面每一个spec.后面的字符直接就作为了该层的名字,没有使用spec的,系统会自动生成,在函数中的命名就是自动生成的,因为无法传递spec
spec.data,spec.label=L.ImageData(source=train_list,batch_size=batch_size,shuffle=True,ntop=2, transform_param=dict(crop_size=112,mirror=False,scale=0.0078125,mean_value=127.5),phase=0) '''
spec.data,spec.label=L.ImageData(source=train_list,batch_size=batch_size,shuffle=True,ntop=2,
transform_param=dict(crop_size=112,mirror=False,scale=0.0078125,mean_value=127.5),include=dict(phase=caffe.TRAIN))
spec.conv1=group_conv(spec.data,kernel_size=3,num_output=32,stride=2)
spec.relu1=after_conv(spec.conv1)
spec.conv2=group_conv(spec.relu1,num_output=32,GROUP=True,kernel_size=3,stride=1)
spec.relu2=after_conv(spec.conv2)
spec.concat1=L.Concat(spec.relu1,spec.relu2,axis=1)
spec.pooling1 = L.Pooling(spec.concat1,pool=P.Pooling.MAX,stride=2,kernel_size=3)
spec.relu3=after_conv(spec.pooling1)
spec.conv3=group_conv(spec.relu3,num_output=64,kernel_size=1,stride=1,pad=0)
spec.relu4=after_conv(spec.conv3)
spec.conv4=group_conv(spec.relu4,num_output=64,GROUP=True,kernel_size=3,stride=1)
spec.relu5=after_conv(spec.conv4)
spec.concat2=L.Concat(spec.pooling1,spec.relu5,axis=1)
spec.relu6=after_conv(spec.concat2)
spec.conv5=group_conv(spec.relu6,num_output=128,kernel_size=1,stride=1,pad=0)
spec.relu7=after_conv(spec.conv5)
spec.conv6=group_conv(spec.relu7,num_output=128,GROUP=True,kernel_size=3,stride=1)
spec.relu8=after_conv(spec.conv6)
spec.concat3=L.Concat(spec.concat2,spec.relu8,axis=1)
spec.pooling2 = L.Pooling(spec.concat3,pool=P.Pooling.MAX,stride=2,kernel_size=3)
spec.relu8=after_conv(spec.pooling2)
spec.conv7=group_conv(spec.relu8,num_output=256,kernel_size=1,stride=1,pad=0)
spec.relu9=after_conv(spec.conv7)
spec.conv8=group_conv(spec.relu9,num_output=256,GROUP=True,kernel_size=3,stride=1)
spec.relu10=after_conv(spec.conv8)
spec.concat4=L.Concat(spec.pooling2,spec.relu10,axis=1)
spec.relu11=after_conv(spec.concat4)
spec.conv9=group_conv(spec.relu11,num_output=512,kernel_size=1,stride=1,pad=0)
spec.relu12=after_conv(spec.conv9)
spec.conv10=group_conv(spec.relu12,num_output=512,GROUP=True,kernel_size=3,stride=1)
spec.relu13=after_conv(spec.conv10)
spec.concat5=L.Concat(spec.concat4,spec.relu13,axis=1)
spec.relu14=after_conv(spec.concat5)
spec.pooling3 = L.Pooling(spec.relu14,pool=P.Pooling.MAX,stride=2,kernel_size=3)
spec.relu15=after_conv(spec.pooling3)
spec.conv11=group_conv(spec.relu15,num_output=1024,kernel_size=1,stride=1,pad=0)
spec.relu16=after_conv(spec.conv11)
spec.conv12=group_conv(spec.relu16,num_output=1024,GROUP=True,kernel_size=3,stride=1)
spec.relu17=after_conv(spec.conv12)
#OUT 7
spec.maxpool=L.Pooling(spec.relu17, pool=P.Pooling.AVE,global_pooling=True)
spec.fc1=L.InnerProduct(spec.maxpool, num_output=1024,weight_filler=dict(type='xavier'))
#relu1=L.ReLU(fc1, in_place=True)
spec.fc2 = L.InnerProduct(spec.fc1, num_output=10000,weight_filler=dict(type='xavier'))
#,phase=0,0对应TRAIN
spec.loss = L.SoftmaxWithLoss(spec.fc2, spec.label,include=dict(phase=caffe.TRAIN))
#acc = L.Accuracy(fc2, label)
#return to_proto(loss, acc,include=dict(phase=TEST))
if include_acc:
return caffe.to_proto(spec.fc1)
else:
spec.acc = L.Accuracy(spec.fc2, spec.label,include=dict(phase=caffe.TEST))
#return spec.to_proto(spec.loss, spec.acc)
return spec.to_proto()