def forward(self, tokens):
net = self.apollo_net
net.f(
layers.NumpyData(self.seed_name,
np.zeros((tokens.shape[0], self.hidden_size))))
for t in range(tokens.shape[1]):
word_name = self.word_name % t
wordvec_name = self.wordvec_name % t
concat_name = self.concat_name % t
lstm_name = self.lstm_name % t
hidden_name = self.hidden_name % t
mem_name = self.mem_name % t
if t == 0:
prev_hidden = self.seed_name
prev_mem = self.seed_name
else:
prev_hidden = self.hidden_name % (t - 1)
prev_mem = self.mem_name % (t - 1)
net.f(layers.NumpyData(word_name, np.asarray(tokens[:, t])))
net.f(
layers.Wordvec(wordvec_name,
self.hidden_size,
len(STRING_INDEX),
bottoms=[word_name],
param_names=[self.wordvec_param_name],
param_lr_mults=[self.param_mult]))
net.f(
layers.Concat(concat_name, bottoms=[prev_hidden,
wordvec_name]))
net.f(
layers.LstmUnit(lstm_name,
bottoms=[concat_name, prev_mem],
param_names=[
self.input_value_param_name,
self.input_gate_param_name,
self.forget_gate_param_name,
self.output_gate_param_name
],
param_lr_mults=[self.param_mult] * 4,
tops=[hidden_name, mem_name],
num_cells=self.hidden_size))
net.f(
layers.InnerProduct(self.ip_name,
len(ANSWER_INDEX),
bottoms=[hidden_name],
param_lr_mults=[self.param_mult] * 2))
net.f(layers.ReLU(self.relu_name, bottoms=[self.ip_name]))
net.f(
layers.Eltwise(self.sum_name,
bottoms=[self.relu_name, self.incoming_name],
operation="SUM"))
def alexnet_layers():
conv_weight_filler = layers.Filler("gaussian", 0.01)
bias_filler0 = layers.Filler("constant", 0.0)
bias_filler1 = layers.Filler("constant", 1.0)
conv_lr_mults = [1.0, 2.0]
conv_decay_mults = [1.0, 0.0]
alexnet_layers = [
layers.Convolution("conv1", bottoms=["data"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_dim=(11, 11),
stride=4, weight_filler=conv_weight_filler, bias_filler=bias_filler0, num_output=96),
layers.ReLU(name="relu1", bottoms=["conv1"], tops=["conv1"]),
layers.Pooling(name="pool1", bottoms=["conv1"], kernel_size=3, stride=2),
layers.LRN(name="norm1", bottoms=["pool1"], local_size=5, alpha=0.0001, beta=0.75),
layers.Convolution(name="conv2", bottoms=["norm1"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_dim=(5, 5),
pad=2, group=2, weight_filler=conv_weight_filler, bias_filler=bias_filler1, num_output=256),
layers.ReLU(name="relu2", bottoms=["conv2"], tops=["conv2"]),
layers.Pooling(name="pool2", bottoms=["conv2"], kernel_size=3, stride=2),
layers.LRN(name="norm2", bottoms=["pool2"], local_size=5, alpha=0.0001, beta=0.75),
layers.Convolution(name="conv3", bottoms=["norm2"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_dim=(3, 3),
pad=1, weight_filler=conv_weight_filler, bias_filler=bias_filler0, num_output=384),
layers.ReLU(name="relu3", bottoms=["conv3"], tops=["conv3"]),
layers.Convolution(name="conv4", bottoms=["conv3"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_dim=(3, 3),
pad=1, group=2, weight_filler=conv_weight_filler, bias_filler=bias_filler1, num_output=384),
layers.ReLU(name="relu4", bottoms=["conv4"], tops=["conv4"]),
layers.Convolution(name="conv5", bottoms=["conv4"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_dim=(3, 3),
pad=1, group=2, weight_filler=conv_weight_filler, bias_filler=bias_filler1, num_output=256),
layers.ReLU(name="relu5", bottoms=["conv5"], tops=["conv5"]),
layers.Pooling(name="pool5", bottoms=["conv5"], kernel_size=3, stride=2),
layers.InnerProduct(name="fc6", bottoms=["pool5"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
weight_filler=layers.Filler("gaussian", 0.005),
bias_filler=bias_filler1, num_output=4096),
layers.ReLU(name="relu6", bottoms=["fc6"], tops=["fc6"]),
layers.Dropout(name="drop6", bottoms=["fc6"], tops=["fc6"], dropout_ratio=0.5),
layers.InnerProduct(name="fc7", bottoms=["fc6"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
weight_filler=layers.Filler("gaussian", 0.005),
bias_filler=bias_filler1, num_output=4096),
layers.ReLU(name="relu7", bottoms=["fc7"], tops=["fc7"]),
layers.Dropout(name="drop7", bottoms=["fc7"], tops=["fc7"], dropout_ratio=0.5),
layers.InnerProduct(name="fc8", bottoms=["fc7"], param_lr_mults=[1.0, 2.0],
param_decay_mults=conv_decay_mults,
weight_filler=layers.Filler("gaussian", 0.01),
bias_filler=bias_filler0, num_output=1000),
layers.SoftmaxWithLoss(name="loss", bottoms=["fc8", "label"]),
]
return alexnet_layers
def forward(self, data):
assert len(data.shape) == 4
batch_size, channels, width, height = data.shape
if self.data_name not in self.apollo_net.blobs.keys():
self.apollo_net.f(layers.NumpyData(self.data_name, data=data))
else:
self.apollo_net.blobs[self.data_name].data[:] = data
if self.proj_size is not None:
self.apollo_net.f(
layers.Convolution(self.proj_name, (1, 1),
self.proj_size,
bottoms=[self.data_name]))
self.apollo_net.f(
layers.ReLU(self.relu_name, bottoms=[self.proj_name]))
def googlenet_layers():
weight_filler = layers.Filler("xavier")
bias_filler = layers.Filler("constant", 0.2)
conv_lr_mults = [1.0, 2.0]
conv_decay_mults = [1.0, 0.0]
googlenet_layers = [
layers.Convolution(name="conv1/7x7_s2",
bottoms=["data"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(7, 7),
stride=2,
pad=3,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="conv1/relu_7x7",
bottoms=["conv1/7x7_s2"],
tops=["conv1/7x7_s2"]),
layers.Pooling(name="pool1/3x3_s2",
bottoms=["conv1/7x7_s2"],
kernel_size=3,
stride=2),
layers.LRN(name="pool1/norm1",
bottoms=["pool1/3x3_s2"],
local_size=5,
alpha=0.0001,
beta=0.75),
layers.Convolution(name="conv2/3x3_reduce",
bottoms=["pool1/norm1"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="conv2/relu_3x3_reduce",
bottoms=["conv2/3x3_reduce"],
tops=["conv2/3x3_reduce"]),
layers.Convolution(name="conv2/3x3",
bottoms=["conv2/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=192),
layers.ReLU(name="conv2/relu_3x3",
bottoms=["conv2/3x3"],
tops=["conv2/3x3"]),
layers.LRN(name="conv2/norm2",
bottoms=["conv2/3x3"],
local_size=5,
alpha=0.0001,
beta=0.75),
layers.Pooling(name="pool2/3x3_s2",
bottoms=["conv2/norm2"],
kernel_size=3,
stride=2),
layers.Convolution(name="inception_3a/1x1",
bottoms=["pool2/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_3a/relu_1x1",
bottoms=["inception_3a/1x1"],
tops=["inception_3a/1x1"]),
layers.Convolution(name="inception_3a/3x3_reduce",
bottoms=["pool2/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=96),
layers.ReLU(name="inception_3a/relu_3x3_reduce",
bottoms=["inception_3a/3x3_reduce"],
tops=["inception_3a/3x3_reduce"]),
layers.Convolution(name="inception_3a/3x3",
bottoms=["inception_3a/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_3a/relu_3x3",
bottoms=["inception_3a/3x3"],
tops=["inception_3a/3x3"]),
layers.Convolution(name="inception_3a/5x5_reduce",
bottoms=["pool2/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=16),
layers.ReLU(name="inception_3a/relu_5x5_reduce",
bottoms=["inception_3a/5x5_reduce"],
tops=["inception_3a/5x5_reduce"]),
layers.Convolution(name="inception_3a/5x5",
bottoms=["inception_3a/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=32),
layers.ReLU(name="inception_3a/relu_5x5",
bottoms=["inception_3a/5x5"],
tops=["inception_3a/5x5"]),
layers.Pooling(name="inception_3a/pool",
bottoms=["pool2/3x3_s2"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_3a/pool_proj",
bottoms=["inception_3a/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=32),
layers.ReLU(name="inception_3a/relu_pool_proj",
bottoms=["inception_3a/pool_proj"],
tops=["inception_3a/pool_proj"]),
layers.Concat(name="inception_3a/output",
bottoms=[
"inception_3a/1x1", "inception_3a/3x3",
"inception_3a/5x5", "inception_3a/pool_proj"
]),
layers.Convolution(name="inception_3b/1x1",
bottoms=["inception_3a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_3b/relu_1x1",
bottoms=["inception_3b/1x1"],
tops=["inception_3b/1x1"]),
layers.Convolution(name="inception_3b/3x3_reduce",
bottoms=["inception_3a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_3b/relu_3x3_reduce",
bottoms=["inception_3b/3x3_reduce"],
tops=["inception_3b/3x3_reduce"]),
layers.Convolution(name="inception_3b/3x3",
bottoms=["inception_3b/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=192),
layers.ReLU(name="inception_3b/relu_3x3",
bottoms=["inception_3b/3x3"],
tops=["inception_3b/3x3"]),
layers.Convolution(name="inception_3b/5x5_reduce",
bottoms=["inception_3a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=32),
layers.ReLU(name="inception_3b/relu_5x5_reduce",
bottoms=["inception_3b/5x5_reduce"],
tops=["inception_3b/5x5_reduce"]),
layers.Convolution(name="inception_3b/5x5",
bottoms=["inception_3b/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=96),
layers.ReLU(name="inception_3b/relu_5x5",
bottoms=["inception_3b/5x5"],
tops=["inception_3b/5x5"]),
layers.Pooling(name="inception_3b/pool",
bottoms=["inception_3a/output"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_3b/pool_proj",
bottoms=["inception_3b/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_3b/relu_pool_proj",
bottoms=["inception_3b/pool_proj"],
tops=["inception_3b/pool_proj"]),
layers.Concat(name="inception_3b/output",
bottoms=[
"inception_3b/1x1", "inception_3b/3x3",
"inception_3b/5x5", "inception_3b/pool_proj"
]),
layers.Pooling(name="pool3/3x3_s2",
bottoms=["inception_3b/output"],
kernel_size=3,
stride=2),
layers.Convolution(name="inception_4a/1x1",
bottoms=["pool3/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=192),
layers.ReLU(name="inception_4a/relu_1x1",
bottoms=["inception_4a/1x1"],
tops=["inception_4a/1x1"]),
layers.Convolution(name="inception_4a/3x3_reduce",
bottoms=["pool3/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=96),
layers.ReLU(name="inception_4a/relu_3x3_reduce",
bottoms=["inception_4a/3x3_reduce"],
tops=["inception_4a/3x3_reduce"]),
layers.Convolution(name="inception_4a/3x3",
bottoms=["inception_4a/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=208),
layers.ReLU(name="inception_4a/relu_3x3",
bottoms=["inception_4a/3x3"],
tops=["inception_4a/3x3"]),
layers.Convolution(name="inception_4a/5x5_reduce",
bottoms=["pool3/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=16),
layers.ReLU(name="inception_4a/relu_5x5_reduce",
bottoms=["inception_4a/5x5_reduce"],
tops=["inception_4a/5x5_reduce"]),
layers.Convolution(name="inception_4a/5x5",
bottoms=["inception_4a/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=48),
layers.ReLU(name="inception_4a/relu_5x5",
bottoms=["inception_4a/5x5"],
tops=["inception_4a/5x5"]),
layers.Pooling(name="inception_4a/pool",
bottoms=["pool3/3x3_s2"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_4a/pool_proj",
bottoms=["inception_4a/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_4a/relu_pool_proj",
bottoms=["inception_4a/pool_proj"],
tops=["inception_4a/pool_proj"]),
layers.Concat(name="inception_4a/output",
bottoms=[
"inception_4a/1x1", "inception_4a/3x3",
"inception_4a/5x5", "inception_4a/pool_proj"
]),
layers.Pooling(name="loss1/ave_pool",
bottoms=["inception_4a/output"],
kernel_size=5,
stride=3,
pool='AVE'),
layers.Convolution(name="loss1/conv",
bottoms=["loss1/ave_pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="loss1/relu_conv",
bottoms=["loss1/conv"],
tops=["loss1/conv"]),
layers.InnerProduct(name="loss1/fc",
bottoms=["loss1/conv"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=1024),
layers.ReLU(name="loss1/relu_fc",
bottoms=["loss1/fc"],
tops=["loss1/fc"]),
layers.Dropout(name="loss1/drop_fc",
bottoms=["loss1/fc"],
tops=["loss1/fc"],
dropout_ratio=0.7),
layers.InnerProduct(name="loss1/classifier",
bottoms=["loss1/fc"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
weight_filler=weight_filler,
bias_filler=layers.CaffeFiller(type="constant",
value=0.0),
num_output=1000),
layers.SoftmaxWithLoss(name="loss1/loss",
bottoms=["loss1/classifier", "label"],
tops=["loss1/loss1"],
loss_weight=0.3),
layers.Convolution(name="inception_4b/1x1",
bottoms=["inception_4a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=160),
layers.ReLU(name="inception_4b/relu_1x1",
bottoms=["inception_4b/1x1"],
tops=["inception_4b/1x1"]),
layers.Convolution(name="inception_4b/3x3_reduce",
bottoms=["inception_4a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=112),
layers.ReLU(name="inception_4b/relu_3x3_reduce",
bottoms=["inception_4b/3x3_reduce"],
tops=["inception_4b/3x3_reduce"]),
layers.Convolution(name="inception_4b/3x3",
bottoms=["inception_4b/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=224),
layers.ReLU(name="inception_4b/relu_3x3",
bottoms=["inception_4b/3x3"],
tops=["inception_4b/3x3"]),
layers.Convolution(name="inception_4b/5x5_reduce",
bottoms=["inception_4a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=24),
layers.ReLU(name="inception_4b/relu_5x5_reduce",
bottoms=["inception_4b/5x5_reduce"],
tops=["inception_4b/5x5_reduce"]),
layers.Convolution(name="inception_4b/5x5",
bottoms=["inception_4b/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_4b/relu_5x5",
bottoms=["inception_4b/5x5"],
tops=["inception_4b/5x5"]),
layers.Pooling(name="inception_4b/pool",
bottoms=["inception_4a/output"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_4b/pool_proj",
bottoms=["inception_4b/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_4b/relu_pool_proj",
bottoms=["inception_4b/pool_proj"],
tops=["inception_4b/pool_proj"]),
layers.Concat(name="inception_4b/output",
bottoms=[
"inception_4b/1x1", "inception_4b/3x3",
"inception_4b/5x5", "inception_4b/pool_proj"
]),
layers.Convolution(name="inception_4c/1x1",
bottoms=["inception_4b/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_4c/relu_1x1",
bottoms=["inception_4c/1x1"],
tops=["inception_4c/1x1"]),
layers.Convolution(name="inception_4c/3x3_reduce",
bottoms=["inception_4b/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_4c/relu_3x3_reduce",
bottoms=["inception_4c/3x3_reduce"],
tops=["inception_4c/3x3_reduce"]),
layers.Convolution(name="inception_4c/3x3",
bottoms=["inception_4c/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=256),
layers.ReLU(name="inception_4c/relu_3x3",
bottoms=["inception_4c/3x3"],
tops=["inception_4c/3x3"]),
layers.Convolution(name="inception_4c/5x5_reduce",
bottoms=["inception_4b/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=24),
layers.ReLU(name="inception_4c/relu_5x5_reduce",
bottoms=["inception_4c/5x5_reduce"],
tops=["inception_4c/5x5_reduce"]),
layers.Convolution(name="inception_4c/5x5",
bottoms=["inception_4c/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_4c/relu_5x5",
bottoms=["inception_4c/5x5"],
tops=["inception_4c/5x5"]),
layers.Pooling(name="inception_4c/pool",
bottoms=["inception_4b/output"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_4c/pool_proj",
bottoms=["inception_4c/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_4c/relu_pool_proj",
bottoms=["inception_4c/pool_proj"],
tops=["inception_4c/pool_proj"]),
layers.Concat(name="inception_4c/output",
bottoms=[
"inception_4c/1x1", "inception_4c/3x3",
"inception_4c/5x5", "inception_4c/pool_proj"
]),
layers.Convolution(name="inception_4d/1x1",
bottoms=["inception_4c/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=112),
layers.ReLU(name="inception_4d/relu_1x1",
bottoms=["inception_4d/1x1"],
tops=["inception_4d/1x1"]),
layers.Convolution(name="inception_4d/3x3_reduce",
bottoms=["inception_4c/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=144),
layers.ReLU(name="inception_4d/relu_3x3_reduce",
bottoms=["inception_4d/3x3_reduce"],
tops=["inception_4d/3x3_reduce"]),
layers.Convolution(name="inception_4d/3x3",
bottoms=["inception_4d/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=288),
layers.ReLU(name="inception_4d/relu_3x3",
bottoms=["inception_4d/3x3"],
tops=["inception_4d/3x3"]),
layers.Convolution(name="inception_4d/5x5_reduce",
bottoms=["inception_4c/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=32),
layers.ReLU(name="inception_4d/relu_5x5_reduce",
bottoms=["inception_4d/5x5_reduce"],
tops=["inception_4d/5x5_reduce"]),
layers.Convolution(name="inception_4d/5x5",
bottoms=["inception_4d/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_4d/relu_5x5",
bottoms=["inception_4d/5x5"],
tops=["inception_4d/5x5"]),
layers.Pooling(name="inception_4d/pool",
bottoms=["inception_4c/output"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_4d/pool_proj",
bottoms=["inception_4d/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=64),
layers.ReLU(name="inception_4d/relu_pool_proj",
bottoms=["inception_4d/pool_proj"],
tops=["inception_4d/pool_proj"]),
layers.Concat(name="inception_4d/output",
bottoms=[
"inception_4d/1x1", "inception_4d/3x3",
"inception_4d/5x5", "inception_4d/pool_proj"
]),
layers.Pooling(name="loss2/ave_pool",
bottoms=["inception_4d/output"],
kernel_size=5,
stride=3,
pool='AVE'),
layers.Convolution(name="loss2/conv",
bottoms=["loss2/ave_pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="loss2/relu_conv",
bottoms=["loss2/conv"],
tops=["loss2/conv"]),
layers.InnerProduct(name="loss2/fc",
bottoms=["loss2/conv"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=1024),
layers.ReLU(name="loss2/relu_fc",
bottoms=["loss2/fc"],
tops=["loss2/fc"]),
layers.Dropout(name="loss2/drop_fc",
bottoms=["loss2/fc"],
tops=["loss2/fc"],
dropout_ratio=0.7),
layers.InnerProduct(name="loss2/classifier",
bottoms=["loss2/fc"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
weight_filler=weight_filler,
bias_filler=layers.CaffeFiller(type="constant",
value=0.0),
num_output=1000),
layers.SoftmaxWithLoss(name="loss2/loss",
bottoms=["loss2/classifier", "label"],
tops=["loss2/loss1"],
loss_weight=0.3),
layers.Convolution(name="inception_4e/1x1",
bottoms=["inception_4d/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=256),
layers.ReLU(name="inception_4e/relu_1x1",
bottoms=["inception_4e/1x1"],
tops=["inception_4e/1x1"]),
layers.Convolution(name="inception_4e/3x3_reduce",
bottoms=["inception_4d/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=160),
layers.ReLU(name="inception_4e/relu_3x3_reduce",
bottoms=["inception_4e/3x3_reduce"],
tops=["inception_4e/3x3_reduce"]),
layers.Convolution(name="inception_4e/3x3",
bottoms=["inception_4e/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=320),
layers.ReLU(name="inception_4e/relu_3x3",
bottoms=["inception_4e/3x3"],
tops=["inception_4e/3x3"]),
layers.Convolution(name="inception_4e/5x5_reduce",
bottoms=["inception_4d/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=32),
layers.ReLU(name="inception_4e/relu_5x5_reduce",
bottoms=["inception_4e/5x5_reduce"],
tops=["inception_4e/5x5_reduce"]),
layers.Convolution(name="inception_4e/5x5",
bottoms=["inception_4e/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_4e/relu_5x5",
bottoms=["inception_4e/5x5"],
tops=["inception_4e/5x5"]),
layers.Pooling(name="inception_4e/pool",
bottoms=["inception_4d/output"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_4e/pool_proj",
bottoms=["inception_4e/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_4e/relu_pool_proj",
bottoms=["inception_4e/pool_proj"],
tops=["inception_4e/pool_proj"]),
layers.Concat(name="inception_4e/output",
bottoms=[
"inception_4e/1x1", "inception_4e/3x3",
"inception_4e/5x5", "inception_4e/pool_proj"
]),
layers.Pooling(name="pool4/3x3_s2",
bottoms=["inception_4e/output"],
kernel_size=3,
stride=2),
layers.Convolution(name="inception_5a/1x1",
bottoms=["pool4/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=256),
layers.ReLU(name="inception_5a/relu_1x1",
bottoms=["inception_5a/1x1"],
tops=["inception_5a/1x1"]),
layers.Convolution(name="inception_5a/3x3_reduce",
bottoms=["pool4/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=160),
layers.ReLU(name="inception_5a/relu_3x3_reduce",
bottoms=["inception_5a/3x3_reduce"],
tops=["inception_5a/3x3_reduce"]),
layers.Convolution(name="inception_5a/3x3",
bottoms=["inception_5a/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=320),
layers.ReLU(name="inception_5a/relu_3x3",
bottoms=["inception_5a/3x3"],
tops=["inception_5a/3x3"]),
layers.Convolution(name="inception_5a/5x5_reduce",
bottoms=["pool4/3x3_s2"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=32),
layers.ReLU(name="inception_5a/relu_5x5_reduce",
bottoms=["inception_5a/5x5_reduce"],
tops=["inception_5a/5x5_reduce"]),
layers.Convolution(name="inception_5a/5x5",
bottoms=["inception_5a/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_5a/relu_5x5",
bottoms=["inception_5a/5x5"],
tops=["inception_5a/5x5"]),
layers.Pooling(name="inception_5a/pool",
bottoms=["pool4/3x3_s2"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_5a/pool_proj",
bottoms=["inception_5a/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_5a/relu_pool_proj",
bottoms=["inception_5a/pool_proj"],
tops=["inception_5a/pool_proj"]),
layers.Concat(name="inception_5a/output",
bottoms=[
"inception_5a/1x1", "inception_5a/3x3",
"inception_5a/5x5", "inception_5a/pool_proj"
]),
layers.Convolution(name="inception_5b/1x1",
bottoms=["inception_5a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=384),
layers.ReLU(name="inception_5b/relu_1x1",
bottoms=["inception_5b/1x1"],
tops=["inception_5b/1x1"]),
layers.Convolution(name="inception_5b/3x3_reduce",
bottoms=["inception_5a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=192),
layers.ReLU(name="inception_5b/relu_3x3_reduce",
bottoms=["inception_5b/3x3_reduce"],
tops=["inception_5b/3x3_reduce"]),
layers.Convolution(name="inception_5b/3x3",
bottoms=["inception_5b/3x3_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(3, 3),
pad=1,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=384),
layers.ReLU(name="inception_5b/relu_3x3",
bottoms=["inception_5b/3x3"],
tops=["inception_5b/3x3"]),
layers.Convolution(name="inception_5b/5x5_reduce",
bottoms=["inception_5a/output"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=48),
layers.ReLU(name="inception_5b/relu_5x5_reduce",
bottoms=["inception_5b/5x5_reduce"],
tops=["inception_5b/5x5_reduce"]),
layers.Convolution(name="inception_5b/5x5",
bottoms=["inception_5b/5x5_reduce"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(5, 5),
pad=2,
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_5b/relu_5x5",
bottoms=["inception_5b/5x5"],
tops=["inception_5b/5x5"]),
layers.Pooling(name="inception_5b/pool",
bottoms=["inception_5a/output"],
kernel_size=3,
stride=1,
pad=1),
layers.Convolution(name="inception_5b/pool_proj",
bottoms=["inception_5b/pool"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
kernel_dim=(1, 1),
weight_filler=weight_filler,
bias_filler=bias_filler,
num_output=128),
layers.ReLU(name="inception_5b/relu_pool_proj",
bottoms=["inception_5b/pool_proj"],
tops=["inception_5b/pool_proj"]),
layers.Concat(name="inception_5b/output",
bottoms=[
"inception_5b/1x1", "inception_5b/3x3",
"inception_5b/5x5", "inception_5b/pool_proj"
]),
layers.Pooling(name="pool5/7x7_s1",
bottoms=["inception_5b/output"],
kernel_size=7,
stride=1,
pool='AVE'),
layers.Dropout(name="pool5/drop_7x7_s1",
bottoms=["pool5/7x7_s1"],
tops=["pool5/7x7_s1"],
dropout_ratio=0.4),
layers.InnerProduct(name="loss3/classifier",
bottoms=["pool5/7x7_s1"],
param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults,
weight_filler=weight_filler,
bias_filler=layers.CaffeFiller(type="constant",
value=0.0),
num_output=1000),
layers.SoftmaxWithLoss(name="loss3/loss3",
bottoms=["loss3/classifier", "label"],
loss_weight=1.0),
]
return googlenet_layers
def net_proto(self):
conv_weight_filler = layers.Filler("gaussian", 0.01)
bias_filler0 = layers.Filler("constant", 0.0)
bias_filler1 = layers.Filler("constant", 0.1)
bias_filler5 = layers.Filler("constant", 0.5)
# same deploy structure as in deploy_demo.prototxt
net_layers = [
# saliency path
layers.Convolution("conv1",
bottoms=["data"],
param_lr_mults=[0, 0],
param_decay_mults=[1, 0],
kernel_dim=(11, 11),
stride=4,
weight_filler=conv_weight_filler,
bias_filler=bias_filler0,
num_output=96),
layers.ReLU(name="relu1", bottoms=["conv1"], tops=["conv1"]),
layers.Pooling(name="pool1",
bottoms=["conv1"],
kernel_size=3,
stride=2),
layers.LRN(name="norm1",
bottoms=["pool1"],
tops=["norm1"],
local_size=5,
alpha=0.0001,
beta=0.75),
layers.Convolution(name="conv2",
bottoms=["norm1"],
param_lr_mults=[0, 0],
param_decay_mults=[1, 0],
kernel_dim=(5, 5),
pad=2,
group=2,
weight_filler=conv_weight_filler,
bias_filler=bias_filler1,
num_output=256),
layers.ReLU(name="relu2", bottoms=["conv2"], tops=["conv2"]),
layers.Pooling(name="pool2",
bottoms=["conv2"],
kernel_size=3,
stride=2),
layers.LRN(name="norm2",
bottoms=["pool2"],
tops=["norm2"],
local_size=5,
alpha=0.0001,
beta=0.75),
layers.Convolution(name="conv3",
bottoms=["norm2"],
param_lr_mults=[0.1, 0.2],
param_decay_mults=[1, 0],
kernel_dim=(3, 3),
pad=1,
weight_filler=conv_weight_filler,
bias_filler=bias_filler0,
num_output=384),
layers.ReLU(name="relu3", bottoms=["conv3"], tops=["conv3"]),
layers.Convolution(name="conv4",
bottoms=["conv3"],
param_lr_mults=[0.1, 0.2],
param_decay_mults=[1, 0],
kernel_dim=(3, 3),
pad=1,
group=2,
weight_filler=conv_weight_filler,
bias_filler=bias_filler1,
num_output=384),
layers.ReLU(name="relu4", bottoms=["conv4"], tops=["conv4"]),
layers.Convolution(name="conv5",
bottoms=["conv4"],
param_lr_mults=[0.1, 0.2],
param_decay_mults=[1, 0],
kernel_dim=(3, 3),
pad=1,
group=2,
weight_filler=conv_weight_filler,
bias_filler=bias_filler1,
num_output=256),
layers.ReLU(name="relu5", bottoms=["conv5"], tops=["conv5"]),
layers.Convolution(name="conv5_red",
bottoms=["conv5"],
param_lr_mults=[1.0, 2.0],
param_decay_mults=[1, 0],
kernel_dim=(1, 1),
weight_filler=conv_weight_filler,
bias_filler=bias_filler1,
num_output=1),
layers.ReLU(name="relu5_red",
bottoms=["conv5_red"],
tops=["conv5_red"]),
# gaze path
layers.Convolution("conv1_face",
bottoms=["face"],
param_lr_mults=[0, 0],
param_decay_mults=[1, 0],
kernel_dim=(11, 11),
stride=4,
weight_filler=conv_weight_filler,
bias_filler=bias_filler0,
num_output=96),
layers.ReLU(name="relu1_face",
bottoms=["conv1_face"],
tops=["conv1_face"]),
layers.Pooling(name="pool1_face",
bottoms=["conv1_face"],
kernel_size=3,
stride=2),
layers.LRN(name="norm1_face",
bottoms=["pool1_face"],
tops=["norm1_face"],
local_size=5,
alpha=0.0001,
beta=0.75),
layers.Convolution(name="conv2_face",
bottoms=["norm1_face"],
param_lr_mults=[0, 0],
param_decay_mults=[1, 0],
kernel_dim=(5, 5),
pad=2,
group=2,
weight_filler=conv_weight_filler,
bias_filler=bias_filler1,
num_output=256),
layers.ReLU(name="relu2_face",
bottoms=["conv2_face"],
tops=["conv2_face"]),
layers.Pooling(name="pool2_face",
bottoms=["conv2_face"],
kernel_size=3,
stride=2),
layers.LRN(name="norm2_face",
bottoms=["pool2_face"],
tops=["norm2_face"],
local_size=5,
alpha=0.0001,
beta=0.75),
layers.Convolution(name="conv3_face",
bottoms=["norm2_face"],
param_lr_mults=[0.2, 0.4],
param_decay_mults=[1, 0],
kernel_dim=(3, 3),
pad=1,
weight_filler=conv_weight_filler,
bias_filler=bias_filler0,
num_output=384),
layers.ReLU(name="relu3_face",
bottoms=["conv3_face"],
tops=["conv3_face"]),
layers.Convolution(name="conv4_face",
bottoms=["conv3_face"],
param_lr_mults=[0.2, 0.4],
param_decay_mults=[1, 0],
kernel_dim=(3, 3),
pad=1,
group=2,
weight_filler=conv_weight_filler,
bias_filler=bias_filler1,
num_output=384),
layers.ReLU(name="relu4_face",
bottoms=["conv4_face"],
tops=["conv4_face"]),
layers.Convolution(name="conv5_face",
bottoms=["conv4_face"],
param_lr_mults=[0.2, 0.4],
param_decay_mults=[1, 0],
kernel_dim=(3, 3),
pad=1,
group=2,
weight_filler=conv_weight_filler,
bias_filler=bias_filler1,
num_output=256),
layers.ReLU(name="relu5_face",
bottoms=["conv5_face"],
tops=["conv5_face"]),
layers.Pooling(name="pool5_face",
bottoms=["conv5_face"],
kernel_size=3,
stride=2),
layers.InnerProduct(name="fc6_face",
bottoms=["pool5_face"],
tops=["fc6_face"],
param_lr_mults=[1, 2],
param_decay_mults=[1, 0],
weight_filler=layers.Filler("gaussian", 0.5),
bias_filler=bias_filler5,
num_output=500),
layers.ReLU(name="relu6_face",
bottoms=["fc6_face"],
tops=["fc6_face"]),
layers.Flatten(name="eyes_grid_flat",
bottoms=["eyes_grid"],
tops=["eyes_grid_flat"]),
layers.Power(name="eyes_grid_mult",
bottoms=["eyes_grid_flat"],
tops=["eyes_grid_mult"],
power=1,
scale=24,
shift=0),
layers.Concat(name="face_input",
bottoms=["fc6_face", "eyes_grid_mult"],
tops=["face_input"],
axis=1),
layers.InnerProduct(name="fc7_face",
bottoms=["face_input"],
tops=["fc7_face"],
param_lr_mults=[1, 2],
param_decay_mults=[1, 0],
weight_filler=layers.Filler("gaussian", 0.01),
bias_filler=bias_filler5,
num_output=400),
layers.ReLU(name="relu7_face",
bottoms=["fc7_face"],
tops=["fc7_face"]),
layers.InnerProduct(name="fc8_face",
bottoms=["fc7_face"],
tops=["fc8_face"],
param_lr_mults=[1, 2],
param_decay_mults=[1, 0],
weight_filler=layers.Filler("gaussian", 0.01),
bias_filler=bias_filler5,
num_output=200),
layers.ReLU(name="relu8_face",
bottoms=["fc8_face"],
tops=["fc8_face"]),
layers.InnerProduct(name="importance_no_sigmoid",
bottoms=["fc8_face"],
tops=["importance_no_sigmoid"],
param_lr_mults=[0.2, 0.0],
param_decay_mults=[1.0, 0.0],
weight_filler=layers.Filler("gaussian", 0.01),
num_output=169),
layers.Sigmoid(name="importance_map_prefilter",
bottoms=["importance_no_sigmoid"],
tops=["importance_map_prefilter"]),
layers.Reshape('importance_map_reshape', (1, 1, 13, 13),
bottoms=['importance_map_prefilter'],
tops=["importance_map_reshape"]),
layers.Convolution(name="importance_map",
bottoms=["importance_map_reshape"],
param_lr_mults=[0.0, 0.0],
param_decay_mults=[1.0, 0.0],
kernel_dim=(3, 3),
pad=1,
stride=1,
weight_filler=conv_weight_filler,
bias_filler=bias_filler1,
num_output=1),
layers.Eltwise(name="fc_7",
bottoms=["conv5_red", "importance_map"],
tops=["fc_7"],
operation="PROD"),
# shifted grids
layers.InnerProduct(name="fc_0_0",
bottoms=["fc_7"],
tops=["fc_0_0"],
param_lr_mults=[1, 2],
param_decay_mults=[1, 0],
weight_filler=layers.Filler("gaussian", 0.01),
bias_filler=bias_filler0,
num_output=25),
layers.InnerProduct(name="fc_1_0",
bottoms=["fc_7"],
tops=["fc_1_0"],
param_lr_mults=[1, 2],
param_decay_mults=[1, 0],
weight_filler=layers.Filler("gaussian", 0.01),
bias_filler=bias_filler0,
num_output=25),
layers.InnerProduct(name="fc_m1_0",
bottoms=["fc_7"],
tops=["fc_m1_0"],
param_lr_mults=[1, 2],
param_decay_mults=[1, 0],
weight_filler=layers.Filler("gaussian", 0.01),
bias_filler=bias_filler0,
num_output=25),
layers.InnerProduct(name="fc_0_1",
bottoms=["fc_7"],
tops=["fc_0_1"],
param_lr_mults=[1, 2],
param_decay_mults=[1, 0],
weight_filler=layers.Filler("gaussian", 0.01),
bias_filler=bias_filler0,
num_output=25),
layers.InnerProduct(name="fc_0_m1",
bottoms=["fc_7"],
tops=["fc_0_m1"],
param_lr_mults=[1, 2],
param_decay_mults=[1, 0],
weight_filler=layers.Filler("gaussian", 0.01),
bias_filler=bias_filler0,
num_output=25),
layers.Reshape('fc_0_0_reshape', (5, 5),
bottoms=['fc_0_0'],
tops=["fc_0_0_reshape"]),
layers.Reshape('fc_1_0_reshape', (5, 5),
bottoms=['fc_1_0'],
tops=["fc_1_0_reshape"]),
layers.Reshape('fc_m1_0_reshape', (5, 5),
bottoms=['fc_m1_0'],
tops=["fc_m1_0_reshape"]),
layers.Reshape('fc_0_1_reshape', (5, 5),
bottoms=['fc_0_1'],
tops=["fc_0_1_reshape"]),
layers.Reshape('fc_0_m1_reshape', (5, 5),
bottoms=['fc_0_m1'],
tops=["fc_0_m1_reshape"])
]
return net_layers