def four_mlp_init(batch):
layer1 = Layer(Type.FC, batch=batch, in_dim=100, out_dim=400)
layer2 = Layer(Type.FC,
batch=batch,
in_dim=400,
out_dim=400,
previous_input=True)
layer3 = Layer(Type.FC,
batch=batch,
in_dim=400,
out_dim=400,
previous_input=True)
layer4 = Layer(Type.FC,
batch=batch,
in_dim=400,
out_dim=10,
previous_input=True)
container_mlp_sample.push_layer(layer1)
container_mlp_sample.push_layer(layer2)
container_mlp_sample.push_layer(layer3)
container_mlp_sample.push_layer(layer4)
def rnn_sa_init(batch, length):
# embedding layer is omitted
N = batch
Ti = length
H = 1024
for i in range(Ti):
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H,
h_dim=H,
no_hidden=True,
previous_input=False)
container_rnn_sa.push_layer(layer_lstm1_bi1)
fc = Layer(Type.FC, batch=N, in_dim=H, out_dim=H)
container_rnn_sa.push_layer(fc)
container_rnn_sa.net_name = 'Sentiment Analysis'
container_rnn_sa.isolated[1] = 4469462
container_rnn_sa.isolated[4] = 4489981
MUT = Mmunit(HEIGHT, WIDTH, DEPTH)
VUT = Vecunit(WIDTH)
UBUF = Buffer(8*MB/4, 358*GB/4, 100, 'UBUF')
WBUF = Buffer(4*MB/4, 358*GB/4, 100, 'WBUF')
ACCQ = Buffer(WIDTH*DEPTH, 358*GB/4, 0, 'ACCQ')
# random container generator
# for sample, just all fc layer instance
container_1 = Container()
container_2 = Container()
container_3 = Container()
container_4 = Container()
layer1 = Layer(Type.FC, batch=200, in_dim=100, out_dim=400)
layer2 = Layer(Type.FC, batch=200, in_dim=400, out_dim=400, previous_input=True)
layer3 = Layer(Type.FC, batch=200, in_dim=400, out_dim=400, previous_input=True)
layer4 = Layer(Type.FC, batch=200, in_dim=400, out_dim=10, previous_input=True)
container_1.push_layer(layer1)
container_1.push_layer(layer2)
container_1.push_layer(layer3)
container_1.push_layer(layer4)
layer1 = Layer(Type.FC, batch=200, in_dim=100, out_dim=400)
layer2 = Layer(Type.FC, batch=200, in_dim=400, out_dim=400, previous_input=True)
layer3 = Layer(Type.FC, batch=200, in_dim=400, out_dim=400, previous_input=True)
layer4 = Layer(Type.FC, batch=200, in_dim=400, out_dim=10, previous_input=True)
container_2.push_layer(layer1)
container_2.push_layer(layer2)
container_2.push_layer(layer3)
def rnn_mt_init(batch, length):
# Assumption: output = input
# Embedding layer is omitted
N = batch
Ti = length
To = length
H = 256 # hidden, for single direction, i.e. 512 total
D = 240 # input
D_HIDDEN = 1024
# encoder
# first layer(256, bidirectional lstm)
for i in range(Ti):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=True,
previous_input=False)
container_rnn_mt.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=True,
previous_input=False)
container_rnn_mt.push_layer(layer_lstm1_bi2)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=False,
previous_input=False)
container_rnn_mt.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=False,
previous_input=False)
container_rnn_mt.push_layer(layer_lstm1_bi2)
# second layer
for i in range(Ti):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H,
no_hidden=True,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H,
no_hidden=False,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
# third layer
for i in range(Ti):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H,
h_dim=H,
no_hidden=True,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H,
h_dim=H,
no_hidden=False,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
# fourth layer
for i in range(Ti):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H,
h_dim=H,
no_hidden=True,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H,
h_dim=H,
no_hidden=False,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
# decoder
for i in range(To):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=True,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=False,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
# second layer
for i in range(To):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H,
no_hidden=True,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H,
no_hidden=False,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
# third layer
for i in range(To):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H,
no_hidden=True,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H,
no_hidden=False,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
# fourth layer
for i in range(To):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H,
no_hidden=True,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H,
no_hidden=False,
previous_input=True)
container_rnn_mt.push_layer(layer_lstm1_bi1)
# attention layer
# Bahdanau attention
# https://github.com/tensorflow/nmt#background-on-the-attention-mechanism
# not accurate
for i in range(To):
layer_t = Layer(Type.GEMM, batch=1, gemm_m=N, gemm_k=H, gemm_n=H)
layer_h = Layer(Type.GEMM, batch=1, gemm_m=N, gemm_k=H, gemm_n=H)
container_rnn_mt.push_layer(layer_t)
container_rnn_mt.push_layer(layer_h)
# fc layer
for i in range(To):
fc = Layer(Type.FC, batch=N, in_dim=H, out_dim=H)
container_rnn_mt.push_layer(fc)
container_rnn_mt.net_name = 'Machine Translation'
container_rnn_mt.isolated[1] = 9911366
container_rnn_mt.isolated[4] = 9911366
def rnn_asr_init(batch, length):
# Assumption: output = int(input/4)
# Embedding layer is omitted
# listener (encoder)
N = batch
Ti = length
To = int(length / 4)
H = 256 # hidden, for single direction, i.e. 512 total
D = 240 # input
D_EMBED = 512
D_HIDDEN = 512
# encoder
# first layer(256, bidirectional lstm)
for i in range(Ti):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=True,
previous_input=False)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=True,
previous_input=False)
container_rnn_asr.push_layer(layer_lstm1_bi2)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=False,
previous_input=False)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=D,
h_dim=H,
no_hidden=False,
previous_input=False)
container_rnn_asr.push_layer(layer_lstm1_bi2)
# second layer
for i in range(int(Ti / 2)):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H * 2,
no_hidden=True,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H * 2,
no_hidden=True,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi2)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H * 2,
no_hidden=False,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=H * 2,
h_dim=H * 2,
no_hidden=False,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi2)
# third layer
for i in range(int(Ti / 4)):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 4,
h_dim=H * 4,
no_hidden=True,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=H * 4,
h_dim=H * 4,
no_hidden=True,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi2)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=H * 4,
h_dim=H * 4,
no_hidden=False,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=H * 4,
h_dim=H * 4,
no_hidden=False,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi2)
# decoder
# first layer
for i in range(To):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D_EMBED,
h_dim=D_HIDDEN,
no_hidden=True,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=D_EMBED,
h_dim=D_HIDDEN,
no_hidden=True,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi2)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D_EMBED,
h_dim=D_HIDDEN,
no_hidden=False,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=D_EMBED,
h_dim=D_HIDDEN,
no_hidden=False,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi2)
# second layer
for i in range(To):
if i == 0:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D_EMBED,
h_dim=D_HIDDEN,
no_hidden=True,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=D_EMBED,
h_dim=D_HIDDEN,
no_hidden=True,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi2)
else:
layer_lstm1_bi1 = Layer(Type.LSTM,
batch=N,
in_dim=D_EMBED,
h_dim=D_HIDDEN,
no_hidden=False,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi1)
layer_lstm1_bi2 = Layer(Type.LSTM,
batch=N,
in_dim=D_EMBED,
h_dim=D_HIDDEN,
no_hidden=False,
previous_input=True)
container_rnn_asr.push_layer(layer_lstm1_bi2)
# attention
attention_score = Layer(Type.GEMM,
batch=N,
gemm_m=To,
gemm_k=D_HIDDEN,
gemm_n=Ti,
previous_input=True)
container_rnn_asr.push_layer(attention_score)
attention_output = Layer(Type.GEMM,
batch=N,
gemm_m=To,
gemm_k=Ti,
gemm_n=D_HIDDEN,
previous_input=True)
container_rnn_asr.push_layer(attention_output)
# mlp
layer1 = Layer(Type.FC, batch=N, in_dim=D_HIDDEN * 2, out_dim=D_HIDDEN)
layer2 = Layer(Type.FC, batch=N, in_dim=D_HIDDEN, out_dim=D)
container_rnn_asr.push_layer(layer1)
container_rnn_asr.push_layer(layer2)
container_rnn_asr.net_name = "Automatic Speech Recognition"
container_rnn_asr.isolated[1] = 13440293
container_rnn_asr.isolated[4] = 13464404
def cnn_alex_init(batch):
conv1 = Layer(Type.CONV,
batch=batch,
in_dim=(224, 224, 3),
kernel_dim=(11, 11),
kernel_num=96,
stride=4,
padding=0)
pool1 = Layer(Type.POOL,
batch=batch,
in_dim=(55, 55, 96),
window_dim=(3, 3),
stride=1,
previous_input=True)
conv2 = Layer(Type.CONV,
batch=batch,
in_dim=(27, 27, 96),
kernel_dim=(5, 5),
kernel_num=256,
stride=1,
padding=2,
previous_input=True)
pool2 = Layer(Type.POOL,
batch=batch,
in_dim=(27, 27, 256),
window_dim=(3, 3),
stride=2,
previous_input=True)
conv3 = Layer(Type.CONV,
batch=batch,
in_dim=(13, 13, 256),
kernel_dim=(3, 3),
kernel_num=384,
stride=1,
padding=1,
previous_input=True)
conv4 = Layer(Type.CONV,
batch=batch,
in_dim=(13, 13, 384),
kernel_dim=(3, 3),
kernel_num=384,
stride=1,
padding=1,
previous_input=True)
conv5 = Layer(Type.CONV,
batch=batch,
in_dim=(13, 13, 384),
kernel_dim=(3, 3),
kernel_num=256,
stride=1,
padding=1,
previous_input=True)
pool5 = Layer(Type.POOL,
batch=batch,
in_dim=(13, 13, 256),
window_dim=(3, 3),
stride=2,
previous_input=True)
fc6 = Layer(Type.FC,
batch=batch,
in_dim=9216,
out_dim=4096,
previous_input=True)
fc7 = Layer(Type.FC,
batch=batch,
in_dim=4096,
out_dim=1000,
previous_input=True)
container_cnn_alex.push_layer(conv1)
container_cnn_alex.push_layer(pool1)
container_cnn_alex.push_layer(conv2)
container_cnn_alex.push_layer(pool2)
container_cnn_alex.push_layer(conv3)
container_cnn_alex.push_layer(conv4)
container_cnn_alex.push_layer(conv5)
container_cnn_alex.push_layer(pool5)
container_cnn_alex.push_layer(fc6)
container_cnn_alex.push_layer(fc7)
container_cnn_alex.net_name = "AlexNet"
container_cnn_alex.isolated[1] = 1697722
container_cnn_alex.isolated[4] = 6473754
def cnn_mobile_init(batch):
conv1 = Layer(Type.CONV,
batch=batch,
in_dim=(224, 224, 3),
kernel_dim=(3, 3),
kernel_num=32,
stride=2,
padding=1,
previous_input=False)
container_cnn_mobile.push_layer(conv1)
convdw2 = Layer(Type.DEPTH,
batch=batch,
in_dim=(112, 112, 32),
kernel_dim=(3, 3),
stride=1,
padding=1,
previous_input=True)
conv2 = Layer(Type.CONV,
batch=batch,
in_dim=(112, 112, 32),
kernel_dim=(1, 1),
kernel_num=64,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw2)
container_cnn_mobile.push_layer(conv2)
convdw3 = Layer(Type.DEPTH,
batch=batch,
in_dim=(112, 112, 64),
kernel_dim=(3, 3),
stride=2,
padding=1,
previous_input=True)
conv3 = Layer(Type.CONV,
batch=batch,
in_dim=(56, 56, 64),
kernel_dim=(1, 1),
kernel_num=128,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw3)
container_cnn_mobile.push_layer(conv3)
convdw4 = Layer(Type.DEPTH,
batch=batch,
in_dim=(56, 56, 128),
kernel_dim=(3, 3),
stride=1,
padding=1,
previous_input=True)
conv4 = Layer(Type.CONV,
batch=batch,
in_dim=(56, 56, 128),
kernel_dim=(1, 1),
kernel_num=128,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw4)
container_cnn_mobile.push_layer(conv4)
convdw5 = Layer(Type.DEPTH,
batch=batch,
in_dim=(56, 56, 128),
kernel_dim=(3, 3),
stride=2,
padding=1,
previous_input=True)
conv5 = Layer(Type.CONV,
batch=batch,
in_dim=(28, 28, 128),
kernel_dim=(1, 1),
kernel_num=256,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw5)
container_cnn_mobile.push_layer(conv5)
convdw6 = Layer(Type.DEPTH,
batch=batch,
in_dim=(28, 28, 256),
kernel_dim=(3, 3),
stride=2,
padding=1,
previous_input=True)
conv6 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 256),
kernel_dim=(1, 1),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw6)
container_cnn_mobile.push_layer(conv6)
convdw7 = Layer(Type.DEPTH,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
stride=1,
padding=1,
previous_input=True)
conv7 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(1, 1),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw7)
container_cnn_mobile.push_layer(conv7)
convdw8 = Layer(Type.DEPTH,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
stride=1,
padding=1,
previous_input=True)
conv8 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(1, 1),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw8)
container_cnn_mobile.push_layer(conv8)
convdw9 = Layer(Type.DEPTH,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
stride=1,
padding=1,
previous_input=True)
conv9 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(1, 1),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw9)
container_cnn_mobile.push_layer(conv9)
convdw10 = Layer(Type.DEPTH,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
stride=1,
padding=1,
previous_input=True)
conv10 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(1, 1),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw10)
container_cnn_mobile.push_layer(conv10)
convdw11 = Layer(Type.DEPTH,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
stride=1,
padding=1,
previous_input=True)
conv11 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(1, 1),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw11)
container_cnn_mobile.push_layer(conv11)
convdw12 = Layer(Type.DEPTH,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
stride=2,
padding=1,
previous_input=True)
conv12 = Layer(Type.CONV,
batch=batch,
in_dim=(7, 7, 512),
kernel_dim=(1, 1),
kernel_num=1024,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw12)
container_cnn_mobile.push_layer(conv12)
convdw13 = Layer(Type.DEPTH,
batch=batch,
in_dim=(7, 7, 1024),
kernel_dim=(3, 3),
stride=1,
padding=1,
previous_input=True)
conv13 = Layer(Type.CONV,
batch=batch,
in_dim=(7, 7, 1024),
kernel_dim=(1, 1),
kernel_num=1024,
stride=1,
padding=1,
previous_input=True)
container_cnn_mobile.push_layer(convdw13)
container_cnn_mobile.push_layer(conv13)
pool = Layer(Type.POOL,
batch=batch,
in_dim=(7, 7, 1024),
window_dim=(7, 7),
stride=1,
previous_input=True)
fc = Layer(Type.FC,
batch=batch,
in_dim=1024,
out_dim=1000,
previous_input=True)
container_cnn_mobile.push_layer(pool)
container_cnn_mobile.push_layer(fc)
container_cnn_mobile.net_name = "MobileNet"
container_cnn_mobile.isolated[1] = 1867013
container_cnn_mobile.isolated[4] = 25908758
def push_inception(batch: int, in_dim: tuple, channel_dim: tuple):
# 1x1
conv1_input = Layer(Type.CONV,
batch=batch,
in_dim=in_dim,
kernel_dim=(1, 1),
kernel_num=channel_dim[0],
stride=1,
padding=0,
previous_input=True)
container_cnn_google.push_layer(conv1_input)
# 1x1 -> 3x3
conv1_no_in = Layer(Type.CONV,
batch=batch,
in_dim=in_dim,
kernel_dim=(1, 1),
kernel_num=channel_dim[1],
stride=1,
padding=0,
previous_input=True)
conv3 = Layer(Type.CONV,
batch=batch,
in_dim=(in_dim[0], in_dim[1], channel_dim[1]),
kernel_dim=(3, 3),
kernel_num=channel_dim[2],
stride=1,
padding=1,
previous_input=True)
container_cnn_google.push_layer(conv1_no_in)
container_cnn_google.push_layer(conv3)
# 1x1 -> 5x5
conv1_no_in = Layer(Type.CONV,
batch=batch,
in_dim=in_dim,
kernel_dim=(1, 1),
kernel_num=channel_dim[3],
stride=1,
padding=0,
previous_input=True)
conv5 = Layer(Type.CONV,
batch=batch,
in_dim=(in_dim[0], in_dim[1], channel_dim[3]),
kernel_dim=(5, 5),
kernel_num=channel_dim[4],
stride=1,
padding=2,
previous_input=True)
container_cnn_google.push_layer(conv1_no_in)
container_cnn_google.push_layer(conv5)
# pool -> 1x1
pool = Layer(Type.POOL,
batch=batch,
in_dim=in_dim,
window_dim=(3, 3),
stride=1,
previous_input=True)
conv1_no_in = Layer(Type.CONV,
batch=batch,
in_dim=in_dim,
kernel_dim=(1, 1),
kernel_num=channel_dim[5],
stride=1,
padding=0,
previous_input=True)
container_cnn_google.push_layer(pool)
container_cnn_google.push_layer(conv1_no_in)
return (in_dim[0], in_dim[1],
channel_dim[0] + channel_dim[2] + channel_dim[4] + channel_dim[5])
def cnn_google_init(batch):
conv1 = Layer(Type.CONV,
batch=batch,
in_dim=(224, 224, 3),
kernel_dim=(7, 7),
kernel_num=64,
stride=2,
padding=3,
previous_input=False)
pool1 = Layer(Type.POOL,
batch=batch,
in_dim=(112, 112, 64),
window_dim=(3, 3),
stride=2)
conv2 = Layer(Type.CONV,
batch=batch,
in_dim=(56, 56, 64),
kernel_dim=(3, 3),
kernel_num=192,
stride=1,
padding=1,
previous_input=True)
pool2 = Layer(Type.POOL,
batch=batch,
in_dim=(56, 56, 192),
window_dim=(3, 3),
stride=2)
container_cnn_google.push_layer(conv1)
container_cnn_google.push_layer(pool1)
container_cnn_google.push_layer(conv2)
container_cnn_google.push_layer(pool2)
next_in = push_inception(batch, (28, 28, 192), (64, 96, 128, 16, 32, 32))
next_in = push_inception(batch, next_in, (128, 128, 192, 32, 96, 64))
pool3 = Layer(Type.POOL,
batch=batch,
in_dim=next_in,
window_dim=(3, 3),
stride=2)
container_cnn_google.push_layer(pool3)
next_in = push_inception(batch, (14, 14, 480), (192, 96, 208, 16, 48, 64))
next_in = push_inception(batch, next_in, (160, 112, 224, 24, 64, 64))
next_in = push_inception(batch, next_in, (128, 128, 256, 24, 64, 64))
next_in = push_inception(batch, next_in, (112, 144, 288, 32, 64, 64))
next_in = push_inception(batch, next_in, (256, 160, 320, 32, 128, 128))
pool4 = Layer(Type.POOL,
batch=batch,
in_dim=next_in,
window_dim=(3, 3),
stride=2)
container_cnn_google.push_layer(pool4)
next_in = push_inception(batch, (7, 7, 832), (256, 160, 320, 32, 128, 128))
next_in = push_inception(batch, next_in, (384, 192, 384, 48, 128, 128))
pool5 = Layer(Type.POOL,
batch=batch,
in_dim=next_in,
window_dim=(7, 7),
stride=1)
container_cnn_google.push_layer(pool5)
fc = Layer(Type.FC,
batch=batch,
in_dim=1024,
out_dim=1000,
previous_input=True)
container_cnn_vgg.push_layer(fc)
container_cnn_google.net_name = "GoogLeNet"
container_cnn_google.isolated[1] = 813951
container_cnn_google.isolated[4] = 11873594
def cnn_vgg_init(batch):
conv1 = Layer(Type.CONV,
batch=batch,
in_dim=(224, 224, 3),
kernel_dim=(3, 3),
kernel_num=64,
stride=1,
padding=1)
conv2 = Layer(Type.CONV,
batch=batch,
in_dim=(224, 224, 64),
kernel_dim=(3, 3),
kernel_num=64,
stride=1,
padding=1,
previous_input=True)
pool2 = Layer(Type.POOL,
batch=batch,
in_dim=(224, 224, 64),
window_dim=(2, 2),
stride=2,
previous_input=True)
container_cnn_vgg.push_layer(conv1)
container_cnn_vgg.push_layer(conv2)
container_cnn_vgg.push_layer(pool2)
conv3 = Layer(Type.CONV,
batch=batch,
in_dim=(112, 112, 64),
kernel_dim=(3, 3),
kernel_num=128,
stride=1,
padding=1,
previous_input=True)
conv4 = Layer(Type.CONV,
batch=batch,
in_dim=(112, 112, 128),
kernel_dim=(3, 3),
kernel_num=128,
stride=1,
padding=1,
previous_input=True)
pool4 = Layer(Type.POOL,
batch=batch,
in_dim=(112, 112, 128),
window_dim=(2, 2),
stride=2,
previous_input=True)
container_cnn_vgg.push_layer(conv3)
container_cnn_vgg.push_layer(conv4)
container_cnn_vgg.push_layer(pool4)
conv5 = Layer(Type.CONV,
batch=batch,
in_dim=(56, 56, 128),
kernel_dim=(3, 3),
kernel_num=256,
stride=1,
padding=1,
previous_input=True)
conv6 = Layer(Type.CONV,
batch=batch,
in_dim=(56, 56, 256),
kernel_dim=(3, 3),
kernel_num=256,
stride=1,
padding=1,
previous_input=True)
conv7 = Layer(Type.CONV,
batch=batch,
in_dim=(56, 56, 256),
kernel_dim=(3, 3),
kernel_num=256,
stride=1,
padding=1,
previous_input=True)
pool7 = Layer(Type.POOL,
batch=batch,
in_dim=(56, 56, 256),
window_dim=(2, 2),
stride=2,
previous_input=True)
container_cnn_vgg.push_layer(conv5)
container_cnn_vgg.push_layer(conv6)
container_cnn_vgg.push_layer(conv7)
container_cnn_vgg.push_layer(pool7)
conv8 = Layer(Type.CONV,
batch=batch,
in_dim=(28, 28, 256),
kernel_dim=(3, 3),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
conv9 = Layer(Type.CONV,
batch=batch,
in_dim=(28, 28, 512),
kernel_dim=(3, 3),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
conv10 = Layer(Type.CONV,
batch=batch,
in_dim=(28, 28, 512),
kernel_dim=(3, 3),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
pool10 = Layer(Type.POOL,
batch=batch,
in_dim=(28, 28, 512),
window_dim=(2, 2),
stride=2,
previous_input=True)
container_cnn_vgg.push_layer(conv8)
container_cnn_vgg.push_layer(conv9)
container_cnn_vgg.push_layer(conv10)
container_cnn_vgg.push_layer(pool10)
conv11 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
conv12 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
conv13 = Layer(Type.CONV,
batch=batch,
in_dim=(14, 14, 512),
kernel_dim=(3, 3),
kernel_num=512,
stride=1,
padding=1,
previous_input=True)
pool13 = Layer(Type.POOL,
batch=batch,
in_dim=(14, 14, 512),
window_dim=(2, 2),
stride=2,
previous_input=True)
pool14 = Layer(Type.POOL,
batch=batch,
in_dim=(7, 7, 512),
window_dim=(2, 2),
stride=2,
previous_input=True)
container_cnn_vgg.push_layer(conv11)
container_cnn_vgg.push_layer(conv12)
container_cnn_vgg.push_layer(conv13)
container_cnn_vgg.push_layer(pool13)
container_cnn_vgg.push_layer(pool14)
fc1 = Layer(Type.FC,
batch=batch,
in_dim=4096,
out_dim=4096,
previous_input=True)
fc2 = Layer(Type.FC,
batch=batch,
in_dim=4096,
out_dim=4096,
previous_input=True)
fc3 = Layer(Type.FC,
batch=batch,
in_dim=4096,
out_dim=1000,
previous_input=True)
container_cnn_vgg.push_layer(fc1)
container_cnn_vgg.push_layer(fc2)
container_cnn_vgg.push_layer(fc3)
container_cnn_vgg.net_name = "VGG16"
container_cnn_vgg.isolated[1] = 6150050
container_cnn_vgg.isolated[4] = 77632498