def resnet_stem(input, bn, model_weight, optimizer):
conv_stem = _conv2d("conv_stem", input, 64, 7, 2)
if bn:
conv_stem = _batch_norm(conv_stem, "conv_stem_bn")
conv_stem = flow.nn.relu(conv_stem)
pool1 = flow.nn.max_pool2d(
conv_stem,
ksize=3,
strides=2,
padding="VALID",
data_format="NCHW",
name="pool1",
)
#最初的卷积层添加model weight
if model_weight == True:
modelWeight.addConv(index="_stem",
dtype=input.dtype,
shape1=(64, input.shape[1], 7, 7),
shape2=(64, ),
optimizer=optimizer)
return pool1
def _conv_block(in_blob,
index,
filters,
conv_times,
optimizer,
model_weight,
bn=True):
conv_block = []
conv_block.insert(0, in_blob)
for i in range(conv_times):
conv_i = conv2d_layer(
name="conv{}".format(index),
input=conv_block[i],
filters=filters[index],
kernel_size=3,
strides=1,
bn=bn,
)
if model_weight == True:
modelWeight.addConv(index=index,
dtype=conv_block[i].dtype,
shape1=(filters[index], conv_block[i].shape[1],
3, 3),
shape2=(filters[index], ),
optimizer=optimizer)
# shape_weight=(filters[index], conv_block[i].shape[1], 3, 3)
# modelWeight.add("conv{}".format(index)+'-weight',conv_block[i].dtype,shape_weight)
# modelWeight.add("conv{}".format(index)+'-bias',conv_block[i].dtype,(filters,))
# modelWeight.add("conv{}".format(index)+'_bn-gamma',conv_block[i].dtype,(filters,))
# modelWeight.add("conv{}".format(index)+'_bn-beta',conv_block[i].dtype,(filters,))
# modelWeight.add("conv{}".format(index)+'_bn-moving_variance',conv_block[i].dtype,(filters,))
# modelWeight.add("conv{}".format(index)+'_bn-moving_mean',conv_block[i].dtype,(filters,))
conv_block.append(conv_i)
index += 1
return conv_block
def residual_block(input, index, i, filter1, filter2, filter3, strides_init,
bn, model_weight, optimizer):
# if strides_init != 1 or block_name == "res2_0":
# #一个conv2d_affine(conv, bn, relu层)
# shortcut = conv2d_affine(input, block_name + "_branch1", 1, 1, filter3, 1, strides_init)
# else:
# shortcut = input
#对输入做变换，使得和三层oncv的输出shape相同，可以相加
shortcut = conv2d_affine(input,
"conv_shortcut" + str(index) + "_" + str(i),
filter3, 3, strides_init, bn)
#shortcut层添加model weight
if model_weight == True:
modelWeight.addConv(index="_shortcut" + str(index) + "_" + str(i),
dtype=input.dtype,
shape1=(filter3, input.shape[1], 3, 3),
shape2=(filter3, ),
optimizer=optimizer)
#三个conv2d_affine(conv, bn, relu层)
bottleneck = bottleneck_transformation(input, filter1, filter2, filter3,
strides_init, bn, model_weight,
optimizer)
# print(bottleneck.shape, shortcut.shape, strides_init, i)
return flow.nn.relu(bottleneck + shortcut)
def bottleneck_transformation(input, filter1, filter2, filter3, strides, bn,
model_weight, optimizer):
global NAME_NUMBER
a = conv2d_affine(
input,
"conv" + str(NAME_NUMBER),
filter1,
1,
1,
bn,
activation="Relu",
)
#添加conv的model weight
if model_weight == True:
modelWeight.addConv(index=NAME_NUMBER,
dtype=input.dtype,
shape1=(filter1, input.shape[1], 1, 1),
shape2=(filter1, ),
optimizer=optimizer)
NAME_NUMBER += 1
b = conv2d_affine(
a,
"conv" + str(NAME_NUMBER),
filter2,
3,
strides,
bn,
activation="Relu",
)
#添加conv的model weight
if model_weight == True:
modelWeight.addConv(index=NAME_NUMBER,
dtype=a.dtype,
shape1=(filter2, a.shape[1], 3, 3),
shape2=(filter2, ),
optimizer=optimizer)
NAME_NUMBER += 1
c = conv2d_affine(b, "conv" + str(NAME_NUMBER), filter3, 1, 1, bn)
#添加conv的model weight
if model_weight == True:
modelWeight.addConv(index=NAME_NUMBER,
dtype=b.dtype,
shape1=(filter3, b.shape[1], 1, 1),
shape2=(filter3, ),
optimizer=optimizer)
NAME_NUMBER += 1
# print(a.shape, b.shape, c.shape, strides)
return c
def lenet(images, cfg, optimizer, trainable=True, need_transpose=False,
training=True, wd=1.0/32768, model_weight=True, bn=True):
if need_transpose:
images = flow.transpose(images, name="transpose", perm=[0, 3, 1, 2])
conv0 = conv2d_layer(name="conv0", input=images, filters=cfg[0], kernel_size=5,
padding="VALID", strides=1, bn=bn)
pool0 = flow.nn.max_pool2d(conv0, 2, 2, "VALID", "NCHW", name="pool0")
conv1 = conv2d_layer(name="conv1", input=pool0, filters=cfg[1], kernel_size=5,
padding="VALID", strides=1, bn=bn)
pool1 = flow.nn.max_pool2d(conv1, 2, 2, "VALID", "NCHW", name="pool1")
pool1 = flow.reshape(pool1, [pool1.shape[0], -1])
# pool1 = flow.reshape(images, [images.shape[0], -1])
dense0 = flow.layers.dense(
inputs=pool1,
units=cfg[2],
activation=flow.nn.relu,
use_bias=True,
kernel_initializer=flow.random_normal_initializer(mean=0, stddev=0.1),
trainable=trainable,
name="dense0")
dense1 = flow.layers.dense(
inputs=dense0,
units=cfg[3],
activation=flow.nn.relu,
use_bias=True,
kernel_initializer=flow.random_normal_initializer(mean=0, stddev=0.1),
trainable=trainable,
name="dense1")
dense2 = flow.layers.dense(
inputs=dense1,
units=cfg[4],
use_bias=True,
kernel_initializer=flow.random_normal_initializer(mean=0, stddev=0.1),
trainable=trainable,
name="dense2")
# flow.watch(fc8)
def getTypeAndShape(inputs,units):
in_shape = inputs.shape
in_num_axes = len(in_shape)
inputs = (flow.reshape(inputs, (-1, in_shape[-1])) if in_num_axes > 2 else inputs)
shape=(units, inputs.shape[1])
dtype=inputs.dtype
return shape,dtype
if model_weight == True:
modelWeight.addConv(index=0, dtype=conv0.dtype,
shape1=(cfg[0], images.shape[1], 5, 5), shape2=(cfg[0],),
optimizer=optimizer)
modelWeight.addConv(index=1, dtype=conv1.dtype,
shape1=(cfg[1], conv0.shape[1], 5, 5), shape2=(cfg[1],),
optimizer=optimizer)
shape_list = []
dtype_list = []
shape_weight, dtype = getTypeAndShape(pool1, cfg[2])
shape_list.append(shape_weight)
dtype_list.append(dtype)
shape_weight, dtype = getTypeAndShape(dense0, cfg[3])
shape_list.append(shape_weight)
dtype_list.append(dtype)
shape_weight, dtype = getTypeAndShape(dense1, cfg[4])
shape_list.append(shape_weight)
dtype_list.append(dtype)
modelWeight.addDense(dtype_old=dtype_list, shape=shape_list,
optimizer=optimizer, dense_num=3)
return dense2