def train_test_model(logdir, hparams):
filter_kernel_2 = json.loads(hparams['filter_kernel_2'])
inputs = keras.Input(shape=(x_train[0].shape[0], x_train[0].shape[1], 1), name='c3_input')
model = layers.Conv2D(filters=int(hparams['filter_1']), kernel_size=int(hparams['kernel_1']), activation='relu')(inputs)
model = layers.MaxPooling2D(pool_size=2)(model)
model = layers.Dropout(0.4)(model)
if int(filter_kernel_2[0]) > 0:
model = layers.Conv2D(filters=int(filter_kernel_2[0]), kernel_size=int(filter_kernel_2[1]), activation='relu')(model)
model = layers.MaxPooling2D(pool_size=2)(model)
model = layers.Dropout(0.4)(model)
model = layers.GlobalAvgPool2D()(model)
model = layers.Dense(hparams['units_1'], activation='relu')(model)
if int(hparams['units_2']) > 0:
model = layers.Dense(int(hparams['units_2']), activation='relu')(model)
model = layers.Dense(1, activation='sigmoid')(model)
model = Model(inputs=inputs, outputs=model, name='c3_model')
model.compile(optimizer=optimizers.Adam(learning_rate=hparams['lr'], decay=0.001), loss='binary_crossentropy', metrics=['accuracy'])
cb = [
callbacks.TensorBoard(log_dir=logdir)
]
history = model.fit(x_train, y_train, validation_data=(x_test, y_test), batch_size=64, epochs=1000, callbacks=cb, verbose=0)
return model, history
def _net(block,
im_width=9,
im_height=9,
im_channel=204,
num_classes=16,
include_top=True):
# tensorflow中的tensor通道排序是NHWC
# (None, 9, 9, 204)
# change
input_image = layers.Input(shape=(im_height, im_width, im_channel),
dtype="float32")
# x = layers.Conv2D(filters=64, kernel_size=7, strides=2, padding="SAME", use_bias=False, name="conv1")(input_image)
x = block()(input_image)
# print("include_top", include_top)
if include_top:
x = layers.GlobalAvgPool2D()(x) # pool + flatten
x = layers.Dense(256, activation='relu')(x)
x = layers.Dropout(0.5)(x)
x = layers.Dense(512, activation='relu')(x)
x = layers.Dropout(0.5)(x)
predict = layers.Dense(16, activation='softmax')(x)
else:
predict = x
model = Model(inputs=input_image, outputs=predict)
return model
def channel_attention(input_xs, reduction_ratio): # input_xs (None,50,50,64)
# 判断输入数据格式，是channels_first还是channels_last
channel_axis = 1 if k.image_data_format() == "channels_first" else 3
# get channel 彩色图片为 3
channel = int(input_xs.shape[channel_axis]) # 64
maxpool_channel = kl.GlobalMaxPooling2D()(input_xs) # (None,64) 全局最大池化
maxpool_channel = kl.Reshape((1, 1, channel))(maxpool_channel) # ( None,1,1,64)
avgpool_channel = kl.GlobalAvgPool2D()(input_xs) # (None,64) 全局平均池化
avgpool_channel = kl.Reshape((1, 1, channel))(avgpool_channel) # (None,1,1,64)
# 权值共享
dense_one = kl.Dense(units=int(channel * reduction_ratio), activation='relu', kernel_initializer='he_normal', use_bias=True, bias_initializer='zeros')
dense_two = kl.Dense(units=int(channel), activation='relu', kernel_initializer='he_normal', use_bias=True, bias_initializer='zeros')
# max path
mlp_1_max = dense_one(maxpool_channel) # (None,1,1,32)
mlp_2_max = dense_two(mlp_1_max) # (None,1,1,64)
mlp_2_max = kl.Reshape(target_shape=(1, 1, int(channel)))(mlp_2_max) # (None,1,1,64)
# avg path
mlp_1_avg = dense_one(avgpool_channel) # (None,1,1,32)
mlp_2_avg = dense_two(mlp_1_avg) # (None,1,1,64)
mlp_2_avg = kl.Reshape(target_shape=(1, 1, int(channel)))(mlp_2_avg) # (None,1,1,64)
channel_attention_feature = kl.Add()([mlp_2_max, mlp_2_avg]) # (None,1,1,64)
channel_attention_feature = kl.Activation('sigmoid')(channel_attention_feature) # (None,1,1,64)
multiply_channel_input = kl.Multiply()([channel_attention_feature, input_xs]) # (None,50,50,64)
return multiply_channel_input
def _resnet(block, blocks_num, im_width=224, im_height=224, num_classes=1000, include_top=True):
# tensorflow中的tensor通道排序是NHWC
# (None, 224, 224, 3)
input_image = layers.Input(shape=(im_height, im_width, 3), dtype="float32")
x = layers.Conv2D(filters=64, kernel_size=7, strides=2,
padding="SAME", use_bias=False, name="conv1")(input_image)
x = layers.BatchNormalization(momentum=0.9, epsilon=1e-5, name="conv1/BatchNorm")(x)
x = layers.ReLU()(x)
x = layers.MaxPool2D(pool_size=3, strides=2, padding="SAME")(x)
x = _make_layer(block, x.shape[-1], 64, blocks_num[0], name="block1")(x)
x = _make_layer(block, x.shape[-1], 128, blocks_num[1], strides=2, name="block2")(x)
x = _make_layer(block, x.shape[-1], 256, blocks_num[2], strides=2, name="block3")(x)
x = _make_layer(block, x.shape[-1], 512, blocks_num[3], strides=2, name="block4")(x)
if include_top:
x = layers.GlobalAvgPool2D()(x) # pool + flatten
x = layers.Dense(num_classes, name="logits")(x)
predict = layers.Softmax()(x)
else:
predict = x
model = Model(inputs=input_image, outputs=predict)
return model
def __init__(self, num_videos, finetuning=False, e_finetuning=None):
super(Discriminator, self).__init__()
self.relu = layers.LeakyReLU()
self.num_videos = num_videos
# in 6*224*224
# self.pad = Padding(224) # out 256*256*6
self.resDown1 = ResBlockDown(6, 64) # out 128*128*64
self.resDown2 = ResBlockDown(64, 128) # out 64*64*128
self.resDown3 = ResBlockDown(128, 256) # out 32*32*256
self.self_att = SelfAttention(256) # out 32*32*256
self.resDown4 = ResBlockDown(256, 512) # out 16*16*512
self.resDown5 = ResBlockDown(512, 512) # out 8*8*512
self.resDown6 = ResBlockDown(512, 512) # out 4*4*512
self.res = ResBlockD(512) # out 512*4*4
# self.sum_pooling = nn.AdaptiveAvgPool2d((1, 1)) # out 1*1*512
self.sum_pooling = layers.GlobalAvgPool2D() # out 1*1*512
# if not finetuning:
# print('Initializing Discriminator weights')
# if not os.path.isdir(self.path_to_Wi):
# os.mkdir(self.path_to_Wi)
# for i in tqdm(range(num_videos)):
# if not os.path.isfile(self.path_to_Wi + '/W_' + str(i) + '/W_' + str(i) + '.tar'):
# w_i = torch.rand(512, 1)
# os.mkdir(self.path_to_Wi + '/W_' + str(i))
# torch.save({'W_i': w_i}, self.path_to_Wi + '/W_' + str(i) + '/W_' + str(i) + '.tar')
self.finetuning = finetuning
self.e_finetuning = e_finetuning
def __init__(self, label_num):
super(DenseNet, self).__init__()
channel_num = 64
# first component
self.layers_lst.extend([
layers.Conv2D(filters=channel_num,
kernel_size=7,
strides=2,
padding='same'),
layers.BatchNormalization(),
layers.MaxPool2D(pool_size=3, strides=2, padding='same')
])
# second: dense and transition block
growth_rate = 32
num_conv_in_dense_block = [4, 4, 4, 4]
for i, num_conv in enumerate(num_conv_in_dense_block):
self.layers_lst.append(
DenseBlock(num_channel=growth_rate, num_conv=num_conv))
channel_num += num_conv * growth_rate
if i != len(num_conv_in_dense_block) - 1:
channel_num = channel_num // 2
self.layers_lst.append(
TransitionBlock(num_channel=channel_num))
# third: final component
self.layers_lst.extend([
layers.BatchNormalization(),
layers.Activation('relu'),
layers.GlobalAvgPool2D(),
layers.Dense(units=label_num)
])
def __init__(self,
channels,
reduction_ratio=16,
num_layers=1,
data_format="channels_last",
**kwargs):
super(ChannelGate, self).__init__(**kwargs)
self.data_format = data_format
mid_channels = channels // reduction_ratio
self.pool = nn.GlobalAvgPool2D(data_format=data_format, name="pool")
self.flatten = nn.Flatten()
self.init_fc = DenseBlock(in_channels=channels,
out_channels=mid_channels,
data_format=data_format,
name="init_fc")
self.main_fcs = SimpleSequential(name="main_fcs")
for i in range(num_layers - 1):
self.main_fcs.children.append(
DenseBlock(in_channels=mid_channels,
out_channels=mid_channels,
data_format=data_format,
name="fc{}".format(i + 1)))
self.final_fc = nn.Dense(units=channels,
input_dim=mid_channels,
name="final_fc")
def __init__(self,
channels,
init_block_channels,
bottleneck,
conv1_stride,
ordinary_init=False,
bends=None,
in_channels=3,
in_size=(224, 224),
classes=1000,
data_format="channels_last",
**kwargs):
super(ResNetD, self).__init__(**kwargs)
self.in_size = in_size
self.classes = classes
self.multi_output = (bends is not None)
self.data_format = data_format
self.features = MultiOutputSequential(name="features")
if ordinary_init:
self.features.add(ResInitBlock(
in_channels=in_channels,
out_channels=init_block_channels,
data_format=data_format,
name="init_block"))
else:
init_block_channels = 2 * init_block_channels
self.features.add(SEInitBlock(
in_channels=in_channels,
out_channels=init_block_channels,
data_format=data_format,
name="init_block"))
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
stage = tf.keras.Sequential(name="stage{}".format(i + 1))
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if ((j == 0) and (i != 0) and (i < 2)) else 1
dilation = (2 ** max(0, i - 1 - int(j == 0)))
stage.add(ResUnit(
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
padding=dilation,
dilation=dilation,
bottleneck=bottleneck,
conv1_stride=conv1_stride,
data_format=data_format,
name="unit{}".format(j + 1)))
in_channels = out_channels
if self.multi_output and ((i + 1) in bends):
stage.do_output = True
self.features.add(stage)
self.features.add(nn.GlobalAvgPool2D(
data_format=data_format,
name="final_pool"))
self.output1 = nn.Dense(
units=classes,
input_dim=in_channels,
name="output1")
def build_model(x, y):
nx = np.empty((len(x), 28, 28, 1))
for i in range(len(x)):
nx[i] = x[i].reshape((28, 28, 1))
sigma_function = 'relu'
model = keras.Sequential([
keras.Input(shape=(28, 28, 1)),
# 28*28*1 -> 28*28*8
layers.Conv2D(8, 2, (1, 1), padding='same', activation=sigma_function),
layers.LayerNormalization(),
# 28*28*8 -> 19*19*16
layers.Conv2D(16, 2, (2, 2), padding='same',
activation=sigma_function),
layers.LayerNormalization(),
# 19*19*16 -> 19*19*32
layers.Conv2D(32, 2, (1, 1), padding='same',
activation=sigma_function),
layers.LayerNormalization(),
layers.GlobalAvgPool2D(),
layers.Dense(10, bias_initializer='one', activation="softmax"),
])
model.summary()
adam = tf.keras.optimizers.Adam()
model.compile(loss='categorical_crossentropy',
optimizer=adam,
metrics=['accuracy'])
return model, nx, y
def _resnet(block,
blocks_num,
im_width=224,
im_height=224,
num_classes=1000,
include_top=True):
#block：如果18和34层传basicblock，如果是其他的就传bottleneck
#blocknum是一个列表，是不同卷积层中残差结构的数量（参见resnet的那张表），比如如果是18层就是[2,2,2,2]
#输入宽度，高度，需要分类的数量
# tensorflow中的tensor通道排序是NHWC
# (None, 224, 224, 3)
input_image = layers.Input(shape=(im_height, im_width, 3), dtype="float32")
x = layers.Conv2D(filters=64,
kernel_size=7,
strides=2,
padding="SAME",
use_bias=False,
name="conv1")(input_image)
x = layers.BatchNormalization(momentum=0.9,
epsilon=1e-5,
name="conv1/BatchNorm")(x)
x = layers.ReLU()(x)
x = layers.MaxPool2D(pool_size=3, strides=2, padding="SAME")(x)
#这里是卷积层1加上第一个池化做完了
#每做一个makelayer，就生成一整个卷积层
#x.shape是上一层输出的shape，[batch,height,weight,channel],所以是channel
#第三个参数是对应层的第一个卷积核的通道数，第四个就是该层中残差模块的数量
x = _make_layer(block, x.shape[-1], 64, blocks_num[0], name="block1")(x)
x = _make_layer(block,
x.shape[-1],
128,
blocks_num[1],
strides=2,
name="block2")(x)
x = _make_layer(block,
x.shape[-1],
256,
blocks_num[2],
strides=2,
name="block3")(x)
x = _make_layer(block,
x.shape[-1],
512,
blocks_num[3],
strides=2,
name="block4")(x)
if include_top:
x = layers.GlobalAvgPool2D()(x) # 平均池化pool + 展平处理flatten
x = layers.Dense(num_classes, name="logits")(x) #全连接
predict = layers.Softmax()(x) #转化为概率分布
else:
predict = x
model = Model(inputs=input_image, outputs=predict)
return model
def __init__(self, data_format="channels_last", **kwargs):
super(GlobalAvgMaxPool2D, self).__init__(**kwargs)
self.axis = get_channel_axis(data_format)
self.avg_pool = nn.GlobalAvgPool2D(data_format=data_format,
name="avg_pool")
self.max_pool = nn.GlobalMaxPool2D(data_format=data_format,
name="max_pool")
def Squeeze_Excitation_Module(input, filters, reduction_ratio, name="SE"):
sq = layers.GlobalAvgPool2D(name=name+"_Squeeze")(input)
ex1 = layers.Dense(filters//reduction_ratio, activation='relu', name=name+"_Excitation_1")(sq)
ex2 = layers.Dense(filters, activation='sigmoid', name=name+"_Excitation_2")(ex1)
ex = layers.Reshape([1, 1, filters], name=name+"_Reshape")(ex2)
out = layers.Multiply(name=name+"_Multiply")([input, ex])
return out
def CNN(num_classes=10, **kwargs):
model = tf.keras.Sequential([
layers.Conv2D(64, 3, 2, 'same', activation=tf.nn.relu, **kwargs),
layers.Conv2D(64, 3, 1, 'same', activation=tf.nn.relu, **kwargs),
layers.Conv2D(128, 3, 2, 'same', activation=tf.nn.relu, **kwargs),
layers.Conv2D(128, 3, 1, 'same', activation=tf.nn.relu, **kwargs),
layers.Conv2D(num_classes, 3, 1, 'same', **kwargs),
layers.GlobalAvgPool2D()
])
return model
def build_vgg16(MyConv2d, ishape, nclass, name):
input = KL.Input(ishape) # 32
x = KL.Conv2D(64, (3, 3),
strides=(1, 1),
padding='same',
activation='relu',
kernel_initializer='he_normal',
name='block1_conv1')(input)
x = MyConv2d(64, (3, 3), name='block1_conv2')(x)
x = KL.MaxPooling2D((2, 2),
strides=(2, 2),
padding='same',
name='block1_pool')(x)
x = MyConv2d(128, (3, 3), name='block2_conv1')(x)
x = MyConv2d(128, (3, 3), name='block2_conv2')(x)
x = KL.MaxPooling2D((2, 2),
strides=(2, 2),
padding='same',
name='block2_pool')(x)
x = MyConv2d(256, (3, 3), name='block3_conv1')(x)
x = MyConv2d(256, (3, 3), name='block3_conv2')(x)
x = MyConv2d(256, (3, 3), name='block3_conv3')(x)
x = KL.MaxPooling2D((2, 2),
strides=(2, 2),
padding='same',
name='block3_pool')(x)
x = MyConv2d(512, (3, 3), name='block4_conv1')(x)
x = MyConv2d(512, (3, 3), name='block4_conv2')(x)
x = MyConv2d(512, (3, 3), name='block4_conv3')(x)
x = KL.MaxPooling2D((2, 2),
strides=(2, 2),
padding='same',
name='block4_pool')(x) # 2
x = MyConv2d(512, (3, 3), name='block5_conv1')(x)
x = MyConv2d(512, (3, 3), name='block5_conv2')(x)
x = MyConv2d(512, (3, 3), name='block5_conv3')(x)
x = KL.MaxPooling2D((2, 2),
strides=(2, 2),
padding='same',
name='block5_pool')(x) # 2*2 -> 1*1
x = KL.GlobalAvgPool2D()(x)
# output = KL.Dense(nclass, activation='softmax', name='predictions')(x)
# XXX in case `mixed_precision` being used
output = KL.Dense(nclass, activation=None, name='predictions')(x)
output = KL.Activation('softmax', dtype='float32')(output)
kmdl = KM.Model(input, output, name=name)
return kmdl
def mobilenetv1(shape, num_classes, activation="relu", alpha=1.0):
"""Obtain a mobilenet V1 model
Args:
shape: the shape of the input tensor
num_classes: the number of outputs
activation: the activation function of each later
alpha: hyper parameter for adjusting the width of convolution layers
"""
model = models.Sequential()
model.add(layers.InputLayer(input_shape=shape))
def add_conv_block(channels, strides, kernel_size=3):
channels = int(channels * alpha)
model.add(
layers.Conv2D(
channels, kernel_size=kernel_size, use_bias=False, padding="same"
)
)
model.add(layers.BatchNormalization())
model.add(layers.Activation(activation))
def add_dw_sep_block(channels, strides):
channels = int(channels * alpha)
model.add(
layers.DepthwiseConv2D(
kernel_size=3, strides=strides, use_bias=False, padding="same"
)
)
model.add(layers.BatchNormalization())
model.add(layers.Activation(activation))
add_conv_block(channels, strides=1, kernel_size=1)
add_conv_block(32, 2)
model_shapes_channel_strides = [
(64, 1),
(128, 2),
(128, 1),
(256, 2),
(256, 1),
(512, 2),
*[(512, 1) for _ in range(5)],
(1024, 2),
(1024, 2),
]
for c, s in model_shapes_channel_strides:
add_dw_sep_block(c, s)
model.add(layers.GlobalAvgPool2D())
model.add(layers.Dense(1000, activation=activation))
model.add(layers.Dense(num_classes, activation="softmax", name="softmax"))
return model
def __init__(self, num_blocks, **kwargs):
super(ResNet, self).__init__(**kwargs)
self.conv = layers.Conv2D(64, kernel_size=7, strides=2, padding='same')
self.bn = layers.BatchNormalization()
self.relu = layers.Activation('relu')
self.mp = layers.MaxPool2D(pool_size=3, strides=2, padding='same')
self.resnet_block1 = ResnetBlock(64, num_blocks[0], first_block=True)
self.resnet_block2 = ResnetBlock(128, num_blocks[1])
self.resnet_block3 = ResnetBlock(256, num_blocks[2])
self.resnet_block4 = ResnetBlock(512, num_blocks[3])
self.gap = layers.GlobalAvgPool2D()
self.fc = layers.Dense(units=10, activation=activations.softmax)
def get_training_model():
model = tf.keras.Sequential([
layers.Conv2D(16, (5, 5), activation="relu", input_shape=(28, 28, 1)),
layers.MaxPooling2D(pool_size=(2, 2)),
layers.Conv2D(32, (5, 5), activation="relu"),
layers.MaxPooling2D(pool_size=(2, 2)),
layers.Dropout(0.2),
layers.GlobalAvgPool2D(),
layers.Dense(128, activation="relu"),
layers.Dense(10, activation="softmax"),
])
return model
def FlipOutCNN(num_classes, **kwargs):
Conv2DFlipOut = tfp.layers.Convolution2DFlipout
model = tf.keras.Sequential([
Conv2DFlipOut(64, 3, 2, 'same', activation=tf.nn.relu, **kwargs),
Conv2DFlipOut(64, 3, 1, 'same', activation=tf.nn.relu, **kwargs),
Conv2DFlipOut(128, 3, 2, 'same', activation=tf.nn.relu, **kwargs),
Conv2DFlipOut(128, 3, 1, 'same', activation=tf.nn.relu, **kwargs),
Conv2DFlipOut(num_classes, 3, 1, 'same', **kwargs),
layers.GlobalAvgPool2D()
])
return model
def ReparamCNN(num_classes, **kwargs):
Conv2DReparam = tfp.layers.Convolution2DReparameterization
model = tf.keras.Sequential([
Conv2DReparam(64, 3, 2, 'same', activation=tf.nn.relu, **kwargs),
Conv2DReparam(64, 3, 1, 'same', activation=tf.nn.relu, **kwargs),
Conv2DReparam(128, 3, 2, 'same', activation=tf.nn.relu, **kwargs),
Conv2DReparam(128, 3, 1, 'same', activation=tf.nn.relu, **kwargs),
Conv2DReparam(num_classes, 3, 1, 'same'),
layers.GlobalAvgPool2D()
])
return model
def __init__(
self,
num_classes,
blocks,
m_layers,
channels,
feature_dim=512,
loss='softmax',
**kwargs
):
super(OSNetTF, self).__init__()
num_blocks = len(blocks)
assert num_blocks == len(m_layers)
assert num_blocks == len(channels) - 1
self.loss = loss
self.channels = channels
# convolutional backbone
self.conv1 = ConvLayerTF(3, channels[0], 7, stride=2, padding="SAME") # original padding=3
# self.maxpool = nn.MaxPool2d(3, stride=2, padding="SAME") # original padding=1
self.maxpool = layers.MaxPool2D(3, strides=2, padding="SAME") # original padding=1
self.conv2 = self._make_layer(
blocks[0],
m_layers[0],
channels[0],
channels[1],
reduce_spatial_size=True,
)
self.conv3 = self._make_layer(
blocks[1],
m_layers[1],
channels[1],
channels[2],
reduce_spatial_size=True
)
self.conv4 = self._make_layer(
blocks[2],
m_layers[2],
channels[2],
channels[3],
reduce_spatial_size=False
)
self.conv5 = Conv1x1TF(channels[3], channels[3])
self.global_avgpool = layers.GlobalAvgPool2D()
# fully connected layer
self.fc = self._construct_fc_layer(
feature_dim, channels[3], dropout_p=None
)
# identity classification layer
# self.classifier = nn.Linear(self.feature_dim, num_classes)
# self.classifier = layers.Dense(num_classes)
self.classifier = layers.Dense(num_classes, kernel_initializer=tf.keras.initializers.RandomNormal(0, 0.01))
def __init__(self,
res_unit,
nunit_per_block,
dilate_config=None,
num_classes=1000,
**kwargs):
super(ResNet, self).__init__(**kwargs)
self.dilation_rate = 1
if dilate_config is None:
dilate_config = [False, False, False, False]
if len(dilate_config) != 4:
raise "`dilate_config` should be None or a 4-element tuple, got %d" % len(
dilate_config)
if dilate_config[0] is not False:
raise "the 1st elemnt of `dilate_config` has to be False"
self.conv = layers.Conv2D(64,
7,
strides=2,
padding="same",
use_bias=False,
name="conv")
self.bn = layers.BatchNormalization(epsilon=1.001e-5, name="bn")
self.relu = layers.Activation("relu", name="relu")
self.maxpool = layers.MaxPool2D(pool_size=3,
strides=2,
padding="same",
name="maxpool")
self.block0 = self._make_block(res_unit,
64,
nunit_per_block[0],
dilate=dilate_config[0],
name="block0")
self.block1 = self._make_block(res_unit,
128,
nunit_per_block[1],
stride=2,
dilate=dilate_config[1],
name="block1")
self.block2 = self._make_block(res_unit,
256,
nunit_per_block[2],
stride=2,
dilate=dilate_config[2],
name="block2")
self.block3 = self._make_block(res_unit,
512,
nunit_per_block[3],
stride=2,
dilate=dilate_config[3],
name="block3")
self.avgpool = layers.GlobalAvgPool2D(name="global_avgpool")
self.dense = layers.Dense(num_classes, name="dense")
def __init__(self,
channels,
init_block_channels,
bottleneck,
conv1_stride,
dilated=False,
in_channels=3,
in_size=(224, 224),
classes=1000,
data_format="channels_last",
**kwargs):
super(ResNetA, self).__init__(**kwargs)
self.in_size = in_size
self.classes = classes
self.data_format = data_format
self.features = SimpleSequential(name="features")
self.features.add(SEInitBlock(
in_channels=in_channels,
out_channels=init_block_channels,
data_format=data_format,
name="init_block"))
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
stage = SimpleSequential(name="stage{}".format(i + 1))
for j, out_channels in enumerate(channels_per_stage):
if dilated:
strides = 2 if ((j == 0) and (i != 0) and (i < 2)) else 1
dilation = (2 ** max(0, i - 1 - int(j == 0)))
else:
strides = 2 if (j == 0) and (i != 0) else 1
dilation = 1
stage.add(ResAUnit(
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
padding=dilation,
dilation=dilation,
bottleneck=bottleneck,
conv1_stride=conv1_stride,
data_format=data_format,
name="unit{}".format(j + 1)))
in_channels = out_channels
self.features.add(stage)
self.features.add(nn.GlobalAvgPool2D(
data_format=data_format,
name="final_pool"))
self.output1 = nn.Dense(
units=classes,
input_dim=in_channels,
name="output1")
def __init__(self,
block,
blocks_num,
num_classes=1000,
include_top=True,
**kwargs):
super(ResNet, self).__init__(**kwargs)
self.include_top = include_top
self.conv1 = layers.Conv2D(filters=64,
kernel_size=7,
strides=2,
padding="SAME",
use_bias=False,
name="conv1")
self.bn1 = layers.BatchNormalization(momentum=0.9,
epsilon=1.001e-5,
name="conv1/BatchNorm")
self.relu1 = layers.ReLU(name="relu1")
self.maxpool1 = layers.MaxPool2D(pool_size=3,
strides=2,
padding="SAME",
name="maxpool1")
self.block1 = self._make_layer(block,
True,
64,
blocks_num[0],
name="block1")
self.block2 = self._make_layer(block,
False,
128,
blocks_num[1],
strides=2,
name="block2")
self.block3 = self._make_layer(block,
False,
256,
blocks_num[2],
strides=2,
name="block3")
self.block4 = self._make_layer(block,
False,
512,
blocks_num[3],
strides=2,
name="block4")
if self.include_top:
self.avgpool = layers.GlobalAvgPool2D(name="avgpool1")
self.fc = layers.Dense(num_classes, name="logits")
self.softmax = layers.Softmax()
def SE_ResNet(height, width, num_class, res_block, blocks_list):
"""
ResNet网络结构，通过传入不同的残差块和重复的次数进行不同层数的ResNet构建
:param height: 网络输入宽度
:param width: 网络输入高度
:param num_class: 分类数量
:param res_block: 残差块单元
:param blocks_list: 每个残差单元重复的次数列表
:return:
"""
input_image = layers.Input(shape=(height, width, 3))
x = layers.Conv2D(filters=64,
kernel_size=7,
strides=2,
padding='SAME',
name='conv1',
use_bias=False)(input_image)
x = layers.BatchNormalization(name="conv1/BatchNorm")(x)
x = layers.ReLU()(x)
x = layers.MaxPool2D(pool_size=3,
strides=2,
padding='SAME',
name="max_pool")(x)
x = resblock_body(res_block, 64, blocks_list[0], strides=1,
name='conv2_x')(x)
x = resblock_body(res_block,
128,
blocks_list[1],
strides=2,
name='conv3_x')(x)
x = resblock_body(res_block,
256,
blocks_list[2],
strides=2,
name='conv4_x')(x)
x = resblock_body(res_block,
512,
blocks_list[3],
strides=2,
name='conv5_x')(x)
x = layers.GlobalAvgPool2D(name='avg_pool')(x)
x = layers.Dense(num_class, name="logits")(x)
outputs = layers.Softmax()(x)
model = models.Model(inputs=input_image, outputs=outputs)
model.summary()
return model
def create_mobilevit(num_classes=5):
inputs = keras.Input((image_size, image_size, 3))
x = layers.Rescaling(scale=1.0 / 255)(inputs)
# Initial conv-stem -> NV2 block.
x = conv_block(x, filters=16)
x = inverted_residual_block(x,
expanded_channels=16 * expansion_factor,
output_channels=16)
# Downsampling with MV2 block.
x = inverted_residual_block(x,
expanded_channels=16 * expansion_factor,
output_channels=24,
strides=2)
x = inverted_residual_block(x,
expanded_channels=24 * expansion_factor,
output_channels=24)
x = inverted_residual_block(x,
expanded_channels=24 * expansion_factor,
output_channels=24)
# First MV2 -> MobileViT block.
x = inverted_residual_block(x,
expanded_channels=24 * expansion_factor,
output_channels=48,
strides=2)
x = mobilevit_block(x, num_blocks=2, projection_dim=64)
# Second MV2 -> MobileViT block.
x = inverted_residual_block(x,
expanded_channels=64 * expansion_factor,
output_channels=64,
strides=2)
x = mobilevit_block(x, num_blocks=4, projection_dim=80)
# Third MV2 -> MobileViT block.
x = inverted_residual_block(x,
expanded_channels=80 * expansion_factor,
output_channels=80,
strides=2)
x = mobilevit_block(x, num_blocks=3, projection_dim=96)
x = conv_block(x, filters=320, kernel_size=1, strides=1)
# Classification head.
x = layers.GlobalAvgPool2D()(x)
outputs = layers.Dense(num_classes, activation="softmax")(x)
return keras.Model(inputs, outputs)
def _resnet(block,
blocks_num,
img_height=224,
img_width=224,
num_class=1000,
include_top=True):
input_image = layers.Input(shape=(img_height, img_width, 3),
dtype='float32')
x = layers.Conv2D(filters=64,
kernel_size=7,
strides=2,
padding='same',
use_bias=False,
name='conv1')(input_image)
x = layers.BatchNormalization(momentum=0.9,
epsilon=1e-5,
name='conv1/BatchNorm')(x)
x = layers.ReLU()(x)
x = layers.MaxPool2D(pool_size=3, strides=2, padding='same')(x)
x = _make_layer(block, x.shape[-1], 64, blocks_num[0], name='block1')(x)
x = _make_layer(block,
x.shape[-1],
128,
blocks_num[1],
strides=2,
name='block2')(x)
x = _make_layer(block,
x.shape[-1],
256,
blocks_num[2],
strides=2,
name='block3')(x)
x = _make_layer(block,
x.shape[-1],
512,
blocks_num[3],
strides=2,
name='block4')(x)
if include_top:
x = layers.GlobalAvgPool2D()(x)
x = layers.Dense(num_class, name='logits')(x)
predict = layers.Softmax()(x)
else:
predict = x
model = Model(inputs=input_image, outputs=predict)
return model
def build_model():
conv_base = VGG16(weights='imagenet', include_top=False, input_shape=(256, 256, 3))
conv_base.trainable = False
gap = layers.GlobalAvgPool2D()(conv_base.layers[-1].output)
dropout_1 = layers.Dropout(0.25)(gap)
dense1 = layers.Dense(512)(dropout_1)
batchNorm = layers.BatchNormalization()(dense1)
activation = layers.Activation(activation='relu')(batchNorm)
dropout_2 = layers.Dropout(0.25)(activation)
dense2 = layers.Dense(128, activation='relu')(dropout_2)
dropout_3 = layers.Dropout(0.25)(dense2)
y = layers.Dense(1, activation='sigmoid')(dropout_3)
model = models.Model(conv_base.inputs, y)
model.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])
return model
def build_model():
conv_base = VGG16(include_top=False, input_shape=(150, 150, 3))
conv_base.trainable = False
x = conv_base.input
gap = layers.GlobalAvgPool2D()(conv_base.output)
dense = layers.Dense(2)(gap)
batch = layers.BatchNormalization()(dense)
act = layers.Activation(activation='relu')(batch)
do = layers.Dropout(0.25)(act)
y = layers.Dense(1, activation='sigmoid')(do)
model = models.Model(x, y)
model.compile(optimizer=optimizers.RMSprop(learning_rate=0.0001, momentum=0.4), loss='binary_crossentropy', metrics=['accuracy'])
return model
def __init__(self, Layer_dim, num_classes=100):
super(ResNet, self).__init__()
self.stem = Sequential([layers.Conv2D(64, (3, 3), strides=1, padding='same'),
layers.BatchNormalization(),
layers.Activation('relu'),
layers.MaxPool2D(pool_size=(2, 2), strides=(1, 1), padding='same')
])
self.layer1 = self.build_resblock(64, Layer_dim[0])
self.layer2 = self.build_resblock(128, Layer_dim[1], stride=2)
self.layer3 = self.build_resblock(256, Layer_dim[2], stride=2)
self.layer4 = self.build_resblock(512, Layer_dim[3], stride=2)
self.avgpool = layers.GlobalAvgPool2D()
self.fc = layers.Dense(num_classes)
def __init__(self,
channels,
init_block_channels,
bottleneck,
dropout_rate=0.0,
in_channels=3,
in_size=(224, 224),
classes=1000,
data_format="channels_last",
**kwargs):
super(ResNeStA, self).__init__(**kwargs)
self.in_size = in_size
self.classes = classes
self.data_format = data_format
self.features = SimpleSequential(name="features")
self.features.add(SEInitBlock(
in_channels=in_channels,
out_channels=init_block_channels,
data_format=data_format,
name="init_block"))
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
stage = SimpleSequential(name="stage{}".format(i + 1))
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
stage.add(ResNeStAUnit(
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
bottleneck=bottleneck,
data_format=data_format,
name="unit{}".format(j + 1)))
in_channels = out_channels
self.features.add(stage)
self.features.add(nn.GlobalAvgPool2D(
data_format=data_format,
name="final_pool"))
self.output1 = SimpleSequential(name="output1")
if dropout_rate > 0.0:
self.output1.add(nn.Dropout(
rate=dropout_rate,
name="output1/dropout"))
self.output1.add(nn.Dense(
units=classes,
input_dim=in_channels,
name="output1/fc"))
