Я пытаюсь перейти с keras на pytorch и получаю противоречивое поведение. Прежде всего, я замечаю, что у меня другое количество параметров, но я также замечаю, что обучение keras сходится намного быстрее и плавнее, чем pytorch. Я сделал что-то не так? Кто-нибудь может помочь? Соответствующий код ниже.
У меня есть следующая простая архитектура в keras:
def identity_block(input_tensor, units):
x = layers.Dense(units, kernel_regularizer=reg)(input_tensor)
x = layers.BatchNormalization()(x)
x = layers.Activation('relu')(x)
x = layers.Dense(units, kernel_regularizer=reg)(x)
x = layers.BatchNormalization()(x)
x = layers.Activation('relu')(x)
x = layers.Dense(units, kernel_regularizer=reg)(x)
x = layers.BatchNormalization()(x)
x = layers.add([x, input_tensor])
x = layers.Activation('relu')(x)
return x
def dens_block(input_tensor, units, reps=2):
x = input_tensor
for _ in range(reps):
x = layers.Dense(units, kernel_regularizer=reg)(x)
x = layers.BatchNormalization()(x)
x = layers.Activation('relu')(x)
x = layers.Dense(units, kernel_regularizer=reg)(x)
x = layers.BatchNormalization()(x)
shortcut = layers.Dense(units, kernel_regularizer=reg)(input_tensor)
shortcut = layers.BatchNormalization()(shortcut)
x = layers.add([x, shortcut])
x = layers.Activation('relu')(x)
return x
def resnet_block(input_tensor, width=16):
x = dens_block(input_tensor, width)
x = identity_block(x, width)
x = identity_block(x, width)
return x
def RegResNet(input_size=8, reps=3, initial_weights=None, width=16, num_gpus=0, lr=1e-4):
input_layer = layers.Input(shape=(input_size,))
x = input_layer
for _ in range(reps):
x = resnet_block(x, width)
x = layers.BatchNormalization()(x)
x = layers.Dense(1, activation=None)(x)
model = models.Model(inputs=input_layer, outputs=x)
return model
Что дает сводку, которую я не смог уместить, и следующий подсчет параметров.
==================================================================================================
Total params: 9,905
Trainable params: 8,913
Non-trainable params: 992
==================================================================================================
Ниже приведен мой pytorch-перевод той же архитектуры:
class _DenseBlock(nn.Module):
def __init__(self, input_size, output_size):
super(_DenseBlock, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.linear = nn.Linear(self.input_size, self.output_size)
self.bnorm = nn.BatchNorm1d(self.output_size)
def forward(self, x):
x = self.linear(x)
x = self.bnorm(x)
x = nn.ReLU()(x)
return x
class DenseBlock(nn.Module):
def __init__(self, input_size, output_size, repetitions=2):
super(DenseBlock, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.repetitions = repetitions
self.dense_blocks = nn.ModuleList([_DenseBlock(self.input_size, self.output_size)]
+ [_DenseBlock(self.output_size, self.output_size)
for _ in range(self.repetitions - 1)])
self.b_norm1 = nn.BatchNorm1d(self.output_size)
self.b_norm2 = nn.BatchNorm1d(self.output_size)
self.linear_1 = nn.Linear(self.output_size, self.output_size)
self.linear_2 = nn.Linear(self.input_size, self.output_size)
def forward(self, x):
identity = x
for l in self.dense_blocks:
x = l(x)
x = self.linear_1(x)
x = self.b_norm1(x)
shortcut = self.linear_2(identity)
shortcut = self.b_norm2(shortcut)
x = shortcut + x
x = nn.ReLU()(x)
return x
class IdentityBlock(nn.Module):
def __init__(self, input_size, output_size):
super(IdentityBlock, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.linear_1 = nn.Linear(self.input_size, self.output_size)
self.linear_2 = nn.Linear(self.input_size, self.output_size)
self.linear_3 = nn.Linear(self.input_size, self.output_size)
self.bnorm_1 = nn.BatchNorm1d(self.input_size)
self.bnorm_2 = nn.BatchNorm1d(self.output_size)
self.bnorm_3 = nn.BatchNorm1d(self.output_size)
def forward(self, x):
input_tensor = x
x = self.linear_1(x)
x = self.bnorm_1(x)
x = nn.ReLU()(x)
x = self.linear_2(x)
x = self.bnorm_2(x)
x = nn.ReLU()(x)
x = self.linear_3(x)
x = self.bnorm_3(x)
x = x + input_tensor
x = nn.ReLU()(x)
return x
class ResnetBlock(nn.Module):
def __init__(self, input_size, output_size):
super(ResnetBlock, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.dense_block = DenseBlock(self.input_size, self.output_size)
self.identity_block_1 = IdentityBlock(self.output_size, self.output_size)
self.identity_block_2 = IdentityBlock(self.output_size, self.output_size)
def forward(self, x):
x = self.dense_block(x)
x = self.identity_block_1(x)
x = self.identity_block_2(x)
return x
class RegResNet(nn.Module):
def __init__(self, input_size, block_width, repititions=3):
super(RegResNet, self).__init__()
self.input_size = input_size
self.repititions = repititions
self.block_width = block_width
self.resnet_blocks = nn.ModuleList([ResnetBlock(self.input_size, self.block_width)]
+ [ResnetBlock(self.block_width, self.block_width)
for _ in range(self.repititions - 1)])
self.bnorm = nn.BatchNorm1d(self.block_width)
self.out = nn.Linear(self.block_width, 1)
def forward(self, x):
for layer in self.resnet_blocks:
x = layer(x)
x = self.bnorm(x)
x = self.out(x)
return x
Используя библиотеку torchsummary, я получил сводку (которая мне не подошла) и следующий счетчик параметров
================================================================
Total params: 8,913
Trainable params: 8,913
Non-trainable params: 0
================================================================
