请预先安装torchinfo:
pip install torchinfo
然后导入库:
from torchinfo import summary
使用resnet进行测试:
import torchvision
model = torchvision.models.resnet152()
summary(model, (1, 3, 224, 224), depth=3)
出现如下界面表示成功:
就可以查看每一层的输出的shape和每一层的参数数量。
查看LSTM参数
from torchinfo import summary
batch, seq_len, insize = 64, 10, 56
input = torch.zeros((batch, seq_len, insize))
model = CLSTM(input_size=insize, channels=[64, 128, 256], hidden_size=150, out_size=1, num_layers=2)
state = model.init_state(batch)
y = model(input, state)
print(model, model(input, state)[0].shape)
summary(model, input.shape)
请注意CLSTM是我自己的模型,可以替换为你自己的LSTM模型,summary传参的时候,第一个参数为模型,第二个参数为input的shape,后面还有一些参数,例如input_data、batch_dim等,可以在源码中查看
那么对于LSTM来说,请将模型的输入的state默认参数设置为None,不然传参的时候会报错,提示传入的参数不够,需要传入state参数。
注意state=none
def forward(self, x: torch.Tensor, state=None):
# input shape:[batch_size,length,in_size]
# input shape of cnn:[batch_size,in_size,length]
out = self.block1(x.permute(0, 2, 1))
out = self.block2(out)
# shape change:[batch_size,in_size,length]->[batch_size,length,in_size]
out = self.block3(out).permute(0, 2, 1)
out, state = self.rnn(out, state)
out = self.dense(out)
return out, state