我正在编写一个模型来为一个学校项目执行图像分类。我有10个类，我在我的模型上批量加载图像：
import torch
import torch.nn as nn
import torch.nn.functional as F
# *****START CODE
class ConvNet(nn.Module):
def __init__(self, in_ch, out_ch):
super(ConvNet, self).__init__()
"""
Number of layers should be exactly same as in the provided JSON.
Do not use any grouping function like Sequential
"""
self.Layer_001 = nn.Conv2d(in_channels=in_ch, out_channels=64, kernel_size=3, padding=1)
self.Layer_002 = nn.ReLU()
self.Layer_003 = nn.MaxPool2d(kernel_size=2,stride=2)
self.Layer_004 = nn.Conv2d(in_channels=64, out_channels=113, kernel_size=3, padding=1)
self.Layer_005 = nn.ReLU()
self.Layer_006 = nn.MaxPool2d(kernel_size=2,stride=2)
self.Layer_007 = nn.Conv2d(in_channels=113, out_channels=248, kernel_size=3, padding=1)
self.Layer_008 = nn.ReLU()
self.Layer_009 = nn.Conv2d(in_channels=248, out_channels=248, kernel_size=3, padding=1)
self.Layer_010 = nn.ReLU()
self.Layer_011 = nn.MaxPool2d(kernel_size=2,stride=2)
self.Layer_012 = nn.Conv2d(in_channels=248, out_channels=519, kernel_size=3, padding=1)
self.Layer_013 = nn.ReLU()
self.Layer_014 = nn.Conv2d(in_channels=519, out_channels=519, kernel_size=3, padding=1)
self.Layer_015 = nn.ReLU()
self.Layer_016 = nn.MaxPool2d(kernel_size=2,stride=2)
self.Layer_017 = nn.Conv2d(in_channels=519, out_channels=519, kernel_size=3, padding=1)
self.Layer_018 = nn.ReLU()
self.Layer_019 = nn.Conv2d(in_channels=519, out_channels=519, kernel_size=3, padding=1)
self.Layer_020 = nn.ReLU()
self.Layer_021 = nn.MaxPool2d(kernel_size=2,stride=2)
self.Layer_022 = nn.AdaptiveAvgPool2d((1,1))
self.Layer_023 = nn.Dropout(p=0.501816987002085)
self.Layer_024 = nn.Linear(in_features=519,out_features=2317)
self.Layer_025 = nn.ReLU()
self.Layer_026 = nn.Linear(in_features=2317, out_features=3018)
self.Layer_027 = nn.Linear(in_features=3018, out_features=3888)
self.Layer_028 = nn.ReLU()
self.Layer_029 = nn.Linear(in_features=3888, out_features=out_ch)
def forward(self, x):
x = self.Layer_001(x)
#print(x.shape)
x = self.Layer_002(x)
#print(x.shape)
x = self.Layer_003(x)
#print(x.shape)
x = self.Layer_004(x)
#print(x.shape)
x = self.Layer_005(x)
#print(x.shape)
x = self.Layer_006(x)
#print(x.shape)
x = self.Layer_007(x)
#print(x.shape)
x = self.Layer_008(x)
#print(x.shape)
x = self.Layer_009(x)
#print(x.shape)
x = self.Layer_009(x)
#print(x.shape)
x = self.Layer_010(x)
#print(x.shape)
x = self.Layer_011(x)
#print(x.shape)
x = self.Layer_012(x)
#print(x.shape)
x = self.Layer_013(x)
#print(x.shape)
x = self.Layer_014(x)
#print(x.shape)
x = self.Layer_015(x)
#print(x.shape)
x = self.Layer_016(x)
#print(x.shape)
x = self.Layer_017(x)
#print(x.shape)
x = self.Layer_018(x)
#print(x.shape)
x = self.Layer_019(x)
#print(x.shape)
x = self.Layer_020(x)
#print(x.shape)
x = self.Layer_021(x)
#print(x.shape)
x = self.Layer_022(x)
#print(x.shape)
x = self.Layer_023(x)
#print(x.shape)
#x = nn.Flatten(x)
##print(x.shape)
x = self.Layer_024(x)
#print(x.shape)
x = self.Layer_025(x)
#print(x.shape)
x = self.Layer_026(x)
#print(x.shape)
x = self.Layer_027(x)
#print(x.shape)
x = self.Layer_028(x)
#print(x.shape)
output = self.Layer_029(x)
print(output.shape)
return output
# *****END CODE当我运行它的时候，我发现了一个层间错误
它将返回以下错误：
RuntimeError: mat1 and mat2 shapes cannot be multiplied (8304x1 and 519x2317)我知道这是一个形状问题，但我正在学习，不明白它发生在哪里......我试图重建这个架构：
'Layer_001': {'input': 3,
'kernel_size': 3,
'output': 64,
'padding': 1,
'type': 'Conv2d'},
'Layer_002': {'type': 'ReLU'},
'Layer_003': {'kernel_size': 2, 'stride': 2, 'type': 'MaxPool2d'},
'Layer_004': {'input': 64,
'kernel_size': 3,
'output': 113,
'padding': 1,
'type': 'Conv2d'},
'Layer_005': {'type': 'ReLU'},
'Layer_006': {'kernel_size': 2, 'stride': 2, 'type': 'MaxPool2d'},
'Layer_007': {'input': 113,
'kernel_size': 3,
'output': 248,
'padding': 1,
'type': 'Conv2d'},
'Layer_008': {'type': 'ReLU'},
'Layer_009': {'input': 248,
'kernel_size': 3,
'output': 248,
'padding': 1,
'type': 'Conv2d'},
'Layer_010': {'type': 'ReLU'},
'Layer_011': {'kernel_size': 2, 'stride': 2, 'type': 'MaxPool2d'},
'Layer_012': {'input': 248,
'kernel_size': 3,
'output': 519,
'padding': 1,
'type': 'Conv2d'},
'Layer_013': {'type': 'ReLU'},
'Layer_014': {'input': 519,
'kernel_size': 3,
'output': 519,
'padding': 1,
'type': 'Conv2d'},
'Layer_015': {'type': 'ReLU'},
'Layer_016': {'kernel_size': 2, 'stride': 2, 'type': 'MaxPool2d'},
'Layer_017': {'input': 519,
'kernel_size': 3,
'output': 519,
'padding': 1,
'type': 'Conv2d'},
'Layer_018': {'type': 'ReLU'},
'Layer_019': {'input': 519,
'kernel_size': 3,
'output': 519,
'padding': 1,
'type': 'Conv2d'},
'Layer_020': {'type': 'ReLU'},
'Layer_021': {'kernel_size': 2, 'stride': 2, 'type': 'MaxPool2d'},
'Layer_022': {'output': 'COMPUTE', 'type': 'AdaptiveAvgPool2d'},
'Layer_023': {'p': 0.501816987002085, 'type': 'Dropout'},
'Layer_024': {'input': 'COMPUTE', 'output': 2317, 'type': 'Linear'},
'Layer_025': {'type': 'ReLU'},
'Layer_026': {'input': 2317, 'output': 'COMPUTE', 'type': 'Linear'},
'Layer_027': {'input': 3018, 'output': 3888, 'type': 'Linear'},
'Layer_028': {'type': 'ReLU'},
'Layer_029': {'input': 3888, 'output': 'COMPUTE', 'type': 'Linear'}型
我认为我的错误来自“Layer_022”：{“输出”：“计算”，“类型”：“AdaptiveAvgPool 2d”}或从“图层_024”上的此位置：{“输入”：“计算”，“输出”：2317，“类型”：'Linear'}但我不确定...我的意思是我真的不知道如何计算这些值，这就是为什么我请求一些帮助：）
我已经尝试将519放在“Layer_022”的输出上：{“输出”：“计算”，“类型”：'AdaptiveAvgPool 2d'}，我还尝试了不同的值，如（2）（2，2）...
1条答案
按热度按时间aiazj4mn1#
你需要放一个
nn.Flatten()进去。你已经在代码中创建了一个flated层，但是你需要像放其他层一样放进去。一个类似的方法是在你的forward调用中调用x = self.Layer_023(x).view([x.shape[0],-1])，以获得[batch x feats]的大小。例如：