请求将指定Keras卷积神经网络转换为PyTorch Sequential定义

Let's fix your PyTorch model step by step, matching the exact architecture and parameter count of your Keras model. First, here's the corrected PyTorch Sequential implementation:

import torch.nn as nn

model = nn.Sequential(
    # Corresponding to Keras Input(shape=(64,64,1))
    # PyTorch expects input shape: (batch_size, channels, height, width)
    nn.Conv2d(in_channels=1, out_channels=32, kernel_size=4),
    nn.ReLU(),
    # Keras MaxPooling2D defaults to stride=pool_size, so this matches
    nn.MaxPool2d(kernel_size=2),
    # Input channel must match the previous layer's output channel (32)
    nn.Conv2d(in_channels=32, out_channels=16, kernel_size=4),
    nn.ReLU(),
    nn.MaxPool2d(kernel_size=2),
    # Keras Dense(10) acts on each spatial position's channels → equivalent to 1x1 conv
    nn.Conv2d(in_channels=16, out_channels=10, kernel_size=1),
    nn.ReLU(),
    # Keras Dense(1) same as above: 1x1 conv to get single output per spatial position
    nn.Conv2d(in_channels=10, out_channels=1, kernel_size=1),
    nn.Sigmoid()
)

Key Fixes & Explanations #

Let's break down where your original code went wrong, and how this matches your Keras model:

1. Incorrect Conv2d Input Channel
   Your second nn.Conv2d(1, 16, 4) had the wrong input channel count. The first Conv2d outputs 32 channels, so the next layer must accept 32 input channels (not 1) to maintain dimension consistency.

2. Understanding Keras Dense Layer Behavior
   In Keras, when a Dense layer receives a 3D tensor ((batch, H, W, C)), it applies the fully connected transformation to each spatial position's channel set independently, preserving the H/W dimensions. This is exactly what a 1x1 convolution does in PyTorch (since PyTorch uses (batch, C, H, W) order). Using nn.Conv2d with kernel_size=1 matches the parameter count and behavior perfectly:

   • Keras Dense(10): 16*10 +10 =170 params → PyTorch Conv2d(16,10,1): (1*1*16 +1)*10=170 params
   • Keras Dense(1): 10*1 +1=11 params → PyTorch Conv2d(10,1,1): (1*1*10 +1)*1=11 params

3. Input Shape Handling
   PyTorch uses channel-first tensor order (batch, channels, H, W) vs Keras' channel-last (batch, H, W, channels). You don't need to explicitly define input shape in Sequential, but you can test with a sample input:

   import torch
   test_input = torch.randn(1, 1, 64, 64)  # (batch=1, channels=1, H=64, W=64)
   output = model(test_input)
   print(output.shape)  # Should be torch.Size([1, 1, 13, 13]) → matches Keras' (None,13,13,1) (just channel order swapped)
   

4. MaxPool2d Stride Confusion
   Keras' MaxPooling2D(pool_size=(2,2)) defaults to stride=pool_size, so it uses a stride of 2. In PyTorch, nn.MaxPool2d(2) also defaults to stride=2, so it's identical to nn.MaxPool2d(2,2)—no need to overcomplicate it!

Parameter Count Verification #

Adding up the params of this PyTorch model gives exactly 8,933, matching your Keras model's total parameter count.

内容的提问来源于stack exchange，提问作者Chris Palmer