PyTorch转Keras的多层感知器(MLP)代码是否等价？代码转换正确性验证咨询

Hey there, let's walk through your MLP conversion and fix the key parts that need adjusting—your core approach is on the right track, but there are a couple of input shape details that need tweaking to match the original PyTorch behavior.

Key Issues in Your Current Keras Code #

• Redundant input_shape parameters: In Keras' Sequential model, you only need to define input_shape on the first layer. All subsequent layers automatically infer their input shape from the previous layer's output—adding input_shape to layer2, layer3, and layer4 is unnecessary and can cause shape mismatches.
• Incorrect input_shape dimension: Your PyTorch nn.Linear(size_in, hidden) expects inputs of shape (batch_size, size_in) (batch dimension first). In Keras, input_shape excludes the batch dimension, so (size_in,) is correct (not (size_in, 1)—that would force a 3D input tensor, which doesn't match your PyTorch setup).

Corrected Keras MLP Implementation #

from tensorflow import keras
from keras import layers

hidden = 128
def mlp(size_in, size_out, act=keras.layers.ReLU):
    return keras.Sequential(
        [
            layers.Dense(hidden, activation=None, name="layer1", input_shape=(size_in,)),
            act(),
            layers.Dense(hidden, activation=None, name="layer2"),
            act(),
            layers.Dense(hidden, activation=None, name="layer3"),
            act(),
            layers.Dense(size_out, activation=None, name="layer4")
        ])

Alternative: Explicit Input Layer (Equivalent) #

If you prefer more clarity (instead of relying on input_shape in the first Dense layer), you can explicitly define an Input layer—this is functionally identical to the code above, and matches the "explicit InputLayer" behavior you mentioned in your question:

from tensorflow import keras
from keras import layers

hidden = 128
def mlp(size_in, size_out, act=keras.layers.ReLU):
    return keras.Sequential(
        [
            layers.Input(shape=(size_in,)),
            layers.Dense(hidden, activation=None, name="layer1"),
            act(),
            layers.Dense(hidden, activation=None, name="layer2"),
            act(),
            layers.Dense(hidden, activation=None, name="layer3"),
            act(),
            layers.Dense(size_out, activation=None, name="layer4")
        ])

Why This Works #

• Both versions accept inputs of shape (None, size_in) (where None represents any batch size), which aligns perfectly with the PyTorch model's input expectations.
• The activation function usage (act()) is correct—this mirrors how you pass the activation class in PyTorch and instantiate it, which is the right pattern in Keras too.

内容的提问来源于stack exchange，提问作者Hirek