Model code different from lua

Hi, I found a few differences from the Lua model.

- ConvBlock's batchnorm has no bias (affine=False) in Lua
- FullyConnected(1x1 conv) has batchnorm(affine=False) in Lua
- The first pooling is max_pool in Lua while pytorch model uses avg_pool

Compare the PyTorch model implementation in this repository with the original Lua model's architecture. Look at the ConvBlock, FullyConnected (1x1 conv) layers, and the first pooling layer. Check the model definition files, likely in the main module or a models/ directory. Verify the parameters for batchnorm (affine flag) and pooling type. Run existing tests or a simple inference script to see if the current behavior matches expectations after adjustments.

Rédigé par le modèle d'indexation à partir du texte de l'issue.