Huge accuracy gap between candle and pytorch on CIFAR-10 simple CNN

[tsubakisakura]

Issue: Huge Accuracy Gap Between Candle and PyTorch on CIFAR-10 CNN

Hi, I've been comparing Candle and PyTorch implementations of a simple CNN on CIFAR-10. Despite trying to replicate the same architecture and training settings, there's a large accuracy gap between the two:

• PyTorch reaches ~82% train accuracy after 20 epochs.
• Candle stays around ~36% train accuracy, and even starts to degrade after epoch 10.

Repositories & Code

I've created a minimal public repository for comparison:
https://github.com/tsubakisakura/compare-candle-pytorch

Implementation

• CNN architecture: 2 conv layers → ReLU → MaxPool → Dropout(0.25) → 2 FC layers
• Loss:
  • PyTorch: nn.CrossEntropyLoss()
  • Candle: candle_nn::loss::cross_entropy(logits, target)
• Optimizer:
  • PyTorch: AdamW(lr=1e-3, weight_decay=0.01)
  • Candle: AdamW with same parameters
• Data preprocessing:
  • PyTorch: use torchvision.datasets
  • Candle: use candle_datasets::vision::cifar::load

Logs

PyTorch

[Epoch 20] Train Loss: 0.4952 Acc: 0.8259 | Test Loss: 0.8723 Acc: 0.7241

Candle

[Epoch 20] Train Loss: 1.7769 Acc: 0.3603 | Test Loss: 1.7041 Acc: 0.3903

Question

I assume Candle is working correctly, but the results differ so much that I'm beginning to doubt whether I may have made a mistake in how I wrote the model or training loop.

Am I missing something in terms of how to correctly express the model or loss function in Candle?
Any suggestions or insight would be greatly appreciated.

Thanks a lot in advance!