CIFAR10 LeNet5 Variation 2: GELU + Dropout Layer
This repository contains our best variation of the original LeNet5 architecture adapted for CIFAR-10, but we will use its architecture and train it on CIFAR-100 this time. The model consists of two convolutional layers followed by two fully connected layers a dropout layer (p=0.5) and a final fully connected layer, using linear (GELU) activations, extending variation 1, and Kaiming uniform initialization. It is trained with a batch size of 32 using the Adam optimizer (learning rate 0.001) and CrossEntropyLoss. In our experiments, this model achieved a test loss of 0.0572 and a top-1 accuracy of 43.08% on CIFAR-100.
Model Details
- Architecture: 2 Convolutional Layers, 2 Fully Connected Layers, 1 Dropout Layer, 1 Final Fully Connected Layer.
- Activations: GELU.
- Weight Initialization: Kaiming Uniform.
- Optimizer: Adam (lr=0.001).
- Loss Function: CrossEntropyLoss.
- Dataset: CIFAR-100.
Usage
Load this model in PyTorch to fine-tune or evaluate on CIFAR-100 using your training and evaluation scripts.
Feel free to update this model card with further training details, benchmarks, or usage examples.