JAX to PyTorch Converted Model (ResNet-10)

It's done in context of porting HIL-SERL paper code (https://hil-serl.github.io/) to Lerobot (https://github.com/Lerobot/lerobot). The HF doesn't have ResNet-10 model, which could be pretty usefult for robotics tasks because of it's small size. This model is converted from JAX to PyTorch, and the weights are preserved.

Model Description

[Brief description of the original model and its purpose]

This model is a PyTorch port of the original JAX implementation. The conversion maintains the original model's architecture and weights while making it accessible to PyTorch users. The original model is from https://github.com/rail-berkeley/hil-serl/blob/7d17d13560d85abffbd45facec17c4f9189c29c0/serl_launcher/serl_launcher/utils/train_utils.py#L103.

Model Details

• Original Framework: JAX
• Target Framework: PyTorch
• Model Architecture: ResNet-10 (4-stage ResNet with basic blocks)
• Original Model: HIL-SERL ResNet-10
• Total Parameters: 4,905,792 (~4.9M parameters)
• Hidden Sizes: [64, 128, 256, 512]
• Input: 3-channel RGB images (128x128)
• Embedding Size: 64

Conversion Process

This model was converted using an automated JAX to PyTorch conversion pipeline, ensuring:

• Weight preservation
• Architecture matching
• Numerical stability

Code

https://github.com/helper2424/resnet10

Usage

from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained("lilkm/resnet10")

Citation

@misc{resnet10,
   title = "Resnet10",
   author = "Eugene Mironov and Khalil Meftah and Adil Zouitine and Michel Aractingi and Ke Wang",
   month = jan,
   year = "2025",
   address = "Online",
   publisher = "Hugging Face",
   url = "https://huggingface.co/helper2424/resnet10",
}