This repo holds the implementation code of the paper:

FishNet: A Versatile Backbone for Image, Region, and Pixel Level Prediction , Shuyang Sun, Jiangmiao Pang, Jianping Shi, Shuai Yi, Wanli Ouyang, NeurIPS 2018.

FishNet was used as a key component for winning the 1st place in COCO Detection Challenge 2018.

Note that the results released here are a bit better than what we have reported in the paper.

• Python 3.6.x
• PyTorch 0.4.0+

Note: We apply weight decay to all weights and biases instead of just the weights of the convolution layers.

To train FishNet-150 with 8 GPUs and batch size 256, simply run

python main.py --config "cfgs/fishnet150.yaml" IMAGENET_ROOT_PATH

Models trained without tricks

Models trained with cosine lr schedule (200 epochs) and label smoothing

To load these models, e.g. FishNet150, you need to first construct your FishNet150 structure like:

from models.network_factory import fishnet150
model = fishnet150()

and then you can load the weights from the pre-trained checkpoint by:

checkpoint = torch.load(model_path)  #  model_path: your checkpoint path, e.g. checkpoints/fishnet150.tar
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])

Note that you do NOT need to decompress the model using the tar command. The model you download from the cloud could be loaded directly.

• Update our arxiv paper.
• Release pre-train models.
• Train the model with more training tricks.

If you find our research useful, please cite the paper:

@inproceedings{sun2018fishnet,
  title={FishNet: A Versatile Backbone for Image, Region, and Pixel Level Prediction},
  author={Sun, Shuyang and Pang, Jiangmiao and Shi, Jianping and Yi, Shuai and Ouyang, Wanli},
  booktitle={Advances in Neural Information Processing Systems},
  pages={760--770},
  year={2018}
}

You can contact Shuyang Sun by sending email to [email protected]