FBNet Pre-trained Weights for a Custom Dataset

facebookresearch / maskrcnn-benchmark

Fast, modular reference implementation of Instance Segmentation and Object Detection algorithms in PyTorch.

MIT License

9.29k stars 2.5k forks source link

FBNet Pre-trained Weights for a Custom Dataset #1050

Open mrteera opened 5 years ago

mrteera commented 5 years ago

❓ Questions and Help

I'm trying to use FBNet as a pre-trained weights for my dataset (ROI_BOX_HEAD: NUM_CLASSES: 3). But it cannot load the weights using this utils _load_c2_pickled_weights() https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/maskrcnn_benchmark/utils/c2_model_loading.py#L192

Do you have any suggestion to load pre-trained FBNet weights for a custom dataset?

Dorozhko-Anton commented 5 years ago

_load_c2_pickled_weights() is used to load only Caffe2 checkpoints in .pkl format https://github.com/facebookresearch/maskrcnn-benchmark/blob/24c8c90efdb7cc51381af5ce0205b23567c3cd21/maskrcnn_benchmark/utils/checkpoint.py#L132-L134

To load pre-trained weights into the model with different ROI_BOX_HEAD.NUM_CLASSES refer to https://github.com/facebookresearch/maskrcnn-benchmark#finetuning-from-detectron-weights-on-custom-datasets

To use trim_detectron_model.py suggested in README.md with .pth file check out #15 , more precisely https://github.com/facebookresearch/maskrcnn-benchmark/issues/15#issuecomment-498575917

mrteera commented 5 years ago

@Dorozhko-Anton Thank you. I successfully convert .pth model file using trim_detectron_model. These are my changed. I'll update again when I train a model with this pre-train weights.

_d =  torch.load(DETECTRON_PATH)

newdict['model'] = removekey(_d['model'],
['module.roi_heads.box.predictor.cls_score.bias', 'module.roi_heads.box.predictor.cls_score.weight', 'module.roi_heads.box.predictor.bbox_pred.bias', 'module.roi_heads.box.predictor.bbox_pred.weight'])