search

sahithchada / NOCS_PyTorch

NOCS reimplementation in PyTorch
MIT License
26 stars 4 forks source link

Question about the pretrained weight #5

Closed colin-de closed 6 months ago

colin-de commented 6 months ago

Hi! Thanks for the great work; I have some questions regarding the pre-trained weight provided in the repo. 

1. mask_rcnn_coco.pth      
2. resnet50_imagenet.pth
3. NOCS_Trained_1.pth      
4. NOCS_Trained_2.pth

what settings/experiments are those weights referring to? Thanks for your assistance!

bharathsivaram10 commented 6 months ago

Hi Colin,

Happy to see you find our code useful. The weights are the following:

1 is the pre-trained weights from which you can start training the network. You can see in train.py

2 is a set of weights for if you want to train just the mask rcnn on COCO data. I don't think we ever used this.

3 & 4 are trained sets of weights you can either use for NOCS testing since we have trained them, or you can load them and continue trying to train.

Hope this helps!

colin-de commented 6 months ago

Thanks for your information. If I'm using a dataset with different categories, is it recommended to train a model like mask_rcnn_coco.pthon new dataset first or can simply start with rand_weight option ?

using the pretrained weight from 1. (rand_weight=True) for training I got the inference result like this pred_coords much worse than the checkpoint from 4. NOCS_Trained_2.pth. pred_coords could you please let me know how to reproduce the training result. Many Thanks!

sahithchada commented 6 months ago

Hi Colin,

The inference results show that the mask head is not producing good segments in the image. Starting with pre-trained weights might ensure that the masks are well-formed. To reproduce the results, you should start with the model pre-trained on COCO data.