open-mmlab / mmfashion

Open-source toolbox for visual fashion analysis based on PyTorch
https://open-mmlab.github.io/
Apache License 2.0
1.26k stars 285 forks source link

While training the Fashion Segmentation model, is there a way to train only bbox detection and turn off segmentation #84

Open kartikwar opened 4 years ago

kartikwar commented 4 years ago

Hi I have a fashion dataset that I want to train. The issue is I have only bbox coordinates with me. I dont want to train segmentation . I am using the configuration deepfashion/mask_rcnn_r50_fpn_15e_deepfashion.py. I am able to train by providing random segmentation points. But the issue is this would impact loss and hence the model (since model is also taking loss for segmentation predictions) so this would affect my detection results as well for the trained model. Is there any way in ehich I can turn off the segmentation mask pipeline in this model. Also if not are there any other alternatives to achieve this ?

Thanks