search

dbolya / yolact

A simple, fully convolutional model for real-time instance segmentation.
MIT License
4.98k stars 1.33k forks source link

Training only object detection layers with YOLACT #451

Open saitarslanboun opened 4 years ago

saitarslanboun commented 4 years ago

Hi,

I have downloaded your code, and customized for my task little bit.

I am using a custom dataset iterator, which does not necessarily include masking annotation. But all the samples in the iterator includes object detection annotations.

In case the batches include only object detection annotation, I don't want to update segmentation layers.

We have the following parameters to back propagate through.

backbone. proto_net. maskiou_net. fpn. prediction_layers.upfeature prediction_layers.0.bbox_layer. prediction_layers.0.conf_layer. prediction_layers.0.mask_layer. semantic_seg_conv.*

From the parameters above, I do not update protocol_net., maskiou_net., prediction_layers.upfeature, and semantic_seg_conv. layers, when I would like to train only object detection layers.

However, I am not sure with the prediction_layers.upfeature* parameters. Should I also update them?

anmolduainter commented 3 years ago

Hi, I also have data containing bbox and bbox+mask and I want to use that. Have you managed to train yolact with this type of data?