Training doesn't converge for Mapillary Vistas Dataset training with MaskRCNN

[Morganh]

@edit_or

Is it possible to train a few last layers of resnet50 in TAO for MaskRCNN?

Refer to “freeze_blocks” in MaskRCNN - NVIDIA Docs

@edit_or

Which software can I use to view annotations (json) and images of dataset?

What do you mean viewing the json file?

@edit_or

Their loss is also around 2.

Below is an example. When train on COCO dataset, the spec is inside this topic.

Poor metric results after retraining maskrcnn using TLT notebook

Update: After finetune the blog’s warmup_steps from 0 to 1000, the AP is 0.31344375, which is closed to blog’s AP( 0.334154785 ). Note that, I am using 8gpus(v100) training. Upload spec file for reference. maskrcnn_blog_finetune.txt (2.1 KB)