Is there a plan to open source the trained weights of MaskRCNN?
hpanwar08 opened this issue · 13 comments
Thank you this large dataset. Is there a plan to open source the trained weights and code in keras or pytorch?
Hi @hpanwar08 Thanks for your interests. Yes we do. We are going to migrate the data to another platform, which also supports sharing models. We are under internal legal assess now. Hopefully we can get approved soon. Please stay tuned :)
Hi @hpanwar08 Thanks for your interests. Yes we do. We are going to migrate the data to another platform, which also supports sharing models. We are under internal legal assess now. Hopefully we can get approved soon. Please stay tuned :)
Looking forward to meeting with faster data-getting method and shared-model :)
@zhxgj HI，is the model ready to publish? How much time did you spend on training this model (what's the setting)? Is there any accelerating skills? Now I try to train the model, but find it may take me many days with only 4GPUs. Thanks.
@zhxgj HI，is the model ready to publish? How much time did you spend on training this model (what's the setting)? Is there any accelerating skills? Now I try to train the model, but find it may take me many days with only 4GPUs. Thanks.
I am training on half of the training data on matterport maskrcnn on 1 GPU, it has been running for days. 1 epoch on half of the training data takes 42 hrs.
@zhxgj HI，is the model ready to publish? How much time did you spend on training this model (what's the setting)? Is there any accelerating skills? Now I try to train the model, but find it may take me many days with only 4GPUs. Thanks.
It took me about a week to train the models on 8 K80 GPUs. The models are still under legal review and unfortunately I do not have timeline right now. But I think it should be soon. I will follow up with our legal team to accelerate things.
Waiting for the pretrained model weights . Hope we will get it soon :)
Waiting for the pretrained model weights . Hope we will get it soon :)
Thanks @monuminu I will keep following up with our legal team
I have trained Mask RCNN using detectron2 on half of the training data (143k images) on a single GPU with resnet101 backbone initialized with coco weights. My results are close to the ones mentioned in the paper.
However model struggles with small objects.
@zhxgj do you have the metrics for small, medium, large object. I am not able to find them in the paper.
I have trained Mask RCNN using detectron2 on half of the training data (143k images) on a single GPU with resnet101 backbone initialized with coco weights. My results are close to the ones mentioned in the paper.
However model struggles with small objects.
@zhxgj do you have the metrics for small, medium, large object. I am not able to find them in the paper.
Hi @hpanwar08, I got similar results on small objects. Below are my more detailed validation results.
INFO json_dataset_evaluator.py: 241: ~~~~ Mean and per-category AP @ IoU=[0.50,0.95] ~~~~
INFO json_dataset_evaluator.py: 242: 91.0
INFO json_dataset_evaluator.py: 250: 91.6
INFO json_dataset_evaluator.py: 250: 84.0
INFO json_dataset_evaluator.py: 250: 88.6
INFO json_dataset_evaluator.py: 250: 96.0
INFO json_dataset_evaluator.py: 250: 94.9
INFO json_dataset_evaluator.py: 251: ~~~~ Summary metrics ~~~~
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.910
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.964
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.944
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.344
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.775
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.944
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.523
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.916
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.927
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.399
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.804
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.956
INFO json_dataset_evaluator.py: 218: Wrote json eval results to: /dccstor/ddig/peter/logs/e2e_mask_rcnn_X-101-64x4d-FPN_1x/model_iter179999/test/medline_val/generalized_rcnn/detection_results.pkl
INFO task_evaluation.py: 62: Evaluating bounding boxes is done!
INFO task_evaluation.py: 105: Evaluating segmentations
INFO json_dataset_evaluator.py: 88: Writing segmentation results json to: /dccstor/ddig/peter/logs/e2e_mask_rcnn_X-101-64x4d-FPN_1x/model_iter179999/test/medline_val/generalized_rcnn/segmentations_medline_val_results.json
Loading and preparing results...
DONE (t=3.53s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *segm*
DONE (t=51.81s).
Accumulating evaluation results...
DONE (t=4.55s).
INFO json_dataset_evaluator.py: 241: ~~~~ Mean and per-category AP @ IoU=[0.50,0.95] ~~~~
INFO json_dataset_evaluator.py: 242: 86.7
INFO json_dataset_evaluator.py: 250: 88.6
INFO json_dataset_evaluator.py: 250: 73.6
INFO json_dataset_evaluator.py: 250: 80.3
INFO json_dataset_evaluator.py: 250: 95.9
INFO json_dataset_evaluator.py: 250: 94.8
INFO json_dataset_evaluator.py: 251: ~~~~ Summary metrics ~~~~
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.867
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.964
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.934
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.297
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.702
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.915
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.508
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.885
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.895
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.351
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.753
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.935
INFO json_dataset_evaluator.py: 137: Wrote json eval results to: /dccstor/ddig/peter/logs/e2e_mask_rcnn_X-101-64x4d-FPN_1x/model_iter179999/test/medline_val/generalized_rcnn/segmentation_results.pkl
INFO task_evaluation.py: 66: Evaluating segmentations is done!
INFO task_evaluation.py: 181: copypaste: Dataset: medline_val
INFO task_evaluation.py: 183: copypaste: Task: box
INFO task_evaluation.py: 186: copypaste: AP,AP50,AP75,APs,APm,APl
INFO task_evaluation.py: 187: copypaste: 0.9100,0.9635,0.9444,0.3436,0.7750,0.9441
INFO task_evaluation.py: 183: copypaste: Task: mask
INFO task_evaluation.py: 186: copypaste: AP,AP50,AP75,APs,APm,APl
INFO task_evaluation.py: 187: copypaste: 0.8667,0.9636,0.9339,0.2971,0.7024,0.9150
Pre-trained Faster-RCNN model and Mask-RCNN model are released.
Any plan to release the weights for detectron2?
Or at least a suggestion about how to use detectron2 with these weights?
Thank you
Pre-trained Faster-RCNN model and Mask-RCNN model are released.
Any plan to release the weights for detectron2?
Or at least a suggestion about how to use detectron2 with these weights?Thank you
We plan to re-train the models on PubLayNet using Detectron2. The NVIDIA driver on our GPU cluster is out-dated and does not support Detectron2. We do not have the permission to upgrade the driver, but we are actively seeking a solution at the moment.
To use Detectron model in Detectron2, I saw someone posted a solution here. I have not tested it, but you can have a try.