The performance of detection in VOC #6
Comments
|
@jancylee Did you directly download the provided pre-trained weights? Please provide your experiment scripts and I will see what's the problem. |
|
I pretrain the model by myself (coco-800ep-resnet50). |
|
Compared to your codes, I didn't change any settings in the pretrained process, and I only change the batchsize and base lr in the detection process. |
|
Please make sure you have followed the instructions in the readme: https://github.com/WXinlong/DenseCL#extracting-backbone-weights You have to 1) extract the backbone weights and 2) convert it to detectron2 format. |
|
I did transform it. By the way, when I train the detection process from scratch (without loading pretrained model), the AP is only 12.8. |
|
It looks like the problems are in your detection experiments, not pre-trained weights. You are suggested to first reproduce the detection results using either random init. or supervised pretrained model, i.e., to make sure you can get the same results with the same settings. |
|
I use 4 gpus, batchsize=4, base lr =0.005, iter=240004, steps = 180004,22000*4 |
|
@jancylee Please directly copy your training config.yaml instead of several parameters, we should make sure you set the correct parameters (e.g. input format: RGB, pixel mean/std and etc). |
|
The config.yaml of VOC07&12 object dection:
|
|
And when I use my preteained model (coco-800ep-resnet50, and the training settings is same to yours) to fine tune the object dection in VOC(the training settings is as above), it's only AP=48.16, compared to AP=51.19(your pretrained model and my fine-tuned object detection) and AP=56.7(result of your paper ). |
|
@jancylee Can you try to use the official model (e.g. mocov2) to reproduce their voc detection performance? https://github.com/open-mmlab/OpenSelfSup/blob/master/docs/MODEL_ZOO.md |
|
Thanks a lot. I know the reason, when I use the batchsize=16, it achieves 56.54(conmpared to the paper: 56.7), but this is based on your pretrained model provided by your github website. While when I train the pretrained model by myself, it only achieves 49.78. I completely use your code and follow your pretraining settings. I don't konw why. By the way, I used the moco-v2 pretrained model to fine tune, it can achieves 53.92, which is almost the same as the your paper results. |
|
I can provide you the training settings later. But the only difference is the workers_per_gpu, I set 8 while you set 4, which is only influence the data loading speed. |
|
@jancylee So the batch size plays a important role to train detection, can you provide the config for training DenseCL? |
2021-03-25 09:31:12,868 - openselfsup - INFO - Environment info:sys.platform: linux
TorchVision: 0.5.0
|
|
@jancylee Could you please upload the model that you pretrained to google cloud / baidu cloud? We will check it for you. |
|
链接：https://pan.baidu.com/s/1tUUzN7UPPKfOoSKHhIC1Bw |
|
@jancylee Have you extracted the backbone weights using tools/extract_backbone_weights.py before fine-tuning object detection? |
|
@jancylee We have tested the model that you train and there is no problem.
Please make sure you have followed the instructions in the readme: |
|
Thank you very much. It was the "extract the backbone weights" worked. |
(8gpus) When I use the pretrained network with coco-800ep-resnet50 to do the detection task with VOC, the "AP" is only 44.76, while you can achieve 56.7. I don't konw why the gap is so large. Note that I change the batchsize from 16 to 8, and as a result, the base lr is set from 0.02 to 0.01.
The text was updated successfully, but these errors were encountered: