Pre-trained models

[xiaomengyc]

Can you please share your trained models? Thank you!

[jwyang]

@xiaomengyc, I will upload the pre-trained model today. stay tuned

[xiaomengyc]

Thanks!

By the way, I also encountered the exactly same problem when compiling the ROI polling layerhere, and I fixed it by changing the line 3 (CUDA_PATH=/usr/local/cuda/) in make.sh to CUDA_PATH=/usr/local/cuda/ and adding the environment variables
export PATH=/usr/local/cuda-8.0/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64:$LD_LIBRARY_PATH

[jwyang]

great, thanks! @xiaomengyc.

I added them into my bashrc, while @g5996706 might do not.

I will add them to make.sh, and remind in readme that these two should be exported.

thanks again!

BTW, I am retraining all the models with the most recent code. I already uploaded a trained resnet101 on pascal-voc (see the link in the table). Now I am training in other settings. All models on pascal voc will be available tonight. COCO might be two days later.

[xiaomengyc]

Thanks a lot!

[Cadene]

Thanks! Do you plan to train on visual genome as well? @jwyang

[jwyang]

@Cadene , yes, I will once I have enough GPUs, :). It should be very soon after I finish the retraining on coco.

[Cadene]

@jwyang I trained a VGG16 on Visual Genome for 20 epochs (it took me one week) and got a mAP of 4.4%.

I could share it. However, I just want to know if someone achieved a better mAP :)

[jwyang]

@Cadene great, how many object categories did you try?

[Cadene]

@jwyang as much as possible (2500 classes)

[jwyang]

@Cadene I guess that performance should be ok, I tried 1600 categories, and got around 10 mAP. Do you want to share your trained model, you can give me a link, I will add it to readme.

[Cadene]

@jwyang http://data.lip6.fr/cadene/faster-rcnn.pytorch/faster_rcnn_1_19_48611.pth

[jwyang]

@Cadene appreciate!

[jwyang]

@Cadene , could you also tell me the detailed hyper parameters for training the model, as shown in the tables?

[Cadene]

@jwyang So it was a VGG16 (not a Resnet101) trained on a single Tesla P100 with the default hyper parameters:

• lr 0.001
• lr_decay 5
• batch_size 4
• max_epoch 20
• 12707MiB / 16276MiB
• around 3h40m per epoch on trainset

(args.set_cfgs = ['ANCHOR_SCALES', '[4, 8, 16, 32]', 'ANCHOR_RATIOS', '[0.5,1,2]', 'MAX_NUM_GT_BOXES', '50'])

[jwyang]

@Cadene got it, thanks!