KaimingHe/deep-residual-networks 5512

Deep Residual Learning for Image Recognition

Faster R-CNN

KaimingHe/resnet-1k-layers 771

Deep Residual Networks with 1K Layers

A TensorFlow re-implementation of Momentum Contrast (MoCo): https://arxiv.org/abs/1911.05722

Caffe fork that supports SPP_net or faster R-CNN

A set of examples around pytorch in Vision, Text, Reinforcement Learning, etc.

issue commentfacebookresearch/moco

as the random guess is log(65536)

Hi @KaimingHe , I reimplemented MoCo with another DL framework and got the training loss stuck at 11.09, which is about log(65536). Is it a flag that the training is already failed?

comment created time in 18 hours

startedKaimingHe/deep-residual-networks

started time in 2 days

startedKaimingHe/deep-residual-networks

started time in 3 days

startedKaimingHe/deep-residual-networks

started time in 3 days

issue closedfacebookresearch/moco

Why the lr in main_lincls is 30?

HI, I wonder why the learning rate is set to 30 when traing classification ? I have never seen such a big learning rate.

`parser.add_argument('--lr', '--learning-rate', default=30., type=float, metavar='LR', help='initial learning rate', dest='lr')`

closed time in 3 days

lzyhhaissue commentfacebookresearch/moco

Why the lr in main_lincls is 30?

Answered in sec 4.1 in the paper.

comment created time in 3 days

issue commentfacebookresearch/moco

AssertionError: Default process group is not initialized

Please provide:

- what you did: do not modify the code and provide the exact command you run
- the full logs you observed

I have the same problem. I did not revise any code, and my running command is:

CUDA_VISIBLE_DEVICES=4,5,6,7 python train_net.py --config-file configs/pascal_voc_R_50_C4_24k_moco.yaml MODEL.WEIGHTS ./output.pkl

Do you have any advice? Very appreciate.

comment created time in 3 days

startedKaimingHe/deep-residual-networks

started time in 3 days

startedKaimingHe/deep-residual-networks

started time in 4 days

startedKaimingHe/deep-residual-networks

started time in 4 days

issue openedfacebookresearch/moco

Why the lr in main_lincls is 30?

HI, I wonder why the learning rate is set to 30 when traing classification ? I have never seen such a big learning rate.

`parser.add_argument('--lr', '--learning-rate', default=30., type=float, metavar='LR', help='initial learning rate', dest='lr')`

created time in 5 days

startedKaimingHe/deep-residual-networks

started time in 5 days

startedKaimingHe/deep-residual-networks

started time in 5 days

startedKaimingHe/deep-residual-networks

started time in 6 days

startedKaimingHe/resnet-1k-layers

started time in 7 days

startedKaimingHe/deep-residual-networks

started time in 7 days

issue openedfacebookresearch/moco

ValueError: Decompressed Data Too Large and acc is strangely low

Thanks for your amazing work! When I run python main_moco.py and main_lincls.py, an error occured:

I searched the error in the Internet and tried to add some lines in front of the python code:

The error is gone and I can run the two python file well. However, the accuracy is strangely low. For example, this is when I run python main_lincls.py with your provided 800-epoch MOCOv2 pre-trained model, which is supposed to get a 71.1 top-1 acc after 100 epochs of linear classify training.

I'm really confused, and I think my env of python, pytorch, torchvision is OK. I have no idea what is going wrong. I hope you could help me. Thank you very much!

created time in 7 days

startedKaimingHe/deep-residual-networks

started time in 7 days

startedKaimingHe/deep-residual-networks

started time in 7 days

startedKaimingHe/deep-residual-networks

started time in 7 days

startedKaimingHe/deep-residual-networks

started time in 8 days