search

XiaohangZhan / cdp

Code for our ECCV 2018 work.
MIT License
454 stars 93 forks source link

Issue on training with ready-made dataset #19

Closed rose-jinyang closed 4 years ago

rose-jinyang commented 4 years ago

Hello Thanks for contributing this paper and code. I had a successful inference with ready-made emore_u200k data and pre-trained CDP model. I tried to retrain CDP model with ready-made emore_l200k in mediator mode. But I met the following issue. image I am using a Tesla V100(GPU memory 16GB) and DDR 128GB. PyTorch version is 1.4.0 and Python version is 3.7 I used the following config.yml. image I thought that this may be due to lack of GPU memory so I decreased the batch_size from 1024 to 8 but the result was same. Please let me know the reason asap. Thanks

rose-jinyang commented 4 years ago

Oh, I fixed the issue. This was due to PyTorch version. I changed Line 96 in mediator.py as the following losses.update(loss.mean().data.cpu().item()) It started to train.

XiaohangZhan commented 4 years ago

Oh, I fixed the issue. This was due to PyTorch version. I changed Line 96 in mediator.py as the following losses.update(loss.mean().data.cpu().item()) It started to train.

Glad to hear that. The PyTorch version for me was 0.3.1.

rose-jinyang commented 4 years ago

Hi I have a question. I'd like to know about threshold in detail. image Is this irrelative to face recognition model ? Is this the Euclidean distance threshold between feature vectors ? Please let me know asap. Thanks

XiaohangZhan commented 4 years ago

Hi, the threshold is the binary classification threshold for MLP. It has nothing to do with the face recognition model. You may have to use a validation set to search this threshold.