lopuhin/kaggle-imet-2019

iMet Collection 2019 - FGVC6

This is PyTorch baseline for https://www.kaggle.com/c/imet-2019-fgvc6/

License is MIT.

kaggle-imet-2019's Issues

Try to use ResNet101

I am trying to use ResNet101 as backbone to train this nerual network.
I use this:
python main.py model_1 --model resnet101 --n-epoch 25
But when i was predicting:
python main.py predict_test model_1

This error happens:
RuntimeError: Error(s) in loading state_dict for ResNet:
Unexpected key(s) in state_dict: "net.layer3.6.conv1.weight", "net.layer3.6.bn1.weight", "net.layer3.6.bn1.bias", "net.layer3.6.bn1.running_mean", "net.layer3.6.bn1.running_var", "net.layer3.6.bn1.num_batches_tracked", "net.layer3.6.conv2.weight", "net.layer3.6.bn2.weight", "net.layer3.6.bn2.bias", "net.layer3.6.bn2.running_mean", "net.layer3.6.bn2.running_var", "net.layer3.6.bn2.num_batches_tracked", "net.layer3.6.conv3.weight"
......

I tried to change this:
net.load_state_dict(torch.load(weights_path)) in models.py
model.load_state_dict(state['model']) in utils.py
to this:
net.load_state_dict(torch.load(weights_path),strict=False)
model.load_state_dict(state['model'],strict=False)
But the score I got was very low. (0.1+)

I don't know how to solve this problem.
Can you please tell me how to change this code?

full cpu usage

When I add nn.DataParallel into the code so that I can use multiple gpus with default worker num, I find that all my cpus are fully loaded. I haven't check the code but I guess it's because of multiprocessing implementations. Please tell me if this is abnormal or just a training feature.

transform to base64 problem

when I run the build.sh file

File "build.py", line 6 def encode_file(path: Path) -> str: ^ SyntaxError: invalid syntax

Sigmoid before BCEWithLogitsLoss

Is the sigmoid activation necessary before the output is fed into BCEWithLogitsLoss? Are there any reasons not using it in your code? Thanks in advance.