Empty supp_dict (image_classification.py)

[andreamigliorati]

Hi, I want to perform image classification and train my own models, I added the train keys to the config file to do that. However I have a problem with the train_one_epoch method. When computing the loss (loss = training_box(sample_batch, targets, supp_dict)) I get an error which appears to be caused by an empty supp_dict.

File "/media/RAID/migliorati/coding_machines/sc2-benchmark-main/script/task/image_classification.py", line 209, in main
    train(teacher_model, student_model, dataset_dict, ckpt_file_path, device, device_ids, distributed, config, args)
  File "/media/RAID/migliorati/coding_machines/sc2-benchmark-main/script/task/image_classification.py", line 160, in train
    train_one_epoch(training_box, aux_module, bottleneck_updated, device, epoch, log_freq)
  File "/media/RAID/migliorati/coding_machines/sc2-benchmark-main/script/task/image_classification.py", line 73, in train_one_epoch
    loss = training_box(sample_batch, targets, supp_dict)

Before being called in loss = training_box(sample_batch, targets, supp_dict), supp_dict is created here:

    for sample_batch, targets, supp_dict in \
            metric_logger.log_every(training_box.train_data_loader, log_freq, header):
        if isinstance(sample_batch, torch.Tensor):
            sample_batch = sample_batch.to(device)

        if isinstance(targets, torch.Tensor):
            targets = targets.to(device)

which in my case gives an empty supp_dict, and therefore a missing key (KeyError: 'backbone.bottleneck_layer'). I am launching my training with bash script/neural_input_compression/ilsvrc2012-image_classification.sh factorized_prior-resnet50 8, and as mentioned I created a train key in the config file so I am not sure what I'm doing wrong. Thanks in advance!

[yoshitomo-matsubara]

Hi @andreamigliorati

Can you share your yaml file here?

[yoshitomo-matsubara]

Please use "Reply" to organize discussions.

Empty supp_dict doesn't matter in this case. You're trying to train Factorized Prior model + ResNet-50 using a configuration to train a modified ResNet-50 model with a non-entropic bottleneck with a pretrained ResNet-50 as a teacher for GHND + BQ (bottleneck quantization).

I recommend that you decide 1) how you want to train 2) what model. Since you'll also want to know about designs of the models, references in the SC2 benchmark paper should also help you. If you just want to train some random models, I suggest starting with config files under supervised_compression/

[andreamigliorati]

Yes of course, here it is. Thanks for the help, again apologies in advance if I'm asking something trivial but I just can't seem to make it work for now.
factorized_prior-resnet50.zip