Error when training a supervised linear classifier on EuroSAT from a pre-trained SeCo model

[chandlerbing65nm]

I encountered error when running this script:
python main_eurosat.py --data_dir data/eurosat --backbone_type pretrain --ckpt_path checkpoints/seco_resnet50_100k.ckpt

Traceback (most recent call last):
  File "main_eurosat.py", line 89, in <module>
    trainer.fit(model, datamodule=datamodule)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py", line 510, in fit
    results = self.accelerator_backend.train()
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/accelerators/accelerator.py", line 57, in train
    return self.train_or_test()
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/accelerators/accelerator.py", line 74, in train_or_test
    results = self.trainer.train()
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py", line 532, in train
    self.run_sanity_check(self.get_model())
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py", line 731, in run_sanity_check
    _, eval_results = self.run_evaluation(max_batches=self.num_sanity_val_batches)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py", line 643, in run_evaluation
    output = self.evaluation_loop.evaluation_step(batch, batch_idx, dataloader_idx)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/trainer/evaluation_loop.py", line 171, in evaluation_step
    output = self.trainer.accelerator_backend.validation_step(args)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py", line 73, in validation_step
    return self._step(self.trainer.model.validation_step, args)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py", line 65, in _step
    output = model_step(*args)
  File "main_eurosat.py", line 39, in validation_step
    loss, acc = self.shared_step(batch)
  File "main_eurosat.py", line 46, in shared_step
    logits = self(x)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "main_eurosat.py", line 29, in forward
    logits = self.classifier(feats)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/torch/nn/modules/linear.py", line 93, in forward
    return F.linear(input, self.weight, self.bias)
  File "/home/chandler_doloriel/.conda/envs/seasonal-contrast/lib/python3.7/site-packages/torch/nn/functional.py", line 1690, in linear
    ret = torch.addmm(bias, input, weight.t())
RuntimeError: mat1 dim 1 must match mat2 dim 0

[xavibou]

Hi,
I encountered the same issue, which comes from the model initialisation in main_eurosat.py. It is hard-coded for in_features=512, corresponding to the dimension of the feature vector provided by a pre-trained resnet18 encoder. In the case of resnet50, this output vector is of dimension 2048. Hence, you can modify the line 83 from:

model = Classifier(backbone, in_features=512, num_classes=datamodule.num_classes)

to

model = Classifier(backbone, in_features=2048, num_classes=datamodule.num_classes)

Hope it helps.
Xavi