inference with 4-channel model about cct HOT 7 CLOSED

I found the cause of the problem. Since I have a 4-channel model, I set pretrained: False in the config file. In model.py the default value for pretrained is True, and thus the weights of ResNet will be loaded. So we have to pass the pretrained value when defining the model

model = models.CCT(num_classes=num_classes, conf=config['model'],pretrained=config['pretrained'],testing=True)

from cct.

Hi @SuzannaLin

It seems from the error that the check point model only has 3 channels and not 4, maybe a double-check of the checkpoint?

from cct.

You mean the checkpoint in the saved/ folder? Where in the script is this loaded?

from cct.

the checkpoint you passed: ABCE_70_30_sup_alti/best_model.pth, either it is not loaded or it is not the current one since in the error you have copying a param with shape torch.Size([64, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([64, 4, 3, 3]).

from cct.

The error occurs here in the inference script:
model = models.CCT(num_classes=num_classes, conf=config['model'], testing=True, pretrained = True)
The checkpoint is only loaded after, right?

from cct.

yes, i think there is just a mismatch between the created model and the checkpoint, one of them still has only 3 channels

from cct.

@SuzannaLin thanks you posting the answer, glad it is working now

from cct.