Imagnet pretrained model: training accuracy to zero

[jordanott]

Using the exact code from the imagenet template, I add the following to the end of the script:

if __name__ == "__main__":
    import pytorch_lightning

    model = ImageNetLightningModel(
        data_path="path/to/imagenet/data",
        pretrained=True,
        workers=8
    )

    trainer = pytorch_lightning.Trainer(
        gpus=4
    )

    trainer.test(model)

    trainer.fit(model)

If I run trainer.test(model) I get the performance I expect from a pretrained imagenet model. However, when running trainer.fit(model) the training accuracy is zero.

Are the weights getting reinitialized when training is started? If so, how do I prevent that from happening?

[rohitgr7]

Are the weights getting reinitialized when training is started? If so, how do I prevent that from happening?

they are not, can you check the loss and training logic inside the training step. Or maybe there is an issue on how you are calculating the accuracy.