Jing Shi

The same question

Solved. For your first question, change `model.load_state_dict(checkpoint['model'].state_dict())` to `model.load_state_dict(checkpoint['model'].state_dict(), strict=False)`, then it can normally loaded. This is due to the higher version of pytorch.

@spandanagella I find that my evaluation result is only 36%. Can you reproduce the evaluation result? It might be the problem introduced by BatchNorm1d, which require the parameter `num_batches_tracked` in...

I compared config1: python3.6 + pytorch0.4.0 and config2: python2.7 + pytorch0.3.1. The config2 gives the reported result but config1 not. I find the data.json generated by `tools/prepro.py` has different list...

@lfdeep , I am also trying to test on my image. It would be helpful if you share your inference code at [email protected]

I also have the same question. Seems that it finetunes all the UNet parameters and the CLIP text encoder parameters.