vis_reconstruction.py question about spark HOT 6 CLOSED

Hi @buxuewushu1314 I just tested the viz code, but did not encounter this problem. Could you double-check whether you download the cnxL384_withdecoder_1kpretrained_spark_style.pth file from https://drive.google.com/file/d/1ZI9Jgtb3fKWE_vDFEly29w-1FWZSNwa0/view?usp=share_link ? I think checking the file will solve the issue.

from spark.

When I load xxx__1kpretrained.pth or xxxx_still_pretraining.pth and want to have a look at the performance of my pretrained model, I encounter this problem too. How can I get my xxxx_withdecoder_1kpretrained_spark_style.pth when i pretraining ？

from spark.

I modified the function：build_spark like below
def build_spark(pretraining_pth,spark_model):
pretrained_state = torch.load(pretraining_pth, map_location='cpu')
spark_model.eval(), [p.requires_grad_(False) for p in spark_model.parameters()]
# load the checkpoint
missing, unexpected = spark_model.load_state_dict(pretrained_state['module'], strict=False)
assert len(missing) == 0, f'load_state_dict missing keys: {missing}'
assert len(unexpected) == 0, f'load_state_dict unexpected keys: {unexpected}'
del pretrained_state
return spark

And then it works using xxxx_still_pretraining.pth
ckpt_file = 'your xxxx_still_pretraining.pth path'
spark = 'your spark_model'
spark = build_spark(ckpt_file,spark)

What's more, you may need to modify the function denorm_for_vis in spark.py : 191 if you use your own dataset

from spark.

@syjabc have you solved the issue through your code? I think the new build_spark is doing the right thing, using torch.load('xxxx_still_pretraining', 'cpu')['module'] as torch.load('xxxx_withdecoder_1kpretrained_spark_style', 'cpu').

from spark.

I think it has been solved. Learned so much from your code, thanks for your excellent work.

from spark.

Also thank you for helping me to refine this visualisation. In the latest commit 1468df8 i modify the notebook for visualizing your own model.

from spark.