About the All-in-one performance

[chensming]

Hi, I am interested in your work. And I reproduce the experiment. And I got the results as followings:

The first line is the result I reproduced, the second line is the result in the paper. I found the task of going to the rain to be much higher than in the thesis, and none of the other tasks lived up to the thesis. Did I set it up wrong somewhere? I changed the batch size from 8 to 4 because of the cuda memory. And I followed issue 4 to prepare the dataset.

[va1shn9v]

Hello, The results reported in the paper and the model checkpoint released are generated by the model trained for 120 epochs, using a batch size of 8. The above results seem to be reported after 50 epochs. Please try to train it longer until 120 epochs. I think the performance on dehazing and denoising would improve with more training. We had observed that batch size of 8 per GPU i.e a total batch size of 32 provided the best result.

[chensming]

Hello, The results reported in the paper and the model checkpoint released are generated by the model trained for 120 epochs, using a batch size of 8. The above results seem to be reported after 50 epochs. Please try to train it longer until 120 epochs. I think the performance on dehazing and denoising would improve with more training. We had observed that batch size of 8 per GPU i.e a total batch size of 32 provided the best result.

oh, thanks. And I also want to know how to get this visualization.

I wonder which of the three prompt generation modules you used to visualize the code？（Since there are 3 prompt components in the model） https://github.com/va1shn9v/PromptIR/blob/20eff7735fd603ba64842bfd282570e70cd46386/net/model.py#L339 https://github.com/va1shn9v/PromptIR/blob/20eff7735fd603ba64842bfd282570e70cd46386/net/model.py#L352 https://github.com/va1shn9v/PromptIR/blob/20eff7735fd603ba64842bfd282570e70cd46386/net/model.py#L364