picsart-ai-research / mi-gan Goto Github PK

Error(s) in loading state_dict for Generator:
Missing key(s) in state_dict: "synthesis.b8.conv1.upsample.insert_zeros.weight", "synthesis.b8.upsample.insert_zeros.weight", "synthesis.b16.conv1.upsample.insert_zeros.weight", "synthesis.b16.upsample.insert_zeros.weight", "synthesis.b32.conv1.upsample.insert_zeros.weight", "synthesis.b32.upsample.insert_zeros.weight", "synthesis.b64.conv1.upsample.insert_zeros.weight", "synthesis.b64.upsample.insert_zeros.weight", "synthesis.b128.conv1.upsample.insert_zeros.weight", "synthesis.b128.upsample.insert_zeros.weight", "synthesis.b256.conv1.upsample.insert_zeros.weight", "synthesis.b256.upsample.insert_zeros.weight".
Unexpected key(s) in state_dict: "synthesis.b8.conv1.upsample.filter_const", "synthesis.b8.upsample.filter_const", "synthesis.b16.conv1.upsample.filter_const", "synthesis.b16.upsample.filter_const", "synthesis.b32.conv1.upsample.filter_const", "synthesis.b32.upsample.filter_const", "synthesis.b64.conv1.upsample.filter_const", "synthesis.b64.upsample.filter_const", "synthesis.b128.conv1.upsample.filter_const", "synthesis.b128.upsample.filter_const", "synthesis.b256.conv1.upsample.filter_const", "synthesis.b256.upsample.filter_const".

Hi, thank you for open sourcing your model and code. I have added this model to my open source project and it can be installed and used in the following way:

pip install iopaint
iopaint start --model migan

open http://localhost:8080/

More info about the tool: https://github.com/Sanster/lama-cleaner/releases/tag/iopaint-1.0.0b2

Without composition, just looking at the output of places 512 model, it looks so horrible, even the inpainted area doesn't seem up-to-par with the paper. Look at this example
Here's the original image

Here's the mask

Here's the direct output (without composition)

Looks so horrible, how did you report such a low FID score?

Hi
Thanks for your work.
I am expressing that when I use the export_inference_model.py file to attempt to convert the pkl file to onnx, I encountered an error. Could you please help me identify the issue?

File "MI-GAN/scripts/export_inference_model.py", line 119, in main
    resume_data = pickle.Unpickler(f).load()
ModuleNotFoundError: No module named 'lib.model_zoo.simpleinpainting'

After I exported the pt model you posted as an onnx model, I found that there were index differences on the same data set. Could you please tell me why?

Hello,

Thank you for the amazing and insightful work! I was trying to train the model on my own set of images. However, I noticed that the number of parameters in the MIGAN generator part of the model that is used to train the model is very different from the one for which the checkpoint was provided (52 M vs 5.9 M respectively). Maybe I missed something here but could you please help me understand this? Thanks!

MI-GAN is a great job, thank you for open source! I would like to ask a few questions:

1. What are your reasons for not using a loss function like L1? Will using this pixel-by-pixel supervised loss affect the final result?
2. The current image resolution on the mobile devices is relatively high, and the model can only handle low-resolution images, such as 512, and the result is blurry. Is there any better solution?
3. The default parameter for model re-parameterization in the code is 4, but the parameter mentioned in the paper is 8. Which number is more appropriate?
   Apologies for the disturbance, and I am looking forward to your response. Thank you very much!

Thank you very much for your work. I plan to train my own data set. The current data set is the original picture 'images' and the labels' masks' made by me, but the masks you gave me are data/places2/train_512, so is the train_512 folder corresponding to the original picture? I see that your mask code is randomly generated, so I would like to ask, how do I train myself with labeled data sets? How to configure the file? How do you put the data set?

Hello and thank you for your amazing work. May I ask if you ever have the intent to release the license. Without it, no one can use your source code to do anything.