High_Resolution_Image_Inpainting
High_Resolution_Image_Inpainting copied to clipboard
Training code
Hi @a-mos Are you going to release the training code? The patch extraction strategy is interesting in the training session.
Hi @a-mos Are you going to release the training code? The patch extraction strategy is interesting in the training session.
Hi! Not sure about the training code as separate machines were used to run the network from the first stage. But I can add the dataset code - https://gist.github.com/a-mos/1eac6772f7e295c8eed993db6406e7f2 with patch extractions.
Hi @a-mos Are you going to release the training code? The patch extraction strategy is interesting in the training session.
Hi! Not sure about the training code as separate machines were used to run the network from the first stage. But I can add the dataset code - https://gist.github.com/a-mos/1eac6772f7e295c8eed993db6406e7f2 with patch extractions.
Thanks for sharing the dataset code. Other questions about the training code:
- How do you train the two stages network? End-to-end or some other training strategies.
- In the training session, the output of the first stage was directly fed into the second stage network or apply the post-processings to them like the testing session.
Hi @a-mos Are you going to release the training code? The patch extraction strategy is interesting in the training session.
Hi! Not sure about the training code as separate machines were used to run the network from the first stage. But I can add the dataset code - https://gist.github.com/a-mos/1eac6772f7e295c8eed993db6406e7f2 with patch extractions.
Thanks for sharing the dataset code. Other questions about the training code:
- How do you train the two stages network? End-to-end or some other training strategies.
- In the training session, the output of the first stage was directly fed into the second stage network or apply the post-processings to them like the testing session.
This was not an end-to-end approach. The output from the neural network from the first stage was pre-calculated separately (coarse_filled_path in the dataset code) and its weights were frozen due to limited computational resources. Post-processing with shifts after the first stage is used both in the train and in the test. The selection of a random patch in the testing session is skipped.
Hi @a-mos Are you going to release the training code? The patch extraction strategy is interesting in the training session.
Hi! Not sure about the training code as separate machines were used to run the network from the first stage. But I can add the dataset code - https://gist.github.com/a-mos/1eac6772f7e295c8eed993db6406e7f2 with patch extractions.
Thanks for sharing the dataset code. Other questions about the training code:
- How do you train the two stages network? End-to-end or some other training strategies.
- In the training session, the output of the first stage was directly fed into the second stage network or apply the post-processings to them like the testing session.
This was not an end-to-end approach. The output from the neural network from the first stage was pre-calculated separately (coarse_filled_path in the dataset code) and its weights were frozen due to limited computational resources. Post-processing with shifts after the first stage is used both in the train and in the test. The selection of a random patch in the testing session is skipped.
Received with thanks.