Evaluation FID in trained model doesn't match the one from the pre-trained model

[arturandre]

Hello, thank you for the code available,

I executed the training script:

bash scripts/train_photosketch_horse_riders.sh

and then I compared the metrics of the trained model with those of the file:

weights/photosketch_horse_riders_aug.pth

using the evaluation script:

run_metrics.py

with the command:

python run_metrics.py --models_list weights/eval_list --output metric_results.csv

I added the trained weights into the weights folder and into the eval_list file so that I could check its FID too.

I very different FIDs for the photosketch_horse_riders_aug.pth (FID ~20.13) and the trained one (FID ~37.07). I assumed that the training script would generate a model similar to the one stored at photosketch_horse_riders_aug.pth.

There is some other different procedure to get a trained model with the same FID as photosketch_horse_riders_aug.pth, which is very close but not quite the same as in the paper (FID ~19.94)?

[PeterWang512]

You would want to find the best iteration recorded in ./checkpoints/<exp_name>/best_iter.txt and use this iteration to get the FID.

[arturandre]

I see, thank you!