Dear authors
First of all, thank you for the wonderful project and for sharing codes and checkpoints
While reproducing the zero-shot retrieval experiments from the Table 6 in your main paper, I observed a huge gap between the reported results in the paper and my reproduced results.
For clarity, here are the results of the zero-shot retrieval performance I obtained:
COCO dataset
The results seem to align closely with the reported performance only in the case of Image-to-Text retrieval on Flickr30k.
I used the CyCLIP checkpoint provided via Google Drive and conducted tests using the Karpathy test split of the COCO and Flickr datasets.
I adapted the code from this repo for retrieval experiments. I hope you can take a quick review on the codes below and help me identify any potential issues.
Additionally, it would be immensely helpful if you could share your codes used for the zero-shot retrieval experiments.
Best regards,
Usage: python test_retrieval.py --dataset coco # or flickr, where
Dear authors First of all, thank you for the wonderful project and for sharing codes and checkpoints
While reproducing the zero-shot retrieval experiments from the Table 6 in your main paper, I observed a huge gap between the reported results in the paper and my reproduced results.
For clarity, here are the results of the zero-shot retrieval performance I obtained: COCO dataset
Flickr dataset
The results seem to align closely with the reported performance only in the case of Image-to-Text retrieval on Flickr30k.
I used the CyCLIP checkpoint provided via Google Drive and conducted tests using the Karpathy test split of the COCO and Flickr datasets.
I adapted the code from this repo for retrieval experiments. I hope you can take a quick review on the codes below and help me identify any potential issues.
Additionally, it would be immensely helpful if you could share your codes used for the zero-shot retrieval experiments.
Best regards,
Usage:
python test_retrieval.py --dataset coco # or flickr
, wheretest_retrieval.py:
src/retrieval.py: