cici-ai-club / 3M

A codebase for multi-style image captions
Apache License 2.0
3 stars 1 forks source link

Pretrained Features? #2

Closed elegieee closed 3 years ago

elegieee commented 3 years ago

Can you provide the Pretrained Features(Dense captions features and ResNext features) ? Thank you!

cici-ai-club commented 3 years ago

Can you provide the Pretrained Features(Dense captions features and ResNext features) ? Thank you!

Hi, elegieee, my original ResNext features, unfortunately, don't get all kept due to some data loss. But ResNext features are easily extracted from ParlAI. And there might be a very slight difference in floating of the current version, compared with my original downloading from them, but should not influence the performance, the command you could use is in ParlAI is: parlai extract_image_feature -im resnext101_32x48d_wsl -t personality_captions -dt train (for the training data and global features, change to test, valid when you need them)

parlai extract_image_feature -im resnext101_32x48d_wsl_spatial -t personality_captions -dt test (this one is for spatial features for test data)

And the text features I already shared in the README of my repo, which is densecap_added1.h5 under the link: https://drive.google.com/drive/u/4/folders/170palQ7QzRsY2ZRyaDTIQAcdHyuVZsFe

elegieee commented 3 years ago

You really help me a lot. Thanks.