pzzhang/VinVL

Pre-extracted image features: COCO 2014 Train/Val Image Features (~50G) is wrong

Opened this issue · 2 comments

gorjanradevski commented

Thanks for taking the time to release the pre-trained models and extracted features for all datasets you've used. However, I noticed that the COCO 2014 Train/Val Image Features (~50G) doesn't contain the features for the COCO Train/Val datasets, but rather for the COCO 2014 Test and COCO 2015 Test. In fact, you can also confirm this from the links themselves:

COCO 2014 Train/Val Image Features: .../coco_X152C4_frcnnbig2_exp168model_0060000model.roi_heads.nm_filter_2_model.roi_heads.score_thresh_0.2/model_0060000/

COCO 2014 Test Image Features:
.../coco_X152C4_frcnnbig2_exp168model_0060000model.roi_heads.nm_filter_2_model.roi_heads.score_thresh_0.2/model_0060000/coco2014test/

COCO 2015 Test Image Features:
.../coco_X152C4_frcnnbig2_exp168model_0060000model.roi_heads.nm_filter_2_model.roi_heads.score_thresh_0.2/model_0060000/coco2015test/

Could you please release the obtained image features for the COCO 2014 Train/Val dataset?

Cheers!

alice-cool commented

Dear schoa

Thanks for taking the time to release the pre-trained models and extracted features for all datasets you've used. However, I noticed that the COCO 2014 Train/Val Image Features (~50G) doesn't contain the features for the COCO Train/Val datasets, but rather for the COCO 2014 Test and COCO 2015 Test. In fact, you can also confirm this from the links themselves:

COCO 2014 Train/Val Image Features: .../coco_X152C4_frcnnbig2_exp168model_0060000model.roi_heads.nm_filter_2_model.roi_heads.score_thresh_0.2/model_0060000/

COCO 2014 Test Image Features: .../coco_X152C4_frcnnbig2_exp168model_0060000model.roi_heads.nm_filter_2_model.roi_heads.score_thresh_0.2/model_0060000/coco2014test/

COCO 2015 Test Image Features: .../coco_X152C4_frcnnbig2_exp168model_0060000model.roi_heads.nm_filter_2_model.roi_heads.score_thresh_0.2/model_0060000/coco2015test/

Could you please release the obtained image features for the COCO 2014 Train/Val dataset?

Cheers!

Dear scholar.
Where download the VinVL pretraining model or it is the same as Oscar. Where download the VinVL new feature for VQA and GQA? I only found that OSCAR's pretraning model weights file, finetuned model weights file and the image features for downstream tasks. But I don't find any the below file for VinVL model.......

prajwalgatti commented

@gorjanradevski
After downloading COCO 2014 Train/Val Image Features, I see that it contains all three parts of data, ie 2014 Test, 2015 Test, as well as 2014 Train+Val combined data.

Here's the list of contents -

./model_0060000/
|-- [4.0K]  coco2014test
|   |-- [452K]  features.lineidx
|   |-- [ 16G]  features.tsv
|   |-- [1.4M]  imageid2idx.json
|   |-- [390K]  predictions.lineidx
|   `-- [517M]  predictions.tsv
|-- [4.0K]  coco2015test
|   |-- [928K]  features.lineidx
|   |-- [ 32G]  features.tsv
|   |-- [2.9M]  imageid2idx.json
|   |-- [774K]  predictions.lineidx
|   `-- [401M]  predictions.tsv
|-- [1.4M]  features.lineidx
|-- [ 49G]  features.tsv
|-- [2.0M]  imageid2idx.json
|-- [1.2M]  predictions.lineidx
`-- [613M]  predictions.tsv