It’s a bit strange that we should share external data and pre-train models. But do you guys have some resources?
Disclosing Multimedia Commons Dataset
Licenses available in: prefix=tools/etc/yfcc100m_dataset.sql (CC0 and CC-BY images only)
These are publicly acessible with the “UNSIGNED” token, see e.g. python - Can I use boto3 anonymously? - Stack Overflow
Also:
https://storage.googleapis.com/openimages/web/index.html
https://www.robots.ox.ac.uk/~vgg/data/pass/
tf_efficientnet_bN_ns, swin and convnext imagenet pretrained from timm package
From my side:
No external data were used.
tf_efficientnet_b5_ns (noisy-student), efficientnetv2_rw_m pretrained from timm package
timm pretrained models: tf_efficientnet_b*ns, convnext, eca_nfnet_
no external data
Already declared but just in case: Pre-trained EfficientNet models from timm
Pre-trained imagenet- TF_EfficientNet, ConvNext, ECA_NFNET, RegNetZ models all from timm.
Assume this isn’t required, but efficientnetv2, seresnet152d, ecaresnet269, and more (timm==0.5.4)
timm pretrained models: tf_efficientnet_b*ns
no external data