Are weights other than imagenet used?

[tristandb8]

Hi,

I'm interested in:
https://huggingface.co/timm/tf_efficientnet_b7.ns_jft_in1k

The efficientnet model is available in keras, but I'm interested in the advanced pretrained weights used here.
When I load kimm.models.EfficientNetB7(), I get
Downloading data from https://github.com/james77777778/keras-image-models/releases/download/0.1.0/efficientnetb7_tf_efficientnet_b7.ns_jft_in1k.keras

Which seems great to me as tf_efficientnet_b7.ns_jft_in1k is what I'm interested in.

In this repo kimm/_src/models/efficientnet.py I see
class EfficientNetB7(EfficientNetVariant):
available_weights = [
(
"imagenet",
EfficientNet.default_origin,
"efficientnetb7_tf_efficientnet_b7.ns_jft_in1k.keras",
)
]

But I can't kimm.models.EfficientNetB7("efficientnetb7_tf_efficientnet_b7.ns_jft_in1k.keras")
It tells me the only weights available are "imagenet"
And I only see
add_model_to_registry(EfficientNetB7, "imagenet")

I'm assuming you loaded EfficientNetB7 with tf_efficientnet_b7.ns_jft_in1k but mislabeled it as imagenet?

Would be helpful to know which weights I'm actually working with.

Thanks for the awesome work by the way.

[james77777778]

Hey @tristandb8
Thanks for your kind words!

Actually, kimm.models.EfficientNetB7() loads "efficientnetb7_tf_efficientnet_b7.ns_jft_in1k". I renamed it to "imagenet" to simplify the naming convention.
You can check this script, which includes a correctness check:
https://github.com/james77777778/keras-image-models/blob/main/tools/convert_efficientnet_from_timm.py

Let me know if you encounter any issues.