efficientnet pretrained model URL is empty

[hjlee9182]

When I use this code

model = geffnet.create_model(
'efficientnet_b5',
num_classes=1000,
in_chans=3,
pretrained=True,
exportable=True)

But this error occured.

efficientnet b0~b3 => okay

efficientnet b4~b7 => URL is empty.

How to solved this problem?

[linhduongtuan]

Ross has not trained efficientnet_b5 by himself yet. You can try using tf_efficientnet_b5. Good luck Linh

…

On 9 Nov 2020, at 17:35, HyunJun Lee ***@***.***> wrote:  When I use this code model = geffnet.create_model( 'efficientnet_b5', num_classes=1000, in_chans=3, pretrained=True, exportable=True) But this error occured. efficientnet b0b3 => okay efficientnet b4b7 => URL is empty. How to solved this problem? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[hjlee9182]

@linhduongtuan
You mean tensorflow hub efficientnet?

[linhduongtuan]

Not really, he converts/ports TF weights to compatible PT weights. Linh

…

On 10 Nov 2020, at 08:52, HyunJun Lee ***@***.***> wrote: @linhduongtuan <https://github.com/linhduongtuan> You mean tensorflow hub efficientnet? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#53 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFKZ27FZSX2C2GQCMQZPIFDSPCMFXANCNFSM4TPFAS4A>.

[hjlee9182]

@linhduongtuan
Oh, I'm so sorry ,I didn't get it exactly.

1. where is converts/ports file? I can't find that in this repository.
   Can you give the link?

2. I'm not correctly know model that is you mentioned, Can you give link that has model?

3. You mean that pretrained TF efficientnet model weight to convert Pytorch weight?

[linhduongtuan]

1/ I don’t really know. You can look more closely eyes in his repos 2/ instead of using model = create_model(‘efficientnet_b5’, pretrained=True), you should use model = create_model(’tf_efficientnet_b5’, pretrained=True). To know where is a weight link, you can go into timm —> models —> efficientnet.py —> url for them 3/ Yes, you will use transfer learning with the pretrained weight from Google which is ported to PT Linh

…

On 10 Nov 2020, at 09:27, HyunJun Lee ***@***.***> wrote: @linhduongtuan <https://github.com/linhduongtuan> Oh, I'm so sorry ,I didn't get it exactly. where is converts/ports file? I can't find that in this repository. Can you give the link? I'm not correctly know model that is you mentioned, Can you give link that has model? You mean that pretrained TF efficientnet model weight to convert Pytorch weight? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#53 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFKZ27FPBGUGA5UNDL4KZZLSPCQHTANCNFSM4TPFAS4A>.

[hjlee9182]

Oh, I understand clearly.
I have one more question.
I see tf_efficientnet is more accurately better than efficientnet.

Did you know the reason that?

[rwightman]

thanks @linhduongtuan

@hjlee9182 all models with the tf_ prefix were converted from original Tensorflow weights, and are available here https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet ... numerous different papers are covered so quite a few different sets of weights. The models without the prefix were trained in PyTorch by me, I've only had resources to train the smaller ones. The models do differ in the padding used. You can research SAME padding in PyTorch if you are curious why that's relevant.