No pretrained weights exist or were found for this model. Using random initialization.

[ouioui199]

Hello,

I've got this print out when importing efficientnetv2_s. Is this normal that we don't have ImagNet pretrained for this efficientnetv2_s backbone ?

[TorbenSDJohansen]

There are indeed pretrained weights for efficientnetv2_s; however, you need to "be more specific" when you create your model.

To access weights trained on ImageNet, you can use model = create_model('tf_efficientnetv2_s.in1k', pretrained=True). To access weights trained on ImageNet21k and then finetuned on ImageNet, you can use model = create_model('tf_efficientnetv2_s.in21k_ft_in1k', pretrained=True).

Note that for older versions of timm the syntax differed slightly, e.g., you would write something like model = create_model('tf_efficientnetv2_s_in21k', pretrained=True) to access weights training on ImageNet21k (i.e., the syntax did not follow the newer "." style).

[TorbenSDJohansen]

**Note**: This is a reply to the following message:

Hello Torben, thank you for your answer. I'm actually using segmentation models of qubvel https://github.com/qubvel/segmentation_models.pytorch. In order to import efficientnetv2_s, I have to precise encoder_name='tu-efficientnetv2_s', and encoder_weights='imagenet'. Normally it returns True for pretrained, but somehow it prints out 'No pretrained weights exist or were found for this model. Using random initialization.'

Sorry for not getting back to you before now. Have you tried replacing "efficientnetv2_s" with "tf_efficientnetv2_s_in21ft1k"? Perhaps that might work. Be sure to let me know!

…

________________________________ Fra: Huy Nguyen ***@***.***> Sendt: 24. marts 2023 14:34 Til: huggingface/pytorch-image-models ***@***.***> Cc: Torben Skov Dyg Johansen ***@***.***>; Comment ***@***.***> Emne: Re: [huggingface/pytorch-image-models] No pretrained weights exist or were found for this model. Using random initialization. (Discussion #1737) Hello Torben, thank you for your answer. I'm actually using segmentation models of qubvel https://github.com/qubvel/segmentation_models.pytorch<https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fqubvel%2Fsegmentation_models.pytorch&data=05%7C01%7Ctsdj%40sam.sdu.dk%7Cad2393cd4dca4a53be3d08db2c6c747a%7C9a97c27db83e4694b35354bdbf18ab5b%7C0%7C0%7C638152616533403080%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=08VIdzbTLggSl1Wq1QiKz6VIo%2B%2BTdUe1hrZDwWt6AwY%3D&reserved=0>. In order to import efficientnetv2_s, I have to precise encoder_name='tu-efficientnetv2_s', and encoder_weights='imagenet'. Normally it returns True for pretrained, but somehow it prints out 'No pretrained weights exist or were found for this model. Using random initialization.' — Reply to this email directly, view it on GitHub<https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fhuggingface%2Fpytorch-image-models%2Fdiscussions%2F1737%23discussioncomment-5418066&data=05%7C01%7Ctsdj%40sam.sdu.dk%7Cad2393cd4dca4a53be3d08db2c6c747a%7C9a97c27db83e4694b35354bdbf18ab5b%7C0%7C0%7C638152616533403080%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=JX%2FdCrKFDNO8Zd77GFHzvQ1M2L5GIJ%2B7hKvScfbsDgA%3D&reserved=0>, or unsubscribe<https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAKKJHKKV7Y3BJBFJCFFWD4LW5WPFFANCNFSM6AAAAAAWGLS6MQ&data=05%7C01%7Ctsdj%40sam.sdu.dk%7Cad2393cd4dca4a53be3d08db2c6c747a%7C9a97c27db83e4694b35354bdbf18ab5b%7C0%7C0%7C638152616533403080%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=SqZu5tNcq2wEs84y%2BV8VKvPp8GKzK75uaxQhKjV56sU%3D&reserved=0>. You are receiving this because you commented.Message ID: ***@***.***>

[rwightman]

It's best to use the API to list models (using pretrained=True to show ones with weights), or look in the model file at the pretrained_cfgs... there are often more than one variant of a similar model arch and not all have weights

On latest main branch / 0.8.x pre-releases, there are no weights for the non TF trained v2_s model, only my own variant (before official weights released) and the TF ones, the TF ones have a hack to match tensorflow paddings where, the rw was my own initial guess at the v2 arch, and the efficientnetv2_s is same as the official arch w/o padding but never trained.

>>> pprint.pprint(timm.list_models('*efficientnet*v2*', pretrained=True))
['efficientnetv2_rw_m.agc_in1k',
 'efficientnetv2_rw_s.ra2_in1k',
 'efficientnetv2_rw_t.ra2_in1k',
 'gc_efficientnetv2_rw_t.agc_in1k',
 'tf_efficientnetv2_b0.in1k',
 'tf_efficientnetv2_b1.in1k',
 'tf_efficientnetv2_b2.in1k',
 'tf_efficientnetv2_b3.in1k',
 'tf_efficientnetv2_b3.in21k',
 'tf_efficientnetv2_b3.in21k_ft_in1k',
 'tf_efficientnetv2_l.in1k',
 'tf_efficientnetv2_l.in21k',
 'tf_efficientnetv2_l.in21k_ft_in1k',
 'tf_efficientnetv2_m.in1k',
 'tf_efficientnetv2_m.in21k',
 'tf_efficientnetv2_m.in21k_ft_in1k',
 'tf_efficientnetv2_s.in1k',
 'tf_efficientnetv2_s.in21k',
 'tf_efficientnetv2_s.in21k_ft_in1k',
 'tf_efficientnetv2_xl.in21k',
 'tf_efficientnetv2_xl.in21k_ft_in1k']