How to use Noisy Weights

[VpkPrasanna]

After seeing many discussion i saw that people mentioned that they use "Noisy" Weights instead of "Image Net" Weights, how to initialise the noisy weights while creating a model ? for Efficient net Model?

[amaarora]

Hi @VpkPrasanna , it has been mentioned in the timm docs here https://rwightman.github.io/pytorch-image-models/models/noisy-student/.

These weights have been ported from Google's official tensorflow weights and you could use tf_efficientnet_b0_ns-tf_efficientnet_b7_ns to use b0-b7 weights.

You could also use tf_efficientnet_l2_ns or tf_efficientnet_l2_ns_475 in model name inside timm.create_model to use these pretrained weights.

[amaarora]

To create EfficientNet B0 with Noisy Student weights (as mentioned in the docs):

import timm 
m = timm.create_model('tf_efficientnet_b0_ns')

[VpkPrasanna]

@amaarora thanks for the response , I am clear now.