Efficientnet number of features are limited to 5?

[msaquib-synsor]

I am trying to extract features from every layer of Efficientnet_b0 which are total "9" features from the end of each network block output, according to Efficientnet_b0 architecture. But only getting 5 features from network block are coming as output ? so I want to know if is it like that mean hard coded or it capable of extract all 9 blocks.
thank you

Efficientnet_b0 architecture for 9 Block feature extraction

[rwightman]

@msaquib-synsor not sure why you'd want 9 block feature extraction? Once you move past the current scheme where each feature map extracted is the deepest available at that feature map resolution it's amost arbitrary and you could just argue you'd want one for every block... it's done this way for feature pyramid use, and also has some degree of consistency across networks since almost all have 5 stride changes.

I am going to add feature extraction for vision transformers and mlp models, which will be more arbitrary and need a different scheme. It might be more useful for this.

For know you'd have to create the model normally (without features_only=True), overwrite the .feature_info attribute on the modwith the module names of all the blocks you want and then wrap in a FeatureHook or FeatureList module depending on whether the features are deep within the network or at block stage outputs only. https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/features.py#L218-L238