[Question] Possible to retrieve layer-wise activations? about transformers.jl HOT 4 OPEN

You can try extracting the actual layers in the classifier head and construct a Flux.Chain and call with Flux.activations. Otherwise, I think a manual loop/calls is probably the simplest.

from transformers.jl.

There is an output_hidden_states configuration that can be set up with HGFConfig:

model_name = "mrm8488/distilroberta-finetuned-financial-news-sentiment-analysis"
cfg = HuggingFace.HGFConfig(load_config(model_name); output_hidden_states = true)
mod = load_model(model_name, "ForSequenceClassification"; config = cfg)

then you can access all layer outputs with mod(a).outputs which is a NTuple{number_layers, @NamedTuple{hidden_state::Array{Float32, 3}}. Another similar configuration is output_attentions that would also include the attentions scores in the named tuples in .outputs.

BTW, if you don't need the sequence classification head, you can simply use load_model(model_name; config = cfg) which would extract the model part without the classification layers.

from transformers.jl.

Amazing, thanks very much for the quick response 👍🏽

(I won't close this since you added the tag for documentation)

from transformers.jl.

Small follow-up question: is it also somehow possible to collect outputs for each layer of the classifier head?

Edit: I realize I can just break down the forward pass into layer-by-layer calls as below, but perhaps there's a more streamline way to do this?

b = clf.layer.layers[1](b).hidden_state |>
        x -> clf.layer.layers[2](x)

from transformers.jl.