HuggingFace TFRobertaModel detailed summary

user_3780142 0
Viewed 63
from transformers import RobertaTokenizer, TFRobertaModel
import tensorflow as tf

tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
model = TFRobertaModel.from_pretrained("roberta-base")

I want a detailed layer summary of this HuggingFace TFRobertaModel() so that I can visualize shapes, layers and customize if needed. However, when I did: model.summary(), it just shows everything in a single layer. I tried digging into it's different attributes, but not able to get a detailed layer summary. Is it possible to do so?

Model: "tf_roberta_model_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
roberta (TFRobertaMainLayer) multiple                  124645632 
=================================================================
Total params: 124,645,632
Trainable params: 124,645,632
Non-trainable params: 0
_________________________________________________________________

Also, there is a related question in HuggingFace forum which hasn't been answered yet.

tensorflow2.0 huggingface-transformers keras tensorflow python
1 Answers

Not exactly a model summary, but you can print the layers like this:

from transformers import RobertaTokenizer, TFRobertaModel
import tensorflow as tf

tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
model = TFRobertaModel.from_pretrained("roberta-base")

def print_layers(l, model):
  for idx, s in enumerate(l.submodules):
    if s.submodules:
      print_layers(s, model)
    print(s)

TFRobertaMainLayer = model.layers[0]   
print_layers(TFRobertaMainLayer, model)

You could also use s.weights to get the weights of each layer.

user_96578610
Related