README.md

This model was pretrained on the bookcorpus dataset using knowledge distillation.

The particularity of this model is that even though it shares the same architecture as BERT, it has a hidden size of 240. Since it has 12 attention heads, the head size (20) is different from the one of the BERT base model (64).

The knowledge distillation was performed using multiple loss functions.

The weights of the model were initialized from scratch.

PS : the tokenizer is the same as the one of the model bert-base-uncased.


To load the model \& tokenizer :

````python
from transformers import AutoModelForMaskedLM, BertTokenizer

model_name = "eli4s/Bert-L12-h240-A12"
model = AutoModelForMaskedLM.from_pretrained(model_name)
tokenizer = BertTokenizer.from_pretrained(model_name)
````

To use it as a masked language model :

````python
import torch

sentence = "Let's have a [MASK]."

model.eval()
encoded_inputs = tokenizer([sentence], padding='longest')
input_ids = torch.tensor(encoded_inputs['input_ids'])
attention_mask = torch.tensor(encoded_inputs['attention_mask'])
output = model(input_ids, attention_mask=attention_mask)

mask_index = input_ids.tolist()[0].index(103)
masked_token = output['logits'][0][mask_index].argmax(axis=-1)
predicted_token = tokenizer.decode(masked_token)

print(predicted_token)
````

Or we can also predict the n most relevant predictions :

````python
top_n = 5

vocab_size = model.config.vocab_size
logits = output['logits'][0][mask_index].tolist()
top_tokens = sorted(list(range(vocab_size)), key=lambda  i:logits[i], reverse=True)[:top_n]

tokenizer.decode(top_tokens)
````