File size: 3,369 Bytes
043ba77 3b729a0 043ba77 98d82d2 043ba77 4b0f874 043ba77 72feb78 043ba77 72feb78 043ba77 586054b 72feb78 043ba77 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 |
---
language: fr
license: mit
datasets:
- oscar
widget:
- text: "J'aime lire les <mask> de SF."
---
DistilCamemBERT
===============
We present a distillation version of the well named [CamemBERT](https://huggingface.co/camembert-base), a RoBERTa French model version, alias DistilCamemBERT. The aim of distillation is to drastically reduce the complexity of the model while preserving the performances. The proof of concept is shown in the [DistilBERT paper](https://arxiv.org/abs/1910.01108) and the code used for the training is inspired by the code of [DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/research_projects/distillation).
Loss function
-------------
The training for the distilled model (student model) is designed to be the closest as possible to the original model (teacher model). To perform this the loss function is composed of 3 parts:
* DistilLoss: a distillation loss which measures the closely probability between the student and teacher outputs with a cross-entropy loss on the MLM task ;
* MLMLoss: a Masked Language Modeling (MLM) task loss to perform the student model with the original task of teacher model ;
* CosineLoss: and finally a cosine embedding loss. This loss function is applied on the last hidden layers of student and teacher models to guarantee a collinearity between us.
The final loss function is a combination of these three loss functions. We use the following ponderation:
Loss = 0.5 DistilLoss + 0.2 MLMLoss + 0.3 CosineLoss
Dataset
-------
To limit the bias between the student and teacher models, the dataset used for the DstilCamemBERT training is the same as the camembert-base training one: OSCAR. The french part of this dataset approximately represents 140 GB on a hard drive disk.
Training
--------
We pre-trained the model on a nVidia Titan RTX during 18 days.
Evaluation results
------------------
| Dataset name | f1-score |
| :----------: | :------: |
| [FLUE](https://huggingface.co/datasets/flue) CLS | 83% |
| [FLUE](https://huggingface.co/datasets/flue) PAWS-X | 77% |
| [FLUE](https://huggingface.co/datasets/flue) XNLI | 68% |
| [wikiner_fr](https://huggingface.co/datasets/Jean-Baptiste/wikiner_fr) NER | 92% |
How to use DistilCamemBERT
--------------------------
Load CamemBERT and its sub-word tokenizer :
```python
from transformers import CamembertModel, CamembertTokenizer
tokeinzer = CamembertTokenizer.from_pretrained("cmarkea/distilcamembert-base")
model = CamembertModel.from_pretrained("cmarkea/distilcamembert-base")
model.eval()
...
```
Filling masks using pipeline :
```python
from transformers import pipeline
model_fill_mask = pipeline("fill-mask", model="cmarkea/distilcamembert-base", tokenizer="cmarkea/distilcamembert-base")
results = model_fill_mask("Le camembert est <mask> :)")
#results
#[{'sequence': '<s> Le camembert est délicieux :)</s>', 'score': 0.3878222405910492, 'token': 7200},
# {'sequence': '<s> Le camembert est excellent :)</s>', 'score': 0.06469205021858215, 'token': 2183},
# {'sequence': '<s> Le camembert est parfait :)</s>', 'score': 0.04534877464175224, 'token': 1654},
# {'sequence': '<s> Le camembert est succulent :)</s>', 'score': 0.04128391295671463, 'token': 26202},
# {'sequence': '<s> Le camembert est magnifique :)</s>', 'score': 0.02425697259604931, 'token': 1509}]
``` |