@@ -17,7 +17,7 @@ This folder contains the original code used to train Distil* as well as examples
...
@@ -17,7 +17,7 @@ This folder contains the original code used to train Distil* as well as examples
## What is Distil*
## What is Distil*
Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distilled-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
We have applied the same method to other Transformer architectures and released the weights:
We have applied the same method to other Transformer architectures and released the weights:
- GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 16.3 compared to 21.1 for **DistilGPT2** (after fine-tuning on the train set).
- GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 16.3 compared to 21.1 for **DistilGPT2** (after fine-tuning on the train set).
...
@@ -57,7 +57,7 @@ Here are the results on the *test* sets for 6 of the languages available in XNLI
...
@@ -57,7 +57,7 @@ Here are the results on the *test* sets for 6 of the languages available in XNLI
This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`.
This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`.
**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0).
**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breaking changes compared to v1.1.0).
Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurrences of each tokens in the data:
Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smooths the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurrences of each tokens in the data:
**Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract.py` and `scripts/extract_distilbert.py` to create a valid initialization checkpoint and use `--student_pretrained_weights` argument to use this initialization for the distilled training!
**Tips:** Starting distilled training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract.py` and `scripts/extract_distilbert.py` to create a valid initialization checkpoint and use `--student_pretrained_weights` argument to use this initialization for the distilled training!
@@ -21,7 +21,7 @@ You can also have a look at this fun *Explain Like I'm Five* introductory [slide
...
@@ -21,7 +21,7 @@ You can also have a look at this fun *Explain Like I'm Five* introductory [slide
One promise of extreme pruning is to obtain extremely small models that can be easily sent (and stored) on edge devices. By setting weights to 0., we reduce the amount of information we need to store, and thus decreasing the memory size. We are able to obtain extremely sparse fine-pruned models with movement pruning: ~95% of the dense performance with ~5% of total remaining weights in the BERT encoder.
One promise of extreme pruning is to obtain extremely small models that can be easily sent (and stored) on edge devices. By setting weights to 0., we reduce the amount of information we need to store, and thus decreasing the memory size. We are able to obtain extremely sparse fine-pruned models with movement pruning: ~95% of the dense performance with ~5% of total remaining weights in the BERT encoder.
In [this notebook](https://github.com/huggingface/transformers/blob/master/examples/movement-pruning/Saving_PruneBERT.ipynb), we showcase how we can leverage standard tools that exist out-of-the-box to efficiently store an extremely sparse question answering model (only 6% of total remaining weights in the encoder). We are able to reduce the memory size of the encoder **from the 340MB (the orignal dense BERT) to 11MB**, without any additional training of the model (every operation is performed *post fine-pruning*). It is sufficiently small to store it on a [91' floppy disk](https://en.wikipedia.org/wiki/Floptical) 📎!
In [this notebook](https://github.com/huggingface/transformers/blob/master/examples/movement-pruning/Saving_PruneBERT.ipynb), we showcase how we can leverage standard tools that exist out-of-the-box to efficiently store an extremely sparse question answering model (only 6% of total remaining weights in the encoder). We are able to reduce the memory size of the encoder **from the 340MB (the original dense BERT) to 11MB**, without any additional training of the model (every operation is performed *post fine-pruning*). It is sufficiently small to store it on a [91' floppy disk](https://en.wikipedia.org/wiki/Floptical) 📎!
While movement pruning does not directly optimize for memory footprint (but rather the number of non-null weights), we hypothetize that further memory compression ratios can be achieved with specific quantization aware trainings (see for instance [Q8BERT](https://arxiv.org/abs/1910.06188), [And the Bit Goes Down](https://arxiv.org/abs/1907.05686) or [Quant-Noise](https://arxiv.org/abs/2004.07320)).
While movement pruning does not directly optimize for memory footprint (but rather the number of non-null weights), we hypothetize that further memory compression ratios can be achieved with specific quantization aware trainings (see for instance [Q8BERT](https://arxiv.org/abs/1910.06188), [And the Bit Goes Down](https://arxiv.org/abs/1907.05686) or [Quant-Noise](https://arxiv.org/abs/2004.07320)).
The National Library of Sweden / KBLab releases three pretrained language models based on BERT and ALBERT. The models are trained on aproximately 15-20GB of text (200M sentences, 3000M tokens) from various sources (books, news, government publications, swedish wikipedia and internet forums) aiming to provide a representative BERT model for Swedish text. A more complete description will be published later on.
The National Library of Sweden / KBLab releases three pretrained language models based on BERT and ALBERT. The models are trained on approximately 15-20GB of text (200M sentences, 3000M tokens) from various sources (books, news, government publications, swedish wikipedia and internet forums) aiming to provide a representative BERT model for Swedish text. A more complete description will be published later on.
The following three models are currently available:
The following three models are currently available:
...
@@ -86,7 +86,7 @@ for token in nlp(text):
...
@@ -86,7 +86,7 @@ for token in nlp(text):
print(l)
print(l)
```
```
Which should result in the following (though less cleanly formated):
Which should result in the following (though less cleanly formatted):
The National Library of Sweden / KBLab releases three pretrained language models based on BERT and ALBERT. The models are trained on aproximately 15-20GB of text (200M sentences, 3000M tokens) from various sources (books, news, government publications, swedish wikipedia and internet forums) aiming to provide a representative BERT model for Swedish text. A more complete description will be published later on.
The National Library of Sweden / KBLab releases three pretrained language models based on BERT and ALBERT. The models are trained on approximately 15-20GB of text (200M sentences, 3000M tokens) from various sources (books, news, government publications, swedish wikipedia and internet forums) aiming to provide a representative BERT model for Swedish text. A more complete description will be published later on.
The following three models are currently available:
The following three models are currently available:
...
@@ -86,7 +86,7 @@ for token in nlp(text):
...
@@ -86,7 +86,7 @@ for token in nlp(text):
print(l)
print(l)
```
```
Which should result in the following (though less cleanly formated):
Which should result in the following (though less cleanly formatted):
