Model Card for Model ID

🍿🎥Welcome to the DistilBERT-DeNiro model card!🎞️📽️

We domain adapt (fine-tune) the DistilBERT base model DistilBERT/distilbert-base-uncased on the IMDB movies dataset for a whole word masked language modeling task.

Model Details

Model Description

The DistilBERT base model is fined-tuned using a custom PyTorch training loop. We supervise a training of DistilBERT for the purpose of masked language modeling on stanfordnlp/imdb, an open source dateset from Stanford NLP available through the 🤗 hub stanfordnlp/imdb. This dataset of movie reviews was concatenated into a single string before being chunked and padded to a size of 256 tokens. We fed these chunks into the model during training with a batch size of 32. After training, the ultimate model is domain adapted to fill [MASK] tokens in an input string with terms and lingo common to movies.

Developed by: John Graham Reynolds
Funded by: Vanderbilt University
Model type: Masked Language Model
Language(s) (NLP): English
Finetuned from model: "DistilBERT/distilbert-base-uncased"

Model Sources [optional]

Repository: https://github.com/johngrahamreynolds/DistilBERT-DeNiro

Uses

Direct Use

In order to query the model effectively, one must pass it a string containing a [MASK] token to be filled. An example is text = "This is a great [MASK]!". The domain-adapted model will attempt to fill the mask with a token relevant to movies, cinema, tv, etc.

How to Use and Query the Model

Use the code below to get started with the model. Users pass a text string detailing a sentence with a [MASK] token. The model will provide options to fill the mask based on the sentence context and its background of knowledge. Note - the DistilBERT base model was trained on a very large general corpus of text. In our training, we have fine-tuned the model on the large IMDB movie review dataset. That is, the model is now accustomed to filling [MASK] tokens with words related to the domain of movies/tv/films. To see the model's afinity for cinematic lingo, it is best to be considerate in one's prompt engineering. Specifically, to most likely generate movie related text, one should ideally pass a masked text string that could reasonably be found in someone's review of a movie. See the example below:


import torch
from transformers import AutoModelForMaskedLM, AutoTokenizer

model = AutoModelForMaskedLM.from_pretrained("MarioBarbeque/DistilBERT-DeNiro").to("cuda")
tokenizer = AutoTokenizer.from_pretrained("distilbert/distilbert-base-uncased")

# Pass a unique string with a [MASK] token for the model to fill
text = "This is a great [MASK]!"

tokenized_text = tokenizer(text, return_tensors="pt").to("cuda")
token_logits = model(**tokenized_text).logits

mask_token_index = torch.where(tokenized_text["input_ids"] == tokenizer.mask_token_id)[1]
mask_token_logits = token_logits[0, mask_token_index, :]

top_5_tokens = torch.topk(mask_token_logits, 5, dim=1).indices[0].tolist()
for token in top_5_tokens:
  print(text.replace(tokenizer.mask_token, tokenizer.decode(token)))

This code outputs the following:

This is a great movie!
This is a great film!
This is a great idea!
This is a great show!
This is a great documentary!

Training Details

Training Data / Preprocessing

The data used comes from the Stanford NLP 🤗 hub. The model card can be found here. This dataset is preprocessed in the following way: The train and test splits are tokenized, concatenated, and chunked into chunks of 256 tokens. We subsequently load the training data into a DataCollator that applies a custom random masking function when batching. We mask of 15% of tokens in each chunk. The evaluation data is masked in its entirety, to remove randomness when evaluating, and passed to a DataCollator with the default collating function.

Training Procedure

The model was trained locally on a single-node with one 16GB Nvidia T4 using 🤗 Transformers, 🤗 Tokenizers, and a custom PyTorch training loop that made use of 🤗 Accelerate.

Training Hyperparameters

Precision: We use FP32 precision, as follows immediately from the precision inhereted for the original "DistilBERT/distilbert-base-uncased" model.
Optimizer: AdamW
Learning Rate: We use a linear learing rate scheduler with an initial learning rate of 5e-5
Batch Size: 32
Number of Training Steps: 2877 steps over the course of 3 epochs

Evaluation / Metrics

We evaluate our masked language model's performance using the perplexity metric, which has a few mathematical defitions. We define the perplexity as the exponential of the cross-entropy. To remove randomness in our metrics, we premask our evaluation dataset with a single masking function. This ensures we are evaluating with respect to the same set of labels each epoch. See the wikipedia links for perplexity and cross-entropy below for more a detailed discussion and various other definitions.

Cross-entropy: https://en.wikipedia.org/wiki/Cross-entropy

Perplexity: https://en.wikipedia.org/wiki/Perplexity

Testing Data, Factors & Metrics

Testing Data

The IMDB dataset from Stanford NLP comes pre-split into training and testing data of 25k reviews each. Our preprocessing, which included the chunking of concatenated, tokenized inputs into chunks of 256 tokens, increased these respective splits by approximately ~5k records each. We apply a single masking function to the evaluation dataset before testing as mentioned above.

Results

We find the following perplexity metrics over 3 training epochs:

epoch	perplexity
0	17.38
1	16.28
2	15.78

Summary

We train this model for the purpose of attempting a local training of a masked language model using both the 🤗 ecosystem and a custom PyTorch training and evaluation loop. We look forward to further fine-tuning this model on more film/actor/cinema related data in order to further improve the model's knowledge and ability in this domain - indeed cinema is one of the author's favorite things.

Environmental Impact

Hardware Type: Nvidia Tesla T4 16GB
Hours used: 1.2
Cloud Provider: Microsoft Azure
Compute Region: EastUS
Carbon Emitted: 0.03 kgCO2

Experiments were conducted using Azure in region eastus, which has a carbon efficiency of 0.37 kgCO$_2$eq/kWh. A cumulative of 1.2 hours of computation was performed on hardware of type T4 (TDP of 70W).

Total emissions are estimated to be 0.03 kgCO$_2$eq of which 100 percents were directly offset by the cloud provider.

Estimations were conducted using the MachineLearning Impact calculator presented in Lacoste et al. (2019).

Hardware

The model was trained locally in an Azure Databricks workspace using a single node with 1 16GB Nvidia T4 GPU for 1.2 GPU Hours.

Software

Training utilized PyTorch, 🤗 Transformers, 🤗 Tokenizers, 🤗 Datasets, 🤗 Accelerate, and more in an Azure Databricks execution environment.

Citations

@article{lacoste2019quantifying, title={Quantifying the Carbon Emissions of Machine Learning}, author={Lacoste, Alexandre and Luccioni, Alexandra and Schmidt, Victor and Dandres, Thomas}, journal={arXiv preprint arXiv:1910.09700}, year={2019} }

MarioBarbeque
/

DistilBERT-DeNiro