Title: Transformers From Scratch Data Created: January 25, 2022
Data From: https://www.statmt.org/wmt12/translation-task.html
In this project, I aim to build a transformer model from scratch based on the "Attention Is All You Need" paper:
https://arxiv.org/pdf/1706.03762.pdf
Using PyTorch, I am going to construct a model that will be able to translate an English sentence to a Spanish sentence.
Below are the following python library requirements to be able to run the project
- PyTorch - 1.8.1
- NumPy - 1.19.5
- Matplotlib - 3.4.2
Note: The library version I used are listed as well to ensure the model can be run successfully on other computers.
To execute this project, download the repository and use the following command to run the project:
python3 main.py
There are many hyperparameters that can be tuned in the model, the hyperparameters can be found at the beginning of the main function in main.py. Please note that the larger the keys, values, embeddings, etc., the more memory the model will require to train.
To train the model, set the flag on line 65 named trainModel to true.
The number of steps can be specified through the numSteps variable in the hyperparameter section of the main function and specifies the number of times the model should be updated.
As the model is being trained, the model will be saved after every 5 steps by default. The number of steps after the model is saved can be specified with the stepsToSave in the Other parameters section of the main function.
Additionally, a graph of the model will be saved at the end of training. The graph shows the relationship between the number of steps and the loss at that current step. By default, the graph will be saved to the file named specified by the lossPlotFileName variable in the File variables section of the main function.
After each step, the model will output some text to the console showing its progress.
- The first line specifies the step number
- The second line is the first 10 words in an English training sentence
- The third line is the first 10 words in a Spanish training sentence (what we want the model to predict)
- The fourth line is the first 10 words the model predict on the English training sentence
The output looks as follows:
When the model is finished training, it will be tested on a couple of sentences. By default, the test sentences are:
- Implementation of the President to enhance the Russian Federation
- This is a test sentence
To use different test sentences, change the array named testData in the Testing the model section of the main function. Each line in the array should be a new test sentence.
To skip training and go straight to testing, the trainModel flag can be changed to False in the Other parameters section of the main function.
By default, the model will be saved to the models/ directory and will be saved to a file named modelCkPt.pt. This path can be changed using the modelSaveName variable in the File variables section of the main function.
As stated in the Training The Model section, the model will be saved every 5 steps by default. This parameter can be changed by specifying the number of steps until the model is saved using the stepsToSave variable in the Other parameters section of the main function.
To load a model, change the loadModel flag in the Other parameters to true and change the modelLoadName variable in the File variables section to the path of the file you want to load in.
NOTE: The hyperparameters at the top of the main function must be the same as the hyperparameters of the model being loaded in
Unfortunately, I do not have the hardware required to train the model on a large dataset, so I trained it on a small sample of the data, which can be found in the data2 directory of this repo. After about 1000 steps, the model pretty much mastered the small dataset, as shown below.
Here's the progress the model made during the beginning of it's training
Here's the progress the same model made during the middle of it's training
Here's the progress the model made at the end of it's training on the training dataset
And here's the translations on a few test sentences
As you can see, the model did really good on the training examples and practically perfectly translated all training sentences. As for the test dataset, the model clearly didn't get the predictions correct, but this problem is due to the small training smaple I used. If it used all the data, the model will likely be able to effectively translate these sentences along with new ones it sees.
Below is a plot showing the progress of the model over 1000 steps where the y axis is the loss and the x axis is the step number.
