Milestone 7 – Executing the training loops

To begin training, just run the following command:

trainer.train()

Figure 4.1 shows an example of the output you can expect to see from the trainer.train() command’s execution:

Figure 4.1 – Sample output from trainer.train() in Google Colab

Each training batch will have an evaluation step that calculates and displays training/validation losses and WER metrics. Depending on your GPU, training could take 5–10 hours. If you run into memory issues, try reducing the batch size and adjusting gradient_accumulation_steps in the declaration of Seq2SeqTrainingArguments.

Because of the parameters we established when declaring Seq2SeqTrainingArguments, our model metrics and performance will be pushed to the Hugging Face Hub with each training iteration. The key parameters driving that push to the Hub are shown here:

from transformers import Seq2SeqTrainingArguments
training_args = Seq2SeqTrainingArguments...