How can I monitor both training and eval loss when finetuning BERT on a GLUE task?

Question

I am running https://github.com/huggingface/transformers/blob/master/examples/run_glue.py to perform finetuning on a binary classification task (CoLA). I'd like to monitor both the training and evaluation losses to prevent overfitting.

Currently the library is at 2.8.0, and I did the install from source.

When I run the example with

python run_glue.py --model_name_or_path bert-base-uncased 
                   --task_name CoLA 
                   --do_train 
                   --do_eval
                   --data_dir my_dir 
                   --max_seq_length 128
                   --per_gpu_train_batch_size 8
                   --per_gpu_eval_batch_size 8 
                   --learning_rate 2e-5
                   --num_train_epochs 3.0
                   --output_dir ./outputs
                   --logging_steps 5

In the stdout logs I see lines with one single value for the loss, such as

{"learning_rate": 3.3333333333333333e-06, "loss": 0.47537623047828675, "step": 25}

By peeking in https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py I see that training and evaluation losses are computed there (looks to me that code was recently refactored).

I have thus replaced https://github.com/huggingface/transformers/blob/abb1fa3f374811ea09d0bc3440d820c50735008d/src/transformers/trainer.py#L314 with

 cr_loss = self._training_step(model, inputs, optimizer)
 tr_loss += cr_loss

and added after line https://github.com/huggingface/transformers/blob/abb1fa3f374811ea09d0bc3440d820c50735008d/src/transformers/trainer.py#L345

logs["training loss"] = cr_loss

with this I get:

0502 14:12:18.644119 23632 summary.py:47] Summary name training loss is illegal; using training_loss instead.  
                          | 4/10  [00:02<00:04,  1.49it/s]  
{"learning_rate":  3.3333333333333333e-06, "loss": 0.47537623047828675, "training loss": 0.5451719760894775, "step": 25}

Is this OK, or am I doing anything wrong here?

What's the best way to monitor in stdout both the averaged training and evaluation loss for a given logging interval during finetuning?

Answer 1

There's likely no change needed in the code if installing a more recent version (I tried 2.9.0 via pip): just fire the finetuning with the additional flag --evaluate_during_training and output will be OK

0506 12:11:30.021593 34540 trainer.py:551] ***** Running Evaluation ***** 
I0506 12:11:30.022596 34540 trainer.py:552]   Num examples = 140  
I0506 12:11:30.023634 34540 trainer.py:553]   Batch size = 8 Evaluation:  
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 18/18 [00:19<00:00,  1.10s/it]  
{"eval_mcc": 0.0, "eval_loss": 0.6600487811697854, "learning_rate": 3.3333333333333333e-06, "loss": 0.50044886469841, "step": 25}

beware that the example scripts change quite frequently, so flags to accomplish this may change names... see also here https://discuss.huggingface.co/t/how-to-monitor-both-train-and-validation-metrics-at-the-same-step/1301