customer-support-chatbot
customer-support-chatbot copied to clipboard
ldulcic
Reproducing results
Following up on discussion from #1
@aitmlouk can you give me more details on how did you train the model? Parameters, number of epochs, batch size, etc. ?
@ldulcic I used the code as it is, in the first step I used the default values from args and the number of epochs vary between 30 and 130 but always no result when it comes to prediction! To speed up the training process I used a training subset from apple support (train ~ 12000 records, val~5000, test~6000). I tried also the default args from pre-trained models and trained more than 40 epochs but always the pre-trained model is not working. How did u get these pre-trained models?
@aitmlouk Well, if you used only a subset of dataset then you will get worse results. Pre-trained models contain args object which has values of all the arguments I used. Also, you can see in predict.py how many epochs it took me to train those models.
Training should take quite a lot of time. I had two GeForce GTX 1080 GPUs → training ran 24-48 hours.
What is you end goal with training this model? What do you want to achieve?
@ldulcic thanks for this clarification, I want to split the dataset into sub-dataset and do distribute training. what you suggest if I want to train the model on small dataset?
@aitmlouk you won't get far with this model if you don't have lot of data. Try fine-tuning BERT.