0

I guess there are better "keywords" than that, especially I'm not sure about the "complexity" word. But I thought of none.

Let's say I have a NLP model with 1. a Embedding Layer 2. a LSTM layer 3. an output layer with n classes.

I would like to know if there is an easy way to estimate the time complexity of training the model ? For exemple, to compare with an embedding dimension of 100 vs 200, with LSTM hidden layer dimension of 128 vs 256 ... for a given sequence length. I suppose I could use %timeit on my train() function (or just on criterion(predicted_label, label).backward())... But if there is a "better way", I'd prefer.

EzrielS
  • 323
  • 1
  • 7
  • What would be the preferred output of such a method? What would you want to 'do' with it? A larger architecture leads to longer training times. A larger batch_size leads to faster training times. Training for more epochs trivially leads to longer training times. There are many factors here. – Robin van Hoorn Feb 24 '23 at 11:22

0 Answers0