Unclear what the column 'E' is outputting in the console output during training

[erotavlas]

At first I thought it was an iteration within each Epoch but now I'm not sure after seeing it complete an entire training. I set my max epochs parameter in the configuration to 25 and this was the result. It looks like there are 25 rows but the Epoch number doesn't correspond to the max value I specified in the configuration. Also I'm not sure how this related to the 'iterations' parameter that we used to use in the spaCy 2.x. Thank s for any clarification.

[polm]

E is for epochs, # is for steps (iterations). Training can end before the max epochs / steps if you've set a patience value. You can read more about how these values interact in #7465.

It looks like your score (right column) didn't improve for 8 * 200 = 1600 steps, which is the default patience value, and should be listed in your config.

Does that make sense?

[erotavlas]

It's kind of confusing. Maybe you should relabel the # column with the actual meaning (i.e. Steps)
I'm not clear what Steps are, but in TensorFlow, patience is based on number of epochs, not steps which also contributes to the confusion here.

[DavidGerva]

@polm (or maybe also @erotavlas) I continue here because I think can be useful for someone else.

I still don't get what a "step" is: In one epoch the earning algorithm works through the entire training dataset once. Right? But what does the algorithm (spacy) is doing in one step?

[adrianeboyd]

A step is running nlp.update for one training batch, with a batch as configured by training.batcher.

[kumarsuraj9450]

I'm still not getting the whole process. Suppose if i have 20k documents in training data and i want to do 5 epochs what should be the configuration for batch size, eval frequency and max steps If patience is set to 0?

Also correct me if I'm wrong but a step means one batch is processed. And one batch means "n" number of data points from training dataset. Like in above code a 100 datapoints from training dataset.

Weights gets updated after each batch or step right?

[adrianeboyd]

To disable everything else and use only max_epochs, set max_steps and patience to 0.

Batch size and eval frequency are independent of the the number of epochs.

Batch sizes aren't necessarily static, they are determined by the [training.batcher] settings in the config.

Eval frequency mainly matters in terms of how often a model-best model is saved and for patience (early stopping). In terms of what is saved as the model-last model, eval frequency doesn't matter. For training for a fixed epoch count, eval frequency could be very high to reduce time spent on eval during training.

Weights are updated after each batch/step.

Please also take a look at the discussion linked above: #7465