0

I tried the solution here: sklearn logistic regression loss value during training

With verbose=0 and verbose=1. loss_history is nothing, and loss_list is empty, although the epoch number and change in loss are still printed in the terminal.

Epoch 1, change: 1.00000000
Epoch 2, change: 0.32949890
Epoch 3, change: 0.19452967
Epoch 4, change: 0.14287635
Epoch 5, change: 0.11357212

I also tried the solution proposed here: how to plot correctly loss curves for training and validation sets?

I.e., training single epochs iteratively. In this case it doesn't train at all, the loss change is always 1:

Epoch 1, change: 1.00000000
max_iter reached after 2 seconds
Epoch 1, change: 1.00000000
max_iter reached after 1 seconds
Epoch 1, change: 1.00000000
max_iter reached after 1 seconds
Epoch 1, change: 1.00000000
max_iter reached after 2 seconds

Is there really no straightforward way to do this? I'm totally open to hacky workarounds as well.

My simple code that I want to plot from:

logreg = LogisticRegression(
   random_state=42,
   C=.001,
   penalty="l1",
   max_iter=500,
   solver="saga",
   n_jobs=8,
   warm_start=True,
   class_weight='balanced', 
   verbose=1)
 
logreg.fit(X_train, y_train)
neverreally
  • 45
  • 6

1 Answers1

0

I only managed a hacky way to solve this, since training iteratively gives different results than a single call to model.fit and the workaround mentioned in other solutions using sys.stdout unfortunately does not work for the LogisticRegression() class, although it does for SGDClassifier().

I run the Python file from the terminal and pipe the output to a file directly:

python3 logreg_train.py > terminal_output.txt

Then one can parse the text file to extract the change in training loss.

Hope this helps someone!

neverreally
  • 45
  • 6