Display Loss Table

A tool to display losses in the console when training a model. I prefer a table to display losses rather than a progress bar.

How to use

1. Create a DisplayLossTable instance.
2. Collect losses and form a table in training.
3. Display losses in every iteration.
4. Call end() to terminate a display procedure.

dt = DisplayLossTable(width=20)
for epoch in range(10):
  for iteration in range(100):
    loss1 = compute_loss1(...)
    loss2 = compute_loss2(...)
    loss_data = {'loss1_name': loss1, 'loss2_name': loss2}
    dt.display(epoch, iteration, loss_data)
  val_loss = validate(...)
  print('val: {}'.format(val_loss))
  dt.end()

Then the console displays:

+--------------------+--------------------+
|epoch: 0            |iter: 99            |
+--------------------+--------------------+
|loss1_name: 0.0396  |loss2_name: 0.2114  |
+--------------------+--------------------+
val: 0.1107722818851471
+--------------------+--------------------+
|epoch: 1            |iter: 99            |
+--------------------+--------------------+
|loss1_name: 0.0847  |loss2_name: 0.3809  |
+--------------------+--------------------+
val: 0.10855870693922043

...

+--------------------+--------------------+
|epoch: 9            |iter: 99            |
+--------------------+--------------------+
|loss1_name: 0.0372  |loss2_name: 0.2032  |
+--------------------+--------------------+
val: 0.1010628417134285

The current table refreshes after an iteration.

Type of loss data

The display method receives a list or a dict as loss data.

Display a Dict

Example 1

...
loss_data = {'a': 1, 'b': 2, 'c': 3}
dt.display(0, 0, loss_data)
...

+--------------------+--------------------+
|epoch: 0            |iter: 0             |
+--------------------+--------------------+--------------------+
|a: 1.0000           |b: 2.0000           |c: 3.0000           |
+--------------------+--------------------+--------------------+

Example 2

...
loss_data = {'a': 1}
dt.display(0, 0, loss_data)
...

+--------------------+--------------------+
|epoch: 0            |iter: 0             |
+--------------------+--------------------+
|a: 1.0000           |
+--------------------+

Display a List

Example 1

...
loss_data = [{'a': 1, 'b': 2, 'c': 3}, {'d': 4}]
dt.display(0, 0, loss_data)
...

+--------------------+--------------------+
|epoch: 0            |iter: 0             |
+--------------------+--------------------+--------------------+
|a: 1.0000           |b: 2.0000           |c: 3.0000           |
+--------------------+--------------------+--------------------+
|d: 4.0000           |
+--------------------+

Example 2

...
loss_data = [{'a': 1}, {'b': 2, 'c': 3, 'd': 4}]
dt.display(0, 0, loss_data)
...

+--------------------+--------------------+
|epoch: 0            |iter: 0             |
+--------------------+--------------------+
|a: 1.0000           |
+--------------------+--------------------+--------------------+
|b: 2.0000           |c: 3.0000           |d: 4.0000           |
+--------------------+--------------------+--------------------+

Why need to call end() at the end of an epoch

When display(...) called, DisplayLossTable records how many rows printed this time and print the same number of \x1b[1A for the next refreshing. At the end of an epoch, we may have to print some validation information. If end() is not called, the information will be print on the first row of the table.

If you do not want to show tables for every epoch, it is unnecessary to call end().