[Fix] Fix bug in tools/analyse_logs.py caused by wrong plot_iter in some cases.

[MengzhangLI]

Motivation

Fix #1426

Current in tools/analyse_logs.py we assert whether certain line in log.json is validation rather than training set by:

  if pre_iter > epoch_logs['iter'][idx]:
      continue

It's OK in many cases like below 32000 (pre_iter) > 250 (epoch_logs['iter'][idx]).

It is in ConvNeXt, whose validation step is 16000. Of course, 16000, 32000, 48000 and so on those values > 250.
Note that 250 is related with number of images in validation sets.

However, in certain cases when (1) valitaion is large and (2) validation step is small, like below:

The valition has 1334 images and validation step is 200. So 200 (pre_iter) < 1334 (epoch_logs['iter'][idx]). And it would generate many 1334 in plot_iters which makes a horizontal line segment in lr figure:

Solution

Using paragraph below to assert whther to keep values in training set.

  if epoch_logs['mode'][idx] == 'train':
      plot_iters.append(epoch_logs['iter'][idx])
      plot_values.append(epoch_logs[metric][idx])

Result

Take logfile.zip from #1426 for example:

LR before:

LR after:

Loss before:

Loss after: