Evaluating error

[TommeyChang]

Evaluating on val_seen env ...
Epoch: [1][1/16] Time 1.776 (1.776) Loss inf (inf)
Epoch: [1][2/16] Time 1.733 (1.754) Loss inf (inf)
Epoch: [1][3/16] Time 1.885 (1.798) Loss inf (inf)
Epoch: [1][4/16] Time 1.867 (1.815) Loss inf (inf)
Epoch: [1][5/16] Time 1.919 (1.836) Loss inf (inf)
Epoch: [1][6/16] Time 1.874 (1.842) Loss inf (inf)
Epoch: [1][7/16] Time 1.779 (1.833) Loss inf (inf)
Epoch: [1][8/16] Time 1.958 (1.849) Loss inf (inf)
Epoch: [1][9/16] Time 1.626 (1.824) Loss inf (inf)
Warning: NaN or Inf found in input tensor.
Epoch: [1][10/16] Time 1.899 (1.832) Loss inf (inf)
Epoch: [1][11/16] Time 1.697 (1.819) Loss inf (inf)
Epoch: [1][12/16] Time 1.781 (1.816) Loss inf (inf)
Epoch: [1][13/16] Time 1.938 (1.826) Loss inf (inf)
Epoch: [1][14/16] Time 1.610 (1.810) Loss inf (inf)
Epoch: [1][15/16] Time 1.358 (1.780) Loss inf (inf)
Epoch: [1][16/16] Time 1.543 (1.765) Loss inf (inf)

[chihyaoma]

Hi,

This is perfectly normal for the regretful agent since we have prevent_oscillation enabled. What this means is that, we block a direction that will lead to oscillation, i.e., a vector for that particular direction is being masked.

However, if that direction is actually the ground-truth the agent should pick (during inference), the computed loss will be effective inf. This will only happen during inference though (it won't affect your training).

However, I haven't seen the warning message before. Perhaps it's a warning from new PyTorch version?

[TommeyChang]

But it seem no warning while training.