TimeProfiler#
Lightning Callback which allows you to measure the time each step takes and log it during the training process.
Logged values#
TimeProfiler logs the following steps:
| Name | Logged by default | Description |
|---|---|---|
| train batch |  |
Time taken by forward, optimizer step, and backward during train step. |
| validation batch | |
Time taken by forward during validation step. |
| train epoch | |
Time taken by train epoch without validation. |
| validation epoch | |
Time taken by validation epoch. |
| avg train downtime* | |
Average downtime in training step. |
| avg val downtime | |
Average downtime in validation step. |
| backward |  |
Time taken by backprop. |
| optimizer step | |
Time taken by optimizer. |
| total train downtime | |
Total downtime in training epoch. |
| total val downtime | |
Total downtime in validation epoch. |
*Downtime - the process during which model does not work (e.g. data loader is working now)
Usage#
from thunder.callbacks import TimeProfiler
from lightning import Trainer
# logs default keys and in addition backward and optimizer step
trainer = Trainer(callbacks=[TimeProfiler("backward", "optimizer step")])