Performance and Bottleneck Profiler¶
Profiling your training run can help you understand if there are any bottlenecks in your code.
Built-in checks¶
PyTorch Lightning supports profiling standard actions in the training loop out of the box, including:
on_epoch_start
on_epoch_end
on_batch_start
tbptt_split_batch
model_forward
model_backward
on_after_backward
optimizer_step
on_batch_end
training_step_end
on_training_end
Enable simple profiling¶
If you only wish to profile the standard actions, you can set profiler=”simple” when constructing your Trainer object.
trainer = Trainer(..., profiler="simple")
The profiler’s results will be printed at the completion of a training fit().
Profiler Report
Action | Mean duration (s) | Total time (s)
-----------------------------------------------------------------
on_epoch_start | 5.993e-06 | 5.993e-06
get_train_batch | 0.0087412 | 16.398
on_batch_start | 5.0865e-06 | 0.0095372
model_forward | 0.0017818 | 3.3408
model_backward | 0.0018283 | 3.4282
on_after_backward | 4.2862e-06 | 0.0080366
optimizer_step | 0.0011072 | 2.0759
on_batch_end | 4.5202e-06 | 0.0084753
on_epoch_end | 3.919e-06 | 3.919e-06
on_train_end | 5.449e-06 | 5.449e-06
Advanced Profiling¶
If you want more information on the functions called during each event, you can use the AdvancedProfiler. This option uses Python’s cProfiler to provide a report of time spent on each function called within your code.
trainer = Trainer(..., profiler="advanced")
or
profiler = AdvancedProfiler()
trainer = Trainer(..., profiler=profiler)
The profiler’s results will be printed at the completion of a training fit(). This profiler report can be quite long, so you can also specify an output_filename to save the report instead of logging it to the output in your terminal. The output below shows the profiling for the action get_train_batch.
Profiler Report
Profile stats for: get_train_batch
4869394 function calls (4863767 primitive calls) in 18.893 seconds
Ordered by: cumulative time
List reduced from 76 to 10 due to restriction <10>
ncalls tottime percall cumtime percall filename:lineno(function)
3752/1876 0.011 0.000 18.887 0.010 {built-in method builtins.next}
1876 0.008 0.000 18.877 0.010 dataloader.py:344(__next__)
1876 0.074 0.000 18.869 0.010 dataloader.py:383(_next_data)
1875 0.012 0.000 18.721 0.010 fetch.py:42(fetch)
1875 0.084 0.000 18.290 0.010 fetch.py:44(<listcomp>)
60000 1.759 0.000 18.206 0.000 mnist.py:80(__getitem__)
60000 0.267 0.000 13.022 0.000 transforms.py:68(__call__)
60000 0.182 0.000 7.020 0.000 transforms.py:93(__call__)
60000 1.651 0.000 6.839 0.000 functional.py:42(to_tensor)
60000 0.260 0.000 5.734 0.000 transforms.py:167(__call__)
You can also reference this profiler in your LightningModule to profile specific actions of interest. If you don’t want to always have the profiler turned on, you can optionally pass a PassThroughProfiler which will allow you to skip profiling without having to make any code changes. Each profiler has a method profile() which returns a context handler. Simply pass in the name of your action that you want to track and the profiler will record performance for code executed within this context.
from pytorch_lightning.profiler import Profiler, PassThroughProfiler
class MyModel(LightningModule):
def __init__(self, profiler=None):
self.profiler = profiler or PassThroughProfiler()
def custom_processing_step(self, data):
with profiler.profile('my_custom_action'):
# custom processing step
return data
profiler = Profiler()
model = MyModel(profiler)
trainer = Trainer(profiler=profiler, max_epochs=1)
PyTorch Profiling¶
Autograd includes a profiler that lets you inspect the cost of different operators inside your model - both on the CPU and GPU.
Find the Pytorch Profiler doc at [PyTorch Profiler](https://pytorch-lightning.readthedocs.io/en/stable/profiler.html)
trainer = Trainer(..., profiler="pytorch")
or
profiler = PyTorchProfiler(...)
trainer = Trainer(..., profiler=profiler)
This profiler works with PyTorch DistributedDataParallel.
If output_filename is provided, each rank will save their profiled operation to their own file.
The profiler’s results will be printed on the completion of a training fit(). This profiler report can be quite long, so you can also specify an output_filename to save the report instead of logging it to the output in your terminal.
This profiler will record only for training_step_and_backward, evaluation_step and test_step functions by default.
The output below shows the profiling for the action training_step_and_backward.
The user can provide PyTorchProfiler(profiled_functions=[...]) to extend the scope of profiled functions.
Note
When using the PyTorch Profiler, wall clock time will not not be representative of the true wall clock time. This is due to forcing profiled operations to be measured synchronously, when many CUDA ops happen asynchronously. It is recommended to use this Profiler to find bottlenecks/breakdowns, however for end to end wall clock time use the SimpleProfiler. # noqa E501
Profiler Report
Profile stats for: training_step_and_backward
--------------------- --------------- --------------- --------------- --------------- ---------------
Name Self CPU total % Self CPU total CPU total % CPU total CPU time avg
--------------------- --------------- --------------- --------------- --------------- ---------------
t 62.10% 1.044ms 62.77% 1.055ms 1.055ms
addmm 32.32% 543.135us 32.69% 549.362us 549.362us
mse_loss 1.35% 22.657us 3.58% 60.105us 60.105us
mean 0.22% 3.694us 2.05% 34.523us 34.523us
div_ 0.64% 10.756us 1.90% 32.001us 16.000us
ones_like 0.21% 3.461us 0.81% 13.669us 13.669us
sum_out 0.45% 7.638us 0.74% 12.432us 12.432us
transpose 0.23% 3.786us 0.68% 11.393us 11.393us
as_strided 0.60% 10.060us 0.60% 10.060us 3.353us
to 0.18% 3.059us 0.44% 7.464us 7.464us
empty_like 0.14% 2.387us 0.41% 6.859us 6.859us
empty_strided 0.38% 6.351us 0.38% 6.351us 3.175us
fill_ 0.28% 4.782us 0.33% 5.566us 2.783us
expand 0.20% 3.336us 0.28% 4.743us 4.743us
empty 0.27% 4.456us 0.27% 4.456us 2.228us
copy_ 0.15% 2.526us 0.15% 2.526us 2.526us
broadcast_tensors 0.15% 2.492us 0.15% 2.492us 2.492us
size 0.06% 0.967us 0.06% 0.967us 0.484us
is_complex 0.06% 0.961us 0.06% 0.961us 0.481us
stride 0.03% 0.517us 0.03% 0.517us 0.517us
--------------------- --------------- --------------- --------------- --------------- ---------------
Self CPU time total: 1.681ms
When running with PyTorchProfiler(emit_nvtx=True). You should run as following:
nvprof --profile-from-start off -o trace_name.prof -- <regular command here>
To visualize the profiled operation, you can either:
Use:
nvvp trace_name.prof
Use:
python -c 'import torch; print(torch.autograd.profiler.load_nvprof("trace_name.prof"))'
-
class
pytorch_lightning.profiler.AdvancedProfiler(output_filename=None, line_count_restriction=1.0)[source] Bases:
pytorch_lightning.profiler.profilers.BaseProfilerThis profiler uses Python’s cProfiler to record more detailed information about time spent in each function call recorded during a given action. The output is quite verbose and you should only use this if you want very detailed reports.
- Parameters
output_filename¶ (
Optional[str]) – optionally save profile results to file instead of printing to std out when training is finished.line_count_restriction¶ (
float) – this can be used to limit the number of functions reported for each action. either an integer (to select a count of lines), or a decimal fraction between 0.0 and 1.0 inclusive (to select a percentage of lines)
-
describe()[source] Logs a profile report after the conclusion of the training run.
-
stop(action_name)[source] Defines how to record the duration once an action is complete.
- Return type
-
class
pytorch_lightning.profiler.BaseProfiler(output_streams=None)[source] Bases:
abc.ABCIf you wish to write a custom profiler, you should inhereit from this class.
-
profile(action_name)[source] Yields a context manager to encapsulate the scope of a profiled action.
Example:
with self.profile('load training data'): # load training data code
The profiler will start once you’ve entered the context and will automatically stop once you exit the code block.
- Return type
-
abstract
stop(action_name)[source] Defines how to record the duration once an action is complete.
- Return type
-
-
class
pytorch_lightning.profiler.PassThroughProfiler[source] Bases:
pytorch_lightning.profiler.profilers.BaseProfilerThis class should be used when you don’t want the (small) overhead of profiling. The Trainer uses this class by default.
Args: output_streams: callable
-
stop(action_name)[source] Defines how to record the duration once an action is complete.
- Return type
-
-
class
pytorch_lightning.profiler.PyTorchProfiler(output_filename=None, enabled=True, use_cuda=False, record_shapes=False, profile_memory=False, group_by_input_shapes=False, with_stack=False, use_kineto=False, use_cpu=True, emit_nvtx=False, export_to_chrome=False, path_to_export_trace=None, row_limit=20, sort_by_key=None, profiled_functions=None, local_rank=None)[source] Bases:
pytorch_lightning.profiler.profilers.BaseProfilerThis profiler uses PyTorch’s Autograd Profiler and lets you inspect the cost of different operators inside your model - both on the CPU and GPU
- Parameters
output_filename¶ (
Optional[str]) – optionally save profile results to file instead of printing to std out when training is finished. When usingddp, each rank will stream the profiled operation to their own file with the extension_{rank}.txtenabled¶ (
bool) – Setting this to False makes this context manager a no-op.use_cuda¶ (
bool) – Enables timing of CUDA events as well using the cudaEvent API. Adds approximately 4us of overhead to each tensor operation.record_shapes¶ (
bool) – If shapes recording is set, information about input dimensions will be collected.profile_memory¶ (
bool) – Whether to report memory usage, default: True (Introduced in PyTorch 1.6.0)group_by_input_shapes¶ (
bool) – Include operator input shapes and group calls by shape.with_stack¶ (
bool) – record source information (file and line number) for the ops (Introduced in PyTorch 1.7.0)use_kineto¶ (
bool) – experimental support for Kineto profiler (Introduced in PyTorch 1.8.0)use_cpu¶ (
bool) – use_kineto=True and can be used to lower the overhead for GPU-only profiling (Introduced in PyTorch 1.8.0)Context manager that makes every autograd operation emit an NVTX range Run:
nvprof --profile-from-start off -o trace_name.prof -- <regular command here>
To visualize, you can either use:
nvvp trace_name.prof torch.autograd.profiler.load_nvprof(path)
export_to_chrome¶ (
bool) – Wether to export the sequence of profiled operators for Chrome. It will generate a.jsonfile which can be read by Chrome.path_to_export_trace¶ (
Optional[str]) – Directory path to export.jsontraces when usingexport_to_chrome=True. By default, it will be save where the file being is being run.row_limit¶ (
int) – Limit the number of rows in a table, 0 is a special value that removes the limit completely.sort_by_key¶ (
Optional[str]) – Keys to sort out profiled tableprofiled_functions¶ (
Optional[List]) – list of profiled functions which will create a context manager on. Any other will be pass through.local_rank¶ (
Optional[int]) – When running in distributed setting, local_rank is used for each process to write to their own file if output_fname is provided.
- Raises
MisconfigurationException – If arg
sort_by_keyis not present inAVAILABLE_SORT_KEYS, or if log file is not a.txtfile.ValueError – If you attempt to stop recording an action which was never started.
-
describe()[source] Logs a profile report after the conclusion of the training run.
-
stop(action_name)[source] Defines how to record the duration once an action is complete.
- Return type
-
class
pytorch_lightning.profiler.SimpleProfiler(output_filename=None, extended=True)[source] Bases:
pytorch_lightning.profiler.profilers.BaseProfilerThis profiler simply records the duration of actions (in seconds) and reports the mean duration of each action and the total time spent over the entire training run.
- Parameters
output_filename¶ (
Optional[str]) – optionally save profile results to file instead of printing to std out when training is finished.
-
describe()[source] Logs a profile report after the conclusion of the training run.
-
stop(action_name)[source] Defines how to record the duration once an action is complete.
- Return type
