Can Swin UNETR be trained on multi GPU?

[diazandr3s]

Hi,

I'm using the Swin UNETR to segment both brain tumors and deformed brain ventricles. I managed to train a Swin UNETR model using a single GPU without major changes from MONAI Label based on this tutorial: https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/swin_unetr_brats21_segmentation_3d.ipynb.

However, I'm getting this error when enabling multi GPU:

[2022-07-26 00:12:05.099]ERROR - Current run is terminating due to exception: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.
[2022-07-26 00:12:05.099]ERROR - Exception: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.
Traceback (most recent call last):
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 840, in _run_once_on_dataset
self.state.output = self._process_function(self, self.state.batch)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/monai/engines/trainer.py", line 211, in _iteration
engine.scaler.scale(engine.state.output[Keys.LOSS]).backward()
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/_tensor.py", line 363, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/init.py", line 173, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/function.py", line 253, in apply
return user_fn(self, *args)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/utils/checkpoint.py", line 146, in backward
torch.autograd.backward(outputs_with_grad, args_with_grad)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/init.py", line 173, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
RuntimeError: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.
[2022-07-26 00:12:05.100]ERROR - Engine run is terminating due to exception: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.
[2022-07-26 00:12:05.100]ERROR - Exception: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.
Traceback (most recent call last):
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 753, in _internal_run
time_taken = self._run_once_on_dataset()
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 854, in _run_once_on_dataset
self._handle_exception(e)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 464, in _handle_exception
self._fire_event(Events.EXCEPTION_RAISED, e)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 421, in _fire_event
func(*first, *(event_args + others), **kwargs)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/monai/handlers/stats_handler.py", line 179, in exception_raised
raise e
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 840, in _run_once_on_dataset
self.state.output = self._process_function(self, self.state.batch)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/monai/engines/trainer.py", line 211, in _iteration
engine.scaler.scale(engine.state.output[Keys.LOSS]).backward()
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/_tensor.py", line 363, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/init.py", line 173, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/function.py", line 253, in apply
return user_fn(self, *args)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/utils/checkpoint.py", line 146, in backward
torch.autograd.backward(outputs_with_grad, args_with_grad)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/init.py", line 173, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
RuntimeError: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.
[2022-07-26 00:12:05.136]ERROR - Current run is terminating due to exception: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.
[2022-07-26 00:12:05.136]ERROR - Engine run is terminating due to exception: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.
Traceback (most recent call last):
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/code.py", line 90, in runcode
exec(code, self.locals)
File "", line 1, in
File "/snap/pycharm-professional/295/plugins/python/helpers/pydev/_pydev_bundle/pydev_umd.py", line 198, in runfile
pydev_imports.execfile(filename, global_vars, local_vars) # execute the script
File "/snap/pycharm-professional/295/plugins/python/helpers/pydev/_pydev_imps/_pydev_execfile.py", line 18, in execfile
exec(compile(contents+"\n", file, 'exec'), glob, loc)
File "/home/andres/Documents/workspace/MONAILabel/sample-apps/radiology/main.py", line 281, in
main()
File "/home/andres/Documents/workspace/MONAILabel/sample-apps/radiology/main.py", line 265, in main
app.train(
File "/home/andres/Documents/workspace/MONAILabel/monailabel/interfaces/app.py", line 411, in train
result = task(request, self.datastore())
File "/home/andres/Documents/workspace/MONAILabel/monailabel/tasks/train/basic_train.py", line 378, in call
torch.multiprocessing.spawn(main_worker, nprocs=world_size, args=(world_size, req, datalist, self))
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/multiprocessing/spawn.py", line 240, in spawn
return start_processes(fn, args, nprocs, join, daemon, start_method='spawn')
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/multiprocessing/spawn.py", line 198, in start_processes
while not context.join():
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/multiprocessing/spawn.py", line 160, in join
raise ProcessRaisedException(msg, error_index, failed_process.pid)
torch.multiprocessing.spawn.ProcessRaisedException:
-- Process 1 terminated with the following error:
Traceback (most recent call last):
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/multiprocessing/spawn.py", line 69, in _wrap
fn(i, *args)
File "/home/andres/Documents/workspace/MONAILabel/monailabel/tasks/train/basic_train.py", line 608, in main_worker
task.train(rank, world_size, request, datastore)
File "/home/andres/Documents/workspace/MONAILabel/monailabel/tasks/train/basic_train.py", line 432, in train
context.trainer.run()
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/monai/engines/trainer.py", line 58, in run
super().run()
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/monai/engines/workflow.py", line 290, in run
super().run(data=self.data_loader, max_epochs=self.state.max_epochs)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 704, in run
return self._internal_run()
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 783, in _internal_run
self._handle_exception(e)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 466, in _handle_exception
raise e
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 753, in _internal_run
time_taken = self._run_once_on_dataset()
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 854, in _run_once_on_dataset
self._handle_exception(e)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 466, in _handle_exception
raise e
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/ignite/engine/engine.py", line 840, in _run_once_on_dataset
self.state.output = self._process_function(self, self.state.batch)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/monai/engines/trainer.py", line 211, in _iteration
engine.scaler.scale(engine.state.output[Keys.LOSS]).backward()
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/_tensor.py", line 363, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/init.py", line 173, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/function.py", line 253, in apply
return user_fn(self, *args)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/utils/checkpoint.py", line 146, in backward
torch.autograd.backward(outputs_with_grad, args_with_grad)
File "/home/andres/anaconda3/envs/monai_env/lib/python3.9/site-packages/torch/autograd/init.py", line 173, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
RuntimeError: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations.
Parameter at index 113 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.

[Nic-Ma]

Hi @tangy5 ,

Could you please help share some comments or best practices here?

Thanks in advance.

[nvahmadi]

Not sure whether this is the source of the problem, but it seems that above code was run with gradient checkpointing. During my experiments in the research-contributions Swin-UNETR SSL repo, I made the experience that multi-GPU only works without gradient checkpointing. Single-GPU works with and without gradient checkpointing. Could be a hint?

[OeslleLucena]

@diazandr3s have you manged to sort that out? I am currently having the same issue when wrapping SwinUNETR using PyTorch lightning

[tangy5]

Hi @diazandr3s , in this repo, it seems the BRATS21 with SwinUNETR works, can you refer to this example for multi-GPU training?
https://github.com/Project-MONAI/research-contributions/tree/main/SwinUNETR/BRATS21

[OeslleLucena]

@tangy5 setting use_checkpoint to false as mentioned link and DDPStrategy(find_unused_parameters=False) fixed my problem.

[tangy5]

Oh, that's great, yeah, the use_checkpoint option impact the testing/inference. DDP training should be good without this.
Thank you for updates.