Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Multi gpu raise error #6

Open
sjtuytc opened this issue Nov 16, 2020 · 0 comments
Open

Multi gpu raise error #6

sjtuytc opened this issue Nov 16, 2020 · 0 comments

Comments

@sjtuytc
Copy link

sjtuytc commented Nov 16, 2020

Hi, in the second training phase, it's fine to use only one gpu, but when using multiple gpus it would run into problems. The layout is as follows. How can I handle this?

2020-11-16 02:03:30.277 | INFO | probnmn.utils.checkpointing:load:156 - Checkpointables not found in file: [] 2020-11-16 02:03:30.335 | INFO | probnmn.utils.checkpointing:load:131 - Loading checkpoint from checkpoints/question_coding_ours/checkpoint_best.pth 2020-11-16 02:03:30.367 | INFO | probnmn.utils.checkpointing:load:153 - optimizer not found in checkpointables. 2020-11-16 02:03:30.368 | INFO | probnmn.utils.checkpointing:load:153 - scheduler not found in checkpointables. 2020-11-16 02:03:30.368 | INFO | probnmn.utils.checkpointing:load:141 - Loading program_generator from checkpoints/question_coding_ours/checkpoint_best.pth 2020-11-16 02:03:30.371 | INFO | probnmn.utils.checkpointing:load:153 - question_reconstructor not found in checkpointables. 2020-11-16 02:03:30.371 | INFO | probnmn.utils.checkpointing:load:156 - Checkpointables not found in file: [] training: 0%| | 0/80000 [00:11<?, ?it/s] Traceback (most recent call last): File "scripts/train.py", line 136, in <module> trainer.step(iteration) File "/localscratch/zelin/batch_soft_reason/baselines/probnmn-clevr/probnmn/trainers/_trainer.py", line 148, in step output_dict = self._do_iteration(batch) File "/localscratch/zelin/batch_soft_reason/baselines/probnmn-clevr/probnmn/trainers/module_training_trainer.py", line 90, in _do_iteration output_dict = self._nmn(batch["image"], pg_output_dict["predictions"], batch["answer"]) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/modules/module.py", line 532, in __call__ result = self.forward(*input, **kwargs) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py", line 153, in forward return self.gather(outputs, self.output_device) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py", line 165, in gather return gather(outputs, output_device, dim=self.dim) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/parallel/scatter_gather.py", line 68, in gather res = gather_map(outputs) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/parallel/scatter_gather.py", line 62, in gather_map for k in out)) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/parallel/scatter_gather.py", line 62, in <genexpr> for k in out)) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/parallel/scatter_gather.py", line 62, in gather_map for k in out)) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/parallel/scatter_gather.py", line 62, in <genexpr> for k in out)) File "/localscratch/ksamel3/anaconda3/envs/soft_reason/lib/python3.7/site-packages/torch/nn/parallel/scatter_gather.py", line 63, in gather_map return type(out)(map(gather_map, zip(*outputs))) TypeError: zip argument #1 must support iteration

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant