Skip to content

Bug: shape mismatch in init_mask when assigning template to X[cmask] #2

Description

@dumbgoos

Bug: shape mismatch in init_mask when assigning template to X[cmask]

Summary
During training, we intermittently hit a shape mismatch when writing template into X[cmask].

Command

python train.py --task struct_prediction  --gpus 0 1 2 3  --wandb_offline 1  --flexible 1  --model_type dyAb  --ex_name dyAb_struct_pred_2  --module_type 1  --batch_size 64

Error (exact)
RuntimeError: shape mismatch: value tensor of shape [4411, 14, 3] cannot be broadcast to indexing result of shape [4413, 14, 3]
[rank: 3] Child process with PID 2759890 terminated with code 1. Forcefully terminating all other processes to avoid zombies 🧟
Killed

Where
.../models/dyMEAN/dyAb_model.py, in init_mask:
X[cmask] = template

Context

  • 4×GPU DDP
  • struct_only=True
  • Occasionally cmask.sum()template.size(0) (diff by a small number).

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions