search

joslefaure / HIT

Official Implementation of our WACV2023 paper: “Holistic Interaction Transformer Network for Action Detection”
https://arxiv.org/abs/2210.12686
58 stars 10 forks source link

CUDA error #54

Open hhh804 opened 1 week ago

hhh804 commented 1 week ago

I tried to run 'test_net.py' file, and I got the error: return self._aggregate(person_boxes, query, person_key, object_key, hand_key, mem_key) File "/media/data4/home/HIT/hit/modeling/roi_heads/action_head/hit_structure.py", line 385, in _aggregate query = I_block(query, person_key) File "/media/data4/home/miniconda3/envs/hit/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, *kwargs) File "/media/data4/home/HIT/hit/modeling/roi_heads/action_head/hit_structure.py", line 131, in forward qk = torch.bmm(query_batch, key_batch) # n, thw, max_other thw RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasSgemmStridedBatched( handle, opa, opb, m, n, k, &alpha, a, lda, stridea, b, ldb, strideb, &beta, c, ldc, stridec, num_batches)

How can I fix it? Thank you.

joslefaure commented 1 week ago

Seems like your cuda version doesn't match your torch version. I recommend using torch=1.7.1 and cuda 10.2

hhh804 commented 1 week ago

My environment is cuda11.3,torch 1.10.0+cu11.3, python 3.7. When I installed pytorch 1.7.1 with cuda 10.2 and run "pip install -e .", I got this error: "nvcc fatal: Unsupported gpu architecture 'compute_86', indicates that CUDA 10.2 does not support the architecture of your NVIDIA GeForce RTX 3070 GPU, which requires compute_86 (architecture for NVIDIA Ampere GPUs)." Help me fix the error, please.