Hello ,thanks for sharing of your work. I' m using MT-DNN on a custom datasets. I'm following the tutorials so I did single task learning and MTL, but I fail to do KD. I don't know if it is necessary a processing of data or others previous steps. I 've simply used
!python train.py --task_def tutorials/tutorial_task_def.yml --data_dir tutorials/bert_base_uncased/ --init_checkpoint="mt_dnn_models/mt_dnn_kd_large_cased.pt" --train_datasets MyTask --test_datasets MytTask --epochs=1 --batch_size=1
Later, I have had an error:
RuntimeError: CUDA error: CUBLAS_STATUS_INTERNAL_ERROR when calling cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc). I'm using Google Colab and I've tried differents batch size.
Hello ,thanks for sharing of your work. I' m using MT-DNN on a custom datasets. I'm following the tutorials so I did single task learning and MTL, but I fail to do KD. I don't know if it is necessary a processing of data or others previous steps. I 've simply used
!python train.py --task_def tutorials/tutorial_task_def.yml --data_dir tutorials/bert_base_uncased/ --init_checkpoint="mt_dnn_models/mt_dnn_kd_large_cased.pt" --train_datasets MyTask --test_datasets MytTask --epochs=1 --batch_size=1
Later, I have had an error:RuntimeError: CUDA error: CUBLAS_STATUS_INTERNAL_ERROR when calling cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)
. I'm using Google Colab and I've tried differents batch size.