Bart长文本训练问题

[YoungChanYY]

我用Bart训练代码，每个训练数据都为：输入文本约1000字符，输出文本长3-5万字符。训练几个epoch后会出错，错误信息如下所示。 但是控制输入和输出的字符长度，比如都为100字符左右，则训练正常，没有报错。

请问一下：Bart模型的输入输出长度有什么要求吗，这应该是内部embedding维度出错了吧。谢谢。

RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling cublasSgemmStridedBatched( handle, opa, opb, m, n, k, &alpha, a, lda, stridea, b, ldb, strideb, &beta, c, ldc, stridec, num_batches)

[YoungChanYY]

出错的位置好像是在predict位置。当取消在训练过程中进行eval处理时，训练得以正常进行。大佬

Traceback (most recent call last): File "train_bart_text2abc.py", line 180, in main() File "train_bart_text2abc.py", line 163, in main model.train_model(train_df, eval_data=eval_df, split_on_space=True, matches=count_matches) File "textgen/seq2seq/bart_seq2seq_model.py", line 452, in train_model kwargs, File "textgen/seq2seq/bart_seq2seq_model.py", line 983, in train kwargs, File "textgen/seq2seq/bart_seq2seq_model.py", line 1153, in eval_model preds = self.predict(to_predict, split_on_space=split_on_space) File "textgen/seq2seq/bart_seq2seq_model.py", line 1310, in predict num_return_sequences=self.args.num_return_sequences, File "/usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(args, kwargs) File "/usr/local/lib/python3.7/dist-packages/transformers/generation/utils.py", line 1400, in generate model_kwargs, File "/usr/local/lib/python3.7/dist-packages/transformers/generation/utils.py", line 2183, in greedy_search output_hidden_states=output_hidden_states, File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1110, in _call_impl return forward_call(input, kwargs) File "/usr/local/lib/python3.7/dist-packages/transformers/models/bart/modeling_bart.py", line 1389, in forward return_dict=return_dict, File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1110, in _call_impl return forward_call(*input, *kwargs) File "/usr/local/lib/python3.7/dist-packages/transformers/models/bart/modeling_bart.py", line 1268, in forward return_dict=return_dict, File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1110, in _call_impl return forward_call(input, kwargs) File "/usr/local/lib/python3.7/dist-packages/transformers/models/bart/modeling_bart.py", line 1124, in forward use_cache=use_cache, File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1110, in _call_impl return forward_call(*input, *kwargs) File "/usr/local/lib/python3.7/dist-packages/transformers/models/bart/modeling_bart.py", line 431, in forward output_attentions=output_attentions, File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1110, in _call_impl return forward_call(input, **kwargs) File "/usr/local/lib/python3.7/dist-packages/transformers/models/bart/modeling_bart.py", line 275, in forward attn_output = torch.bmm(attn_probs, value_states) RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling cublasSgemmStridedBatched( handle, opa, opb, m, n, k, &alpha, a, lda, stridea, b, ldb, strideb, &beta, c, ldc, stridec, num_batches) ../aten/src/ATen/native/cuda/Indexing.cu:650: indexSelectSmallIndex: block: [0,0,0], thread: [0,0,0] Assertion srcIndex < srcSelectDimSize failed.

[shibing624]

我看看evaluate的逻辑

[YoungChanYY]

多谢。

我看到另一处地方，应该有些问题： 在textgen/seq2seq/bart_seq2seq_utils.py的preprocess_data_bart(data)函数中，对target_ids 数据处理的问题和建议如下，大佬看看对不对。谢谢！

def preprocess_data_bart(data): input_text, target_text, tokenizer, args = data ...... target_ids = tokenizer.batch_encode_plus( [target_text],

max_length=args.max_seq_length, #原代码

    max_length=args.max_length,            #建议代码
    padding="max_length",
    return_tensors="pt",
    truncation=True,
)

[shibing624]

对的，fixed： https://github.com/shibing624/textgen/commit/7a0be5931234262165d32fc0f915af822e0b1665

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.(由于长期不活动，机器人自动关闭此问题，如果需要欢迎提问)