self.random_sample和 self.beam_search返回值tensor未统一转化numpy，导致两函数同时调用报错

[ocyisheng]

提问时请尽可能提供如下信息：

基本信息

• 你使用的操作系统: win0
• 你使用的Python版本: 3.10
• 你使用的Pytorch版本: 2.1
• 你使用的bert4torch版本:
• 你加载的预训练模型:

核心代码

# 请在此处贴上你的核心代码

def generate(self, passage, topk=1, topp=0.95):
token_ids, segment_ids = tokenizer.encode(passage, maxlen=max_p_len)
a_ids = self.random_sample([token_ids, segment_ids], n=1, topp=topp)[0] # 基于随机采样
token_ids += list(a_ids)
segment_ids += [1] * len(a_ids)
q_ids = self.beam_search([token_ids, segment_ids], topk=topk)[0] # 基于beam search
return (tokenizer.decode(q_ids.cpu().numpy()), tokenizer.decode(a_ids.cpu().numpy()))

输出信息

# 请在此处贴上你的调试输出

Traceback (most recent call last):
File "F:\project\code\python\bert4torch\examples\seq2seq\task_question_answer_generation_by_seq2seq.py", line 191, in
valid(valid_data)
File "F:\project\code\python\bert4torch\examples\seq2seq\task_question_answer_generation_by_seq2seq.py", line 165, in valid
q,a = qag.generate(d[0])
File "F:\project\code\python\bert4torch\examples\seq2seq\task_question_answer_generation_by_seq2seq.py", line 145, in generate
q_ids = self.beam_search([token_ids, segment_ids], topk=topk)[0] # 基于beam search
File "F:\project\code\python\bert4torch\bert4torch\generation.py", line 375, in beam_search
inputs = self._trans2tensors(inputs)
File "F:\project\code\python\bert4torch\bert4torch\generation.py", line 197, in trans2tensors
input_new = torch.tensor(sequence_padding(input, value=self.pad_token_id, mode=self.pad_mode), dtype=torch.long, device=self.device)
File "F:\project\code\python\bert4torch\bert4torch\snippets\data_process.py", line 196, in sequence_padding
length = np.max([np.shape(x)[:seq_dims] for x in inputs], axis=0)
File "F:\project\code\python\bert4torch\bert4torch\snippets\data_process.py", line 196, in
length = np.max([np.shape(x)[:seq_dims] for x in inputs], axis=0)
File "D:\Develop\anaconda\envs\NovelSpeech\lib\site-packages\numpy\core\fromnumeric.py", line 2024, in shape
result = asarray(a).shape
File "D:\Develop\anaconda\envs\NovelSpeech\lib\site-packages\torch_tensor.py", line 1030, in array
return self.numpy()
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

自我尝试

此处请贴上你的自我尝试过程
def generate(self, passage, topk=1, topp=0.95):
token_ids, segment_ids = tokenizer.encode(passage, maxlen=max_p_len)
a_ids = self.random_sample([token_ids, segment_ids], n=1, topp=topp)[0] # 基于随机采样
# return tokenizer.decode(a_ids.cpu().numpy())
a_ids = a_ids.cpu().numpy() #这里需要转化成numpy
token_ids += list(a_ids)
segment_ids += [1] * len(a_ids)
q_ids = self.beam_search([token_ids, segment_ids], topk=topk)[0] # 基于beam search
# return tokenizer.decode(q_ids.cpu().numpy())
return (tokenizer.decode(q_ids.cpu().numpy()), tokenizer.decode(a_ids))#

[ocyisheng]

你使用的bert4torch版本:git仓库直接拉取

[Tongjilibo]

这样操作应该是可以的，已经修改了