RuntimeError: The pointer[tensor] is null.

[jiuniZZzz]

Describe the bug/ 问题描述 (Mandatory / 必填)
GPU环境 TrOCR预训练模型微调 求梯度时报错 RuntimeError: The pointer[tensor] is null.

• Hardware Environment(Ascend/GPU/CPU) / 硬件环境:

Modelarts CPU 8核32G
GPU Tnt004 16G

• Software Environment / 软件环境 (Mandatory / 必填):
  -- MindSpore version (e.g., 1.7.0.Bxxx) : MindSpore 2.2.14
  -- Python version (e.g., Python 3.7.5) :Python 3.9.19
  -- OS platform and distribution (e.g., Linux Ubuntu 16.04):ubuntu18.04
  -- GCC/Compiler version (if compiled from source):7.5

• Excute Mode / 执行模式 (Mandatory / 必填)(PyNative/Graph):

/mode pynative

To Reproduce / 重现步骤 (Mandatory / 必填)
model<mindnlp.transformers.models.vision_encoder_decoder.modeling_vision_encoder_decoder.VisionEncoderDecoderModel object>
dataset<mindspore.dataset.engine.datasets.BatchDataset object >
运行代码
'''
for epoch in range(1):
# train
model.train()
train_loss = 0.0
for bacth in tqdm(eval_dataloader.create_dict_iterator()):
pixel_values = bacth['pixel_values']
labels = bacth['labels']
def compute_loss(pixel_values, labels):
outputs = model(pixel_values=pixel_values, labels=labels)
loss = outputs.loss
return loss
grad_fn = mindspore.value_and_grad(fn=compute_loss, weights=model.parameters())
loss, grads = grad_fn(pixel_values, labels)
optimizer.step(loss)
'''
Expected behavior / 预期结果 (Mandatory / 必填)
正常运行

Screenshots/ 日志 / 截图 (Mandatory / 必填)
If applicable, add screenshots to help explain your problem.

Additional context / 备注 (Optional / 选填)
Add any other context about the problem here.

[lvyufeng]

要用mindspore2.3

[lvyufeng]

复现代码提供完整的

[jiuniZZzz]

用2.3版本的mindspore CPU 训练会报错如下

GPU太贵了

df = pd.read_fwf('./IAM/gt_test.txt', header=None)
df.rename(columns={0: "file_name", 1: "text"}, inplace=True)
del df[2]
# some file names end with jp instead of jpg, let's fix this
df['file_name'] = df['file_name'].apply(lambda x: x + 'g' if x.endswith('jp') else x)
# print(df.head())

train_df, test_df = train_test_split(df, test_size=0.2)
# we reset the indices to start from zero
train_df.reset_index(drop=True, inplace=True)
test_df.reset_index(drop=True, inplace=True)


class IAMDataset():
    def __init__(self, root_dir, df, processor, max_target_length=128):
        super(IAMDataset).__init__()
        self.root_dir = root_dir
        self.df = df
        self.processor = processor
        self.max_target_length = max_target_length

    def __len__(self):
        return len(self.df)

    def __getitem__(self, idx):
        # get file name + text
        file_name = self.df['file_name'][idx]
        text = self.df['text'][idx]
        # prepare image (i.e. resize + normalize)
        image = Image.open(self.root_dir + '/' + file_name).convert("RGB")
        pixel_values = self.processor(image, return_tensors="ms").pixel_values
        # add labels (input_ids) by encoding the text
        labels = self.processor.tokenizer(text,
                                          padding="max_length",
                                          max_length=self.max_target_length).input_ids
        # important: make sure that PAD tokens are ignored by the loss function
        labels = [label if label != self.processor.tokenizer.pad_token_id else -100 for label in labels]

        # encoding = {: , : }
        return pixel_values.squeeze(), Tensor(labels)


model_path1 = "C:\\Users\\virgo\\Documents\\GitHub\\trocr-base-printed"
model_path2 = "C:\\Users\\virgo\\Documents\\GitHub\\trocr-base-stage1"
processor = TrOCRProcessor.from_pretrained(model_path1)
train_dataset = IAMDataset(root_dir='./IAM/image',
                           df=train_df,
                           processor=processor)
eval_dataset = IAMDataset(root_dir='./IAM/image',
                          df=test_df,
                          processor=processor)
train_dataloader = GeneratorDataset(train_dataset, column_names=["pixel_values", "labels"], shuffle=True).batch(4)
eval_dataloader = GeneratorDataset(eval_dataset, column_names=["pixel_values", "labels"], shuffle=True)
model = VisionEncoderDecoderModel.from_pretrained(model_path2)
# set special tokens used for creating the decoder_input_ids from the labels
model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
model.config.pad_token_id = processor.tokenizer.pad_token_id
# make sure vocab size is set correctly
model.config.vocab_size = model.config.decoder.vocab_size

# set beam search parameters
model.config.eos_token_id = processor.tokenizer.sep_token_id
model.config.max_length = 64
model.config.early_stopping = True
model.config.no_repeat_ngram_size = 3
model.config.length_penalty = 2.0
model.config.num_beams = 4

cer_metric = evaluate.load("cer")


def compute_cer(pred_ids, label_ids):
    pred_str = processor.batch_decode(pred_ids, skip_special_tokens=True)
    label_ids[label_ids == -100] = processor.tokenizer.pad_token_id
    label_str = processor.batch_decode(label_ids, skip_special_tokens=True)

    cer = cer_metric.compute(predictions=pred_str, references=label_str)

    return cer


optimizer = AdamW(model.parameters(), lr=5e-5)

for epoch in range(1):
    # train
    model.train()
    train_loss = 0.0
    for bacth in tqdm(eval_dataloader.create_dict_iterator()):
        pixel_values = bacth['pixel_values']
        labels = bacth['labels']
        def compute_loss(pixel_values, labels):
            outputs = model(pixel_values=pixel_values, labels=labels)
            loss = outputs.loss
            return loss
        grad_fn = mindspore.value_and_grad(fn=compute_loss, weights=model.parameters())
        loss, grads = grad_fn(pixel_values, labels)
        optimizer.step(loss)

[lvyufeng]

数据类型有问题