Bartmodelinferenceresultsafterconvertingfromhugginfacetoonnx

将huggingface中的Bart模型转换为ONNX格式，以便在C++等其他语言的应用程序中进行推理。以下是转换后的结果的代码示例：

import torch
import transformers
from transformers import BartModel, BartTokenizer

model_name = "facebook/bart-large"
model = BartModel.from_pretrained(model_name)
tokenizer = BartTokenizer.from_pretrained(model_name)
dummy_input_ids = torch.zeros([1, 1024], dtype=torch.long)
dummy_input_mask = torch.ones([1, 1024], dtype=torch.long)
dummy_input = (dummy_input_ids, dummy_input_mask)

# Export to ONNX
torch.onnx.export(model, dummy_input, "bart_large.onnx")

# Load ONNX model
import onnxruntime
ort_session = onnxruntime.InferenceSession("bart_large.onnx")

# Generate predictions
inputs = tokenizer("Hello World", return_tensors="pt")
outputs = model(**inputs)
ort_inputs = {'input_ids': inputs['input_ids'].cpu().numpy(),
              'attention_mask': inputs['attention_mask'].cpu().numpy()}
ort_outputs = ort_session.run(None, ort_inputs)

# Compare results
print(outputs.last_hidden_state.shape)
print(ort_outputs[0].shape)

通过执行上述代码，可以将huggingface中的Bart模型转换为ONNX格式，然后使用ONNXruntime库进行推理。最后，将输出与原始PyTorch模型的输出进行比较，确保结果相同。