BEV模型编译报错

1. 芯片型号：J5

3. 问题定位：模型转换、板端部署

4. 问题具体描述：

你好。我在对bev_ipm_efficientnetb0_multitask_nuscenes这个例子修改并做模型编译(PC)时遇到了问题

在scripts下运行命令：

python3 tools/compile_perf.py --config configs/bev/bev_ipm_efficientnetb0_multitask_nuscenes.py --out-dir ./tmp_models/compile --opt 3

在deploy_inputs中直接传入torch.randn 类型的img和points

comile的配置如下

compile_dir = os.path.join(ckpt_dir, "compile")

compile_cfg = dict(

  march=march,

  name=task_name + "_model",

  out_dir=compile_dir,

  hbm=os.path.join(compile_dir, "model.hbm"),

  layer_details=True,

  input_source="pyramid",

  opt="O3",

)

直接将 "points"传入view_transformer.py中的WrappingTransformer->_spatial_transfom函数

转换时报错

WARNING: model has multiple input tensors that are ('arg0[img]', 'arg0[points]')

INFO: launch 4 threads for optimization

WARNING: the model has 2 inputs, but given 1 sources types. all inputs will be set as the same input type

Network Architecture Error:

Expected data shape range of model input tensor from pyramid "arg0[points]" is [[1, 4096],[1, 65536],[1, 65536],[1, 3]], but the data shape is [2,256,2,256]