【正点原子i.MX93开发板试用连载体验】为什么模型不能运行在NPU上

 本文最早发表于电子发烧友论坛：【新提醒】【正点原子i.MX93开发板试用连载体验】基于深度学习的语音本地控制 - 正点原子学习小组 - 电子技术论坛 - 广受欢迎的专业电子论坛! (elecfans.com)

昨天提到要使模型运行的NPU上，必须先将其量化。如果对没有量化的模型使用vela工具进行转换，工具会给出警告，所生成的模型仍然是只能运行在CPU上，而无法运行在NPU上的。 下面就是用vela工具对simple_audio_model_numpy.tflite文件进行转换的结果。   root@atk-imx93:~/shell/simple# vela simple_audio_model_numpy.tflite Warning: Unsupported TensorFlow Lite semantics for RESIZE_BILINEAR 'sequential_2/resizing_2/resize/ResizeBilinear;StatefulPartitionedCall/sequential_2/resizing_2/resize/ResizeBilinear'. Placing on CPU instead - Input(s), Output and Weight tensors must have quantization parameters    Op has tensors with missing quantization parameters: input_3, sequential_2/resizing_2/resize/ResizeBilinear;StatefulPartitionedCall/sequential_2/resizing_2/resize/ResizeBilinear Warning: Unsupported TensorFlow Lite semantics for CONV_2D 'sequential_2/conv2d_4/Relu;StatefulPartitionedCall/sequential_2/conv2d_4/Relu;sequential_2/conv2d_4/BiasAdd;StatefulPartitionedCall/sequential_2/conv2d_4/BiasAdd;sequential_2/conv2d_4/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_4/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_4/BiasAdd/ReadVariableOp2'. Placing on CPU instead - Input(s), Output and Weight tensors must have quantization parameters    Op has tensors with missing quantization parameters: sequential_2/resizing_2/resize/ResizeBilinear;StatefulPartitionedCall/sequential_2/resizing_2/resize/ResizeBilinear, sequential_2/conv2d_4/Relu;StatefulPartitionedCall/sequential_2/conv2d_4/Relu;sequential_2/conv2d_4/BiasAdd;StatefulPartitionedCall/sequential_2/conv2d_4/BiasAdd;sequential_2/conv2d_4/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_4/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_4/BiasAdd/ReadVariableOp1_reshape, sequential_2/conv2d_4/Relu;StatefulPartitionedCall/sequential_2/conv2d_4/Relu;sequential_2/conv2d_4/BiasAdd;StatefulPartitionedCall/sequential_2/conv2d_4/BiasAdd;sequential_2/conv2d_4/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_4/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_4/BiasAdd/ReadVariableOp2 Warning: Unsupported TensorFlow Lite semantics for CONV_2D 'sequential_2/conv2d_5/Relu;StatefulPartitionedCall/sequential_2/conv2d_5/Relu;sequential_2/conv2d_5/BiasAdd;StatefulPartitionedCall/sequential_2/conv2d_5/BiasAdd;sequential_2/conv2d_5/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_5/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_5/BiasAdd/ReadVariableOp'. Placing on CPU instead - Input(s), Output and Weight tensors must have quantization parameters    Op has tensors with missing quantization parameters: sequential_2/conv2d_4/Relu;StatefulPartitionedCall/sequential_2/conv2d_4/Relu;sequential_2/conv2d_4/BiasAdd;StatefulPartitionedCall/sequential_2/conv2d_4/BiasAdd;sequential_2/conv2d_4/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_4/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_4/BiasAdd/ReadVariableOp2, sequential_2/conv2d_5/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_5/Conv2D_reshape, sequential_2/conv2d_5/Relu;StatefulPartitionedCall/sequential_2/conv2d_5/Relu;sequential_2/conv2d_5/BiasAdd;StatefulPartitionedCall/sequential_2/conv2d_5/BiasAdd;sequential_2/conv2d_5/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_5/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_5/BiasAdd/ReadVariableOp Warning: Unsupported TensorFlow Lite semantics for MAX_POOL_2D 'sequential_2/max_pooling2d_2/MaxPool;StatefulPartitionedCall/sequential_2/max_pooling2d_2/MaxPool'. Placing on CPU instead - Input(s), Output and Weight tensors must have quantization parameters    Op has tensors with missing quantization parameters: sequential_2/conv2d_5/Relu;StatefulPartitionedCall/sequential_2/conv2d_5/Relu;sequential_2/conv2d_5/BiasAdd;StatefulPartitionedCall/sequential_2/conv2d_5/BiasAdd;sequential_2/conv2d_5/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_5/Conv2D;StatefulPartitionedCall/sequential_2/conv2d_5/BiasAdd/ReadVariableOp, sequential_2/max_pooling2d_2/MaxPool;StatefulPartitionedCall/sequential_2/max_pooling2d_2/MaxPool Warning: Unsupported TensorFlow Lite semantics for RESHAPE 'sequential_2/flatten_2/Reshape;StatefulPartitionedCall/sequential_2/flatten_2/Reshape'. Placing on CPU instead - Input(s), Output and Weight tensors must have quantization parameters    Op has tensors with missing quantization parameters: sequential_2/max_pooling2d_2/MaxPool;StatefulPartitionedCall/sequential_2/max_pooling2d_2/MaxPool, sequential_2/flatten_2/Reshape;StatefulPartitionedCall/sequential_2/flatten_2/Reshape Warning: Unsupported TensorFlow Lite semantics for FULLY_CONNECTED 'sequential_2/dense_4/Relu;StatefulPartitionedCall/sequential_2/dense_4/Relu;sequential_2/dense_4/BiasAdd;StatefulPartitionedCall/sequential_2/dense_4/BiasAdd'. Placing on CPU instead - Input(s), Output and Weight tensors must have quantization parameters    Op has tensors with missing quantization parameters: sequential_2/flatten_2/Reshape;StatefulPartitionedCall/sequential_2/flatten_2/Reshape, sequential_2/dense_4/MatMul;StatefulPartitionedCall/sequential_2/dense_4/MatMul_reshape, sequential_2/dense_4/Relu;StatefulPartitionedCall/sequential_2/dense_4/Relu;sequential_2/dense_4/BiasAdd;StatefulPartitionedCall/sequential_2/dense_4/BiasAdd Warning: Unsupported TensorFlow Lite semantics for FULLY_CONNECTED 'Identity'. Placing on CPU instead - Input(s), Output and Weight tensors must have quantization parameters    Op has tensors with missing quantization parameters: sequential_2/dense_4/Relu;StatefulPartitionedCall/sequential_2/dense_4/Relu;sequential_2/dense_4/BiasAdd;StatefulPartitionedCall/sequential_2/dense_4/BiasAdd, sequential_2/dense_5/MatMul;StatefulPartitionedCall/sequential_2/dense_5/MatMul_reshape, Identity Network summary for simple_audio_model_numpy Accelerator configuration               Ethos_U65_256 System configuration                 internal-default Memory mode                          internal-default Accelerator clock                                1000 MHz CPU operators = 7 (100.0%) NPU operators = 0 (0.0%) Neural network macs                                 0 MACs/batch Network Tops/s                                    nan Tops/s NPU cycles                                          0 cycles/batch SRAM Access cycles                                  0 cycles/batch DRAM Access cycles                                  0 cycles/batch On-chip Flash Access cycles                         0 cycles/batch Off-chip Flash Access cycles                        0 cycles/batch Total cycles                                        0 cycles/batch Batch Inference time                 0.00 ms,     nan inferences/s (batch size 1) Warning: Could not write the following attributes to RESHAPE 'sequential_2/flatten_2/Reshape;StatefulPartitionedCall/sequential_2/flatten_2/Reshape' ReshapeOptions field: new_shape 这个错误信息明确指出Vela不支持 TensorFlow Lite 对特定操作的支持问题。具体来说，这个警告说明了：量化参数缺失 ，错误信息指出，涉及的输入、输出和权重张量必须具有量化参数，但在这个操作中，某些张量（如 input_3 和 sequential_2/resizing_2/resize/ResizeBilinear）缺失了这些量化参数。由于不支持，相关的操作将被放置在 CPU 上执行，而不是利用可能存在的更高效的硬件加速（NPU）。 我们使用netron.app可以查看一下模型文件。

从中可以看到input_3是float32类型的。

而查看被vela支持的模型，可以看到其输入参数已经被量化，是int8类型的。

如果我们想利用i.MX 93的NPU能力就需要先对模型文件进行量化。当然如果觉得i.MX 93的CPU推理能力已经够用了，此步骤也可以省略。