【AI计算机视觉应用 ruyistudio量化】使用ruyistudio量化自己训练的mobilenetv2权重生成WK失败

floor(2048/(ALIGN16(w)/16 * ALIGN16(c)/16)) + pad up + pad down >= kernel_h
而Global Pooling的kernel_w和kernel_h为输入的input height 和 input width,所以容易超出规格，报上面的错误。
可以按照input channel 切分，规避此问题。
举例：
左边是原prototxt，右边是切分后的prototxt
| 原prototxt | 切分后prototxt |
|-----------|-------------|
|    ![输入图片说明](https://foruda.gitee.com/images/1678673230561457925/67eec4c5_7774751.png "屏幕截图")       |       ![输入图片说明](https://foruda.gitee.com/images/1678673782577398158/a25c6ba4_7774751.png "屏幕截图")      |

原prototxt
```
layer {
	name: "data"
	type: "Input"
	top: "data"
	input_param {
		shape {
			dim: 1
			dim: 3
			dim: 56
			dim: 56
		}
	}
}

layer {
	name: "conv1"
	type: "Convolution"
	bottom: "data"
	top: "conv1"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	convolution_param {
		num_output: 256
		bias_term: false
		pad: 0
		kernel_size: 1
		stride: 1
	}
}

layer {
	name: "pooling"
	type: "Pooling"
	bottom: "conv1"
	top: "pooling"
	pooling_param {
		pool: AVE
		global_pool: true
	}
}

layer {
	name: "scale1"
	type: "Scale"
	bottom: "pooling"
	top: "scale"
}

```

切分后的prototxt
```

layer {
	name: "data"
	type: "Input"
	top: "data"
	input_param {
		shape {
			dim: 1
			dim: 3
			dim: 56
			dim: 56
		}
	}
}

layer {
	name: "conv1"
	type: "Convolution"
	bottom: "data"
	top: "conv1"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	convolution_param {
		num_output: 256
		bias_term: false
		pad: 1
		kernel_size: 1
		stride: 1
	}
}

layer {
	name: "slice1"
	type: "Slice"
	bottom: "conv1"
	top: "slice1_1"
	top: "slice1_2"
	slice_param {
		axis: 1
		scale_point: 128
	}
}

layer {
	name: "pooling1"
	type: "Pooling"
	bottom: "slice1_1"
	top: "pooling1"
	pooling_param {
		pool: AVE
		global_pool: true
	}
}

layer {
	name: "pooling2"
	type: "Pooling"
	bottom: "slice1_2"
	top: "pooling2"
	pooling_param {
		pool: AVE
		global_pool: true
	}
}

layer {
	name: "concat1"
	type: "Concat"
	bottom: "pooling1"
	bottom: "pooling2"
	top: "concat1"
	concat_param {
		axis: 1
	}
}

layer {
	name: "scale1"
	type: "Scale"
	bottom: "concat1"
	top: "scale"
}

```

HiSpark/HiSpark_NICU2023

内容风险标识

评论 (2)