量化感知训练场景下，出现数据类型不一致的解决方法
问题描述
混合精度训练场景下，用户做量化感知训练，调用create_quant_retrain_model接口之后，脚本提示如下信息：
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 |
case_info: ['AMCT7_MINDSPORE_CREATE_QUANT_CONFIG_037', 'cascade_conv', '1,10,32,32', 10, 10, '3,3', '1,1', 'same', '0,0,0,0', '1,1', 1, 1, 0, '-1,1,-1,1,-1,1', 'uniform,uniform,uniform'] in_channels: 10 out_channels: 10 kernel_size: [3, 3] stride: (1, 1) pad_mode: same padding: (0, 0, 0, 0) dilation: (1, 1) group: 1 has_bias: True uniform uniform 0 [-1.0, 1.0, -1.0, 1.0] [1, 10, 32, 32] cascade_conv (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) case_info: ['AMCT7_MINDSPORE_CREATE_QUANT_CONFIG_037', 'cascade_conv', '1,10,32,32', 10, 10, '3,3', '1,1', 'same', '0,0,0,0', '1,1', 1, 1, 0, '-1,1,-1,1,-1,1', 'uniform,uniform,uniform'] in_channels: 10 out_channels: 10 kernel_size: [3, 3] stride: (1, 1) pad_mode: same padding: (0, 0, 0, 0) dilation: (1, 1) group: 1 has_bias: True uniform uniform 0 [-1.0, 1.0, -1.0, 1.0] [1, 10, 32, 32] cascade_conv (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) (-1.0, 1.0) For 'Conv2D', type of `w` should be same as `x`, but `x` with type Tensor[Float16] and `w` with type Ref[Tensor(F32)]. |
可能原因
调用create_quant_retrain_model接口之后，用户原始网络中的nn.Dense会被替换为DenseQatBlock对象，nn.Conv2D被替换为Conv2dQatBlock对象。在调用量化感知训练脚本时，需要先导入上述两个对象。
解决方法
修改用户量化感知训练的脚本，先导入上述两个对象。
如果用户原始脚本为：
1 2 3 4 5 6 |
def switch_precision(net, data_type, config): if config.platform == "Ascend": net.to_float(data_type) for _, cell in net.cells_and_names(): if isinstance(cell, (nn.Dense, )): cell.to_float(mstype.float32) |
需要修改为：
1 2 3 4 5 6 7 8 |
from amct_mindspore.cells.dense_qat_cell import DenseQatBlock from amct_mindspore.cells.conv2d_qat_cell import Conv2dQatBlock def switch_precision(net, data_type, config): if config.platform == "Ascend": net.to_float(data_type) for _, cell in net.cells_and_names(): if isinstance(cell, (nn.Dense, DenseQatBlock)): cell.to_float(mstype.float32) |
父主题： FAQ
