开发环境：
- MegEngine 1.3.1
- Python 3.7
- Jupyter Notebook
问题详述：
使用 module_stats（）函数统计量化后模型的参数量和计算量，代码如下：
import megengine as mge
import ipynb_importer
from model import MCNN
from megengine.utils import module_stats
mcnn = MCNN()
# print(mcnn)
checkpoint = mge.load('mcnn_net_quantized.pkl')
mcnn.load_state_dict(checkpoint, strict=False)
# 指定输入shape
input_shape = (1, 3, 1200, 1600)
# float model.
total_params, total_flops = module_stats.module_stats(
mcnn, input_shape, log_params=True, log_flops=True
)
print("params: {} flops: {}".format(total_params, total_flops))
报错如下：
RuntimeError Traceback (most recent call last)
<ipython-input-7-46a892a93210> in <module>
14 # float model.
15 total_params, total_flops = module_stats.module_stats(
---> 16 mcnn, input_shape, log_params=True, log_flops=True
17 )
18 print("params: {} flops: {}".format(total_params, total_flops))
~/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/utils/module_stats.py in module_stats(model, input_size, bar_length_max, log_params, log_flops)
373 inputs = [zeros(in_size, dtype=np.float32) for in_size in input_size]
374 with adjust_stats(model, training=False) as model:
--> 375 model(*inputs)
376
377 for h in hooks:
~/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/module/module.py in __call__(self, *inputs, **kwargs)
146 inputs = modified_inputs
147
--> 148 outputs = self.forward(*inputs, **kwargs)
149
150 for hook in self._forward_hooks.values():
~/workspace/fishIfeed/cejc/model.ipynb in forward(self, x)
~/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/module/module.py in __call__(self, *inputs, **kwargs)
146 inputs = modified_inputs
147
--> 148 outputs = self.forward(*inputs, **kwargs)
149
150 for hook in self._forward_hooks.values():
~/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/module/sequential.py in forward(self, inp)
95 # avoid layer_values as a name prefix, see Module.__getattribute__
96 for layer in [getattr(self, key) for key in self.layer_keys]:
---> 97 inp = layer(inp)
98 return inp
~/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/module/module.py in __call__(self, *inputs, **kwargs)
146 inputs = modified_inputs
147
--> 148 outputs = self.forward(*inputs, **kwargs)
149
150 for hook in self._forward_hooks.values():
~/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/module/conv.py in forward(self, inp)
399
400 def forward(self, inp):
--> 401 return self.calc_conv(inp, self.weight, self.bias)
402
403
~/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/module/conv.py in calc_conv(self, inp, weight, bias)
395 self.groups,
396 self.conv_mode,
--> 397 self.compute_mode,
398 )
399
~/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/functional/nn.py in conv2d(inp, weight, bias, stride, padding, dilation, groups, conv_mode, compute_mode)
167 )
168 inp, weight = utils.convert_inputs(inp, weight)
--> 169 (output,) = apply(op, inp, weight)
170 if bias is not None:
171 output += bias
RuntimeError: assertion `src.dtype.enumv() == filter.dtype.enumv()' failed at /home/code/dnn/src/common/convolution.cpp:597: megdnn::ConvolutionBase<Parameter>::CanonizedFilterMeta megdnn::ConvolutionBase<Parameter>::deduce_layout_fwd(const megdnn::TensorLayout&, const megdnn::TensorLayout&, megdnn::TensorLayout&) const [with Parameter = megdnn::param::Convolution]
extra message: src={1(5760000),3(1920000),1200(1600),1600(1) Float32}, filter={16(243),3(81),9(9),9(1) QuantizedS8}, dst={ Float32}, is_nchw=1, is_xcorr=1, pad_h=4, pad_w=4, stride_h=1, stride_w=1, dilate_h=1, dilate_w=1
backtrace:
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(_ZN3mgb13MegBrainErrorC1ERKSs+0x4a) [0x7f48bc4b612a]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(+0x28128c1) [0x7f48bc4f48c1]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(_ZN6megdnn12ErrorHandler15on_megdnn_errorERKSs+0x14) [0x7f48bc890294]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(_ZN6megdnn12ErrorHandler15on_megdnn_errorEPKc+0x22) [0x7f48bc891862]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(_ZN6megdnn15__assert_fail__EPKciS1_S1_S1_z+0x190) [0x7f48bc908280]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(_ZNK6megdnn15ConvolutionBaseINS_5param11ConvolutionEE17deduce_layout_fwdERKNS_12TensorLayoutES6_RS4_+0x583) [0x7f48bc8a4be3]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(_ZN6megdnn18ConvolutionForward13deduce_layoutERKNS_12TensorLayoutES3_RS1_+0x1b) [0x7f48bc89932b]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(_ZNK3mgb3opr18ConvolutionForward20get_output_var_shapeERKN6megdnn11SmallVectorINS2_11TensorShapeELj4EEERS5_+0xda) [0x7f48bc62341a]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/lib/libmegengine_export.so(_ZN3mgb2cg5mixin24OutshapePureByInshapeOpr10infer_descEmRN6megdnn11TensorShapeERKNS0_12static_infer6InpValE+0x19d) [0x7f48bc525ebd]
/home/megstudio/.miniconda/envs/xuan/lib/python3.7/site-packages/megengine/core/_imperative_rt.cpython-37m-x86_64-linux-gnu.so(+0x215313) [0x7f4924f05313]
针对上述使用module_stats（）函数无法统计量化之后int8模型的计算量问题，请问应该怎样统计量化之后模型的flops呢?