Wrong rounding-type for Pooling layers after converting a MXNet model to IR #24
Description
In the original MXNet model, the pooling layer is defined with ceil_mode=False (resulting in pooling_convention=valid in the json file). But in the xml file of the converted IR, the rounding-type is "ceil". This causes the network producing incorrect results after CNNNetwork::reshape is called (which probably re-calculates the shape of the blobs) even if the input width and height are not changed.
op definition in MXNet json file:
{
"op": "Pooling",
"name": "pool_fwd",
"attrs": {
"global_pool": "False",
"kernel": "(3, 3)",
"pad": "(1, 1)",
"pool_type": "max",
"pooling_convention": "valid",
"stride": "(2, 2)"
},
"inputs": [[4, 0, 0]]
}
Layer in xml file of the IR:
<layer id="3" name="pool_fwd" precision="FP32" type="Pooling">
<data exclude-pad="false" kernel-x="3" kernel-y="3" pad-b="1" pad-r="1" pad-x="1" pad-y="1" pool-method="max" rounding-type="ceil" stride="1,1,2,2" stride-x="2" stride-y="2"/>
<input>
<port id="0">
<dim>1</dim>
<dim>32</dim>
<dim>112</dim>
<dim>112</dim>
</port>
</input>
<output>
<port id="1">
<dim>1</dim>
<dim>32</dim>
<dim>56</dim>
<dim>56</dim>
</port>
</output>
</layer>
This problem seems to be caused by the following code in mo.front.mxnet.extractors.pooling.py:
pooling_conv = attrs.str("pooling_convention", 'valid')
if pooling_conv:
data["pooling_convention"] = pooling_conv
data["rounding_type"] = 'ceil'
I have modified the code as follows and then the network is correct:
pooling_conv = attrs.str("pooling_convention", 'valid')
if pooling_conv:
data["pooling_convention"] = pooling_conv
if pooling_conv == 'valid':
data["rounding_type"] = 'floor'
else:
data["rounding_type"] = 'ceil'