Hi, I have been trying to perform QAT on FPENet, however when I try to prepare the model I encounter a KeyError deep down as part of prepare_qat_pt2e the specific error is:
KeyError: (getitem_389, adaptive_avg_pool2d)
I have successfully converted other models with adaptive average pooling layers so I am unsure why this one fails.
I am using PyTorch 2.3.0, Torchvision 0.18.0
The FPENet code I am using is from this repo.
Here is the exact code I am trying to run for prototyping.
Thanks in advance!
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch._export import capture_pre_autograd_graph
from torch._export import dynamic_dim
from torch.ao.quantization.quantizer.xnnpack_quantizer import (
XNNPACKQuantizer,
get_symmetric_quantization_config,
)
from torch.ao.quantization.quantize_pt2e import (
prepare_qat_pt2e,
convert_pt2e,
)
def conv3x3(in_planes, out_planes, stride=1, padding=1, dilation=1, groups=1, bias=False):
"""3x3 convolution with padding"""
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=padding, dilation=dilation, groups=groups,bias=bias)
def conv1x1(in_planes, out_planes, stride=1, bias=False):
"""1x1 convolution"""
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=bias)
class SEModule(nn.Module):
def __init__(self, channels, reduction=16):
super(SEModule, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.fc1 = nn.Conv2d(channels, channels // reduction, kernel_size=1, padding=0)
self.relu = nn.ReLU(inplace=True)
self.fc2 = nn.Conv2d(channels // reduction, channels, kernel_size=1, padding=0)
self.sigmoid = nn.Sigmoid()
def forward(self, input):
x = self.avg_pool(input)
x = self.fc1(x)
x = self.relu(x)
x = self.fc2(x)
x = self.sigmoid(x)
return input * x
class FPEBlock(nn.Module):
def __init__(self, inplanes, outplanes, dilat, downsample=None, stride=1, t=1, scales=4, se=False, norm_layer=None):
super(FPEBlock, self).__init__()
if inplanes % scales != 0:
raise ValueError('Planes must be divisible by scales')
if norm_layer is None:
norm_layer = nn.BatchNorm2d
bottleneck_planes = inplanes * t
self.conv1 = conv1x1(inplanes, bottleneck_planes, stride)
self.bn1 = norm_layer(bottleneck_planes)
self.conv2 = nn.ModuleList([conv3x3(bottleneck_planes // scales, bottleneck_planes // scales,
groups=(bottleneck_planes // scales),dilation=dilat[i],
padding=1*dilat[i]) for i in range(scales)])
self.bn2 = nn.ModuleList([norm_layer(bottleneck_planes // scales) for _ in range(scales)])
self.conv3 = conv1x1(bottleneck_planes, outplanes)
self.bn3 = norm_layer(outplanes)
self.relu = nn.ReLU(inplace=True)
self.se = SEModule(outplanes) if se else None
self.downsample = downsample
self.stride = stride
self.scales = scales
def forward(self, x):
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
xs = torch.chunk(out, self.scales, 1)
ys = []
for s in range(self.scales):
if s == 0:
ys.append(self.relu(self.bn2[s](self.conv2[s](xs[s]))))
else:
ys.append(self.relu(self.bn2[s](self.conv2[s](xs[s] + ys[-1]))))
out = torch.cat(ys, 1)
out = self.conv3(out)
out = self.bn3(out)
if self.se is not None:
out = self.se(out)
if self.downsample is not None:
identity = self.downsample(identity)
out += identity
out = self.relu(out)
return out
class MEUModule(nn.Module):
def __init__(self, channels_high, channels_low, channel_out):
super(MEUModule, self).__init__()
self.conv1x1_low = nn.Conv2d(channels_low, channel_out, kernel_size=1, bias=False)
self.bn_low = nn.BatchNorm2d(channel_out)
self.sa_conv = nn.Conv2d(1, 1, kernel_size=1, bias=False)
self.conv1x1_high = nn.Conv2d(channels_high, channel_out, kernel_size=1, bias=False)
self.bn_high = nn.BatchNorm2d(channel_out)
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.ca_conv = nn.Conv2d(channel_out, channel_out, kernel_size=1, bias=False)
self.sa_sigmoid = nn.Sigmoid()
self.ca_sigmoid = nn.Sigmoid()
self.relu = nn.ReLU(inplace=True)
def forward(self, fms_high, fms_low):
"""
:param fms_high: High level Feature map. Tensor.
:param fms_low: Low level Feature map. Tensor.
"""
_, _, h, w = fms_low.shape
#
fms_low = self.conv1x1_low(fms_low)
fms_low= self.bn_low(fms_low)
sa_avg_out = self.sa_sigmoid(self.sa_conv(torch.mean(fms_low, dim=1, keepdim=True)))
#
fms_high = self.conv1x1_high(fms_high)
fms_high = self.bn_high(fms_high)
ca_avg_out = self.ca_sigmoid(self.relu(self.ca_conv(self.avg_pool(fms_high))))
#
fms_high_up = F.interpolate(fms_high, size=(h,w), mode='bilinear', align_corners=False)
fms_sa_att = sa_avg_out * fms_high_up
#
fms_ca_att = ca_avg_out * fms_low
out = fms_ca_att + fms_sa_att
return out
class FPENet(nn.Module):
def __init__(self, classes=19, zero_init_residual=False,
width=16, scales=4, se=False, norm_layer=None):
super(FPENet, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
outplanes = [int(width * 2 ** i) for i in range(3)] # planes=[16,32,64]
self.block_num = [1,3,9]
self.dilation = [1,2,4,8]
self.inplanes = outplanes[0]
self.conv1 = nn.Conv2d(3, outplanes[0], kernel_size=3, stride=2, padding=1,bias=False)
self.bn1 = norm_layer(outplanes[0])
self.relu = nn.ReLU(inplace=True)
self.layer1 = self._make_layer(FPEBlock, outplanes[0], self.block_num[0], dilation=self.dilation,
stride=1, t=1, scales=scales, se=se, norm_layer=norm_layer)
self.layer2 = self._make_layer(FPEBlock, outplanes[1], self.block_num[1], dilation=self.dilation,
stride=2, t=4, scales=scales, se=se, norm_layer=norm_layer)
self.layer3 = self._make_layer(FPEBlock, outplanes[2], self.block_num[2], dilation=self.dilation,
stride=2, t=4, scales=scales, se=se, norm_layer=norm_layer)
self.meu1 = MEUModule(64,32,64)
self.meu2 = MEUModule(64,16,32)
# Projection layer
self.project_layer = nn.Conv2d(32, classes, kernel_size = 1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
if zero_init_residual:
for m in self.modules():
if isinstance(m, FPEBlock):
nn.init.constant_(m.bn3.weight, 0)
def _make_layer(self, block, planes, blocks, dilation, stride=1, t=1, scales=4, se=False, norm_layer=None):
if norm_layer is None:
norm_layer = nn.BatchNorm2d
downsample = None
if stride != 1 or self.inplanes != planes:
downsample = nn.Sequential(
conv1x1(self.inplanes, planes, stride),
norm_layer(planes),
)
layers = []
layers.append(block(self.inplanes, planes, dilat=dilation, downsample=downsample, stride=stride, t=t, scales=scales, se=se,
norm_layer=norm_layer))
self.inplanes = planes
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes, dilat=dilation, scales=scales, se=se, norm_layer=norm_layer))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x_1 = self.layer1(x)
x_2_0 = self.layer2[0](x_1)
x_2_1 = self.layer2[1](x_2_0)
x_2_2 = self.layer2[2](x_2_1)
x_2 = x_2_0 + x_2_2
x_3_0 = self.layer3[0](x_2)
x_3_1 = self.layer3[1](x_3_0)
x_3_2 = self.layer3[2](x_3_1)
x_3_3 = self.layer3[3](x_3_2)
x_3_4 = self.layer3[4](x_3_3)
x_3_5 = self.layer3[5](x_3_4)
x_3_6 = self.layer3[6](x_3_5)
x_3_7 = self.layer3[7](x_3_6)
x_3_8 = self.layer3[8](x_3_7)
x_3 = x_3_0 + x_3_8
x2 = self.meu1(x_3, x_2)
x1 = self.meu2(x2, x_1)
output = self.project_layer(x1)
# Bilinear interpolation x2
output = F.interpolate(output,scale_factor=2, mode = 'bilinear', align_corners=False)
return output
if __name__ == '__main__':
_model = FPENet(classes=3).to("cpu")
example_inputs = (torch.rand(1, 3, 224, 224),)
exported_model = capture_pre_autograd_graph(_model, example_inputs)
quantizer = XNNPACKQuantizer()
quantizer.set_global(get_symmetric_quantization_config(is_qat=True))
model = prepare_qat_pt2e(exported_model, quantizer)