import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from .utils import PriorBox
from ..box_utils import batched_decode
class FEM(nn.Module):
def __init__(self, channel_size):
super(FEM, self).__init__()
self.cs = channel_size
self.cpm1 = nn.Conv2d(self.cs, 256, kernel_size=3, padding=1)
self.cpm2 = nn.Conv2d(self.cs, 256, kernel_size=3, dilation=2, padding=2)
self.cpm3 = nn.Conv2d(256, 128, kernel_size=3, padding=1)
self.cpm4 = nn.Conv2d(256, 128, kernel_size=3, dilation=2, padding=2)
self.cpm5 = nn.Conv2d(128, 128, kernel_size=3, padding=1)
def forward(self, x):
x1_1 = self.cpm1(x).relu()
x1_2 = self.cpm2(x).relu()
x2_1 = self.cpm3(x1_2).relu()
x2_2 = self.cpm4(x1_2).relu()
x3_1 = self.cpm5(x2_2).relu()
return torch.cat([x1_1, x2_1, x3_1], dim=1)
class SSD(nn.Module):
"""Single Shot Multibox Architecture
The network is composed of a base VGG network followed by the
added multibox conv layers. Each multibox layer branches into
1) conv2d for class conf scores
2) conv2d for localization predictions
3) associated priorbox layer to produce default bounding
boxes specific to the layer's feature map size.
See: https://arxiv.org/pdf/1512.02325.pdf for more details.
Args:
phase: (string) Can be "test" or "train"
size: input image size
base: VGG16 layers for input, size of either 300 or 500
extras: extra layers that feed to multibox loc and conf layers
head: "multibox head" consists of loc and conf conv layers
"""
def __init__(self, cfg):
super(SSD, self).__init__()
self.num_classes = 2 # Background and face
self.cfg = cfg
resnet = torchvision.models.resnet152(pretrained=False)
self.layer1 = nn.Sequential(
resnet.conv1, resnet.bn1, resnet.relu,
resnet.maxpool, resnet.layer1)
self.layer2 = nn.Sequential(resnet.layer2)
self.layer3 = nn.Sequential(resnet.layer3)
self.layer4 = nn.Sequential(resnet.layer4)
self.layer5 = nn.Sequential(
nn.Conv2d(2048, 512, kernel_size=1),
nn.BatchNorm2d(512),
nn.ReLU(inplace=True),
nn.Conv2d(512, 512, kernel_size=3, padding=1, stride=2),
nn.BatchNorm2d(512),
nn.ReLU(inplace=True)
)
self.layer6 = nn.Sequential(
nn.Conv2d(512, 128, kernel_size=1),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2),
nn.BatchNorm2d(256),
nn.ReLU(inplace=True)
)
output_channels = [256, 512, 1024, 2048, 512, 256]
# Feature Pyramid Network
fpn_in = output_channels
self.latlayer3 = nn.Conv2d(fpn_in[3], fpn_in[2], kernel_size=1)
self.latlayer2 = nn.Conv2d(fpn_in[2], fpn_in[1], kernel_size=1)
self.latlayer1 = nn.Conv2d(fpn_in[1], fpn_in[0], kernel_size=1)
self.smooth3 = nn.Conv2d(fpn_in[2], fpn_in[2], kernel_size=1)
self.smooth2 = nn.Conv2d(fpn_in[1], fpn_in[1], kernel_size=1)
self.smooth1 = nn.Conv2d(fpn_in[0], fpn_in[0], kernel_size=1)
# Feature enhance module
cpm_in = output_channels
self.cpm3_3 = FEM(cpm_in[0])
self.cpm4_3 = FEM(cpm_in[1])
self.cpm5_3 = FEM(cpm_in[2])
self.cpm7 = FEM(cpm_in[3])
self.cpm6_2 = FEM(cpm_in[4])
self.cpm7_2 = FEM(cpm_in[5])
head = pa_multibox(output_channels, self.cfg['mbox'], self.num_classes)
self.loc = nn.ModuleList(head[0])
self.conf = nn.ModuleList(head[1])
# Testing scenario
self.softmax = nn.Softmax(dim=-1)
# Cache to stop computing new priors per fowrard pass
self.prior_cache = {
}
def init_priors(self, feature_maps, image_size):
# Hacky key system, but works....
key = ".".join([str(item) for i in range(len(feature_maps)) for item in feature_maps[i]]) + \
"," + ".".join([str(_) for _ in image_size])
if key in self.prior_cache:
return self.prior_cache[key].clone()
priorbox = PriorBox(self.cfg, image_size, feature_maps)
prior = priorbox.forward()
self.prior_cache[key] = prior.clone()
return prior
def forward(self, x, confidence_threshold, nms_threshold):
"""Applies network layers and ops on input image(s) x.
Args:
x: input image or batch of images. Shape: [batch,3,300,300].
Return:
Depending on phase:
test:
Variable(tensor) of output class label predictions,
confidence score, and corresponding location predictions for
each object detected. Shape: [batch,topk,7]
train:
list of concat outputs from:
1: confidence layers, Shape: [batch*num_priors,num_classes]
2: localization layers, Shape: [batch,num_priors*4]
3: priorbox layers, Shape: [2,num_priors*4]
"""
image_size = [x.shape[2], x.shape[3]]
loc = list()
conf = list()
# ResNet152
conv3_3_x = self.layer1(x)
conv4_3_x = self.layer2(conv3_3_x)
conv5_3_x = self.layer3(conv4_3_x)
fc7_x = self.layer4(conv5_3_x)
conv6_2_x = self.layer5(fc7_x)
conv7_2_x = self.layer6(conv6_2_x)
# FPN
lfpn3 = self._upsample_product(
self.latlayer3(fc7_x), self.smooth3(conv5_3_x))
lfpn2 = self._upsample_product(
self.latlayer2(lfpn3), self.smooth2(conv4_3_x))
lfpn1 = self._upsample_product(
self.latlayer1(lfpn2), self.smooth1(conv3_3_x))
conv5_3_x = lfpn3
conv4_3_x = lfpn2
conv3_3_x = lfpn1
sources = [
self.cpm3_3(conv3_3_x),
self.cpm4_3(conv4_3_x),
self.cpm5_3(conv5_3_x),
self.cpm7(fc7_x),
self.cpm6_2(conv6_2_x),
self.cpm7_2(conv7_2_x)]
# Feature Enhance Module
# apply multibox head to source layers
featuremap_size = []
for (x, l, c) in zip(sources, self.loc, self.conf):
featuremap_size.append([x.shape[2], x.shape[3]])
loc.append(l(x).permute(0, 2, 3, 1).contiguous())
# Max in out
len_conf = len(conf)
out = self.mio_module(c(x), len_conf)
conf.append(out.permute(0, 2, 3, 1).contiguous())
# Progressive Anchor
mbox_num = self.cfg['mbox'][0]
face_loc = torch.cat([
o[:, :, :, :4*mbox_num].contiguous().view(o.size(0), -1)
for o in loc], dim=1)
face_conf = torch.cat([
o[:, :, :, :2*mbox_num].contiguous().view(o.size(0), -1)
for o in conf], dim=1)
# Test Phase
self.priors = self.init_priors(featuremap_size, image_size)
self.priors = self.priors.to(face_conf.device)
conf_preds = face_conf.view(
face_conf.size(0), -1, self.num_classes).softmax(dim=-1)
face_loc = face_loc.view(face_loc.size(0), -1, 4)
boxes = batched_decode(
face_loc, self.priors,
self.cfg["variance"]
)
scores = conf_preds.view(-1, self.priors.shape[0], 2)[:, :, 1:]
output = torch.cat((boxes, scores), dim=-1)
return output
def mio_module(self, each_mmbox, len_conf):
chunk = torch.chunk(each_mmbox, each_mmbox.shape[1], 1)
bmax = torch.max(torch.max(chunk[0], chunk[1]), chunk[2])
if len_conf == 0:
out = torch.cat([bmax, chunk[3]], dim=1)
else:
out = torch.cat([chunk[3], bmax], dim=1)
if len(chunk) == 6:
out = torch.cat([out, chunk[4], chunk[5]], dim=1)
elif len(chunk) == 8:
out = torch.cat(
[out, chunk[4], chunk[5], chunk[6], chunk[7]], dim=1)
return out
def _upsample_product(self, x, y):
'''Upsample and add two feature maps.
Args:
x: (Variable) top feature map to be upsampled.
y: (Variable) lateral feature map.
Returns:
(Variable) added feature map.
Note in PyTorch, when input size is odd, the upsampled feature map
with `F.upsample(..., scale_factor=2, mode='nearest')`
maybe not equal to the lateral feature map size.
e.g.
original input size: [N,_,15,15] ->
conv2d feature map size: [N,_,8,8] ->
upsampled feature map size: [N,_,16,16]
So we choose bilinear upsample which supports arbitrary output sizes.
'''
# Deprecation warning. align_corners=False default in 0.4.0, but in 0.3.0 it was True
# Original code was written in 0.3.1, I guess this is correct.
return y * F.interpolate(
x, size=y.shape[2:], mode="bilinear", align_corners=True)
class DeepHeadModule(nn.Module):
def __init__(self, input_channels, output_channels):
super().__init__()
self._input_channels = input_channels
self._output_channels = output_channels
self._mid_channels = min(self._input_channels, 256)
self.conv1 = nn.Conv2d(
self._input_channels, self._mid_channels, kernel_size=3, padding=1)
self.conv2 = nn.Conv2d(
self._mid_channels, self._mid_channels, kernel_size=3, padding=1)
self.conv3 = nn.Conv2d(
self._mid_channels, self._mid_channels, kernel_size=3, padding=1)
self.conv4 = nn.Conv2d(
self._mid_channels, self._output_channels, kernel_size=1,)
def forward(self, x):
out = self.conv1(x).relu()
out = self.conv2(out).relu()
out = self.conv3(out).relu()
out = self.conv4(out)
return out
def pa_multibox(output_channels, mbox_cfg, num_classes):
loc_layers = []
conf_layers = []
for k, v in enumerate(output_channels):
input_channels = 512
if k == 0:
loc_output = 4
conf_output = 2
elif k == 1:
loc_output = 8
conf_output = 4
else:
loc_output = 12
conf_output = 6
loc_layers += [
DeepHeadModule(input_channels, mbox_cfg[k] * loc_output)]
conf_layers += [
DeepHeadModule(input_channels, mbox_cfg[k] * (2+conf_output))]
return (loc_layers, conf_layers)
