def __init__(self, config):
super(ResNet50Conv5ROIFeatureExtractor, self).__init__()
resolution = config.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
scales = config.MODEL.ROI_BOX_HEAD.POOLER_SCALES
sampling_ratio = config.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO
pooler = Pooler(
output_size=(resolution, resolution),
scales=scales,
sampling_ratio=sampling_ratio,
)
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=config.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=config.MODEL.RESNETS.RES5_DILATION)
self.pooler = pooler
self.head = head
def __init__(self, config, in_channels, RCNN_top=None):
super(ResNet50Conv5ROIFeatureExtractorFlatten, self).__init__()
resolution = config.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
scales = config.MODEL.ROI_BOX_HEAD.POOLER_SCALES
sampling_ratio = config.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO
pooler = Pooler(
output_size=(resolution, resolution),
scales=scales,
sampling_ratio=sampling_ratio,
)
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=config.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=config.MODEL.RESNETS.RES5_DILATION)
self.pooler = pooler
self.head = head
hidden_channels = head.out_channels
use_gn = config.MODEL.ROI_BOX_HEAD.USE_GN
self.out_channels = config.MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM
self.avgpool = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Sequential(
make_fc(hidden_channels, self.out_channels, use_gn), nn.ReLU())
def __init__(self, config, select_num=10, class_num=5, is_cat=True):
super(Network, self).__init__()
self.detector = build_detection_model(config)
self.detector.eval()
self.detector.load_state_dict(torch.load(config.MODEL.WEIGHT)['model'],
strict=False)
# output_dir = config.OUTPUT_DIR
# checkpointer = DetectronCheckpointer(config, self.detector, save_dir=output_dir)
# ckpt = config.MODEL.WEIGHT
# _ = checkpointer.load(ckpt, use_latest=True)
for count, child in enumerate(self.detector.children()):
count += 1
print(count)
print(child)
if count < 4:
for param in child.parameters():
param.requires_grad = False
self.selector = Selector()
self.select_num = select_num
self.sftmax = nn.Softmax(dim=0)
self.is_cat = is_cat
# ResNet Head
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
self.head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=config.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=config.MODEL.RESNETS.RES5_DILATION)
self.avgpool_glob = nn.AdaptiveAvgPool2d(output_size=14)
# self.conv_glob = nn.Conv2d(1024, 2048, 2, stride=2)
# self.relu_glob = nn.PReLU(2048)
if self.is_cat:
self.conv1 = nn.Conv2d(2048, 512, 1)
# self.relu1 = nn.PReLU(512)
self.relu1 = nn.ReLU(inplace=True)
self.fc1 = nn.Linear(512 * (self.select_num + 1),
512) # select top 5
# self.fc1_relu = nn.PReLU(512)
self.fc1_relu = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(512, class_num)
else:
self.conv1 = nn.Conv2d(4096, 512, 1, 1)
# self.relu1 = nn.PReLU(512)
self.relu1 = nn.ReLU(inplace=True)
self.fc1 = nn.Linear(512, 512)
# self.fc1_relu = nn.PReLU(512)
self.fc1_relu = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(512, class_num)
self.avgpool = nn.AdaptiveAvgPool2d(output_size=1)
self.dropout = nn.Dropout()
def __init__(self,
config,
select_num=10,
class_num=4,
side=False,
is_cat=False):
super(Predictor, self).__init__()
self.selector = Selector()
self.select_num = select_num
self.sftmax = nn.Softmax(dim=0)
self.side = side
self.is_cat = is_cat
# ResNet Head
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
self.head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
dilation=config.MODEL.RESNETS.RES5_DILATION)
self.head.load_state_dict(
torch.load('/data6/SRIP19_SelfDriving/Outputs/layer4.pth')
) #TODO: how to avoid this? Maybe we can involve the weights into model_final.pth
self.avgpool_glob = nn.AdaptiveAvgPool2d(output_size=14)
if self.is_cat:
self.avg1 = nn.AdaptiveAvgPool3d(output_size=1)
self.avg2 = nn.AdaptiveAvgPool2d(output_size=1)
self.fc1 = nn.Linear(4096, 100)
self.relu1 = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(100, class_num)
self.drop = nn.Dropout(p=0.25)
if self.side:
self.fc_side1 = nn.Linear(4096, 100)
self.relu_side1 = nn.ReLU(inplace=True)
self.fc_side2 = nn.Linear(100, 21)
else:
self.avg = nn.AdaptiveAvgPool2d(output_size=1)
self.fc2 = nn.Linear(2048, class_num)
self.drop = nn.Dropout(p=0.25)
if self.side:
self.fc_side = nn.Linear(2048, 21)
def __init__(self, config, in_channels):
super(ResNet50Conv5ROIFeatureExtractor, self).__init__()
pooler = Pooler(config.MODEL.ROI_BOX_HEAD)
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=config.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=config.MODEL.RESNETS.RES5_DILATION)
self.pooler = pooler
self.head = head
self.out_channels = head.out_channels
def __init__(self, config, in_channels):
super(ResNet50Conv5ROIFeatureExtractor, self).__init__()
resolution = config.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
scales = config.MODEL.ROI_BOX_HEAD.POOLER_SCALES
sampling_ratio = config.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO
pooler = Pooler(
output_size=(resolution, resolution),
scales=scales,
sampling_ratio=sampling_ratio,
)
# ResNet50Conv5 input size is 1024, which is the default for backbone
# without FPN. For FPN structure, the in_channels will be 256. So we
# need to add a transit conv layer to make it compatible.
# The corresponding predictor should be FastRCNNPredictor.
if in_channels != 1024:
self.trans_conv = nn.Conv2d(in_channels,
1024,
kernel_size=3,
padding=1)
torch.nn.init.normal_(self.trans_conv.weight, std=0.01)
torch.nn.init.constant_(self.trans_conv.bias, 0)
else:
self.trans_conv = None
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=config.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=config.MODEL.RESNETS.RES5_DILATION)
self.pooler = pooler
self.head = head
self.out_channels = head.out_channels
def __init__(self, cfg):
super(ResNet50Conv5ROIFeatureExtractor, self).__init__()
self.cfg = cfg
self.in_channels = 2048
self.out_channels = cfg.REID.OUT_CHANNELS
self.avgpool = nn.AdaptiveAvgPool2d(1)
self.fc = make_fc(self.in_channels, self.out_channels)
if self.cfg.REID.USE_DIFF_FEAT:
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
head = resnet.ResNetHead(
block_module=cfg.MODEL.RESNETS.TRANS_FUNC,
stages=(stage,),
num_groups=cfg.MODEL.RESNETS.NUM_GROUPS,
width_per_group=cfg.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=cfg.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=cfg.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=cfg.MODEL.RESNETS.RES5_DILATION
)
self.head = head
def __init__(self, config, select_num=5, class_num=5, is_cat=True):
super(Predictor, self).__init__()
self.selector = Selector()
self.select_num = select_num
self.sftmax = nn.Softmax(dim=0)
self.is_cat = is_cat
# ResNet Head
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
self.head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=config.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=config.MODEL.RESNETS.RES5_DILATION)
self.avgpool_glob = nn.AdaptiveAvgPool2d(output_size=14)
# self.conv_glob = nn.Conv2d(1024, 2048, 2, stride=2)
# self.relu_glob = nn.PReLU(2048)
if self.is_cat:
self.conv1 = nn.Conv2d(2048, 512, 1)
self.relu1 = nn.ReLU(inplace=True)
self.fc1 = nn.Linear(512 * (self.select_num + 1),
512) # select top 5
self.fc1_relu = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(512, class_num)
else:
self.conv1 = nn.Conv2d(4096, 512, 1, 1)
self.relu1 = nn.ReLU(inplace=True)
self.fc1 = nn.Linear(512)
self.fc1_relu = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(512, class_num)
self.avgpool = nn.AdaptiveAvgPool2d(output_size=1)
self.dropout = nn.Dropout()
def __init__(self, config):
super(MultiLabelCls, self).__init__()
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
self.head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=config.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=config.MODEL.RESNETS.RES5_DILATION)
self.pooler = nn.AdaptiveAvgPool2d(7)
self.avgpool = nn.AdaptiveAvgPool2d(1)
self.num_classes = config.MODEL.ROI_BOX_HEAD.NUM_CLASSES - 1 # 80
# 1024 for both c4 and FPN architecture
num_inputs = config.MODEL.RESNETS.BACKBONE_OUT_CHANNELS * 2
self.classifier = nn.Linear(num_inputs, self.num_classes)
nn.init.normal_(self.classifier.weight, mean=0, std=0.01)
nn.init.constant_(self.classifier.bias, 0)
self.softmax = nn.Softmax()
def __init__(self, config, in_channels):
super(ResNet50Conv5ROIFeatureExtractor, self).__init__()
resolution = config.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
scales = config.MODEL.ROI_BOX_HEAD.POOLER_SCALES
sampling_ratio = config.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO
pooler = Pooler(
output_size=(resolution, resolution),
scales=scales,
sampling_ratio=sampling_ratio,
)
stage = resnet.StageSpec(index=4, block_count=3, return_features=False)
head = resnet.ResNetHead(
block_module=config.MODEL.RESNETS.TRANS_FUNC,
stages=(stage, ),
num_groups=config.MODEL.RESNETS.NUM_GROUPS,
width_per_group=config.MODEL.RESNETS.WIDTH_PER_GROUP,
stride_in_1x1=config.MODEL.RESNETS.STRIDE_IN_1X1,
stride_init=None,
res2_out_channels=config.MODEL.RESNETS.RES2_OUT_CHANNELS,
dilation=config.MODEL.RESNETS.RES5_DILATION)
model_parameters = head.state_dict()
if '.pkl' in cfg.MODEL.VG.RESNET_PARAMS_FILE:
with open(cfg.MODEL.VG.RESNET_PARAMS_FILE, 'rb') as f:
if torch._six.PY3:
pretrained_paras = pickle.load(f,
encoding="latin1")['blobs']
else:
pretrained_paras = pickle.load(f)['blobs']
stages = ["1.2", "2.3", "3.5", "4.2"]
pretrained_paras = _rename_weights_for_resnet(
pretrained_paras, stages)
pretrained_dict = {}
model_parameters.keys()
for k, v in pretrained_paras.items():
if k in list(model_parameters.keys()):
print(k)
pretrained_dict[k] = v
else:
pretrained_paras = torch.load(
cfg.MODEL.VG.RESNET_PARAMS_FILE)['model']
pretrained_new = {}
keys_list = list(pretrained_paras.keys())
for keys in keys_list:
pretrained_new[keys] = pretrained_paras[keys].cpu()
pretrained_paras = pretrained_new
pretrained_dict = {}
model_parameters.keys()
print('loading pascal pretrained weights')
for key in key_words:
new_key = "module.roi_heads.box.feature_extractor.head." + key
if new_key in list(pretrained_paras.keys()):
pretrained_dict[key] = pretrained_paras[new_key]
print(key, '------>', new_key)
model_parameters.update(pretrained_dict)
head.load_state_dict(model_parameters)
print('loading bottom up attention feature done')
self.pooler = pooler
self.head = head
self.out_channels = head.out_channels
spatial_dim = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
if cfg.MODEL.VG.SPATIAL_FEAT:
self.spatial_transform = nn.Sequential(
nn.Linear(2 * spatial_dim * spatial_dim, 256), nn.ReLU())
