def add_ResNet_roi_conv5_head_for_masks(
model, blob_in, dim_in, spatial_scale, preprefix='_[mask]_',
dilation=1, shared=False):
assert not shared, \
'Using shared ResNet stage not supported (temporarily)'
model.RoIFeatureTransform(
blob_in,
blob_out=preprefix + 'pool5',
blob_rois='mask_rois',
method=cfg.MRCNN.ROI_XFORM_METHOD,
resolution=cfg.MRCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.MRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
stride_init = int(cfg.MRCNN.ROI_XFORM_RESOLUTION / 7) # by default: 2
if not shared:
s, dim_in = ResNet.add_stage(
model, preprefix + 'res5', preprefix + 'pool5',
3, dim_in, 2048, 512, dilation, stride_init=stride_init)
else:
s, dim_in = ResNet.add_stage_shared(
model, preprefix, 'res5', preprefix + 'pool5',
3, dim_in, 2048, 512, dilation, stride_init=stride_init)
return s, 2048, spatial_scale
def add_ResNet_roi_conv5_head_for_keypoints(model, blob_in, dim_in,
spatial_scale):
"""Add a ResNet "conv5" / "stage5" head for Mask R-CNN keypoint prediction.
"""
model.RoIFeatureTransform(
blob_in,
'_[pose]_pool5',
blob_rois='keypoint_rois',
method=cfg.KRCNN.ROI_XFORM_METHOD,
resolution=cfg.KRCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.KRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
# Using the prefix '_[pose]_' to 'res5' enables initializing the head's
# parameters using pretrained 'res5' parameters if given (see
# utils.net.initialize_gpu_0_from_weights_file)
s, dim_in = ResNet.add_stage(model,
'_[pose]_res5',
'_[pose]_pool5',
3,
dim_in,
2048,
512,
cfg.KRCNN.DILATION,
stride_init=int(
cfg.KRCNN.ROI_XFORM_RESOLUTION / 7))
return s, 2048
def ResNet_roi_conv5_head_for_masks(dim_in):
"""ResNet "conv5" / "stage5" head for predicting masks."""
dilation = cfg.MRCNN.DILATION
stride_init = cfg.MRCNN.ROI_XFORM_RESOLUTION // 7 # by default: 2
module, dim_out = ResNet.add_stage(dim_in, 2048, 512, 3, dilation,
stride_init)
return module, dim_out
def add_ResNet_roi_conv5_head_for_masks(model, blob_in, dim_in, spatial_scale):
"""Add a ResNet "conv5" / "stage5" head for predicting masks."""
model.RoIFeatureTransform(
blob_in,
blob_out='_[mask]_pool5',
blob_rois='mask_rois',
method=cfg.MRCNN.ROI_XFORM_METHOD,
resolution=cfg.MRCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.MRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
dilation = cfg.MRCNN.DILATION
stride_init = int(cfg.MRCNN.ROI_XFORM_RESOLUTION / 7) # by default: 2
s, dim_in = ResNet.add_stage(model,
'_[mask]_res5',
'_[mask]_pool5',
3,
dim_in,
2048,
512,
dilation,
stride_init=stride_init)
return s, 2048
def add_ResNet_roi_conv5_head_for_keypoints(
model, blob_in, dim_in, spatial_scale
):
"""Add a ResNet "conv5" / "stage5" head for Mask R-CNN keypoint prediction.
"""
model.RoIFeatureTransform(
blob_in,
'_[pose]_pool5',
blob_rois='keypoint_rois',
method=cfg.KRCNN.ROI_XFORM_METHOD,
resolution=cfg.KRCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.KRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale
)
# Using the prefix '_[pose]_' to 'res5' enables initializing the head's
# parameters using pretrained 'res5' parameters if given (see
# utils.net.initialize_gpu_0_from_weights_file)
s, dim_in = ResNet.add_stage(
model,
'_[pose]_res5',
'_[pose]_pool5',
3,
dim_in,
2048,
512,
cfg.KRCNN.DILATION,
stride_init=int(cfg.KRCNN.ROI_XFORM_RESOLUTION / 7)
)
return s, 2048
def add_ResNet_roi_conv5_head_for_masks(model, blob_in, dim_in, spatial_scale):
"""Add a ResNet "conv5" / "stage5" head for predicting masks."""
model.RoIFeatureTransform(
blob_in,
blob_out='_[mask]_pool5',
blob_rois='mask_rois',
method=cfg.MRCNN.ROI_XFORM_METHOD,
resolution=cfg.MRCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.MRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale
)
dilation = cfg.MRCNN.DILATION
stride_init = int(cfg.MRCNN.ROI_XFORM_RESOLUTION / 7) # by default: 2
s, dim_in = ResNet.add_stage(
model,
'_[mask]_res5',
'_[mask]_pool5',
3,
dim_in,
2048,
512,
dilation,
stride_init=stride_init
)
return s, 2048
def detectron_weight_mapping(self):
detectron_weight_mapping, orphan_in_detectron = \
ResNet.residual_stage_detectron_mapping(self.res5, 'res5', 3, 5)
detectron_weight_mapping.update({
'upconv5.weight': 'conv5_mask_w',
'upconv5.bias': 'conv5_mask_b'
})
return detectron_weight_mapping, orphan_in_detectron
def detectron_weight_mapping(self):
detectron_weight_mapping, orphan_in_detectron = \
ResNet.residual_stage_detectron_mapping(self.res5, 'res5', 3, 5)
detectron_weight_mapping.update({
'upconv5.weight': 'conv5_mask_w',
'upconv5.bias': 'conv5_mask_b'
})
return detectron_weight_mapping, orphan_in_detectron
def detectron_weight_mapping(self):
detectron_weight_mapping, orphan_in_detectron = \
ResNet.residual_stage_detectron_mapping(self.res5, 'res5', 3, 5)
# Assign None for res5 modules, do not load from or save to checkpoint
for k in detectron_weight_mapping:
detectron_weight_mapping[k] = None
detectron_weight_mapping.update({
'upconv5.weight': 'conv5_mask_w',
'upconv5.bias': 'conv5_mask_b'
})
return detectron_weight_mapping, orphan_in_detectron
def detectron_weight_mapping(self):
detectron_weight_mapping, orphan_in_detectron = \
ResNet.residual_stage_detectron_mapping(self.res5, 'res5', 3, 5)
# Assign None for res5 modules, do not load from or save to checkpoint
for k in detectron_weight_mapping:
detectron_weight_mapping[k] = None
detectron_weight_mapping.update({
'upconv5.weight': 'conv5_mask_w',
'upconv5.bias': 'conv5_mask_b'
})
return detectron_weight_mapping, orphan_in_detectron
def __init__(self, dim_in, roi_xform_func, spatial_scale):
super().__init__()
self.dim_in = dim_in
self.roi_xform = roi_xform_func
self.spatial_scale = spatial_scale
self.dim_out = cfg.MRCNN.DIM_REDUCED
self.res5, dim_out = ResNet_roi_conv5_head_for_masks(dim_in)
self.upconv5 = nn.ConvTranspose2d(dim_out, self.dim_out, 2, 2, 0)
# Freeze all bn (affine) layers in resnet!!!
self.res5.apply(lambda m: ResNet.freeze_params(m)
if isinstance(m, mynn.AffineChannel2d) else None)
self._init_weights()
def add_ResNet_roi_conv5_head_for_keypoints(
model, blob_in, dim_in, spatial_scale):
model.RoIFeatureTransform(
blob_in, '_[pose]_pool5',
blob_rois='keypoint_rois',
method=cfg.KRCNN.ROI_XFORM_METHOD,
resolution=cfg.KRCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.KRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
s, dim_in = ResNet.add_stage(
model, '_[pose]_res5', '_[pose]_pool5',
3, dim_in, 2048, 512, cfg.KRCNN.DILATION,
stride_init=int(cfg.KRCNN.ROI_XFORM_RESOLUTION / 7))
return s, 2048, spatial_scale
def __init__(self, dim_in, roi_xform_func, spatial_scale):
super().__init__()
self.dim_in = dim_in
self.roi_xform = roi_xform_func
self.spatial_scale = spatial_scale
self.dim_out = cfg.MRCNN.DIM_REDUCED
self.res5, dim_out = ResNet_roi_conv5_head_for_masks(dim_in)
self.upconv5 = nn.ConvTranspose2d(dim_out, self.dim_out, 2, 2, 0)
# Freeze all bn (affine) layers in resnet!!!
self.res5.apply(
lambda m: ResNet.freeze_params(m)
if isinstance(m, mynn.AffineChannel2d) else None)
self._init_weights()
def add_ResNet_roi_conv5_head_for_keypoints(model, blob_in, dim_in,
spatial_scale):
model.RoIFeatureTransform(
blob_in,
'_[pose]_pool5',
blob_rois='keypoint_rois',
method=cfg.KRCNN.ROI_XFORM_METHOD,
resolution=cfg.KRCNN.ROI_XFORM_RESOLUTION,
sampling_ratio=cfg.KRCNN.ROI_XFORM_SAMPLING_RATIO,
spatial_scale=spatial_scale)
s, dim_in = ResNet.add_stage(model,
'_[pose]_res5',
'_[pose]_pool5',
3,
dim_in,
2048,
512,
cfg.KRCNN.DILATION,
stride_init=int(
cfg.KRCNN.ROI_XFORM_RESOLUTION / 7))
return s, 2048, spatial_scale
def ResNet_roi_conv5_head_for_masks(dim_in):
"""ResNet "conv5" / "stage5" head for predicting masks."""
dilation = cfg.MRCNN.DILATION
stride_init = cfg.MRCNN.ROI_XFORM_RESOLUTION // 7 # by default: 2
module, dim_out = ResNet.add_stage(dim_in, 2048, 512, 3, dilation, stride_init)
return module, dim_out
x = Variable(torch.Tensor(x)).to(device)
''' Load Model '''
model_save_type = ["all", "state_dict", "ckp"] # 학습 재개를 위해서 ckp 저장 권장 // else state_dict
model_name = "./modeling/ResNet"
option = 2
if option == 0:
# 모델 클래스는 어딘가에 반드시 선언되어 있어야 합니다
model_path = model_name + '_all.pt'
model = torch.load(model_path)
elif option == 1:
model_path = model_name + '_state_dict.pt'
model = ResNet().to(device)
model.load_state_dict(torch.load(model_path))
elif option == 2:
# model, optimizer 초기화
model = ResNet(ResidualBlock, [2, 2, 2]).to(device)
#optimizer = optim.Adam(model.parameters(), lr=lr)
model_path = model_name + '_ckp.tar'
checkpoint = torch.load(model_path, map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
#optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
last_epoch = checkpoint['last_epoch']
loss = checkpoint['loss']
''' Audio Test '''
