def forward(self, x):
"""
Args:
- x: The convOut from a layer in the backbone network
Size: [batch_size, in_channels, conv_h, conv_w])
Returns a tuple (bbox_coords, class_confs, mask_output, prior_boxes) with sizes
- bbox_coords: [batch_size, conv_h*conv_w*num_priors, 4]
- class_confs: [batch_size, conv_h*conv_w*num_priors, num_classes]
- mask_output: [batch_size, conv_h*conv_w*num_priors, mask_dim]
- prior_boxes: [conv_h*conv_w*num_priors, 4]
"""
# In case we want to use another module's layers
src = self if self.parent[0] is None else self.parent[0]
conv_h = x.size(2)
conv_w = x.size(3)
if cfg.extra_head_net is not None:
x = src.upfeature(x)
if cfg.use_prediction_module:
# The two branches of PM design (c)
a = src.block(x)
b = src.conv(x)
b = src.bn(b)
b = F.relu(b)
# TODO: Possibly switch this out for a product
x = a + b
bbox_x = src.bbox_extra(x)
conf_x = src.conf_extra(x)
mask_x = src.mask_extra(x)
bbox = src.bbox_layer(bbox_x).permute(0, 2, 3, 1).contiguous().view(
x.size(0), -1, 4)
conf = src.conf_layer(conf_x).permute(0, 2, 3, 1).contiguous().view(
x.size(0), -1, self.num_classes)
if cfg.eval_mask_branch:
mask = src.mask_layer(mask_x).permute(
0, 2, 3, 1).contiguous().view(x.size(0), -1, self.mask_dim)
else:
mask = torch.zeros(x.size(0),
bbox.size(1),
self.mask_dim,
device=bbox.device)
if cfg.use_instance_coeff:
inst = src.inst_layer(x).permute(0, 2, 3, 1).contiguous().view(
x.size(0), -1, cfg.num_instance_coeffs)
# See box_utils.decode for an explanation of this
if cfg.use_yolo_regressors:
bbox[:, :, :2] = torch.sigmoid(bbox[:, :, :2]) - 0.5
bbox[:, :, 0] /= conv_w
bbox[:, :, 1] /= conv_h
if cfg.eval_mask_branch:
if cfg.mask_type == mask_type.direct:
mask = torch.sigmoid(mask)
elif cfg.mask_type == mask_type.lincomb:
mask = cfg.mask_proto_coeff_activation(mask)
if cfg.mask_proto_coeff_gate:
gate = src.gate_layer(x).permute(0, 2, 3,
1).contiguous().view(
x.size(0), -1,
self.mask_dim)
mask = mask * torch.sigmoid(gate)
priors = self.make_priors(conv_h, conv_w)
preds = {'loc': bbox, 'conf': conf, 'mask': mask, 'priors': priors}
if cfg.use_instance_coeff:
preds['inst'] = inst
return preds
def forward(self, x):
"""
Args:
- x: The convOut from a layer in the backbone network
Size: [batch_size, in_channels, conv_h, conv_w]) -- channel first type
Returns a tuple (bbox_coords, class_confs, mask_output, prior_boxes) with sizes
- bbox_coords: [batch_size, conv_h*conv_w*num_priors, 4]
- class_confs: [batch_size, conv_h*conv_w*num_priors, num_classes]
- mask_output: [batch_size, conv_h*conv_w*num_priors, mask_dim]
- prior_boxes: [conv_h*conv_w*num_priors, 4]
"""
#x->outs[0] #(n, 256, 69, 69)
# outs[1] #(n, 256, 34, 34)
# outs[2] #(n, 256, 17, 17)
# outs[3] #(n, 256, 8, 8)
# outs[4] #(n, 256, 4, 4)
# 2번째 호출부터 parent가짐.
# In case we want to use another module's layers
src = self if self.parent[0] is None else self.parent[0]
conv_h = x.size(2)
conv_w = x.size(3)
#'extra_head_net': [(256, 3, {'padding': 1})],
if cfg.extra_head_net is not None:
x = src.upfeature(x)
# False
if cfg.use_prediction_module:
# The two branches of PM design (c)
a = src.block(x)
b = src.conv(x)
b = src.bn(b)
b = F.relu(b)
# TODO: Possibly switch this out for a product
x = a + b
# lambda(x): return x
bbox_x = src.bbox_extra(x)
conf_x = src.conf_extra(x)
mask_x = src.mask_extra(x)
# H*W pixels 각각에서 9개의 prior box생성, 각 prior boxes는 4개의 좌표와 num_classes개의 confidence value 가짐.
# bbox_coords: [batch_size, conv_h*conv_w*num_priors, 4]
# class_confs: [batch_size, conv_h*conv_w*num_priors, num_classes]
bbox = src.bbox_layer(bbox_x).permute(0, 2, 3, 1).contiguous().view(
x.size(0), -1, 4)
conf = src.conf_layer(conf_x).permute(0, 2, 3, 1).contiguous().view(
x.size(0), -1, self.num_classes)
#True
if cfg.eval_mask_branch:
# mask_output: [batch_size, conv_h*conv_w*num_priors, mask_dim]
mask = src.mask_layer(mask_x).permute(
0, 2, 3, 1).contiguous().view(x.size(0), -1, self.mask_dim)
else:
mask = torch.zeros(x.size(0),
bbox.size(1),
self.mask_dim,
device=bbox.device)
# False
if cfg.use_mask_scoring:
score = src.score_layer(x).permute(0, 2, 3, 1).contiguous().view(
x.size(0), -1, 1)
# False
if cfg.use_instance_coeff:
inst = src.inst_layer(x).permute(0, 2, 3, 1).contiguous().view(
x.size(0), -1, cfg.num_instance_coeffs)
# See box_utils.decode for an explanation of this
if cfg.use_yolo_regressors:
bbox[:, :, :2] = torch.sigmoid(bbox[:, :, :2]) - 0.5
bbox[:, :, 0] /= conv_w
bbox[:, :, 1] /= conv_h
if cfg.eval_mask_branch:
if cfg.mask_type == mask_type.direct:
mask = torch.sigmoid(mask)
#Here
elif cfg.mask_type == mask_type.lincomb:
mask = cfg.mask_proto_coeff_activation(mask) #tanh(mask)
if cfg.mask_proto_coeff_gate:
gate = src.gate_layer(x).permute(0, 2, 3,
1).contiguous().view(
x.size(0), -1,
self.mask_dim)
mask = mask * torch.sigmoid(gate)
# False
if cfg.mask_proto_split_prototypes_by_head and cfg.mask_type == mask_type.lincomb:
mask = F.pad(mask,
(self.index * self.mask_dim,
(self.num_heads - self.index - 1) * self.mask_dim),
mode='constant',
value=0)
priors = self.make_priors(conv_h, conv_w, x.device)
preds = {'loc': bbox, 'conf': conf, 'mask': mask, 'priors': priors}
# False
if cfg.use_mask_scoring:
preds['score'] = score
# False
if cfg.use_instance_coeff:
preds['inst'] = inst
return preds
