smart bias bug fix

This commit is contained in:
Glenn Jocher 2020-04-03 12:42:09 -07:00
parent 41a002e798
commit 682c2b27e7
1 changed files with 3 additions and 2 deletions

View File

@ -77,11 +77,12 @@ def create_modules(module_defs, img_size):
elif mdef['type'] == 'yolo':
yolo_index += 1
stride = [32, 16, 8, 4, 2][yolo_index] # P3-P7 stride
layers = mdef['from'] if 'from' in mdef else []
modules = YOLOLayer(anchors=mdef['anchors'][mdef['mask']], # anchor list
nc=mdef['classes'], # number of classes
img_size=img_size, # (416, 416)
yolo_index=yolo_index, # 0, 1, 2...
layers=mdef['from'] if 'from' in mdef else [], # output layers
layers=layers, # output layers
stride=stride)
# Initialize preceding Conv2d() bias (https://arxiv.org/pdf/1708.02002.pdf section 3.3)
@ -89,7 +90,7 @@ def create_modules(module_defs, img_size):
bo = -4.5 #  obj bias
bc = math.log(1 / (modules.nc - 0.99)) # cls bias: class probability is sigmoid(p) = 1/nc
j = l[yolo_index] if 'from' in mdef else -1
j = layers[yolo_index] if 'from' in mdef else -1
bias_ = module_list[j][0].bias # shape(255,)
bias = bias_[:modules.no * modules.na].view(modules.na, -1) # shape(3,85)
bias[:, 4] += bo - bias[:, 4].mean() # obj