scripts/libraries: Simplify YOLO post-processing using keepdims.

This commit is contained in:
Kwabena W. Agyeman 2025-01-22 17:20:04 -08:00
parent 6d21326233
commit 979b4e8ee6

View File

@ -74,14 +74,13 @@ class yolo_v2_postprocess:
def __init__(self, threshold=0.6, anchors=None, nms_threshold=0.1, nms_sigma=0.1): def __init__(self, threshold=0.6, anchors=None, nms_threshold=0.1, nms_sigma=0.1):
self.threshold = threshold self.threshold = threshold
if anchors is not None: self.anchors = anchors
self.anchors = anchors if self.anchors is None:
else:
self.anchors = np.array([[0.98830, 3.36060], self.anchors = np.array([[0.98830, 3.36060],
[2.11940, 5.37590], [2.11940, 5.37590],
[3.05200, 9.13360], [3.05200, 9.13360],
[5.55170, 9.30660], [5.55170, 9.30660],
[9.72600, 11.1422]], dtype=np.float) [9.72600, 11.1422]])
self.anchors_len = len(self.anchors) self.anchors_len = len(self.anchors)
self.nms_threshold = nms_threshold self.nms_threshold = nms_threshold
self.nms_sigma = nms_sigma self.nms_sigma = nms_sigma
@ -97,8 +96,8 @@ class yolo_v2_postprocess:
return a - (b * (a // b)) return a - (b * (a // b))
def softmax(x): def softmax(x):
e_x = np.exp(x - np.max(x)) e_x = np.exp(x - np.max(x, axis=1, keepdims=True))
return e_x / np.sum(e_x) return e_x / np.sum(e_x, axis=1, keepdims=True)
# Reshape the output to a 2D array # Reshape the output to a 2D array
colum_outputs = outputs[0].reshape((oh * ow * self.anchors_len, colum_outputs = outputs[0].reshape((oh * ow * self.anchors_len,
@ -121,18 +120,13 @@ class yolo_v2_postprocess:
bb_anchors = mod(score_indices, self.anchors_len) bb_anchors = mod(score_indices, self.anchors_len)
# Get the anchor box information # Get the anchor box information
bb_a_array = [self.anchors[i] for i in bb_anchors.tolist()] bb_a_array = np.take(self.anchors, bb_anchors, axis=0)
bb_a_array = np.array(bb_a_array)
# Get the score information # Get the score information
bb_scores = sigmoid(bb[:, _YOLO_V2_SCORE]) bb_scores = sigmoid(bb[:, _YOLO_V2_SCORE])
# Get the class information # Get the class information
bb_classes = [] bb_classes = np.argmax(softmax(bb[:, _YOLO_V2_CLASSES:]), axis=1)
for i in range(len(score_indices)):
s = softmax(bb[i, _YOLO_V2_CLASSES:])
bb_classes.append(np.argmax(s))
bb_classes = np.array(bb_classes, dtype=np.uint16)
# Compute the bounding box information # Compute the bounding box information
x_center = (bb_cols + sigmoid(bb[:, _YOLO_V2_TX])) / ow x_center = (bb_cols + sigmoid(bb[:, _YOLO_V2_TX])) / ow
@ -192,8 +186,7 @@ class yolo_v5_postprocess:
bb_scores = bb[:, _YOLO_V5_SCORE] bb_scores = bb[:, _YOLO_V5_SCORE]
# Get the class information # Get the class information
bb_classes = [np.argmax(bb[x, _YOLO_V5_CLASSES:]) for x in range(bb.shape[0])] bb_classes = np.argmax(bb[:, _YOLO_V5_CLASSES:], axis=1)
bb_classes = np.array(bb_classes, dtype=np.uint16)
# Compute the bounding box information # Compute the bounding box information
x_center = bb[:, _YOLO_V5_CX] x_center = bb[:, _YOLO_V5_CX]