-
Notifications
You must be signed in to change notification settings - Fork 85
/
loss.py
373 lines (319 loc) · 17.6 KB
/
loss.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
import math
import numpy as np
import tensorflow as tf
def broadcast_iou(box_1, box_2):
""" 计算最终iou
:param box_1:
:param box_2:
:return: [batch_size, grid, grid, anchors, num_gt_box]
"""
# box_1: (..., (x1, y1, x2, y2))
# box_2: (N, (x1, y1, x2, y2))
# broadcast boxes
box_1 = tf.expand_dims(box_1, -2)
box_2 = tf.expand_dims(box_2, 0)
# new_shape: (..., N, (x1, y1, x2, y2))
new_shape = tf.broadcast_dynamic_shape(tf.shape(box_1), tf.shape(box_2))
box_1 = tf.broadcast_to(box_1, new_shape)
box_2 = tf.broadcast_to(box_2, new_shape)
int_w = tf.maximum(tf.minimum(box_1[..., 2], box_2[..., 2]) -
tf.maximum(box_1[..., 0], box_2[..., 0]), 0)
int_h = tf.maximum(tf.minimum(box_1[..., 3], box_2[..., 3]) -
tf.maximum(box_1[..., 1], box_2[..., 1]), 0)
int_area = int_w * int_h
box_1_area = (box_1[..., 2] - box_1[..., 0]) * \
(box_1[..., 3] - box_1[..., 1])
box_2_area = (box_2[..., 2] - box_2[..., 0]) * \
(box_2[..., 3] - box_2[..., 1])
return int_area / (box_1_area + box_2_area - int_area)
def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7):
""" 计算iou
:param box1:
:param box2:
:param x1y1x2y2:
:param GIoU:
:param DIoU:
:param CIoU:
:param eps:
:return:
"""
# box2 = box2.T
# Get the coordinates of bounding boxes
if x1y1x2y2: # x1, y1, x2, y2 = box1
b1_x1, b1_y1, b1_x2, b1_y2 = box1[:, 0], box1[:, 1], box1[:, 2], box1[:, 3]
b2_x1, b2_y1, b2_x2, b2_y2 = box2[:, 0], box2[:, 1], box2[:, 2], box2[:, 3]
else: # transform from xywh to xyxy
b1_x1, b1_x2 = box1[:, 0] - box1[:, 2] / 2, box1[:, 0] + box1[:, 2] / 2
b1_y1, b1_y2 = box1[:, 1] - box1[:, 3] / 2, box1[:, 1] + box1[:, 3] / 2
b2_x1, b2_x2 = box2[:, 0] - box2[:, 2] / 2, box2[:, 0] + box2[:, 2] / 2
b2_y1, b2_y2 = box2[:, 1] - box2[:, 3] / 2, box2[:, 1] + box2[:, 3] / 2
# Intersection area
inter = (tf.minimum(b1_x2, b2_x2) - tf.maximum(b1_x1, b2_x1)) * \
(tf.minimum(b1_y2, b2_y2) - tf.maximum(b1_y1, b2_y1))
# Union Area
w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
union = w1 * h1 + w2 * h2 - inter + eps
iou = inter / union
if GIoU or DIoU or CIoU:
# 这里计算得到一个最小的边框, 这个边框刚好能将b1,b2包住
cw = tf.maximum(b1_x2, b2_x2) - tf.minimum(b1_x1, b2_x1)
ch = tf.maximum(b1_y2, b2_y2) - tf.minimum(b1_y1, b2_y1)
if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared
rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 +
(b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center distance squared
if DIoU:
return iou - rho2 / c2 # DIoU
elif CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
# with torch.no_grad():)
v = (4 / math.pi ** 2) * tf.pow(tf.atan(w2 / h2) - tf.atan(w1 / h1), 2)
# with torch.no_grad():
# alpha = v / (v - iou + (1 + eps))
alpha = tf.stop_gradient(v / (v - iou + (1 + eps)))
return iou - (rho2 / c2 + v * alpha) # CIoU
else: # GIoU https://arxiv.org/pdf/1902.09630.pdf
c_area = cw * ch + eps # convex area
return iou - (c_area - union) / c_area # GIoU
else:
return iou
class ComputeLoss:
"""yolov5损失计算"""
def __init__(self, image_shape, anchors, anchor_masks, num_class,
box_loss_gain=0.05, class_loss_gain=0.5, obj_loss_gain=1.0,
anchor_ratio_thres=4, only_best_anchor=True, balanced_rate=20,
iou_ignore_thres=0.5, layer_balance=[4., 1.0, 0.4]):
self.image_shape = image_shape
self.anchors = anchors
self.anchor_masks = anchor_masks
self.num_class = num_class
self.box_loss_gain = box_loss_gain
self.class_loss_gain = class_loss_gain
self.obj_loss_gain = obj_loss_gain
self.anchor_ratio_thres = anchor_ratio_thres
self.only_best_anchor = only_best_anchor
self.balanced_rate = balanced_rate
self.iou_ignore_thres = iou_ignore_thres
self.layer_balance = layer_balance
def _transform_expand_target(self, gt_box_class_anchor, grid_size):
# y_true: [batch, boxes, (x1, y1, x2, y2, class, best_anchor)]
batch, num_boxes, _ = np.shape(gt_box_class_anchor)
# y_true_out: [N, grid, grid, anchors, [x1, y1, x2, y2, obj, class]]
y_true_out = np.zeros((batch, grid_size, grid_size, len(self.anchor_masks[0]), 6), dtype=np.float32)
for i in np.arange(batch):
for j in np.arange(num_boxes):
# 这里如果是padding的数据则跳过
if gt_box_class_anchor[i][j][2] == 0:
continue
# 计算中心点
box = gt_box_class_anchor[i, j, 0:4]
box_xy = (box[..., 0:2] + box[..., 2:4]) / 2
anchor_idx = int(gt_box_class_anchor[i, j, 5])
# 计算目标边框和anchor的宽高比
w = box[2] - box[0]
h = box[3] - box[1]
target_anchor = self.anchors[anchor_idx]
w_ratio = w / target_anchor[0]
h_ratio = h / target_anchor[1]
w_ratio_bool = (w_ratio > 1 / self.anchor_ratio_thres) and \
(w_ratio < self.anchor_ratio_thres)
h_ratio_bool = (h_ratio > 1 / self.anchor_ratio_thres) and \
(h_ratio < self.anchor_ratio_thres)
# 最大iou的anchor同时需要满足比例在[0.25, 4]之间
if w_ratio_bool and h_ratio_bool:
grid_xy = np.array(box_xy // (1 / grid_size), np.int32)
# 0,1,2 % 3 = 0,1,2 3,4,5 % 3 = 0,1,2 6,7,8 % 3 = 0,1,2
best_anchor_id = anchor_idx % len(self.anchor_masks[0])
if self.only_best_anchor:
y_true_out[i, grid_xy[0], grid_xy[1], best_anchor_id, :] = \
[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]
else:
y_true_out[i, grid_xy[0], grid_xy[1], :, :] = \
np.array([[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]])
# 扩展更多正样本
xy = box_xy * grid_size
inv_xy = grid_size - xy
if (xy[0] % 1 < 0.5) and (xy[0] > 1.):
jxy = xy - np.array([0.5, 0])
grid_jxy = np.array(jxy, np.int32)
if self.only_best_anchor:
y_true_out[i, grid_jxy[0], grid_jxy[1], best_anchor_id, :] = \
[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]
else:
y_true_out[i, grid_jxy[0], grid_jxy[1], :, :] = \
np.array([[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]])
if (xy[1] % 1 < 0.5) and (xy[1] > 1.):
kxy = xy - np.array([0, 0.5])
grid_kxy = np.array(kxy, np.int32)
if self.only_best_anchor:
y_true_out[i, grid_kxy[0], grid_kxy[1], best_anchor_id, :] = \
[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]
else:
y_true_out[i, grid_kxy[0], grid_kxy[1], :, :] = \
np.array([[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]])
if (inv_xy[0] % 1 < 0.5) and (inv_xy[0] > 1.):
inv_lxy = xy + np.array([0.5, 0])
grid_lxy = np.array(inv_lxy, np.int32)
if self.only_best_anchor:
y_true_out[i, grid_lxy[0], grid_lxy[1], best_anchor_id, :] = \
[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]
else:
y_true_out[i, grid_lxy[0], grid_lxy[1], :, :] = \
np.array([[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]])
if (inv_xy[1] % 1 < 0.5) and (inv_xy[1] > 1.):
inv_mxy = xy + np.array([0, 0.5])
grid_mxy = np.array(inv_mxy, np.int32)
if self.only_best_anchor:
y_true_out[i, grid_mxy[0], grid_mxy[1], best_anchor_id, :] = \
[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]
else:
y_true_out[i, grid_mxy[0], grid_mxy[1], :, :] = \
np.array([[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]])
return y_true_out
def build_targets(self, predicts, gt_boxes, gt_classes):
"""
:param predicts: [3, batch, grid, grid, anchors, 5+num_class]
:param gt_boxes: [batch, num_box, (x1, y1, x2, y2)]
:param gt_classes: [batch, num_box]
:return [3, batch, grid, grid, anchors, 6(x1, y1, x2, y2, obj, class)]
"""
gt_classes = np.expand_dims(gt_classes, axis=-1)
# [batch, num_box, (w, h)]
box_wh = gt_boxes[..., 2:4] - gt_boxes[..., 0:2]
# [batch, num_box, 3, (w, h)]
box_wh = np.tile(np.expand_dims(box_wh, axis=-2), (1, 1, len(self.anchor_masks[0]), 1))
# [batch, num_box, 3]
box_area = box_wh[..., 0] * box_wh[..., 1]
targets = []
for i, predict in enumerate(predicts):
predict = predict.numpy()
grid_size = predict.shape[1]
cur_anchor_ids = self.anchor_masks[i]
cur_anchors = self.anchors[cur_anchor_ids]
# (3, )
anchor_area = cur_anchors[..., 0] * cur_anchors[..., 1]
# 计算iou, 沿用v3的做法, 只要iou最大的anchor
intersection = np.minimum(box_wh[..., 0], cur_anchors[..., 0]) * \
np.minimum(box_wh[..., 1], cur_anchors[..., 1])
iou = intersection / (box_area + anchor_area - intersection)
anchor_idx = np.array(cur_anchor_ids[np.argmax(iou, axis=-1)], dtype=np.float32)
anchor_idx = np.expand_dims(anchor_idx, axis=-1)
# 拼接最后的结果
# [batch, num_box, (x1, y1, x2, y2, class, best_anchor_id)]
gt_box_class_anchor = np.concatenate([gt_boxes, gt_classes, anchor_idx], axis=-1)
target = self._transform_expand_target(gt_box_class_anchor, grid_size)
targets.append(target)
return targets
def __call__(self, predicts, gt_boxes, gt_classes):
"""
:param predicts: [3, batch, grid, grid, anchors, 5+num_class]
:param gt_boxes
:param gt_classes
:return
"""
loss_xy = 0.0
loss_wh = 0.0
loss_box = 0.0
loss_obj = 0.0
loss_cls = 0.0
targets = self.build_targets(predicts, gt_boxes, gt_classes)
for i, predict in enumerate(predicts):
batch = predict.shape[0]
grid_size = predict.shape[1]
# ----------------- 这里处理预测数据 --------------------------
pred_xy, pred_wh, pred_obj, pred_cls = tf.split(predict, (2, 2, 1, self.num_class), axis=-1)
# [batch, grid, grid, anchors, 2]
pred_xy = 2 * tf.sigmoid(pred_xy) - 0.5
# [batch, grid, grid, anchors, 2]
pred_wh = (tf.sigmoid(pred_wh) * 2) ** 2 * self.anchors[self.anchor_masks[i]]
# [batch, grid, grid, anchors, 4]
pred_xywh = tf.concat((pred_xy, pred_wh), axis=-1)
# [batch, grid, grid, anchors, 1]
pred_obj = tf.sigmoid(pred_obj)
# [batch, grid, grid, anchors, num_class]
pred_cls = tf.keras.layers.Softmax()(pred_cls)
grid_y, grid_x = tf.meshgrid(tf.range(grid_size), tf.range(grid_size))
grid = tf.expand_dims(tf.stack([grid_x, grid_y], axis=-1), axis=2)
# 这里xy从偏移量转成具体中心点坐标, 并且做了归一化, anchors在传进来前也做了归一化
pred_grid_xy = (pred_xy + tf.cast(grid, tf.float32)) / tf.cast(grid_size, tf.float32)
pred_x1y1 = pred_grid_xy - pred_wh / 2
pred_x2y2 = pred_grid_xy + pred_wh / 2
x1, y1 = tf.split(pred_x1y1, (1, 1), axis=-1)
x2, y2 = tf.split(pred_x2y2, (1, 1), axis=-1)
x1 = tf.minimum(tf.maximum(x1, 0.), self.image_shape[1])
y1 = tf.minimum(tf.maximum(y1, 0.), self.image_shape[0])
x2 = tf.minimum(tf.maximum(x2, 0.), self.image_shape[1])
y2 = tf.minimum(tf.maximum(y2, 0.), self.image_shape[0])
pred_box = tf.concat([x1, y1, x2, y2], axis=-1)
# ----------------- 这里处理target数据 --------------------------
# [batch, grid, grid, anchors, 6(x1, y1, x2, y2, obj, class)]
target = targets[i]
true_box, true_obj, true_cls = tf.split(target, (4, 1, 1), axis=-1)
true_xy = (true_box[..., 0:2] + true_box[..., 2:4]) / 2
true_wh = true_box[..., 2:4] - true_box[..., 0:2]
# 计算true_box的平移缩放量
# [batch_size, grid, grid, anchors, 2]
true_xy = true_xy * tf.cast(grid_size, tf.float32) - tf.cast(grid, tf.float32)
# 4. calculate all masks
# [batch_size, grid, grid, anchors]
obj_mask = tf.squeeze(true_obj, -1)
positive_num = tf.cast(tf.reduce_sum(obj_mask), tf.int32) + 1
negative_num = self.balanced_rate * positive_num
# print(grid_size)
# print(tf.where(obj_mask > 0))
# print(tf.boolean_mask(pred_xy, obj_mask > 0))
# print(tf.boolean_mask(true_xy, obj_mask > 0))
# print("--------------")
# ignore false positive when iou is over threshold
# [batch_size, grid, grid, anchors, num_gt_box] => [batch_size, grid, grid, anchors, 1]
best_iou = tf.map_fn(
lambda x: tf.reduce_max(broadcast_iou(x[0], tf.boolean_mask(
x[1], tf.cast(x[2], tf.bool))), axis=-1),
(pred_box, true_box, obj_mask),
tf.float32)
# [batch_size, grid, grid, anchors, 1]
ignore_mask = tf.cast(best_iou < self.iou_ignore_thres, tf.float32)
# 这里做了下样本均衡.
ignore_num = tf.cast(tf.reduce_sum(ignore_mask), tf.int32)
if ignore_num > negative_num:
neg_inds = tf.random.shuffle(tf.where(ignore_mask))[:negative_num]
neg_inds = tf.expand_dims(neg_inds, axis=1)
ones = tf.ones(tf.shape(neg_inds)[0], tf.float32)
ones = tf.expand_dims(ones, axis=1)
# 更新mask
ignore_mask = tf.zeros_like(ignore_mask, tf.float32)
ignore_mask = tf.tensor_scatter_nd_add(ignore_mask, neg_inds, ones)
# 5. calculate all losses
# [batch_size, grid, grid, anchors]
box_loss_scale = 2 - true_wh[..., 0] * true_wh[..., 1]
xy_loss = obj_mask * box_loss_scale * tf.reduce_sum(tf.square(true_xy - pred_xy), axis=-1)
# [batch_size, grid, grid, anchors]
wh_loss = obj_mask * box_loss_scale * tf.reduce_sum(tf.square(true_wh - pred_wh), axis=-1)
iou = bbox_iou(tf.boolean_mask(pred_box, obj_mask > 0),
tf.boolean_mask(true_box, obj_mask > 0), CIoU=True)
box_loss = (1. - iou)
# obj_loss = binary_crossentropy(true_obj, pred_obj)
conf_focal = tf.pow(obj_mask - tf.squeeze(pred_obj, -1), 2)
# indices = tf.where(true_obj > 0)
# true_obj = tf.tensor_scatter_nd_add(true_obj, indices, iou)
obj_loss = tf.keras.losses.binary_crossentropy(true_obj, pred_obj)
obj_loss = conf_focal * (obj_mask * obj_loss + (1 - obj_mask) * ignore_mask * obj_loss)
# obj_loss = obj_mask * obj_loss + (1 - obj_mask) * ignore_mask * obj_loss
# obj_loss = tf.keras.losses.binary_crossentropy(true_obj, pred_obj)
# 这里除了正样本会计算损失, 负样本低于一定置信的也计算损失
# obj_loss = obj_mask * obj_loss + (1 - obj_mask) * ignore_mask * obj_loss
# TODO: use binary_crossentropy instead
# class_loss = obj_mask * sparse_categorical_crossentropy(true_class_idx, pred_class)
class_loss = obj_mask * tf.keras.losses.sparse_categorical_crossentropy(true_cls, pred_cls)
# 6. sum over (batch, gridx, gridy, anchors) => (batch, 1)
loss_xy += tf.reduce_mean(xy_loss) * batch
loss_wh += tf.reduce_mean(wh_loss) * batch
if tf.size(iou) > 0:
loss_box += tf.reduce_mean(box_loss) * batch * self.box_loss_gain
loss_obj += tf.reduce_mean(obj_loss) * self.layer_balance[i] * batch * self.obj_loss_gain
loss_cls += tf.reduce_mean(class_loss) * batch * self.class_loss_gain
# return xy_loss + wh_loss + obj_loss + class_loss
return loss_xy, loss_wh, loss_box, loss_obj, loss_cls
# return loss_xy, loss_wh, loss_xy+loss_wh, loss_obj, loss_cls
if __name__ == "__main__":
pass