Skip to content

Commit af124dc

Browse files
heavengateheavengate
committed
fix API error
1 parent c945ffa commit af124dc

File tree

4 files changed

+43
-27
lines changed

4 files changed

+43
-27
lines changed

‎paddle/fluid/API.spec

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -324,7 +324,7 @@ paddle.fluid.layers.generate_mask_labels ArgSpec(args=['im_info', 'gt_classes',
324324
paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
325325
paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None))
326326
paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
327-
paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'gtscore', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample', 'label_smooth', 'name'], varargs=None, keywords=None, defaults=(None,))
327+
paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'gtscore', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample', 'use_label_smooth', 'name'], varargs=None, keywords=None, defaults=(True, None,))
328328
paddle.fluid.layers.multiclass_nms ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None))
329329
paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
330330
paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1))

‎paddle/fluid/operators/yolov3_loss_op.h

Lines changed: 34 additions & 21 deletions
Original file line numberDiff line numberDiff line change
@@ -121,13 +121,13 @@ template <typename T>
121121
static void CalcBoxLocationLoss(T* loss, const T* input, Box<T> gt,
122122
std::vector<int> anchors, int an_idx,
123123
int box_idx, int gi, int gj, int grid_size,
124-
int input_size, int stride) {
124+
int input_size, int stride, T score) {
125125
T tx = gt.x * grid_size - gi;
126126
T ty = gt.y * grid_size - gj;
127127
T tw = std::log(gt.w * input_size / anchors[2 * an_idx]);
128128
T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]);
129129

130-
T scale = 2.0 - gt.w * gt.h;
130+
T scale = (2.0 - gt.w * gt.h) * score;
131131
loss[0] += SCE<T>(input[box_idx], tx) * scale;
132132
loss[0] += SCE<T>(input[box_idx + stride], ty) * scale;
133133
loss[0] += L1Loss<T>(input[box_idx + 2 * stride], tw) * scale;
@@ -138,13 +138,14 @@ template <typename T>
138138
static void CalcBoxLocationLossGrad(T* input_grad, const T loss, const T* input,
139139
Box<T> gt, std::vector<int> anchors,
140140
int an_idx, int box_idx, int gi, int gj,
141-
int grid_size, int input_size, int stride) {
141+
int grid_size, int input_size, int stride,
142+
T score) {
142143
T tx = gt.x * grid_size - gi;
143144
T ty = gt.y * grid_size - gj;
144145
T tw = std::log(gt.w * input_size / anchors[2 * an_idx]);
145146
T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]);
146147

147-
T scale = 2.0 - gt.w * gt.h;
148+
T scale = (2.0 - gt.w * gt.h) * score;
148149
input_grad[box_idx] = SCEGrad<T>(input[box_idx], tx) * scale * loss;
149150
input_grad[box_idx + stride] =
150151
SCEGrad<T>(input[box_idx + stride], ty) * scale * loss;
@@ -157,23 +158,24 @@ static void CalcBoxLocationLossGrad(T* input_grad, const T loss, const T* input,
157158
template <typename T>
158159
static inline void CalcLabelLoss(T* loss, const T* input, const int index,
159160
const int label, const int class_num,
160-
const int stride, const T pos, const T neg) {
161+
const int stride, const T pos, const T neg,
162+
T score) {
161163
for (int i = 0; i < class_num; i++) {
162164
T pred = input[index + i * stride];
163-
loss[0] += SCE<T>(pred, (i == label) ? pos : neg);
165+
loss[0] += SCE<T>(pred, (i == label) ? pos : neg) * score;
164166
}
165167
}
166168

167169
template <typename T>
168170
static inline void CalcLabelLossGrad(T* input_grad, const T loss,
169171
const T* input, const int index,
170172
const int label, const int class_num,
171-
const int stride, const T pos,
172-
const T neg) {
173+
const int stride, const T pos, const T neg,
174+
T score) {
173175
for (int i = 0; i < class_num; i++) {
174176
T pred = input[index + i * stride];
175177
input_grad[index + i * stride] =
176-
SCEGrad<T>(pred, (i == label) ? pos : neg) * loss;
178+
SCEGrad<T>(pred, (i == label) ? pos : neg) * score * loss;
177179
}
178180
}
179181

@@ -187,8 +189,12 @@ static inline void CalcObjnessLoss(T* loss, const T* input, const T* objness,
187189
for (int k = 0; k < h; k++) {
188190
for (int l = 0; l < w; l++) {
189191
T obj = objness[k * w + l];
190-
if (obj > -0.5) {
191-
loss[i] += SCE<T>(input[k * w + l], obj);
192+
if (obj > 1e-5) {
193+
// positive sample: obj = mixup score
194+
loss[i] += SCE<T>(input[k * w + l], 1.0) * obj;
195+
} else if (obj > -0.5) {
196+
// negetive sample: obj = 0
197+
loss[i] += SCE<T>(input[k * w + l], 0.0);
192198
}
193199
}
194200
}
@@ -209,8 +215,11 @@ static inline void CalcObjnessLossGrad(T* input_grad, const T* loss,
209215
for (int k = 0; k < h; k++) {
210216
for (int l = 0; l < w; l++) {
211217
T obj = objness[k * w + l];
212-
if (obj > -0.5) {
213-
input_grad[k * w + l] = SCEGrad<T>(input[k * w + l], obj) * loss[i];
218+
if (obj > 1e-5) {
219+
input_grad[k * w + l] =
220+
SCEGrad<T>(input[k * w + l], 1.0) * obj * loss[i];
221+
} else if (obj > -0.5) {
222+
input_grad[k * w + l] = SCEGrad<T>(input[k * w + l], 0.0) * loss[i];
214223
}
215224
}
216225
}
@@ -315,7 +324,7 @@ class Yolov3LossKernel : public framework::OpKernel<T> {
315324

316325
if (best_iou > ignore_thresh) {
317326
int obj_idx = (i * mask_num + j) * stride + k * w + l;
318-
obj_mask_data[obj_idx] = static_cast<T>(-1.0);
327+
obj_mask_data[obj_idx] = static_cast<T>(-1);
319328
}
320329
// TODO(dengkaipeng): all losses should be calculated if best IoU
321330
// is bigger then truth thresh should be calculated here, but
@@ -357,20 +366,20 @@ class Yolov3LossKernel : public framework::OpKernel<T> {
357366
int mask_idx = GetMaskIndex(anchor_mask, best_n);
358367
gt_match_mask_data[i * b + t] = mask_idx;
359368
if (mask_idx >= 0) {
369+
T score = gt_score_data[i * b + t];
360370
int box_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
361371
an_stride, stride, 0);
362372
CalcBoxLocationLoss<T>(loss_data + i, input_data, gt, anchors, best_n,
363-
box_idx, gi, gj, h, input_size, stride);
373+
box_idx, gi, gj, h, input_size, stride, score);
364374

365-
T score = gt_score_data[i * b + t];
366375
int obj_idx = (i * mask_num + mask_idx) * stride + gj * w + gi;
367376
obj_mask_data[obj_idx] = score;
368377

369378
int label = gt_label_data[i * b + t];
370379
int label_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
371380
an_stride, stride, 5);
372381
CalcLabelLoss<T>(loss_data + i, input_data, label_idx, label,
373-
class_num, stride, label_pos, label_neg);
382+
class_num, stride, label_pos, label_neg, score);
374383
}
375384
}
376385
}
@@ -387,6 +396,7 @@ class Yolov3LossGradKernel : public framework::OpKernel<T> {
387396
auto* input = ctx.Input<Tensor>("X");
388397
auto* gt_box = ctx.Input<Tensor>("GTBox");
389398
auto* gt_label = ctx.Input<Tensor>("GTLabel");
399+
auto* gt_score = ctx.Input<Tensor>("GTScore");
390400
auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
391401
auto* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));
392402
auto* objness_mask = ctx.Input<Tensor>("ObjectnessMask");
@@ -418,6 +428,7 @@ class Yolov3LossGradKernel : public framework::OpKernel<T> {
418428
const T* input_data = input->data<T>();
419429
const T* gt_box_data = gt_box->data<T>();
420430
const int* gt_label_data = gt_label->data<int>();
431+
const T* gt_score_data = gt_score->data<T>();
421432
const T* loss_grad_data = loss_grad->data<T>();
422433
const T* obj_mask_data = objness_mask->data<T>();
423434
const int* gt_match_mask_data = gt_match_mask->data<int>();
@@ -429,22 +440,24 @@ class Yolov3LossGradKernel : public framework::OpKernel<T> {
429440
for (int t = 0; t < b; t++) {
430441
int mask_idx = gt_match_mask_data[i * b + t];
431442
if (mask_idx >= 0) {
443+
T score = gt_score_data[i * b + t];
432444
Box<T> gt = GetGtBox(gt_box_data, i, b, t);
433445
int gi = static_cast<int>(gt.x * w);
434446
int gj = static_cast<int>(gt.y * h);
435447

436448
int box_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
437449
an_stride, stride, 0);
438-
CalcBoxLocationLossGrad<T>(
439-
input_grad_data, loss_grad_data[i], input_data, gt, anchors,
440-
anchor_mask[mask_idx], box_idx, gi, gj, h, input_size, stride);
450+
CalcBoxLocationLossGrad<T>(input_grad_data, loss_grad_data[i],
451+
input_data, gt, anchors,
452+
anchor_mask[mask_idx], box_idx, gi, gj, h,
453+
input_size, stride, score);
441454

442455
int label = gt_label_data[i * b + t];
443456
int label_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
444457
an_stride, stride, 5);
445458
CalcLabelLossGrad<T>(input_grad_data, loss_grad_data[i], input_data,
446459
label_idx, label, class_num, stride, label_pos,
447-
label_neg);
460+
label_neg, score);
448461
}
449462
}
450463
}

‎python/paddle/fluid/layers/detection.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -482,7 +482,7 @@ def yolov3_loss(x,
482482
raise TypeError("Attr anchor_mask of yolov3_loss must be list or tuple")
483483
if not isinstance(class_num, int):
484484
raise TypeError("Attr class_num of yolov3_loss must be an integer")
485-
if not isinstance(use_label_smooth, int):
485+
if not isinstance(use_label_smooth, bool):
486486
raise TypeError("Attr ues_label_smooth of yolov3 must be a bool value")
487487
if not isinstance(ignore_thresh, float):
488488
raise TypeError(

‎python/paddle/fluid/tests/unittests/test_yolov3_loss_op.py

Lines changed: 7 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -142,7 +142,7 @@ def YOLOv3Loss(x, gtbox, gtlabel, gtscore, attrs):
142142
ty = gtbox[i, j, 1] * w - gj
143143
tw = np.log(gtbox[i, j, 2] * input_size / mask_anchors[an_idx][0])
144144
th = np.log(gtbox[i, j, 3] * input_size / mask_anchors[an_idx][1])
145-
scale = 2.0 - gtbox[i, j, 2] * gtbox[i, j, 3]
145+
scale = (2.0 - gtbox[i, j, 2] * gtbox[i, j, 3]) * gtscore[i, j]
146146
loss[i] += sce(x[i, an_idx, gj, gi, 0], tx) * scale
147147
loss[i] += sce(x[i, an_idx, gj, gi, 1], ty) * scale
148148
loss[i] += l1loss(x[i, an_idx, gj, gi, 2], tw) * scale
@@ -152,11 +152,14 @@ def YOLOv3Loss(x, gtbox, gtlabel, gtscore, attrs):
152152

153153
for label_idx in range(class_num):
154154
loss[i] += sce(x[i, an_idx, gj, gi, 5 + label_idx], label_pos
155-
if label_idx == gtlabel[i, j] else label_neg)
155+
if label_idx == gtlabel[i, j] else
156+
label_neg) * gtscore[i, j]
156157

157158
for j in range(mask_num * h * w):
158-
if objness[i, j] >= 0:
159-
loss[i] += sce(pred_obj[i, j], objness[i, j])
159+
if objness[i, j] > 0:
160+
loss[i] += sce(pred_obj[i, j], 1.0) * objness[i, j]
161+
elif objness[i, j] == 0:
162+
loss[i] += sce(pred_obj[i, j], 0.0)
160163

161164
return (loss, objness.reshape((n, mask_num, h, w)).astype('float32'), \
162165
gt_matches.astype('int32'))

0 commit comments

Comments
 (0)