✨ [New] validation code!
Browse files- yolo/tools/drawer.py +1 -1
- yolo/tools/solver.py +50 -4
- yolo/utils/bounding_box_utils.py +47 -1
yolo/tools/drawer.py
CHANGED
@@ -37,7 +37,7 @@ def draw_bboxes(
|
|
37 |
font = ImageFont.load_default(30)
|
38 |
|
39 |
for bbox in bboxes:
|
40 |
-
class_id, x_min, y_min, x_max, y_max = bbox
|
41 |
if scaled_bbox:
|
42 |
x_min = x_min * width
|
43 |
x_max = x_max * width
|
|
|
37 |
font = ImageFont.load_default(30)
|
38 |
|
39 |
for bbox in bboxes:
|
40 |
+
class_id, x_min, y_min, x_max, y_max, *conf = bbox
|
41 |
if scaled_bbox:
|
42 |
x_min = x_min * width
|
43 |
x_max = x_max * width
|
yolo/tools/solver.py
CHANGED
@@ -30,6 +30,12 @@ class ModelTrainer:
|
|
30 |
self.progress = ProgressTracker(cfg.name, save_path, cfg.use_wandb)
|
31 |
self.num_epochs = cfg.task.epoch
|
32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
if getattr(train_cfg.ema, "enabled", False):
|
34 |
self.ema = ExponentialMovingAverage(model, decay=train_cfg.ema.decay)
|
35 |
else:
|
@@ -89,9 +95,7 @@ class ModelTrainer:
|
|
89 |
epoch_loss = self.train_one_epoch(dataloader)
|
90 |
self.progress.finish_one_epoch()
|
91 |
|
92 |
-
|
93 |
-
if (epoch + 1) % 5 == 0:
|
94 |
-
self.save_checkpoint(epoch, f"checkpoint_epoch_{epoch+1}.pth")
|
95 |
|
96 |
|
97 |
class ModelTester:
|
@@ -100,7 +104,7 @@ class ModelTester:
|
|
100 |
self.device = device
|
101 |
self.progress = ProgressTracker(cfg, save_path, cfg.use_wandb)
|
102 |
|
103 |
-
self.anchor2box = AnchorBoxConverter(cfg, device)
|
104 |
self.nms = cfg.task.nms
|
105 |
self.save_path = save_path
|
106 |
|
@@ -125,3 +129,45 @@ class ModelTester:
|
|
125 |
else:
|
126 |
raise e
|
127 |
dataloader.stop()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
self.progress = ProgressTracker(cfg.name, save_path, cfg.use_wandb)
|
31 |
self.num_epochs = cfg.task.epoch
|
32 |
|
33 |
+
validation_dataloader = create_dataloader(cfg.task.validation.data, cfg.dataset, cfg.task.validation.task)
|
34 |
+
anchor2box = AnchorBoxConverter(cfg.model, cfg.image_size, device)
|
35 |
+
self.validator = ModelValidator(
|
36 |
+
cfg.task.validation, model, save_path, device, self.progress, anchor2box, validation_dataloader
|
37 |
+
)
|
38 |
+
|
39 |
if getattr(train_cfg.ema, "enabled", False):
|
40 |
self.ema = ExponentialMovingAverage(model, decay=train_cfg.ema.decay)
|
41 |
else:
|
|
|
95 |
epoch_loss = self.train_one_epoch(dataloader)
|
96 |
self.progress.finish_one_epoch()
|
97 |
|
98 |
+
self.validator.solve()
|
|
|
|
|
99 |
|
100 |
|
101 |
class ModelTester:
|
|
|
104 |
self.device = device
|
105 |
self.progress = ProgressTracker(cfg, save_path, cfg.use_wandb)
|
106 |
|
107 |
+
self.anchor2box = AnchorBoxConverter(cfg.model, cfg.image_size, device)
|
108 |
self.nms = cfg.task.nms
|
109 |
self.save_path = save_path
|
110 |
|
|
|
129 |
else:
|
130 |
raise e
|
131 |
dataloader.stop()
|
132 |
+
|
133 |
+
|
134 |
+
class ModelValidator:
|
135 |
+
def __init__(
|
136 |
+
self,
|
137 |
+
validation_cfg: ValidationConfig,
|
138 |
+
model: YOLO,
|
139 |
+
save_path: str,
|
140 |
+
device,
|
141 |
+
progress: ProgressTracker,
|
142 |
+
anchor2box,
|
143 |
+
validation_dataloader,
|
144 |
+
):
|
145 |
+
self.model = model
|
146 |
+
self.device = device
|
147 |
+
self.progress = progress
|
148 |
+
self.save_path = save_path
|
149 |
+
|
150 |
+
self.anchor2box = anchor2box
|
151 |
+
self.nms = validation_cfg.nms
|
152 |
+
self.validdataloader = validation_dataloader
|
153 |
+
|
154 |
+
def solve(self):
|
155 |
+
# logger.info("🧪 Start Validation!")
|
156 |
+
self.model.eval()
|
157 |
+
|
158 |
+
iou_thresholds = torch.arange(0.5, 1.0, 0.05)
|
159 |
+
map_all = []
|
160 |
+
self.progress.start_one_epoch(len(self.validdataloader))
|
161 |
+
for data, targets in self.validdataloader:
|
162 |
+
data, targets = data.to(self.device), targets.to(self.device)
|
163 |
+
with torch.no_grad():
|
164 |
+
raw_output = self.model(data)
|
165 |
+
predict, _ = self.anchor2box(raw_output[0][3:], with_logits=True)
|
166 |
+
|
167 |
+
nms_out = bbox_nms(predict, self.nms)
|
168 |
+
for idx, predict in enumerate(nms_out):
|
169 |
+
map_value = calculate_map(predict, targets[idx], iou_thresholds)
|
170 |
+
map_all.append(map_value[0])
|
171 |
+
self.progress.one_batch(mapp=torch.Tensor(map_all).mean())
|
172 |
+
|
173 |
+
self.progress.finish_one_epoch()
|
yolo/utils/bounding_box_utils.py
CHANGED
@@ -297,6 +297,7 @@ def bbox_nms(predicts: Tensor, nms_cfg: NMSConfig):
|
|
297 |
cls_val, cls_idx = cls_dist.max(dim=-1, keepdim=True)
|
298 |
valid_mask = cls_val > nms_cfg.min_confidence
|
299 |
valid_cls = cls_idx[valid_mask].float()
|
|
|
300 |
valid_box = bbox[valid_mask.repeat(1, 1, 4)].view(-1, 4)
|
301 |
|
302 |
batch_idx, *_ = torch.where(valid_mask)
|
@@ -305,7 +306,52 @@ def bbox_nms(predicts: Tensor, nms_cfg: NMSConfig):
|
|
305 |
for idx in range(predicts.size(0)):
|
306 |
instance_idx = nms_idx[idx == batch_idx[nms_idx]]
|
307 |
|
308 |
-
predict_nms = torch.cat(
|
|
|
|
|
309 |
|
310 |
predicts_nms.append(predict_nms)
|
311 |
return predicts_nms
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
297 |
cls_val, cls_idx = cls_dist.max(dim=-1, keepdim=True)
|
298 |
valid_mask = cls_val > nms_cfg.min_confidence
|
299 |
valid_cls = cls_idx[valid_mask].float()
|
300 |
+
valid_con = cls_val[valid_mask].float()
|
301 |
valid_box = bbox[valid_mask.repeat(1, 1, 4)].view(-1, 4)
|
302 |
|
303 |
batch_idx, *_ = torch.where(valid_mask)
|
|
|
306 |
for idx in range(predicts.size(0)):
|
307 |
instance_idx = nms_idx[idx == batch_idx[nms_idx]]
|
308 |
|
309 |
+
predict_nms = torch.cat(
|
310 |
+
[valid_cls[instance_idx][:, None], valid_con[instance_idx][:, None], valid_box[instance_idx]], dim=-1
|
311 |
+
)
|
312 |
|
313 |
predicts_nms.append(predict_nms)
|
314 |
return predicts_nms
|
315 |
+
|
316 |
+
|
317 |
+
def calculate_map(predictions, ground_truths, iou_thresholds):
|
318 |
+
# TODO: Refactor this block
|
319 |
+
device = predictions.device
|
320 |
+
n_preds = predictions.size(0)
|
321 |
+
n_gts = (ground_truths[:, 0] != -1).sum()
|
322 |
+
ground_truths = ground_truths[:n_gts]
|
323 |
+
aps = []
|
324 |
+
|
325 |
+
ious = calculate_iou(predictions[:, 2:], ground_truths[:, 1:]) # [n_preds, n_gts]
|
326 |
+
|
327 |
+
for threshold in iou_thresholds:
|
328 |
+
tp = torch.zeros(n_preds, device=device)
|
329 |
+
fp = torch.zeros(n_preds, device=device)
|
330 |
+
|
331 |
+
max_iou, max_indices = torch.max(ious, dim=1)
|
332 |
+
above_threshold = max_iou >= threshold
|
333 |
+
matched_classes = predictions[:, 0] == ground_truths[max_indices, 0]
|
334 |
+
tp[above_threshold & matched_classes] = 1
|
335 |
+
fp[above_threshold & ~matched_classes] = 1
|
336 |
+
fp[max_iou < threshold] = 1
|
337 |
+
|
338 |
+
_, indices = torch.sort(predictions[:, 1], descending=True)
|
339 |
+
tp = tp[indices]
|
340 |
+
fp = fp[indices]
|
341 |
+
|
342 |
+
tp_cumsum = torch.cumsum(tp, dim=0)
|
343 |
+
fp_cumsum = torch.cumsum(fp, dim=0)
|
344 |
+
|
345 |
+
precision = tp_cumsum / (tp_cumsum + fp_cumsum + 1e-6)
|
346 |
+
recall = tp_cumsum / (n_gts + 1e-6)
|
347 |
+
|
348 |
+
recall_thresholds = torch.arange(0, 1, 0.1)
|
349 |
+
precision_at_recall = torch.zeros_like(recall_thresholds)
|
350 |
+
for i, r in enumerate(recall_thresholds):
|
351 |
+
precision_at_recall[i] = precision[recall >= r].max().item() if torch.any(recall >= r) else 0
|
352 |
+
|
353 |
+
ap = precision_at_recall.mean()
|
354 |
+
aps.append(ap)
|
355 |
+
|
356 |
+
mean_ap = torch.mean(torch.stack(aps))
|
357 |
+
return mean_ap, aps
|