2. Entrenament de models utilitzant transfer learning
En aquesta lliçó, aprendrem com entrenar un model de detecció d’objectes utilitzant transfer learning. Aquestes tècniques redueixen l’esforç computacional i milloren el rendiment del model, fins i tot amb datasets petits.
Primer, afegim el mòdul d’ajuda:
addpath('help-module');
Entrenament
Paràmetres YOLO necessaris
Els paràmetres següents són obligatoris quan es configura una sessió d’entrenament de YOLO:
- configFile (str) : Ruta al fitxer de configuració, normalment en format
.yaml, que defineix l’estructura del dataset i les classes. Aquest fitxer ha d’incloure les rutes als conjunts d’entrenament, validació i, opcionalment, test, així com la llista de noms de classes. Com a exemple, el fitxer que utilitzarem en aquest tutorial és data.yaml, que ja conté tota l’estructura definida. - baseModel (str): Aquest paràmetre especifica quin model s’utilitzarà per a l’entrenament. Permet continuar entrenant un model entrenat prèviament o començar l’entrenament a partir d’un model YOLO preentrenat. Pots triar entre diferents versions de YOLO, tasques, per exemple detecció d’objectes, segmentació o classificació, i mides de model. Per a aquests tutorials, utilitzarem YOLOv8 per a detecció d’objectes. Els models preentrenats disponibles, ordenats de més petit a més gran, són: yolov8n.pt, yolov8s.pt, yolov8m.pt, yolov8l.pt, yolov8x.pt. En aquestes lliçons es recomana utilitzar models més petits, per exemple
yolov8n.ptoyolov8s.pt, per tenir un entrenament i una inferència més ràpids. MaxEpochs(int): Aquest paràmetreestableix el nombre màxim d’èpoques per a l’entrenament, on una època és una passada completa per totes les imatges del dataset d’entrenament. Augmentar aquest valor normalment millora el rendiment del model, però també allarga el temps d’entrenament.ImageSize(list[int]):La mida, en píxels, a la qual es redimensionaran totes les imatges abans de l’entrenament. Les mides més grans poden millorar la precisió, però requereixen més recursos computacionals.
En el codi següent, definim els valors d’aquests paràmetres necessaris. Utilitza aquests paràmetres per accelerar l’execució:
configFile = fullfile(pwd, 'datasets', 'fruits_3_4998', 'data.yaml');
baseModel = 'yolov8s-oiv7.pt';
options.MaxEpochs = 1;
options.ImageSize = [256 256 3];
Visualització de les capes de YOLO
El codi següent permet visualitzar l’arquitectura i els paràmetres d’un model YOLOv8:
det = yolov8ObjectDetector2('yolov8s');
Pretrained yolov8s network already exists.
% Analitzar el model carregat
analyzeNetwork(det.Network);
Aquesta comanda obre una finestra nova que mostra totes les capes i paràmetres del model YOLO que s’està utilitzant. Tingues en compte que les diferents variants de YOLOv8, per exemple yolov8s, yolov8m, etc., no només difereixen en el nombre de paràmetres, sinó també lleugerament en la seva complexitat arquitectònica.
En el cas de YOLOv8, el model normalment consta de 23 blocs principals. Aquests es poden agrupar en tres components principals:
- Backbone, primers 10 blocs: Responsable d’extreure característiques de la imatge d’entrada.
- Neck, 12 blocs següents: Combina característiques a diferents escales per millorar el rendiment de la detecció.
- Head, últim bloc: Responsable de predir bounding boxes, etiquetes de classe i puntuacions de confiança.
Paràmetre de congelació de capes per al transfer learning
Un dels principals avantatges dels models YOLO és que venen preentrenats. Entrenar un model de detecció d’objectes des de zero normalment requereix milions d’imatges i una potència computacional considerable. Aquest problema es pot resoldre utilitzant transfer learning, que consisteix a agafar un model que ja ha estat entrenat amb un dataset gran i reentrenar-lo lleugerament utilitzant un conjunt molt més petit de dades noves.
El transfer learning és especialment efectiu en visió per computador perquè les xarxes neuronals convolucionals (CNN), que són la base de YOLO i de tots els models de detecció d’objectes, aprenen característiques de manera jeràrquica. Les capes inicials detecten característiques de baix nivell, com vores i textures, mentre que les capes més profundes capturen patrons més complexos i estructures d’objectes. Aquesta jerarquia fa que les capes inicials siguin altament reutilitzables en diferents tasques.
Normalment, les capes inicials responsables de l’extracció de característiques, el backbone, es congelen durant l’entrenament, cosa que significa que els seus pesos no s’actualitzen. Això permet que el model conservi les característiques visuals generals apreses a partir de datasets grans. Mentrestant, les capes posteriors es fine-tune amb el nou dataset per adaptar-se a la tasca específica.
En alguns escenaris, tot el model es pot fine-tune utilitzant una taxa d’aprenentatge més baixa, cosa que li permet adaptar-se gradualment a les dades noves mentre es minimitza el risc de sobreescriure coneixement preentrenat valuós.
Per això utilitzem YOLOv8. Ve preentrenat en datasets grans com COCO (per exemple, yolov8n.pt, yolov8s.pt) o Open Image V7 (per exemple, yolov8n-oiv7.pt, yolov8s-oiv7.pt). Utilitzar aquests models redueix significativament el temps d’entrenament i millora el rendiment, fins i tot amb datasets més petits.
Triar quines capes congelar
La decisió de quantes capes de YOLOv8 congelar depèn de com de semblant sigui el teu dataset objectiu a les dades preentrenades. YOLOv8 té 22 capes congelables: les primeres 10 extreuen característiques genèriques, com vores, textures i formes, mentre que les capes posteriors s’especialitzen en detecció. Com que les nostres classes, apple, orange i pear, estan ben representades a COCO i Open Images V7, els dominis coincideixen força. Això ens permet congelar la majoria de capes per mantenir intactes les característiques generals i fine-tune només les capes de més alt nivell per a les nostres fruites, cosa que accelera l’entrenament i redueix el risc de sobreajust.
Per congelar capes específiques durant l’entrenament, utilitzem el paràmetre freeze de la manera següent:
optionalArgs = {'freeze', 15};
Hi ha paràmetres addicionals disponibles per configurar el model, que s’explicaran a la Lliçó 4: Experimentació.
Iniciar l’entrenament
A continuació, entrenem el model YOLO amb els paràmetres que hem definit anteriorment. Això pot trigar aproximadament 2 minuts.
[yolov8Det, results] = trainYOLOv8ObjectDetector( ...
configFile, ...
baseModel, ...
options, ...
optionalArgs{:});
ans =
PythonEnvironment with properties:
Version: "3.11"
Executable: "C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\win64\python\python.exe"
Library: "C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\win64\python\python311.dll"
Home: "C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\win64\python"
Status: Terminated
ExecutionMode: OutOfProcess
False
is instance freeze
New https://pypi.org/project/ultralytics/8.3.174 available Update with 'pip install -U ultralytics'
Ultralytics YOLOv8.2.66 Python-3.11.5 torch-2.7.0+cpu CPU (11th Gen Intel Core(TM) i5-1135G7 2.40GHz)
□[34m□[1mengine\trainer: □[0mtask=detect, mode=train, model=yolov8s-oiv7.pt, data=C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\data.yaml, epochs=1, time=None, patience=100, batch=16, imgsz=256, save=True, save_period=-1, cache=False, device=None, workers=8, project=None, name=first, exist_ok=False, pretrained=True, optimizer=auto, verbose=True, seed=0, deterministic=True, single_cls=False, rect=False, cos_lr=False, close_mosaic=10, resume=False, amp=True, fraction=1.0, profile=False, freeze=15, multi_scale=False, overlap_mask=True, mask_ratio=4, dropout=0.0, val=True, split=val, save_json=False, save_hybrid=False, conf=None, iou=0.7, max_det=300, half=False, dnn=False, plots=True, source=None, vid_stride=1, stream_buffer=False, visualize=False, augment=False, agnostic_nms=False, classes=None, retina_masks=False, embed=None, show=False, save_frames=False, save_txt=False, save_conf=False, save_crop=False, show_labels=True, show_conf=True, show_boxes=True, line_width=None, format=torchscript, keras=False, optimize=False, int8=False, dynamic=False, simplify=False, opset=None, workspace=4, nms=False, lr0=0.01, lrf=0.01, momentum=0.937, weight_decay=0.0005, warmup_epochs=3.0, warmup_momentum=0.8, warmup_bias_lr=0.1, box=7.5, cls=0.5, dfl=1.5, pose=12.0, kobj=1.0, label_smoothing=0.0, nbs=64, hsv_h=0.015, hsv_s=0.7, hsv_v=0.4, degrees=0.0, translate=0.1, scale=0.5, shear=0.0, perspective=0.0, flipud=0.0, fliplr=0.5, bgr=0.0, mosaic=1.0, mixup=0.0, copy_paste=0.0, auto_augment=randaugment, erasing=0.4, crop_fraction=1.0, cfg=None, tracker=botsort.yaml, save_dir=C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first
Overriding model.yaml nc=601 with nc=3
from n params module arguments
0 -1 1 928 ultralytics.nn.modules.conv.Conv [3, 32, 3, 2]
1 -1 1 18560 ultralytics.nn.modules.conv.Conv [32, 64, 3, 2]
2 -1 1 29056 ultralytics.nn.modules.block.C2f [64, 64, 1, True]
3 -1 1 73984 ultralytics.nn.modules.conv.Conv [64, 128, 3, 2]
4 -1 2 197632 ultralytics.nn.modules.block.C2f [128, 128, 2, True]
5 -1 1 295424 ultralytics.nn.modules.conv.Conv [128, 256, 3, 2]
6 -1 2 788480 ultralytics.nn.modules.block.C2f [256, 256, 2, True]
7 -1 1 1180672 ultralytics.nn.modules.conv.Conv [256, 512, 3, 2]
8 -1 1 1838080 ultralytics.nn.modules.block.C2f [512, 512, 1, True]
9 -1 1 656896 ultralytics.nn.modules.block.SPPF [512, 512, 5]
10 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
11 [-1, 6] 1 0 ultralytics.nn.modules.conv.Concat [1]
12 -1 1 591360 ultralytics.nn.modules.block.C2f [768, 256, 1]
13 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
14 [-1, 4] 1 0 ultralytics.nn.modules.conv.Concat [1]
15 -1 1 148224 ultralytics.nn.modules.block.C2f [384, 128, 1]
16 -1 1 147712 ultralytics.nn.modules.conv.Conv [128, 128, 3, 2]
17 [-1, 12] 1 0 ultralytics.nn.modules.conv.Concat [1]
18 -1 1 493056 ultralytics.nn.modules.block.C2f [384, 256, 1]
19 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
20 [-1, 9] 1 0 ultralytics.nn.modules.conv.Concat [1]
21 -1 1 1969152 ultralytics.nn.modules.block.C2f [768, 512, 1]
22 [15, 18, 21] 1 2117209 ultralytics.nn.modules.head.Detect [3, [128, 256, 512]]
YOLOv8s summary: 225 layers, 11,136,761 parameters, 11,136,745 gradients, 28.7 GFLOPs
Transferred 349/355 items from pretrained weights
□[34m□[1mTensorBoard: □[0mStart with 'tensorboard --logdir C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first', view at http://localhost:6006/
Freezing layer 'model.0.conv.weight'
Freezing layer 'model.0.bn.weight'
Freezing layer 'model.0.bn.bias'
Freezing layer 'model.1.conv.weight'
Freezing layer 'model.1.bn.weight'
Freezing layer 'model.1.bn.bias'
Freezing layer 'model.2.cv1.conv.weight'
Freezing layer 'model.2.cv1.bn.weight'
Freezing layer 'model.2.cv1.bn.bias'
Freezing layer 'model.2.cv2.conv.weight'
Freezing layer 'model.2.cv2.bn.weight'
Freezing layer 'model.2.cv2.bn.bias'
Freezing layer 'model.2.m.0.cv1.conv.weight'
Freezing layer 'model.2.m.0.cv1.bn.weight'
Freezing layer 'model.2.m.0.cv1.bn.bias'
Freezing layer 'model.2.m.0.cv2.conv.weight'
Freezing layer 'model.2.m.0.cv2.bn.weight'
Freezing layer 'model.2.m.0.cv2.bn.bias'
Freezing layer 'model.3.conv.weight'
Freezing layer 'model.3.bn.weight'
Freezing layer 'model.3.bn.bias'
Freezing layer 'model.4.cv1.conv.weight'
Freezing layer 'model.4.cv1.bn.weight'
Freezing layer 'model.4.cv1.bn.bias'
Freezing layer 'model.4.cv2.conv.weight'
Freezing layer 'model.4.cv2.bn.weight'
Freezing layer 'model.4.cv2.bn.bias'
Freezing layer 'model.4.m.0.cv1.conv.weight'
Freezing layer 'model.4.m.0.cv1.bn.weight'
Freezing layer 'model.4.m.0.cv1.bn.bias'
Freezing layer 'model.4.m.0.cv2.conv.weight'
Freezing layer 'model.4.m.0.cv2.bn.weight'
Freezing layer 'model.4.m.0.cv2.bn.bias'
Freezing layer 'model.4.m.1.cv1.conv.weight'
Freezing layer 'model.4.m.1.cv1.bn.weight'
Freezing layer 'model.4.m.1.cv1.bn.bias'
Freezing layer 'model.4.m.1.cv2.conv.weight'
Freezing layer 'model.4.m.1.cv2.bn.weight'
Freezing layer 'model.4.m.1.cv2.bn.bias'
Freezing layer 'model.5.conv.weight'
Freezing layer 'model.5.bn.weight'
Freezing layer 'model.5.bn.bias'
Freezing layer 'model.6.cv1.conv.weight'
Freezing layer 'model.6.cv1.bn.weight'
Freezing layer 'model.6.cv1.bn.bias'
Freezing layer 'model.6.cv2.conv.weight'
Freezing layer 'model.6.cv2.bn.weight'
Freezing layer 'model.6.cv2.bn.bias'
Freezing layer 'model.6.m.0.cv1.conv.weight'
Freezing layer 'model.6.m.0.cv1.bn.weight'
Freezing layer 'model.6.m.0.cv1.bn.bias'
Freezing layer 'model.6.m.0.cv2.conv.weight'
Freezing layer 'model.6.m.0.cv2.bn.weight'
Freezing layer 'model.6.m.0.cv2.bn.bias'
Freezing layer 'model.6.m.1.cv1.conv.weight'
Freezing layer 'model.6.m.1.cv1.bn.weight'
Freezing layer 'model.6.m.1.cv1.bn.bias'
Freezing layer 'model.6.m.1.cv2.conv.weight'
Freezing layer 'model.6.m.1.cv2.bn.weight'
Freezing layer 'model.6.m.1.cv2.bn.bias'
Freezing layer 'model.7.conv.weight'
Freezing layer 'model.7.bn.weight'
Freezing layer 'model.7.bn.bias'
Freezing layer 'model.8.cv1.conv.weight'
Freezing layer 'model.8.cv1.bn.weight'
Freezing layer 'model.8.cv1.bn.bias'
Freezing layer 'model.8.cv2.conv.weight'
Freezing layer 'model.8.cv2.bn.weight'
Freezing layer 'model.8.cv2.bn.bias'
Freezing layer 'model.8.m.0.cv1.conv.weight'
Freezing layer 'model.8.m.0.cv1.bn.weight'
Freezing layer 'model.8.m.0.cv1.bn.bias'
Freezing layer 'model.8.m.0.cv2.conv.weight'
Freezing layer 'model.8.m.0.cv2.bn.weight'
Freezing layer 'model.8.m.0.cv2.bn.bias'
Freezing layer 'model.9.cv1.conv.weight'
Freezing layer 'model.9.cv1.bn.weight'
Freezing layer 'model.9.cv1.bn.bias'
Freezing layer 'model.9.cv2.conv.weight'
Freezing layer 'model.9.cv2.bn.weight'
Freezing layer 'model.9.cv2.bn.bias'
Freezing layer 'model.12.cv1.conv.weight'
Freezing layer 'model.12.cv1.bn.weight'
Freezing layer 'model.12.cv1.bn.bias'
Freezing layer 'model.12.cv2.conv.weight'
Freezing layer 'model.12.cv2.bn.weight'
Freezing layer 'model.12.cv2.bn.bias'
Freezing layer 'model.12.m.0.cv1.conv.weight'
Freezing layer 'model.12.m.0.cv1.bn.weight'
Freezing layer 'model.12.m.0.cv1.bn.bias'
Freezing layer 'model.12.m.0.cv2.conv.weight'
Freezing layer 'model.12.m.0.cv2.bn.weight'
Freezing layer 'model.12.m.0.cv2.bn.bias'
Freezing layer 'model.22.dfl.conv.weight'
C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\win64\python\Lib\site-packages\ultralytics\engine\trainer.py:268: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
self.scaler = torch.cuda.amp.GradScaler(enabled=self.amp)
□[34m□[1mtrain: □[0mScanning C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\labels.cache... 1019 images, 136 backgrounds, 13 corrupt: 100%|##########| 1019/1019 [00:00<?, ?it/s]
□[34m□[1mtrain: □[0mScanning C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\labels.cache... 1019 images, 136 backgrounds, 13 corrupt: 100%|##########| 1019/1019 [00:00<?, ?it/s]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\0510fe79-68bd-4b4a-951f-dab86920adeb.png: ignoring corrupt image/label: negative label values [ -0.047368]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\0eaaca45-2f14-4cef-87e4-25c02c6bf3b2.png: ignoring corrupt image/label: negative label values [ -0.018722]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\15aef5fc-3df2-4f0e-8dd8-95ea8c730407.png: ignoring corrupt image/label: negative label values [ -0.022958]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\175a13d3-1a18-428a-a577-48807a5871fe.png: ignoring corrupt image/label: negative label values [ -0.083108]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\23232dd4-2cb4-4229-b335-98862a76323a.png: ignoring corrupt image/label: negative label values [ -0.049029]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\3d14a064-5447-4257-a19d-ab909a56da01.png: ignoring corrupt image/label: negative label values [ -0.016604]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\55c845f7-6a6e-4b05-9ad6-c50cb78e9874.png: ignoring corrupt image/label: negative label values [ -0.063713]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\757484cc-6ad5-4ad9-b03b-b0b4ce7ef7ee.png: ignoring corrupt image/label: negative label values [ -0.054816]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\7bbc8b1d-26fb-4268-8d72-6c10bdedce90.png: ignoring corrupt image/label: negative label values [ -0.046767]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\8b961909-3110-4bc8-b08f-f6bb6eee123a.png: ignoring corrupt image/label: negative label values [ -0.02139]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\d402070d-cb8b-4109-aea2-7b33f01cadad.png: ignoring corrupt image/label: negative label values [ -0.015335]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\f8891e69-28dd-4fc8-8a2e-4d1ad88145b8.png: ignoring corrupt image/label: negative label values [ -0.059821]
□[34m□[1mtrain: □[0mWARNING ⚠️ C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\train\images\fef31132-3b8d-48f2-9493-1e522dde091c.png: ignoring corrupt image/label: negative label values [ -0.075973]
C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\win64\python\Lib\site-packages\torch\utils\data\dataloader.py:665: UserWarning: 'pin_memory' argument is set as true but no accelerator is found, then device pinned memory won't be used.
warnings.warn(warn_msg)
□[34m□[1mval: □[0mScanning C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\val\labels.cache... 144 images, 0 backgrounds, 0 corrupt: 100%|##########| 144/144 [00:00<?, ?it/s]
□[34m□[1mval: □[0mScanning C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\val\labels.cache... 144 images, 0 backgrounds, 0 corrupt: 100%|##########| 144/144 [00:00<?, ?it/s]
Plotting labels to C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\labels.jpg...
□[34m□[1moptimizer:□[0m 'optimizer=auto' found, ignoring 'lr0=0.01' and 'momentum=0.937' and determining best 'optimizer', 'lr0' and 'momentum' automatically...
□[34m□[1moptimizer:□[0m AdamW(lr=0.001429, momentum=0.9) with parameter groups 57 weight(decay=0.0), 64 weight(decay=0.0005), 63 bias(decay=0.0)
□[34m□[1mTensorBoard: □[0mmodel graph visualization added ✅
Image sizes 256 train, 256 val
Using 0 dataloader workers
Logging results to □[1mC:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first□[0m
Starting training for 1 epochs...
Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size
0%| | 0/63 [00:00<?, ?it/s]
1/1 0G 0.9298 3.627 0.9719 69 256: 0%| | 0/63 [00:01<?, ?it/s]
1/1 0G 0.9298 3.627 0.9719 69 256: 2%|1 | 1/63 [00:01<01:51, 1.79s/it]
1/1 0G 0.9627 3.689 0.9906 73 256: 2%|1 | 1/63 [00:04<01:51, 1.79s/it]
1/1 0G 0.9627 3.689 0.9906 73 256: 3%|3 | 2/63 [00:04<02:13, 2.19s/it]
1/1 0G 0.9275 3.788 1.002 58 256: 3%|3 | 2/63 [00:06<02:13, 2.19s/it]
1/1 0G 0.9275 3.788 1.002 58 256: 5%|4 | 3/63 [00:06<02:13, 2.22s/it]
1/1 0G 0.9068 3.739 0.9966 54 256: 5%|4 | 3/63 [00:09<02:13, 2.22s/it]
1/1 0G 0.9068 3.739 0.9966 54 256: 6%|6 | 4/63 [00:09<02:21, 2.39s/it]
1/1 0G 0.9173 3.631 1.002 57 256: 6%|6 | 4/63 [00:12<02:21, 2.39s/it]
1/1 0G 0.9173 3.631 1.002 57 256: 8%|7 | 5/63 [00:12<02:29, 2.57s/it]
1/1 0G 0.9202 3.549 1.009 67 256: 8%|7 | 5/63 [00:14<02:29, 2.57s/it]
1/1 0G 0.9202 3.549 1.009 67 256: 10%|9 | 6/63 [00:14<02:23, 2.51s/it]
1/1 0G 0.9254 3.444 1.02 66 256: 10%|9 | 6/63 [00:16<02:23, 2.51s/it]
1/1 0G 0.9254 3.444 1.02 66 256: 11%|#1 | 7/63 [00:16<02:09, 2.31s/it]
1/1 0G 0.923 3.352 1.02 73 256: 11%|#1 | 7/63 [00:18<02:09, 2.31s/it]
1/1 0G 0.923 3.352 1.02 73 256: 13%|#2 | 8/63 [00:18<01:59, 2.16s/it]
1/1 0G 0.9153 3.242 1.015 65 256: 13%|#2 | 8/63 [00:20<01:59, 2.16s/it]
1/1 0G 0.9153 3.242 1.015 65 256: 14%|#4 | 9/63 [00:20<01:52, 2.08s/it]
1/1 0G 0.9033 3.136 1.01 52 256: 14%|#4 | 9/63 [00:21<01:52, 2.08s/it]
1/1 0G 0.9033 3.136 1.01 52 256: 16%|#5 | 10/63 [00:21<01:47, 2.02s/it]
1/1 0G 0.9005 3.025 1.012 66 256: 16%|#5 | 10/63 [00:23<01:47, 2.02s/it]
1/1 0G 0.9005 3.025 1.012 66 256: 17%|#7 | 11/63 [00:24<01:44, 2.02s/it]
1/1 0G 0.8889 2.926 1.007 63 256: 17%|#7 | 11/63 [00:25<01:44, 2.02s/it]
1/1 0G 0.8889 2.926 1.007 63 256: 19%|#9 | 12/63 [00:25<01:40, 1.97s/it]
1/1 0G 0.8769 2.819 1.004 59 256: 19%|#9 | 12/63 [00:27<01:40, 1.97s/it]
1/1 0G 0.8769 2.819 1.004 59 256: 21%|## | 13/63 [00:27<01:36, 1.93s/it]
1/1 0G 0.8719 2.757 1.003 40 256: 21%|## | 13/63 [00:29<01:36, 1.93s/it]
1/1 0G 0.8719 2.757 1.003 40 256: 22%|##2 | 14/63 [00:29<01:35, 1.96s/it]
1/1 0G 0.8685 2.685 1.003 84 256: 22%|##2 | 14/63 [00:31<01:35, 1.96s/it]
1/1 0G 0.8685 2.685 1.003 84 256: 24%|##3 | 15/63 [00:31<01:34, 1.97s/it]
1/1 0G 0.86 2.605 0.9977 76 256: 24%|##3 | 15/63 [00:33<01:34, 1.97s/it]
1/1 0G 0.86 2.605 0.9977 76 256: 25%|##5 | 16/63 [00:33<01:31, 1.94s/it]
1/1 0G 0.8428 2.519 0.9916 64 256: 25%|##5 | 16/63 [00:35<01:31, 1.94s/it]
1/1 0G 0.8428 2.519 0.9916 64 256: 27%|##6 | 17/63 [00:35<01:27, 1.91s/it]
1/1 0G 0.8388 2.441 0.992 55 256: 27%|##6 | 17/63 [00:37<01:27, 1.91s/it]
1/1 0G 0.8388 2.441 0.992 55 256: 29%|##8 | 18/63 [00:37<01:24, 1.88s/it]
1/1 0G 0.84 2.373 0.9888 79 256: 29%|##8 | 18/63 [00:39<01:24, 1.88s/it]
1/1 0G 0.84 2.373 0.9888 79 256: 30%|### | 19/63 [00:39<01:21, 1.85s/it]
1/1 0G 0.835 2.31 0.9859 53 256: 30%|### | 19/63 [00:40<01:21, 1.85s/it]
1/1 0G 0.835 2.31 0.9859 53 256: 32%|###1 | 20/63 [00:40<01:19, 1.84s/it]
1/1 0G 0.8379 2.257 0.9843 93 256: 32%|###1 | 20/63 [00:42<01:19, 1.84s/it]
1/1 0G 0.8379 2.257 0.9843 93 256: 33%|###3 | 21/63 [00:42<01:17, 1.84s/it]
1/1 0G 0.8338 2.203 0.9858 67 256: 33%|###3 | 21/63 [00:44<01:17, 1.84s/it]
1/1 0G 0.8338 2.203 0.9858 67 256: 35%|###4 | 22/63 [00:44<01:14, 1.83s/it]
1/1 0G 0.8331 2.148 0.9844 68 256: 35%|###4 | 22/63 [00:46<01:14, 1.83s/it]
1/1 0G 0.8331 2.148 0.9844 68 256: 37%|###6 | 23/63 [00:46<01:13, 1.83s/it]
1/1 0G 0.8254 2.099 0.9816 49 256: 37%|###6 | 23/63 [00:48<01:13, 1.83s/it]
1/1 0G 0.8254 2.099 0.9816 49 256: 38%|###8 | 24/63 [00:48<01:10, 1.82s/it]
1/1 0G 0.8228 2.055 0.9807 74 256: 38%|###8 | 24/63 [00:49<01:10, 1.82s/it]
1/1 0G 0.8228 2.055 0.9807 74 256: 40%|###9 | 25/63 [00:49<01:09, 1.82s/it]
1/1 0G 0.8168 2.01 0.977 68 256: 40%|###9 | 25/63 [00:51<01:09, 1.82s/it]
1/1 0G 0.8168 2.01 0.977 68 256: 41%|####1 | 26/63 [00:51<01:05, 1.77s/it]
1/1 0G 0.8111 1.967 0.9749 82 256: 41%|####1 | 26/63 [00:53<01:05, 1.77s/it]
1/1 0G 0.8111 1.967 0.9749 82 256: 43%|####2 | 27/63 [00:53<01:02, 1.74s/it]
1/1 0G 0.8055 1.927 0.9738 76 256: 43%|####2 | 27/63 [00:54<01:02, 1.74s/it]
1/1 0G 0.8055 1.927 0.9738 76 256: 44%|####4 | 28/63 [00:54<01:00, 1.72s/it]
1/1 0G 0.8027 1.896 0.9747 55 256: 44%|####4 | 28/63 [00:56<01:00, 1.72s/it]
1/1 0G 0.8027 1.896 0.9747 55 256: 46%|####6 | 29/63 [00:56<00:58, 1.72s/it]
1/1 0G 0.8 1.864 0.9742 72 256: 46%|####6 | 29/63 [00:58<00:58, 1.72s/it]
1/1 0G 0.8 1.864 0.9742 72 256: 48%|####7 | 30/63 [00:58<00:56, 1.72s/it]
1/1 0G 0.7944 1.829 0.9732 56 256: 48%|####7 | 30/63 [01:00<00:56, 1.72s/it]
1/1 0G 0.7944 1.829 0.9732 56 256: 49%|####9 | 31/63 [01:00<00:55, 1.73s/it]
1/1 0G 0.7917 1.795 0.9737 64 256: 49%|####9 | 31/63 [01:02<00:55, 1.73s/it]
1/1 0G 0.7917 1.795 0.9737 64 256: 51%|##### | 32/63 [01:02<00:58, 1.90s/it]
1/1 0G 0.7873 1.764 0.9697 69 256: 51%|##### | 32/63 [01:04<00:58, 1.90s/it]
1/1 0G 0.7873 1.764 0.9697 69 256: 52%|#####2 | 33/63 [01:04<01:01, 2.04s/it]
1/1 0G 0.787 1.738 0.9682 52 256: 52%|#####2 | 33/63 [01:06<01:01, 2.04s/it]
1/1 0G 0.787 1.738 0.9682 52 256: 54%|#####3 | 34/63 [01:06<00:59, 2.04s/it]
1/1 0G 0.7849 1.713 0.9681 63 256: 54%|#####3 | 34/63 [01:08<00:59, 2.04s/it]
1/1 0G 0.7849 1.713 0.9681 63 256: 56%|#####5 | 35/63 [01:08<00:52, 1.88s/it]
1/1 0G 0.7815 1.692 0.967 71 256: 56%|#####5 | 35/63 [01:10<00:52, 1.88s/it]
1/1 0G 0.7815 1.692 0.967 71 256: 57%|#####7 | 36/63 [01:10<00:50, 1.86s/it]
1/1 0G 0.7792 1.668 0.966 77 256: 57%|#####7 | 36/63 [01:11<00:50, 1.86s/it]
1/1 0G 0.7792 1.668 0.966 77 256: 59%|#####8 | 37/63 [01:11<00:45, 1.74s/it]
1/1 0G 0.7767 1.641 0.9655 55 256: 59%|#####8 | 37/63 [01:13<00:45, 1.74s/it]
1/1 0G 0.7767 1.641 0.9655 55 256: 60%|###### | 38/63 [01:13<00:42, 1.69s/it]
1/1 0G 0.7738 1.619 0.9645 59 256: 60%|###### | 38/63 [01:14<00:42, 1.69s/it]
1/1 0G 0.7738 1.619 0.9645 59 256: 62%|######1 | 39/63 [01:14<00:40, 1.68s/it]
1/1 0G 0.7722 1.597 0.9632 79 256: 62%|######1 | 39/63 [01:16<00:40, 1.68s/it]
1/1 0G 0.7722 1.597 0.9632 79 256: 63%|######3 | 40/63 [01:16<00:41, 1.81s/it]
1/1 0G 0.7683 1.577 0.9616 80 256: 63%|######3 | 40/63 [01:18<00:41, 1.81s/it]
1/1 0G 0.7683 1.577 0.9616 80 256: 65%|######5 | 41/63 [01:18<00:39, 1.78s/it]
1/1 0G 0.7645 1.554 0.9601 59 256: 65%|######5 | 41/63 [01:20<00:39, 1.78s/it]
1/1 0G 0.7645 1.554 0.9601 59 256: 67%|######6 | 42/63 [01:20<00:36, 1.72s/it]
1/1 0G 0.7644 1.541 0.9603 42 256: 67%|######6 | 42/63 [01:21<00:36, 1.72s/it]
1/1 0G 0.7644 1.541 0.9603 42 256: 68%|######8 | 43/63 [01:21<00:34, 1.72s/it]
1/1 0G 0.7588 1.523 0.9583 59 256: 68%|######8 | 43/63 [01:24<00:34, 1.72s/it]
1/1 0G 0.7588 1.523 0.9583 59 256: 70%|######9 | 44/63 [01:24<00:36, 1.93s/it]
1/1 0G 0.7553 1.506 0.956 58 256: 70%|######9 | 44/63 [01:27<00:36, 1.93s/it]
1/1 0G 0.7553 1.506 0.956 58 256: 71%|#######1 | 45/63 [01:27<00:41, 2.31s/it]
1/1 0G 0.7536 1.489 0.955 48 256: 71%|#######1 | 45/63 [01:30<00:41, 2.31s/it]
1/1 0G 0.7536 1.489 0.955 48 256: 73%|#######3 | 46/63 [01:30<00:42, 2.51s/it]
1/1 0G 0.7496 1.472 0.9532 58 256: 73%|#######3 | 46/63 [01:32<00:42, 2.51s/it]
1/1 0G 0.7496 1.472 0.9532 58 256: 75%|#######4 | 47/63 [01:32<00:38, 2.38s/it]
1/1 0G 0.7499 1.458 0.9536 60 256: 75%|#######4 | 47/63 [01:34<00:38, 2.38s/it]
1/1 0G 0.7499 1.458 0.9536 60 256: 76%|#######6 | 48/63 [01:34<00:34, 2.28s/it]
1/1 0G 0.7497 1.442 0.9519 60 256: 76%|#######6 | 48/63 [01:36<00:34, 2.28s/it]
1/1 0G 0.7497 1.442 0.9519 60 256: 78%|#######7 | 49/63 [01:36<00:31, 2.24s/it]
1/1 0G 0.7487 1.431 0.9503 57 256: 78%|#######7 | 49/63 [01:38<00:31, 2.24s/it]
1/1 0G 0.7487 1.431 0.9503 57 256: 79%|#######9 | 50/63 [01:38<00:28, 2.19s/it]
1/1 0G 0.7491 1.417 0.9517 49 256: 79%|#######9 | 50/63 [01:40<00:28, 2.19s/it]
1/1 0G 0.7491 1.417 0.9517 49 256: 81%|######## | 51/63 [01:40<00:25, 2.11s/it]
1/1 0G 0.7476 1.407 0.9512 62 256: 81%|######## | 51/63 [01:42<00:25, 2.11s/it]
1/1 0G 0.7476 1.407 0.9512 62 256: 83%|########2 | 52/63 [01:42<00:23, 2.09s/it]
1/1 0G 0.7443 1.391 0.9499 54 256: 83%|########2 | 52/63 [01:44<00:23, 2.09s/it]
1/1 0G 0.7443 1.391 0.9499 54 256: 84%|########4 | 53/63 [01:44<00:20, 2.04s/it]
1/1 0G 0.7434 1.377 0.9505 65 256: 84%|########4 | 53/63 [01:46<00:20, 2.04s/it]
1/1 0G 0.7434 1.377 0.9505 65 256: 86%|########5 | 54/63 [01:46<00:18, 2.07s/it]
1/1 0G 0.7416 1.365 0.9503 60 256: 86%|########5 | 54/63 [01:49<00:18, 2.07s/it]
1/1 0G 0.7416 1.365 0.9503 60 256: 87%|########7 | 55/63 [01:49<00:16, 2.07s/it]
1/1 0G 0.7406 1.354 0.9508 61 256: 87%|########7 | 55/63 [01:51<00:16, 2.07s/it]
1/1 0G 0.7406 1.354 0.9508 61 256: 89%|########8 | 56/63 [01:51<00:14, 2.11s/it]
1/1 0G 0.7386 1.341 0.9497 66 256: 89%|########8 | 56/63 [01:53<00:14, 2.11s/it]
1/1 0G 0.7386 1.341 0.9497 66 256: 90%|######### | 57/63 [01:53<00:12, 2.07s/it]
1/1 0G 0.737 1.33 0.9496 57 256: 90%|######### | 57/63 [01:55<00:12, 2.07s/it]
1/1 0G 0.737 1.33 0.9496 57 256: 92%|#########2| 58/63 [01:55<00:10, 2.09s/it]
1/1 0G 0.7359 1.32 0.9494 79 256: 92%|#########2| 58/63 [01:57<00:10, 2.09s/it]
1/1 0G 0.7359 1.32 0.9494 79 256: 94%|#########3| 59/63 [01:57<00:08, 2.08s/it]
1/1 0G 0.7359 1.309 0.9487 83 256: 94%|#########3| 59/63 [01:59<00:08, 2.08s/it]
1/1 0G 0.7359 1.309 0.9487 83 256: 95%|#########5| 60/63 [01:59<00:06, 2.19s/it]
1/1 0G 0.7341 1.3 0.9478 76 256: 95%|#########5| 60/63 [02:01<00:06, 2.19s/it]
1/1 0G 0.7341 1.3 0.9478 76 256: 97%|#########6| 61/63 [02:01<00:04, 2.18s/it]
1/1 0G 0.732 1.288 0.9469 64 256: 97%|#########6| 61/63 [02:04<00:04, 2.18s/it]
1/1 0G 0.732 1.288 0.9469 64 256: 98%|#########8| 62/63 [02:04<00:02, 2.20s/it]
1/1 0G 0.7312 1.277 0.946 70 256: 98%|#########8| 62/63 [02:06<00:02, 2.20s/it]
1/1 0G 0.7312 1.277 0.946 70 256: 100%|##########| 63/63 [02:06<00:00, 2.12s/it]
1/1 0G 0.7312 1.277 0.946 70 256: 100%|##########| 63/63 [02:06<00:00, 2.00s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 0%| | 0/5 [00:00<?, ?it/s]
Class Images Instances Box(P R mAP50 mAP50-95): 20%|## | 1/5 [00:03<00:14, 3.68s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 40%|#### | 2/5 [00:06<00:09, 3.17s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 60%|###### | 3/5 [00:09<00:06, 3.04s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 80%|######## | 4/5 [00:12<00:02, 2.98s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 100%|##########| 5/5 [00:13<00:00, 2.40s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 100%|##########| 5/5 [00:13<00:00, 2.73s/it]
all 144 426 0.856 0.824 0.905 0.722
1 epochs completed in 0.041 hours.
Optimizer stripped from C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\weights\last.pt, 22.5MB
Optimizer stripped from C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\weights\best.pt, 22.5MB
Validating C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\weights\best.pt...
Ultralytics YOLOv8.2.66 Python-3.11.5 torch-2.7.0+cpu CPU (11th Gen Intel Core(TM) i5-1135G7 2.40GHz)
YOLOv8s summary (fused): 168 layers, 11,126,745 parameters, 0 gradients, 28.4 GFLOPs
Class Images Instances Box(P R mAP50 mAP50-95): 0%| | 0/5 [00:00<?, ?it/s]
Class Images Instances Box(P R mAP50 mAP50-95): 20%|## | 1/5 [00:02<00:09, 2.27s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 40%|#### | 2/5 [00:04<00:06, 2.22s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 60%|###### | 3/5 [00:06<00:04, 2.27s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 80%|######## | 4/5 [00:08<00:02, 2.20s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 100%|##########| 5/5 [00:10<00:00, 1.82s/it]
Class Images Instances Box(P R mAP50 mAP50-95): 100%|##########| 5/5 [00:10<00:00, 2.01s/it]
all 144 426 0.856 0.823 0.905 0.722
apple 69 153 0.792 0.944 0.933 0.829
orange 51 84 0.947 0.833 0.948 0.777
pear 72 189 0.829 0.692 0.834 0.561
Speed: 0.5ms preprocess, 58.1ms inference, 0.0ms loss, 1.0ms postprocess per image
Results saved to □[1mC:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first□[0m
Ultralytics YOLOv8.2.66 Python-3.11.5 torch-2.7.0+cpu CPU (11th Gen Intel Core(TM) i5-1135G7 2.40GHz)
YOLOv8s summary (fused): 168 layers, 11,126,745 parameters, 0 gradients, 28.4 GFLOPs
□[34m□[1mPyTorch:□[0m starting from 'C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\weights\best.pt' with input shape (1, 3, 256, 256) BCHW and output shape(s) (1, 7, 1344) (21.4 MB)
□[34m□[1mONNX:□[0m starting export with onnx 1.16.1 opset 14...
□[34m□[1mONNX:□[0m export success ✅ 1.2s, saved as 'C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\weights\best.onnx' (42.5 MB)
Export complete (3.3s)
Results saved to □[1mC:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\weights□[0m
Predict: yolo predict task=detect model=C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\weights\best.onnx imgsz=256
Validate: yolo val task=detect model=C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\runs\detect\first\weights\best.onnx imgsz=256 data=C:\Users\Noel Nathan\Desktop\Universidad\8tQuadrimestre\tfg\from-code-to-robot\computer_vision\datasets\fruits_3_4998\data.yaml
Visualize: https://netron.app
Continuar un entrenament
Si el procés d’entrenament s’interromp per qualsevol motiu, es pot reprendre utilitzant l’opció resume. Aquesta funcionalitat carrega els pesos de l’últim model desat i també restaura l’estat de l’optimitzador, el scheduler de la taxa d’aprenentatge i el número d’època. Això permet continuar el procés d’entrenament de manera fluida des del punt on s’havia deixat.
model_folder = fullfile(pwd, 'runs', 'detect', 'train3', 'weights', 'best.pt');
yolov8Det = trainYOLOv8ObjectDetector( ...
configFile, ...
model_folder, ...
options, ...
'resume', true ...
);
Per a més detalls, consulta la secció Resume Training de la documentació oficial.
Reentrenar un model
Alternativament, pots continuar entrenant un model ja entrenat utilitzant-lo com a punt de partida en comptes d’un model base com yolov8n.pt o yolov8s.pt. En aquest cas, simplement carregues els pesos del model entrenat prèviament.
Tanmateix, a diferència de l’opció resume, aquest enfocament no restaura l’estat de l’optimitzador, la planificació de la taxa d’aprenentatge ni el número d’època. Aquests paràmetres es reinicien, i l’entrenament comença des de l’època 1 amb les opcions d’entrenament actuals.
Aquest mètode és útil quan vols fine-tune encara més un model o adaptar-lo a un nou dataset, però sense continuar exactament des del punt on va quedar l’entrenament anterior.
model_folder = fullfile(pwd, 'runs', 'detect', 'train1', 'weights', 'best.pt');
yolov8Det = trainYOLOv8ObjectDetector( ...
configFile, ...
baseModel, ...
options
optionalArgs{:});
Carregar un model
Quan s’entrena un model YOLO, els pesos entrenats normalment es desen al directori runs/detect/train@/weights, on @ és un comptador incremental, per exemple train1, train2, etc., per evitar sobreescriure execucions anteriors.
Si aquesta carpeta no es crea, és probable que la configuració de YOLO no s’hagi configurat correctament durant el pas 1. Installation.md. Per resoldre-ho, torna a executar aquesta secció i assegura’t que tot estigui configurat correctament.
Dins de la carpeta weights, trobaràs diversos fitxers de model, incloent-hi last.pt, best.pt i best.onnx. La funció utils.loadModel sempre carrega el model best.onnx de la carpeta seleccionada.
Per tant, en el codi següent, has de carregar el model del primer entrenament:
modelPath = fullfile(pwd, 'runs', 'detect', 'first', 'weights');
configFile = fullfile(pwd, 'datasets', 'fruits_3_4998', 'data.yaml');
yolov8Det = utils.loadModel(modelPath, configFile);
Predicció
En aquesta secció, utilitzem el model que hem entrenat anteriorment per detectar fruites en les imatges següents.
utils.detectAndDisplayImage(yolov8Det, 'pear.jpg');
utils.detectAndDisplayImage(yolov8Det, 'apple.jpg');
utils.detectAndDisplayImage(yolov8Det, 'orange.jpg');
Com s’observa en els resultats, el model produeix un nombre excessiu de deteccions, cosa que dona lloc a una sortida de baixa precisió.
Tractarem aquest tema en detall a la pròxima lliçó.