2.Training models using transfer learning
In this lesson, we will learn how to train an object detection model using transfer learning. These techniques reduce computational effort and improve model performance, even with small datasets.
First, we add the help module:
addpath('help-module');
Training
Required YOLO Parameters
The following parameters are mandatory when configuring a YOLO training session:
- configFile (str) : Path to the configuration file (usually in
.yamlformat) that defines the dataset structure and classes. This file should include paths to training, validation, and optional test sets, as well as the list of class names. As an example the file we will use in this tutorial is data.yaml which already contains all the defined structure. - baseModel (str): This parameter specifies which model will be used for training. It allows you to either continue training a previously trained model or start training from a pre-trained YOLO model.You can choose among different YOLO versions, tasks (e.g., object detection, segmentation, classification), and model sizes. For these tutorials, we will use YOLOv8 for object detection. The available pre-trained models, ordered from smallest to largest, are: yolov8n.pt, yolov8s.pt, yolov8m.pt, yolov8l.pt, yolov8x.pt. It is recommended in these lessons to use smaller models (e.g.,
yolov8n.ptoryolov8s.pt) for faster training and inference speed. MaxEpochs(int): This parametersets the maximum number of epochs for training, where an epoch is one complete pass over all the images in the training dataset. Increasing this value usually improves model performance but also extends training time.ImageSize(list[int]):The size (in pixels) that all images will be resized to before training. Larger sizes may improve accuracy but require more computational resources.
In the following code, we define the values for these required parameters. Use these parameters to speed up execution:
configFile = fullfile(pwd, 'datasets', 'fruits_3_4998', 'data.yaml');
baseModel = 'yolov8s-oiv7.pt';
options.MaxEpochs = 1;
options.ImageSize = [256 256 3];
YOLO layers visualization
The following code allows you to visualize the architecture and parameters of a YOLOv8 model:
det = yolov8ObjectDetector2('yolov8s');
Pretrained yolov8s network already exists.
% Analyze loaded model
analyzeNetwork(det.Network);
This command opens a new window displaying all the layers and parameters of the YOLO model being used. Keep in mind that different YOLOv8 variants (e.g., yolov8s, yolov8m, etc.) differ not only in the number of parameters but also slightly in their architectural complexity.
In the case of YOLOv8, the model typically consists of 23 main blocks. These can be grouped into three main components:
- Backbone (first 10 blocks): Responsible for extracting features from the input image.
- Neck (next 12 blocks): Combines features at different scales to enhance detection performance.
- Head (last block): Responsible for predicting bounding boxes, class labels, and confidence scores.
Freezing Layer Parameter for Transfer Learning
One of the main advantages of YOLO models is that they come pre-trained. Training an object detection model from scratch usually requires millions of images and considerable computational power. This problem can be solved using transfer learning, which involves taking a model that has already been trained on a large dataset and slightly retraining it using a much smaller set of new data.
Transfer learning is particularly effective in computer vision because convolutional neural networks (CNNs), which are the foundation of YOLO (and all object detection models), learn features in a hierarchical manner. The initial layers detect low-level features such as edges and textures, while deeper layers capture more complex patterns and object structures. This hierarchy makes the early layers highly reusable across different tasks.
Typically, the early layers responsible for feature extraction (the backbone) are frozen during training, meaning their weights are not updated. This allows the model to retain the general visual features learned from large datasets. Meanwhile, the later layers are fine-tuned on the new dataset to adapt to the specific task at hand.
In some scenarios, the entire model may be fine-tuned using a lower learning rate, allowing it to gradually adapt to the new data while minimizing the risk of overwriting valuable pre-trained knowledge.
This is why we use YOLOv8. It comes pre-trained on large datasets such as COCO (e.g., yolov8n.pt, yolov8s.pt) or Open Image V7 (e.g., yolov8n-oiv7.pt, yolov8s-oiv7.pt). Using these models significantly reduces training time and improves performance, even with smaller datasets.
Choosing which layers to freeze
The choice of how many YOLOv8 layers to freeze depends on how similar your target dataset is to the pretrained data. YOLOv8 has 22 freeze-able layers: the first 10 extract generic features (edges, textures, shapes), while later layers specialize in detection. Since our classes (apple, orange, pear) are well represented in COCO and Open Images V7, the domains closely match. This allows us to freeze most layers to keep general features intact and fine-tune only the higher-level layers for our fruits, which speeds training and lowers overfitting risk.
To freeze specific layers during training, we use the freeze parameter as follows:
optionalArgs = {'freeze', 15};
There are additional parameters available to configure the model, which will be explained in Lesson 4: Experimentation.
Starting training
Next, we train the YOLO model with the parameters we defined earlier. This may take approximately 2 minutes.
[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
Continue a train
If the training process is interrupted for any reason, it can be resumed using the resume option. This feature loads the weights from the last saved model and also restores the optimizer state, learning rate scheduler, and epoch number. This allows you to continue the training process seamlessly from where it was left off.
model_folder = fullfile(pwd, 'runs', 'detect', 'train3', 'weights', 'best.pt');
yolov8Det = trainYOLOv8ObjectDetector( ...
configFile, ...
model_folder, ...
options, ...
'resume', true ...
);
For more details, refer to the Resume Training section in the official documentation.
Retrain a model
Alternatively, you can continue training an already trained model by using it as a starting point instead of a base model like yolov8n.pt or yolov8s.pt. In this case, you simply load the weights from the previously trained model.
However, unlike the resume option, this approach does not restore the optimizer state, learning rate schedule, or epoch number. These parameters are reset, and training starts from epoch 1 with the current training options.
This method is useful when you want to fine-tune a model further or adapt it to a new dataset, but without continuing exactly where the previous training left off.
model_folder = fullfile(pwd, 'runs', 'detect', 'train1', 'weights', 'best.pt');
yolov8Det = trainYOLOv8ObjectDetector( ...
configFile, ...
baseModel, ...
options
optionalArgs{:});
Load a model
When training a YOLO model, the trained weights are typically saved in the runs/detect/train@/weights directory, where @ is an incrementing counter (e.g., train1, train2, etc.) to avoid overwriting previous runs.
If this folder is not created, it is likely that the YOLO settings were not properly configured during the 1. Installation.md step. To resolve this, re-run that section and ensure everything is correctly set up.
Inside the weights folder, you’ll find several model files, including last.pt, best.pt, and best.onnx. The function utils.loadModel always loads the best.onnx model from the selected folder.
Therefore, in the following code, you should load the model from the first training run:
modelPath = fullfile(pwd, 'runs', 'detect', 'first', 'weights');
configFile = fullfile(pwd, 'datasets', 'fruits_3_4998', 'data.yaml');
yolov8Det = utils.loadModel(modelPath, configFile);
Prediction
In this section, we use the model we trained earlier to detect fruits in the following images.
utils.detectAndDisplayImage(yolov8Det, 'pear.jpg');
utils.detectAndDisplayImage(yolov8Det, 'apple.jpg');
utils.detectAndDisplayImage(yolov8Det, 'orange.jpg');
As observed in the results, the model produces an excessive number of detections, leading to a low-precision output.
We will discuss this topic in detail in the next lesson.