Module 5 workflow tool
This page turns the hardhat detection notebook into a guided learning workflow. It helps you understand the dataset, labels, training process, and testing flow here in the browser, while the real YOLO training and detection still happen in Colab.
Workflow stages
The underlying model and dataset stay the same as the notebook. The difference is that each stage is explained clearly before you ever need to look at the code.
This module uses YOLOv5, a deep learning object detection model. It does more than classify an image: it finds object locations and predicts labels such as helmet and head.
YOLOv5 is practical for teaching because it is widely used, supports pretrained weights, and makes it easier to show the full pipeline from annotation to training, validation, and testing.
In traditional machine learning, people usually choose the features first. In deep learning, the model learns the useful visual features directly from many labeled images.
best.pt or last.pt1
Use the Google Drive file ID to bring the dataset into Colab.
2
Split images into training, validation, and testing folders.
3
Change XML annotations into YOLO text labels.
4
Use a pretrained YOLO model to learn hardhat detection.
5
Monitor loss, precision, recall, and mAP during training.
6
Run inference on unseen images and inspect predictions.
Data setup
This step is for preparing the dataset, choosing the split, and generating the Colab cells. The explanation stays visible, and the code is grouped into smaller blocks so it is easier to follow.
These settings update the generated Colab cells below. If you change the file ID, split ratios, epochs, batch size, image size, or experiment name, the code blocks are rewritten automatically.
These code blocks update live when you change the workflow settings on the left.
This cell downloads and unzips the dataset in Colab.
This cell organizes files and converts XML labels into YOLO text labels.
This cell writes the YAML file that tells YOLO where the data lives.
This cell installs YOLOv5 and starts model training.
This cell checks model quality and runs inference on unseen images.
Interactive preview
The main workflow uses Google Drive and Colab. This section is an optional visual aid for understanding dataset structure, annotations, labels, and the train-validation-test split.
0
Files
0
Images
0
Labels/XML
Optionally upload a folder to generate a dataset tree.
No folder loaded yet.
80%
Train
10%
Validation
10%
Test
The dataset will be divided into train, validation, and test subsets.
Upload a sample image to preview the workflow visually.
Upload an annotation file to preview object names and coordinates.Upload both an image and XML file to generate YOLO labels.Model training
This final step explains what the model is doing while it learns, what the main metrics mean, and how the training curves usually behave over time.
dataset/
images/
annotations/
train/images
train/labels
val/images
val/labels
test/images
test/labelsValidation checks model quality during training. It helps us compare learning progress without using the final test images.
Testing happens after training. It uses unseen images so you can judge how well the model generalizes.
Training process
During training, the model repeatedly compares its predictions to the labeled training data, calculates error, updates its weights, and then checks progress on the validation set.
A
A small group of training images and labels is loaded into the model.
B
The model predicts object boxes, labels, and confidence values.
C
The prediction is compared with the true annotation to measure error.
D
The optimizer changes model parameters to reduce error next time.
E
The model is checked on validation data to monitor generalization.
F
This cycle repeats over many epochs until performance improves enough.
One epoch means the model has seen the whole training dataset once. If training uses 10 epochs, the model goes through the training data 10 times.
Batch size tells YOLO how many images to process at one time before updating the model weights. Larger batches use more memory; smaller batches are lighter but may train more slowly.
Image size is the resolution YOLO uses during training and testing. Larger sizes may preserve more detail, but they also require more time and memory.
Used directly to teach the model from labeled examples.
Used during training to check whether the model is improving without overfitting.
Used only after training to estimate final performance on unseen data.
After training
When training ends, YOLOv5 saves the learned model and the training history. These output files are what you use later for validation, testing, and inference on new images.
best.ptThis is the most important output file. It stores the version of the model that performed best on the validation set during training.
last.ptThis stores the model from the final training epoch. It is useful if you want to resume training or compare the final epoch to the best validation checkpoint.
YOLOv5 also saves charts and metrics such as loss, precision, recall, and mAP so you can inspect how learning changed across epochs.
runs/train/yolo_hardhat_exp/
weights/best.pt
weights/last.pt
results.png
results.csv
confusion_matrix.pngThe exact folder name depends on the experiment name you choose during training.
Every time the model sees labeled examples like this, it adjusts its internal weights so it becomes better at recognizing the visual patterns linked to helmet and head.
Model weights are the learned numbers inside the neural network. During training, YOLO keeps adjusting them so important visual patterns like helmet shape and head boundaries matter more, while unhelpful background details matter less.
Using the real Module 5 image, we can imagine the model assigning stronger learned weight values to useful clues and weaker values to less useful ones.
The model learns that rounded helmet boundaries are highly useful for the helmet class.
The model learns that head position and shape help separate head from background clutter.
These patterns can help, but usually matter less than the direct object shape.
Unhelpful background texture should get much weaker importance after training.
These values are illustrative, not extracted from the actual YOLOv5 checkpoint. They are shown here to help connect real image features to the idea of learned weights.
The model starts with generic pretrained knowledge, but it is not yet specialized for this hardhat dataset.
The weights are adjusted after each batch so predictions move closer to the labeled answers.
The saved weight files contain the learned version of the model that can now be reused for validation, testing, and detection.
Run a browser-side demo that simulates the training curves from your current settings.
0
Epoch
0.00
mAP@0.5
0.00
Precision
0.00
Recall
0.00
Train loss
0.00
Val loss
Testing
Demo only: this browser preview illustrates what testing on unseen images looks like, but it does not run the trained YOLO model. Upload a new image to see example boxes and confidence scores here, then use the Colab testing cell to generate the real detection results.
Upload a new image to begin testing.