har_trees: Update README with instructions on custom data

Author: jonnor

examples/har_trees/README.md

@@ -11,7 +11,7 @@ The same kind of approach can also be applied to animals (*Animal Activity Recog
 which enables tracking of pets, lifestock and wild animals.
 This has been used for many kinds of animals - such as cats, dogs, diary cattle.
 
-The same approach can be used for simple gesture recognition.
+The same approach can be used for simple gesture recognition, at least for repetitive gestures.
 
 ## Status
 Working. Tested running on ESP32 with MicroPython 1.24.
@@ -22,6 +22,9 @@ The dataset used is rather simple, and may not reflect the data you get from you
 - which will lead to poor classifications.
 For a real world usage you should probably replace the dataset with your own data, collected on your own device.
 
+At the bottom on the README there are some instructions and tools for collecting your own data,
+and training a custom model on such a dataset.
+
 ## Machine Learning pipeline
 
 This example uses an approach based on the paper
@@ -118,8 +121,8 @@ mpremote har_live.py
 
 ## Run training
 
-This will train a new model.
-Uses CPython on the PC.
+This will train a new model for the HAR UCI dataset.
+You need to have Python (CPython) installed on the PC.
 
 Install requirements
 ```
@@ -139,4 +142,85 @@ python har_train.py
 This will output a new model (named `_trees.csv`).
 Can then be deployed to device following the steps above.
 
+## Recording motion data for custom tasks
+
+To learn a classifier for your own custom tasks, you will need to: 1) record data, 2) label the data, 3) run training with custom data.
+This example provides some basic tools to assist with this process.
+
+Recording data. Requires a M5StickC PLUS 2.
+Before you do this, make sure to **first run live classification example** (to get the dependencies).
+
+```
+mpremote run har_record.py
+```
+
+Alternatively: Copy the program to device.
+This way it will run even if there is no USB device connected.
+```
+mpremote cp har_record.py main.py
+mpremote reset
+```
+
+To get good timestamps on the recorded files, rembember to set the RTC clock.
+```
+mpremote rtc --set
+```
+
+When the recording program runs, the device should show a screen which allows to select between different classes.
+Clicking the big button by the screen allows selecting class.
+Holding the big button down allows to start recording.
+The red LED will light up while recording.
+
+This allows to coarsely label the classes, which is helpful to sort and annotate them later.
+To adjust the classes, make changes to `har_record.py`.
+
+The files are placed in the `data/` folder on the internal FLASH filesystem.
+
+To copy the files over to your computer, for building a dataset, use
+```
+mkdir ./data/raw/mydata1/
+mpremote cp -r :./har_record ./data/raw/mydata1/
+```
+
+## Labeling recorded motion data
+
+To label the motion data, we can use [Label Studio](https://labelstud.io/).
+You can use their hosted version, or run it in Docker, or install it via `pip`.
+
+There is a ready-made Label Studio task defintion provided, 
+in the file `labeling/har_classify_config.xml`.
+
+There is a tool designed to convert the .npy files from the `har_record.py` application
+into .csv files understood by Label Studio.
+
+```
+python har_data2labelstudio.py
+```
+
+When you have started Label Studio, create a new Project, and Import the .CSV files as data.
+Then you can label each piece of data.
+
+
+Then use this tool to combine the sensor data files with the labels.
+```
+python har_labelstudio2dataset.py
+```
+
+This will produce a dataset as a .parquet file, which can be used with `har_train.py`.
+
+
+## Train model on custom dataset
+
+
+Add your dataset definition to `dataset_config` in `har_train.py`, with a unique name (example `mydata1`).
+
+Then you can run the training process.
+
+```
+python har_train.py --dataset mydata1
+```
+
+It should output a new model (named `_trees.csv`).
+This model can be deployed to device following the steps above.
+