Stack

• PEFT/Lora for training existing model from HuggingFace
• Fastapi and websocket: render IoT Signal and failure/anomaly detection
• Pytest(unit-test)
• Frontend: websocket, HTML and highchartJS index.html

Diagram

Demo

• Red line indicates the anomaly/failure detection point. I'm supposed to break it down for 2 lines rather than single one like this
• Other lines are device signals
• Data is streamed near realtime via websocket

Installation

git clone https://github.com/weburnit/llm-iot
cd llm-iot

pip install -r requirements.txt
pip install -e .

Train

python main.py --train-files=sample_data/iot_pmfp_data.feather,sample_data/iot_pmfp_labels.feather
               --metadata-file=sample_data/metadata.json
               --train-base-model=google/flan-t5-base
               --trained-new-name=iot-device
               --train=True

Start

We train the new model iot-device from google/flan-t5-base, time to start the service from there

python aitomic/main.py --train-base-model=google/flan-t5-base
               --trained-new-name=iot-device
               --train=False

Mock MQTT by using feather data

python aitomic/main.py --mqtt=true --train-files=sample_data/iot_pmfp_data.feather,sample_data/iot_pmfp_labels.feather

Training model with IOT device signal

Data point for training only relies on these features(vibration,rotate,pressure, volt, failure, anomaly)

Here is a summary of the main aspects of the Trainer class:

I have trained 30% of sample data which can be found on Google Drive - trained model

I used google/flan-t5-based

Initialization:

At the start, the Trainer is initialized with a model, tokenizer, and datasets (all set to None). There are also placeholders for LoRA configurations and paths to data files.

Data Loading and Preparation:

The load_data and prepare_dataset functions are used to load the data from specified feather files and a metadata file. The data is then prepared by splitting it into training, validation, and test sets.

Model Training:

The train function first loads the data, then prepares the model for k-bit training. The function also configures a training directory, sets up the training arguments, and creates a transformers.Trainer instance. Finally, the model is trained and the result is returned.

Model Generation:

The generate function uses the loaded model and tokenizer to generate a response to a given prompt.

From the class I have learnt about:

• How to load and prepare data for a sequence-to-sequence machine learning model.
• How to train a model using the Transformers library and additional optimization techniques by peft and lora

Missing part

• Dont't adopt accelerate to speed up training
• Can not train full data from provided sample. Google Colab has been crashed 3 times, that costs me 3 days for nothing.
• Lack of ReactJS skill to complete nicer frontend. I can understand highchart and adopt it in Pure JS