This project is a about sending good night messages just before your loved one(s) goes to sleep, by predicting with PyTorch the time they will go to sleep based on the Telegram accesses timestamps and/or last seen time.
I started this for fun some time ago. After moving to Korea, I and a girl I had a (sort of) relationship with decided to just be friends due to the distance and not to send each other sweet things like the good night message. Since I didn't want to break the deal, I thought: "Why not let an AI send the good night message? She will receive it and I technically wasn't the one sending it!" So, here is how the idea was born. Nothing too complex, just an algorithm collecting the access timestamps, detecting the time when she goes to sleep and wake up, make some inferences with machine learning about tomorrow's time to go to bed, send a message just before that and... voilà! As you can imagine, it took me some time to figure out all the details of this project, but at the end it is kind of working. Was it something really useful? Probably not. But who cares, it was fun! 😄
This project could be used to track the usage patterns of users without consent: therefore, even though this project is for fun, use at your own discretion and most importantly, citing Butcher in The Boys: "Don't be a c*nt!"
The following .py executables can be either used as daemons or executed normally. I recommend trying to first executing them with your account or someone you know (you don't want to send random messages around, do you?) and see how they work out; and finally converting them to system daemons so that they can be completely automatic.
The variable good_nighter indicates the person you want to send the good night message to, you can change it data/good_nighter.txt. Remember to use the correct naming according to telegram, with underscores _ indicating a space (i.e. Federico_Berto).
Collects all the accesses and saves them in the file data/LastSeenDataset.csv. The following are variables to be tweaked:
polling_time = 30.0 # Send a request to the Telegram server for the status
telegram_path = "/usr/bin/telegram-cli" # telegram-cli installation
pubkey_path = "/home/[YOUR_USER]/tg/server.pub"
Reads the data and reorganizes creates a with the following features to train the neural network with src/data_miner.py:
- Last seen time (arguably the most important)
- Wake up time
- Number of Telegram accesses during the previous day
- Public holiday presence in the following day (using the holidays library)
- Day of the week
Then, we use a PyTorch model to train the model for predicting the next bedtime, such as:
- MPL: Multi Layer Perceptron, simple deep neural network with hidden layer
- RNN: Recurrent Neural Network, more suitable for time series
- LSTM: Long Short Term Memory, an advancement of RNN
- Transformer: currently (2020) state of the art, but OP for this task
- ... Other
Currently, we are using by default the MLP. You may want to try and make a different model work.
After the training, the executable saves the predicted time under data/prediction.txt.
Variables to tweak:
num_features = 5 # You may want to improve the dataset for extra features
min_hour = 21 # Minimum hour for sleep detection
max_hour = 5 # Maximum hour for sleep detection
train_window = 3 # Sequence length of past days
local_holidays = holidays.Italy(prov='BO') # Get the holidays in your local area ^^
EPOCHS = 500 # Training epochs
batch_size = 16 # Training batch size
You can also give a look at the models and change them for better performance or to reduce the overfitting (you can also tweak the GoodNightDataset.noisy() function for this purpose).
Note that the implementation is quite rough and does not include a validation set, given the small size of the data, so we can just use our intuition to see if the model is performing well on new data. Besides, I still have a long way to go in Machine Learning (yeah, I'm a noob) but I'm here to learn and your suggestions and contributions to this project will be highly appreciated 😉
Reads the data/prediction.txt file and sends a message before the time to go to bed, hoping that our model is good enough so that the person reads the message at the right time.
The messages are stored under data/messages.txt and you should modify them with your custom ones!
Variables to tweak:
polling_time = 300 # seconds to between reading of the file
advance_time = 30 # Minutes we send the message in advance with respect to the prediction, so that the good nighter reads themessage
telegram_path = "/usr/bin/telegram-cli"
pubkey_path = "/home/[YOUR_USERNAME]/tg/server.pub"
Clone this repo
git clone https://github.com/Juju-botu/good-night-ml.git
cd good-night-ml
Install the required packages; here is an incomplete list of the most important ones
- torch (see guide below for ARM version)
- pytg
- numpy
- json
- datetime
- csv
- holidays
In order to add data_collector.py and sender.py as system daemons, first create a file under /etc/systemd/system (the following example is for the data_collector.py, then change the names for sender.py):
sudo touch data-collector@goodnightml.service
Then copy and paste the following inside changing the parameters:
sudo nano data-collector@goodnightml.service
[Unit]
Description=Data collector for good-night-ml
[Service]
Type=simple
User=[YOUR_USER]
WorkingDirectory=[INSTALLATION_PATH]/good-night-ml
Restart=always
ExecStart=/usr/bin/python3 data_collector.py
[Install]
WantedBy=multi-user.target
Try first if the script is working by starting it (if you are testing, remember to change the good_nighter!):
sudo systemctl start data-collector@goodnightml.service
sudo systemctl status data-collector@goodnightml.service
If the status is active, it is working then stop the service
sudo systemctl stop data-collector@goodnightml.service
Enable the service to start automatically at reboot via the following command:
sudo systemctl enable data-collector@goodnightml.service
Now you can easily check the service status by doing:
sudo service data-collector@goodnightml.service status/start/stop
The predictor.py does not need to always run in the background, so we can add it to crontab and run it as a periodic task: we just need one prediction (or set of predictions, if we are tracking multiple users) about the next time to go to sleep for each day:
sudo crontab -e
Add the following at the end of the file with your time (you may follow this guide as an easy reference)
0 18 * * * /usr/bin/python3 [YOUR_PATH]/good-night-ml/predictor.py
Installing PyTorch on my Raspberry Pi 4B was not that easy for me, so I decided to include the instructions for installing it below with some comments. You can find the original Stackoverflow post and credits go to that writer.
First of all, if you have a Raspberry PI 3 or lower, you need to increase the SWAP, since the build is a RAM eater.
If you have a RBPi 4 or higher with more than 3GB of RAM, skip this step.
Modify the file /etc/dphys-swapfile :
CONF_SWAPFILE=2048M
Then call the following command to update changes.
sudo dphys-swapfile setup
Install the following packages:
sudo apt install build-essential make cmake git python3-pip libatlas-base-dev
Libtorch needs CMake>=3.15 to be built properly, check cmake version with cmake --version
If it's lower than 3.15, follow the following commands to build a newer version and remove the previous one:
wget https://github.com/Kitware/CMake/releases/download/v3.18.0-rc1/cmake-3.18.0-rc1.tar.gz
tar -xzf cmake-3.18.0-rc1.tar.gz
cd cmake<version>
mkdir build
cd build
cmake ..
make
sudo make install
sudo apt remove cmake
sudo ln -s /usr/local/bin/cmake /usr/bin/cmake
sudo ldconfig
Don't forget to increase the SWAP to 2048M if you don't have 3GB or RAM.
Getting all needed libraries:
sudo apt-get update
sudo apt-get install build-essential tk-dev libncurses5-dev libncursesw5-dev libreadline6-dev libdb5.3-dev libgdbm-dev libsqlite3-dev libssl-dev libbz2-dev libexpat1-dev liblzma-dev zlib1g-dev
Getting PyTorch sources (change --branch=release/x.y according to the version you want to install):
git clone --recursive https://github.com/pytorch/pytorch --branch=release/1.7
cd pytorch
Init all the submodules :
git submodule update --init --recursive
git submodule update --remote third_party/protobuf # To prevent a bug I had
Getting all needed libraries:
sudo apt-get update
sudo apt-get install build-essential tk-dev libncurses5-dev libncursesw5-dev libreadline6-dev libdb5.3-dev libgdbm-dev libsqlite3-dev libssl-dev libbz2-dev libexpat1-dev liblzma-dev zlib1g-dev
Setting up environment variables for the build.
Add the following lines to the ~/.bashrc file.
export NO_CUDA=1
export NO_DISTRIBUTED=1
export NO_MKLDNN=1
export NO_NNPACK=1
export NO_QNNPACK=1
Log in as root, and use the .bashrc file to setup the environment variables
sudo su
source /home/<user>/.bashrc
Install python dependencies
pip3 install setuptools pyyaml numpy
Build and install PyTorch, time to grab a ☕, it make take a while (took ~ 2 hours on my RPi 4B 4GB ram!)
Don't forget the -E that forces the environment variables to be used.
sudo -E python3 setup.py install
Check the installation worked:
cd
python3
import torch
torch.__version__
In your CMakeLists.txt :
cmake_minimum_required(VERSION 2.6)
project(projectName)
set(CMAKE_PREFIX_PATH "/home/pi/pytorch/torch") # Adding the directory where torch as been installed
set(CMAKE_CXX_STANDARD 14) # C14 required to compile Torch
set(CMAKE_CXX_STANDARD_REQUIRED TRUE)
add_compile_definitions(_GLIBCXX_USE_CXX11_ABI=0) # Torch is compiled with CXX11_ABI, so your program needs to be also, or you may have conflicts in some libraries (such as GTest for example)
# Specifying we are using pthread for UNIX systems.
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS} -pthread -Wall")
find_package(Torch REQUIRED)
if(NOT Torch_FOUND)
message(FATAL_ERROR "Pytorch Not Found!")
endif(NOT Torch_FOUND)
message(STATUS "Pytorch status :")
message(STATUS " libraries: ${TORCH_LIBRARIES}")
message(STATUS " Torch Flags: ${TORCH_CXX_FLAGS}")
# Program executable
add_executable(projectName <sources>)
target_link_libraries(projectName PRIVATE pthread dl util ${TORCH_LIBRARIES})