20 lines
1.6 KiB
Markdown
20 lines
1.6 KiB
Markdown
great. realtime.py works. now let's examine and contunue with our 500m NN in a NN folder with different modules - first module will be around 100m Convolutional NN that is historically used for image recognition with great success by detecting features on multiple levels - deep NN. create the NN class and integrated RL pipeline that will use historical data to retrospectively identify buy/sell opportunities and use that to train the module. use the data from realtime.py (add easy to use realtime data interface if existing functions are not convenient enough)
|
|
create a new main file in the NN folder for our new MoE model. we'll use one main NN module that will orchestrate data flows. our CNN module should have training and inference pipelines implemented internally, but the orchestrator will get the realtime data and forward it. use a common interface. another module later will be Transformer module that will take as input raw data from the latest hidden layers of the CNN where high end features are learned as well as the output, which will be BUY/HOLD/SELL signals as well as key support/resistance trend lines
|
|
|
|
|
|
# setup:
|
|
setup_env.bat
|
|
|
|
# run:
|
|
python -m NN.start_tensorboard --logdir=NN/models/saved/logs
|
|
# and
|
|
run_nn.py / run_nn_in_conda.bat / run_pytorch_nn.bat
|
|
|
|
# Train a CNN model
|
|
python -m NN.main --mode train --symbol BTC/USDT --timeframes 1h 4h --model-type cnn --epochs 100
|
|
|
|
# Make predictions with a trained model
|
|
python -m NN.main --mode predict --symbol BTC/USDT --timeframe 1h --model-type cnn
|
|
|
|
# Run real-time analysis
|
|
python -m NN.main --mode realtime --symbol BTC/USDT --timeframe 1h --inference-interval 60 |