popov/gogo2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

multi pair inputs, wip, working training??

Browse Source
This commit is contained in:
Dobromir Popov
2025-03-29 01:56:45 +02:00
parent b313f70cd7
commit 0b2000e3e7
9 changed files with 686 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
_notes.md
View File

@@ -10,7 +10,7 @@ python main.py --mode live --symbol BTC/USDT --timeframe 1m --use-websocket --da
& 'C:\Users\popov\miniforge3\python.exe' 'c:\Users\popov\.cursor\extensions\ms-python.debugpy-2024.6.0-win32-x64\bundled\libs\debugpy\adapter/../..\debugpy\launcher' '51766' '--' 'main.py' '--mode' 'live' '--demo' 'false' '--symbol' 'ETH/USDT' '--timeframe' '1m' '--leverage' '50'
python -c "import sys; sys.path.append('f:/projects/gogo2'); from NN.realtime_main import main; main()" --mode train --model-type cnn --framework pytorch
ensure we use GPU if available to train faster. during training we need to have RL loop that looks at streaming data, and retrospective backtesting/training on predictions. sincr the start of the traing we're only loosing. implement robust penalty and analysis when closing a loosing trade and improve the reward function.
@@ -29,4 +29,11 @@ C:\Users\popov\miniforge3\Lib\site-packages\torch\amp\grad_scaler.py:132: UserWa
2025-03-10 12:11:30,927 - INFO - Starting training on device: cpu
2025-03-10 12:11:30,928 - ERROR - Training failed: 'TradingEnvironment' object has no attribute 'initialize_price_predictor'
2025-03-10 12:11:30,928 - INFO - Exchange connection closed
Backend tkagg is interactive backend. Turning interactive mode on.
Backend tkagg is interactive backend. Turning interactive mode on.
remodel our NN architecture. we should support up to 3 pairs simultaniously. so input can be 3 pairs: each pair will have up to 5 timeframes 1s(ticks, unspecified length), 1m, 1h, 1d + one additionall. we should normalize them in a way that preserves the relations between them (one price should be normalized to the same value across all tieframes). additionally to the 5 features OHLCV we will add up to 20 additional features for various technical indcators. 1s timeframe will be streamed in realtime. the MOE model should handle all that. we still need to access latest of the CNN hidden layers in the MOe model so we can extract learned features recognition
.
now let's run our "NN Training Pipeline" debug config. for now we start with single pair - BTC/USD. later we'll add up to 3 pairs for context. the NN will always have only 1 "main" pair - where the buy/sell actions are applied and which price prediction is calculater for each frame. we'll also try to predict the next local extrema that will help us be profitable