)idZddlZddlZddlmZddlmZddlZddl Z ddl m Z m Z m Z mZeej dz ZGddZy) u SPY Optimizer — Deep Learning Inference (CPU) Charge le modèle LSTM+Attention entraîné sur GPU pour l'inference en production. Ce module fonctionne SUR LE SERVEUR (CPU-only). Le modèle est entraîné sur le PC (GPU) et transféré via deploy_model.sh. Usage: from deep_inference import DeepPredictor predictor = DeepPredictor() result = predictor.predict(klines_df, timestamp_ms, surge_type="FLASH_SURGE") # → {"probability": 0.72, "signal": "BUY", "confidence": 42.0, "threshold": 0.5} N)Path)Optional)SurgePredictorbuild_sequence_featuresN_SEQUENCE_FEATURESSURGE_TYPE_MAPmodelsc zeZdZdZd deefdZdZej dddde d ed e fd Z d e fd Z y) DeepPredictoru Inference CPU pour le modèle LSTM+Attention entraîné sur GPU. Thread-safe pour utilisation dans market_spy.py. N models_dirc|r t|nt|_d|_d|_d|_i|_d|_|jy)N?<F) r MODELS_DIRr model thresholdseq_lenmetadata is_loaded_load)selfr s ?/home/ubuntu/crypto_trading_bot/spy_optimizer/deep_inference.py__init__zDeepPredictor.__init__&s>.8$z*j    c|jdz }|jdz }|jdz }|jrpt|5}tj||_ddd|j j dd|_|j j dd|_|jrj tjj t|d |_ |jjd |_td |j y|jr tj|d d}t%|j dt&|j dd|j ddd}|j)|d|j||_ |j dd|_|j dd|_d |_td|j ytd|jdy#1swYxYw#t"$r}td |Yd}~)d}~wwxYw#t"$r}td|Yd}~hd}~wwxYw)u,Charge le modèle TorchScript ou checkpoint.zsurge_predictor_gpu.ptz"surge_predictor_gpu_checkpoint.pthzsurge_predictor_gpu_meta.jsonNrrrrcpu) map_locationTu' ✅ Deep model loaded (TorchScript): u# ⚠️ TorchScript load failed: F)r weights_only n_features hidden_size num_layersr)rr r"dropoutmodel_state_dictu& ✅ Deep model loaded (checkpoint): u" ⚠️ Checkpoint load failed: u ℹ️ No GPU model found in z. Deep inference disabled.)r existsopenjsonloadrgetrrtorchjitstrrevalrprintname Exceptionrrload_state_dict)rts_path ckpt_path meta_pathfeckptrs rrzDeepPredictor._load/s$//$<<OO&JJ OO&EE     i -A $ !   -!]]..{C@DN==,,Y;DL >>  A"YY^^CLu^M  !!%? ~NO     @zz)%eT&#xx 6IJ $ <#xx a8  %%d+=&>? " !%+s!;#xx 26 !%>y~~>NOP 00AA[\]K - - A;A3?@@ A* @:1#>?? @s> H'?A(H49CI'H14 I=II I:"I55I: klines_dfz pd.DataFrame timestamp_ms surge_typereturnc|jsddd|jddSt|||j}|ddd|jddSt j |tj jd}tj|d}t j |gtj}t j|d d d }t|jtjjr!|j||j!}n%|j||} | d j!}||jk\rdnd} t#||jz t%|jdz dz} t'| d} t)|| t)| t)|jddS)uo Prédit si un signal de surge sera rentable. Args: klines_df: DataFrame des klines 1m pour le symbole timestamp_ms: timestamp en ms du moment d'analyse surge_type: type de surge détecté ("FLASH_SURGE", etc.) Returns: dict avec probability, signal, confidence, threshold, model_type rBUYrnone) probabilitysignal confidencer model_type deep_no_data)dtypegg$@g$)nanposinfneginfr@SKIPg{Gz?dlstm_attention)rrrrr+tensorfloat32 unsqueezerr*long nan_to_num isinstancerr, ScriptModuleitemabsmaxminfloat) rr9r:r;seqx surge_idxsurge_tproboutputrArBs rpredictzDeepPredictor.predict_s$~~"!^^$  &it||L ;"!^^,   LLEMM 2 <  !> >  >  > > @ $ rr )rgr(ospathlibrtypingrnumpynpr+ deep_modelrrrr__file__parentrr rkrrrtsJ   (^ " "X - G G r