Upload Quasar_axrvi_ranker.py

Quasar_axrvi_ranker.py

@@ -1,3 +1,10 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Fri Apr  3 21:26:48 2026
+
+@author: taten
+"""
+
 # -*- coding: utf-8 -*-
 """
 Created on Fri Apr  3 13:46:18 2026
@@ -2411,10 +2418,26 @@ def crps_loss(quantiles: torch.Tensor, quantile_levels: torch.Tensor,
               y_true: torch.Tensor) -> torch.Tensor:
     """
     Continuous Ranked Probability Score (proper scoring rule for distributions).
+
+    Supports two calling conventions:
+      • 2-D quantiles (B, Q)    — e.g. selected-asset slice, y_true shape (B,)
+      • 3-D quantiles (B, N, Q) — e.g. full asset roster,    y_true shape (B,)
+
+    The previous implementation used `y_true.unsqueeze(-1)` → (B, 1) for both
+    cases.  When quantiles is 3-D, PyTorch left-pads (B, 1) to (1, B, 1) before
+    broadcasting against (B, N, Q), producing a dim-1 mismatch of B vs N
+    (e.g. "size 4 must match 9 at non-singleton dimension 1").  Fixed below by
+    branching on ndim.
     """
-    y_exp = y_true.unsqueeze(-1)
-    errors = y_exp - quantiles
-    tau = quantile_levels.view(1, 1, -1)
+    if quantiles.dim() == 2:
+        # (B, Q) — selected-asset slice or standalone call
+        y_exp = y_true.unsqueeze(-1)         # (B, 1)
+        tau   = quantile_levels.view(1, -1)  # (1, Q)
+    else:
+        # (B, N, Q) — full asset roster
+        y_exp = y_true.view(-1, 1, 1)          # (B, 1, 1)
+        tau   = quantile_levels.view(1, 1, -1)  # (1, 1, Q)
+    errors  = y_exp - quantiles
     pinball = torch.where(errors >= 0, tau * errors, (tau - 1) * errors)
     return pinball.mean()
 
@@ -3474,8 +3497,12 @@ class HybridTrainer:
             l_gate = out.get("gate_entropy_loss", torch.tensor(0.0, device=seq_t.device))
 
             # L_crps — distributional calibration (CRPS proper scoring rule) [v8]
+            # Use sel_q (B, n_quantiles) — the selected asset's quantile predictions —
+            # not the full (B, N, n_quantiles) tensor.  Calibrating every asset's
+            # distribution against a single scalar reward is semantically incorrect;
+            # only the asset that was actually traded has a matching observed return.
             ql_levels = self.model.distributional.quantile_levels
-            l_crps = crps_loss(quantiles, ql_levels, rewards)
+            l_crps = crps_loss(sel_q, ql_levels, rewards)
 
             # L_rent — regime-router entropy regularisation [v8]
             r_probs    = out["regime_probs"] + 1e-8              # (B, N, n_regimes)
@@ -4550,7 +4577,7 @@ class QuasarAXRVIBridge:
         reward_strategy: str  = "simple",
         hub_ws_url:      str  = os.environ.get("QUASAR_HUB_URL", "ws://localhost:7860/ws/subscribe"),
         enable_logging:  bool = True,
-        checkpoint_dir:  str  = "./Ranker4",   # folder for full-state checkpoints
+        checkpoint_dir:  str  = "./Ranker5",   # folder for full-state checkpoints
         resume:          bool = False,         # start afresh by default; set True to resume
     ):
         self.config          = config or AssetRankerConfig()
@@ -6204,7 +6231,7 @@ class RankerCheckpointManager:
                On load, missing local files are pulled from the repo first.
                Requires HF_TOKEN env-var with write permission.
 
-    Fallback : Local disk  ./Ranker4/
+    Fallback : Local disk  ./Ranker5/
                Used when HF_TOKEN is absent or upload/download fails.
                All saves still succeed locally even without network access.
 
@@ -6258,7 +6285,7 @@ class RankerCheckpointManager:
 
     Usage
     ─────
-    mgr = RankerCheckpointManager(checkpoint_dir="./Ranker4",
+    mgr = RankerCheckpointManager(checkpoint_dir="./Ranker5",
                                    save_interval_seconds=1800)
     mgr.load(bridge)              # once after initialize()
     mgr.maybe_save(bridge)        # call frequently; respects save_interval_seconds
@@ -6274,7 +6301,7 @@ class RankerCheckpointManager:
 
     def __init__(
         self,
-        checkpoint_dir:         str   = "./Ranker4",
+        checkpoint_dir:         str   = "./Ranker5",
         save_interval_seconds:  float = 1800.0,   # 30 minutes
     ):
         import pathlib
@@ -6328,7 +6355,7 @@ class RankerCheckpointManager:
             )
             logger.info(
                 f"[RankerCheckpoint] ☁️  Uploaded {local_path.name} → "
-                f"hf://{self.HF_REPO_ID}/Ranker4/{local_path.name}"
+                f"hf://{self.HF_REPO_ID}/Ranker5/{local_path.name}"
             )
             return True
         except Exception as exc:
@@ -6359,7 +6386,7 @@ class RankerCheckpointManager:
             )
             logger.info(
                 f"[RankerCheckpoint] ⬇️  Downloaded {filename} from "
-                f"hf://{self.HF_REPO_ID}/Ranker4/"
+                f"hf://{self.HF_REPO_ID}/Ranker5/"
             )
             return True
         except Exception as exc:
@@ -6900,7 +6927,7 @@ async def run_live_trading_system(
     hub_ws_url:      str  = "ws://localhost:7860/ws/subscribe",
     enable_logging:  bool = True,
     shreve_config:   Optional[ShreveConfig] = None,
-    checkpoint_dir:  str  = "./Ranker4",
+    checkpoint_dir:  str  = "./Ranker5",
     resume:          bool = False,   # start fresh by default
 ) -> None:
     config = AssetRankerConfig(
@@ -7212,8 +7239,8 @@ def _parse_args():
                         help="[S6/S8] Trade horizon τ in seconds (default 60)")
     parser.add_argument("--martingale-epsilon", type=float, default=0.05,
                         help="[S7] Gate E martingale deviation threshold (default 0.05)")
-    parser.add_argument("--checkpoint-dir", default="./Ranker4",
-                        help="Directory for full-state checkpoints (default ./Ranker4)")
+    parser.add_argument("--checkpoint-dir", default="./Ranker5",
+                        help="Directory for full-state checkpoints (default ./Ranker5)")
     parser.add_argument("--resume", action="store_true",
                         help="Resume training from the latest saved checkpoint (default: start fresh)")
     return parser.parse_args(filtered)