0.7

2026-07-03 01:11:00 +08:00
parent 6c0875e832
commit fda254e573
71 changed files with 2837 additions and 719 deletions
@@ -0,0 +1,102 @@
+"""read_progress.py — 读取当前训练进度并做回测分析"""
+import sys, torch
+sys.path.insert(0, '.')
+
+ckpt = torch.load('checkpoints/ckpt_step_0480.pt', map_location='cpu', weights_only=False)
+h       = ckpt['training_history']
+tokens  = ckpt['best_formula']
+best_sc = ckpt['best_score']
+
+from model_core.vocab import FORMULA_VOCAB
+names = FORMULA_VOCAB.token_names
+n = len(h['step'])
+
+vals  = h['val_score']
+bests = h['best_score']
+rews  = h['avg_reward']
+ents  = h['entropy']
+ics   = h['ic_mean']
+
+print(f"=== 当前训练进度 step={ckpt['step']} ===")
+print(f"全局最优 BestScore : {best_sc:.4f}")
+print(f"末步批次 ValScore  : {vals[-1]:+.4f}")
+print(f"末步 Entropy       : {ents[-1]:.3f}")
+print(f"末步 IC            : {ics[-1]:+.5f}")
+print(f"最优公式 tokens    : {tokens}")
+print(f"最优公式 解读      : {' -> '.join(names[t] for t in tokens)}")
+print()
+
+print("=== BestScore 跳变点 ===")
+prev = bests[0]
+print(f"  step  0: {prev:.4f}")
+for i in range(1, n):
+    if bests[i] > prev + 0.05:
+        print(f"  step{h['step'][i]:4d}: {prev:.4f} -> {bests[i]:.4f}  (+{bests[i]-prev:.4f})")
+        prev = bests[i]
+
+print()
+print("=== 分阶段统计 ===")
+for s, e in [(0,100),(100,200),(200,300),(300,400),(400,481)]:
+    sv  = [vals[i] for i in range(n) if s <= h['step'][i] < e]
+    sr  = [rews[i] for i in range(n) if s <= h['step'][i] < e]
+    se  = [ents[i] for i in range(n) if s <= h['step'][i] < e]
+    sic = [ics[i]  for i in range(n) if s <= h['step'][i] < e]
+    if sv:
+        print(f"  {s:3d}-{e}: val_mean={sum(sv)/len(sv):+.3f}"
+              f"  val_max={max(sv):+.3f}"
+              f"  H={sum(se)/len(se):.2f}"
+              f"  IC={sum(sic)/len(sic):+.5f}")
+
+for i, v in enumerate(vals):
+    if v > 0:
+        print(f"\n首次 val>0: step {h['step'][i]} (val={v:+.4f})")
+        break
+
+# ── 交易频率分析 ──────────────────────────────────────────────────
+print()
+print("=== 最优公式交易频率 ===")
+from data_pipeline.fetcher import MT5DataFetcher
+from data_pipeline.data_manager import MT5DataManager
+from model_core.vm import StackVM
+
+with MT5DataFetcher() as fetcher:
+    mgr = MT5DataManager(fetcher)
+    mgr.load()
+    feat = mgr.feat_tensor
+    syms = mgr.symbols
+
+vm = StackVM()
+factor = vm.execute(tokens, feat)
+pos    = torch.sign(torch.tanh(factor))
+T      = pos.shape[1]
+print(f"数据长度 T={T} bars (H1 交集对齐)")
+print()
+
+total_trades = 0
+for i, sym in enumerate(syms):
+    p = pos[i]
+    trades, runs, cur_len, cur_dir, prev = 0, [], 0, 0, 0
+    for v in p.tolist():
+        vi = int(v)
+        if vi != 0:
+            if vi == cur_dir:
+                cur_len += 1
+            else:
+                if cur_len > 0: runs.append(cur_len)
+                cur_dir, cur_len = vi, 1
+                if prev != vi: trades += 1
+        else:
+            if cur_len > 0: runs.append(cur_len)
+            cur_dir, cur_len = 0, 0
+        prev = vi
+    if cur_len > 0: runs.append(cur_len)
+    avg_hold = sum(runs) / len(runs) if runs else 0
+    zero_pct = (p == 0).float().mean().item() * 100
+    long_pct = (p == 1).float().mean().item() * 100
+    total_trades += trades
+    print(f"  {sym}: {trades}笔  每100bar={trades/T*100:.1f}笔  "
+          f"均持仓={avg_hold:.1f}bar  空仓={zero_pct:.0f}%  "
+          f"多{long_pct:.0f}%/空{100-zero_pct-long_pct:.0f}%")
+
+avg100 = total_trades / (T * len(syms)) * 100
+print(f"\n  三品种合计: 每100bar={avg100:.1f}笔  每天~{avg100*24/100:.1f}笔")