Mortal

2025-10-07 20:30:03 +08:00
commit b7a7d7404a
441 changed files with 23367 additions and 0 deletions
--- a/mortal/one_vs_three.py
+++ b/mortal/one_vs_three.py
@@ -0,0 +1,103 @@
+import prelude
+
+import numpy as np
+import torch
+import secrets
+import os
+from model import Brain, DQN
+from engine import MortalEngine
+from libriichi.arena import OneVsThree
+from config import config
+
+def main():
+    cfg = config['1v3']
+    games_per_iter = cfg['games_per_iter']
+    seeds_per_iter = games_per_iter // 4
+    iters = cfg['iters']
+    log_dir = cfg['log_dir']
+    use_akochan = cfg['akochan']['enabled']
+
+    if (key := cfg.get('seed_key', -1)) == -1:
+        key = secrets.randbits(64)
+
+    if use_akochan:
+        os.environ['AKOCHAN_DIR'] = cfg['akochan']['dir']
+        os.environ['AKOCHAN_TACTICS'] = cfg['akochan']['tactics']
+    else:
+        state = torch.load(cfg['champion']['state_file'], weights_only=True, map_location=torch.device('cpu'))
+        cham_cfg = state['config']
+        version = cham_cfg['control'].get('version', 1)
+        conv_channels = cham_cfg['resnet']['conv_channels']
+        num_blocks = cham_cfg['resnet']['num_blocks']
+        mortal = Brain(version=version, conv_channels=conv_channels, num_blocks=num_blocks).eval()
+        dqn = DQN(version=version).eval()
+        mortal.load_state_dict(state['mortal'])
+        dqn.load_state_dict(state['current_dqn'])
+        if cfg['champion']['enable_compile']:
+            mortal.compile()
+            dqn.compile()
+        engine_cham = MortalEngine(
+            mortal,
+            dqn,
+            is_oracle = False,
+            version = version,
+            device = torch.device(cfg['champion']['device']),
+            enable_amp = cfg['champion']['enable_amp'],
+            enable_rule_based_agari_guard = cfg['champion']['enable_rule_based_agari_guard'],
+            name = cfg['champion']['name'],
+        )
+
+    state = torch.load(cfg['challenger']['state_file'], weights_only=True, map_location=torch.device('cpu'))
+    chal_cfg = state['config']
+    version = chal_cfg['control'].get('version', 1)
+    conv_channels = chal_cfg['resnet']['conv_channels']
+    num_blocks = chal_cfg['resnet']['num_blocks']
+    mortal = Brain(version=version, conv_channels=conv_channels, num_blocks=num_blocks).eval()
+    dqn = DQN(version=version).eval()
+    mortal.load_state_dict(state['mortal'])
+    dqn.load_state_dict(state['current_dqn'])
+    if cfg['challenger']['enable_compile']:
+        mortal.compile()
+        dqn.compile()
+    engine_chal = MortalEngine(
+        mortal,
+        dqn,
+        is_oracle = False,
+        version = version,
+        device = torch.device(cfg['challenger']['device']),
+        enable_amp = cfg['challenger']['enable_amp'],
+        enable_rule_based_agari_guard = cfg['challenger']['enable_rule_based_agari_guard'],
+        name = cfg['challenger']['name'],
+    )
+
+    seed_start = 10000
+    for i, seed in enumerate(range(seed_start, seed_start + seeds_per_iter * iters, seeds_per_iter)):
+        print('-' * 50)
+        print('#', i)
+        env = OneVsThree(
+            disable_progress_bar = False,
+            log_dir = log_dir,
+        )
+        if use_akochan:
+            rankings = env.ako_vs_py(
+                engine = engine_chal,
+                seed_start = (seed, key),
+                seed_count = seeds_per_iter,
+            )
+        else:
+            rankings = env.py_vs_py(
+                challenger = engine_chal,
+                champion = engine_cham,
+                seed_start = (seed, key),
+                seed_count = seeds_per_iter,
+            )
+        rankings = np.array(rankings)
+        avg_rank = rankings @ np.arange(1, 5) / rankings.sum()
+        avg_pt = rankings @ np.array([90, 45, 0, -135]) / rankings.sum()
+        print(f'challenger rankings: {rankings} ({avg_rank}, {avg_pt}pt)')
+
+if __name__ == '__main__':
+    try:
+        main()
+    except KeyboardInterrupt:
+        pass