Fix dataset-not-detected and mouse-recording training crash

CHANGELOG.md

@@ -5,6 +5,19 @@ All notable changes to BOT-MMORPG-AI will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+### Fixed
+- Recordings no longer disappear from the Train tab: the bundled
+  `versions/0.01/1-collect_data.py` now honors the `--out` argument the
+  Tauri shell passes, so data lands in `datasets/<game>/<name>/` where
+  the dataset scanner looks (issues #57, #60, #63, #65).
+- Training no longer crashes with
+  `Target size (... 39) must be the same as input size (... 29)` when
+  mouse recording is enabled. `2-train_model.py` now auto-detects the
+  output-head size from the dataset's action vector (29 without mouse,
+  39 with), and `--num-actions` defaults to `0` (auto) (issue #64).
+
 ## [1.0.0] - 2026-02-09
 
 ### Added

docs/installer/04-bug-index.md

@@ -208,6 +208,38 @@ The PowerShell driver script computes the wrong driver path under PROD.
   not `$INSTDIR\resources\drivers\`).
 - **See:** `05-case-studies.md` Bug #5 (`9d8a16f`).
 
+### "Recording saved" but no dataset appears in the Train tab
+
+The Rust shell spawns `1-collect_data.py` with
+`--out <data_root>/datasets/<gid>/<name>` so the recording lands where
+`modelhub/tauri.py::_scan_datasets_fs` looks. The **legacy**
+`versions/0.01/1-collect_data.py` had no argument parser, so it ignored
+`--out` entirely and wrote to a bare `datasets/` folder (driven only by
+`BOTMMO_OUTPUT_DIR`). Result: recording "succeeds" but the file lands
+outside `datasets/<gid>/` and never shows up in the UI.
+
+- **File:** `versions/0.01/1-collect_data.py` (`_resolve_output_dir`)
+- **Fix:** parse `--out` (with `parse_known_args` so future flags don't
+  crash the recorder) and use it as the output dir, falling back to
+  `BOTMMO_OUTPUT_DIR` then the default. Keeps the writer, the session
+  bookkeeping, and `_scan_datasets_fs` all pointed at the same path.
+- **Issues:** #57, #60, #63, #65.
+
+### Training crash: `Target size (... 39) must be the same as input size (... 29)`
+
+`BCEWithLogitsLoss` failed in `2-train_model.py` because the dataset's
+action vector was 39 wide (29 keyboard+gamepad **plus** 10 mouse values
+when mouse recording is on) while the model's output head was hard-sized
+to 29 via the `--num-actions` default.
+
+- **File:** `versions/0.01/2-train_model.py` (`GameplayDataset.num_actions`,
+  `main()` head-size resolution)
+- **Fix:** auto-detect the head size from the dataset
+  (`self.actions.shape[1]`) and default `--num-actions` to `0` (auto). A
+  positive `--num-actions` still overrides for advanced use. The model
+  head now always matches the recorded action width (29 or 39).
+- **Issue:** #64.
+
 ## UI-layer symptoms (frontend can't reach the backend)
 
 ### Notification "Dismiss" / "×" / "Later" buttons appear inert

versions/0.01/1-collect_data.py

@@ -2,6 +2,7 @@
 import sys
 import time
 import signal
+import argparse
 from pathlib import Path
 
 import cv2
@@ -188,8 +189,35 @@ def save_chunk(path: Path, data: list) -> None:
     print(f"[SAVED] {path} ({len(data)} samples)")
 
 
-def main():
-    out_path, idx = next_available_file(OUTPUT_DIR, FILE_PREFIX)
+def _resolve_output_dir(argv=None) -> Path:
+    """Decide where to write the .npy chunks.
+
+    Priority (highest first):
+      1. ``--out PATH`` on the command line. The Tauri shell spawns us
+         with ``--out <data_root>/datasets/<gid>/<name>`` so the
+         recording lands exactly where the Train tab's scanner
+         (``modelhub/tauri.py::_scan_datasets_fs``) looks. Before this
+         was honored, the flag was silently ignored and recordings
+         landed in a bare ``datasets/`` folder that never showed up in
+         the UI -- the "dataset not appearing after recording" bug
+         (issues #57, #60, #63, #65).
+      2. ``BOTMMO_OUTPUT_DIR`` environment variable (legacy launcher).
+      3. The default ``datasets/`` relative to the working directory.
+
+    Unknown args are tolerated (parse_known_args) so a future caller can
+    pass extra flags without crashing the recorder mid-launch.
+    """
+    parser = argparse.ArgumentParser(add_help=False)
+    parser.add_argument("--out", "--output", dest="out", default=None)
+    args, _unknown = parser.parse_known_args(argv)
+    if args.out:
+        return Path(args.out)
+    return OUTPUT_DIR
+
+
+def main(argv=None):
+    output_dir = _resolve_output_dir(argv)
+    out_path, idx = next_available_file(output_dir, FILE_PREFIX)
 
     # Mouse recording support (issues #21, #33, #36)
     mouse_enabled = os.getenv("BOTMMO_CAPTURE_MOUSE", "").lower() == "true"
@@ -202,7 +230,7 @@ def main():
     print(f"[Collector] CaptureRegion: {CAPTURE_REGION}")
     print(f"[Collector] Resolution   : {RESOLUTION}")
     print(f"[Collector] Mouse Record : {'ENABLED' if mouse_enabled else 'DISABLED'}")
-    print(f"[Collector] Output Dir   : {OUTPUT_DIR.resolve()}")
+    print(f"[Collector] Output Dir   : {output_dir.resolve()}")
     print(f"[Collector] Output File  : {out_path.name}")
     print("=================================================")
     if not mouse_enabled:
@@ -346,7 +374,7 @@ def _handle_sigint(signum, frame):
                 save_chunk(out_path, training_data)
                 training_data = []
                 idx += 1
-                out_path = OUTPUT_DIR / f"{FILE_PREFIX}-{idx}.npy"
+                out_path = output_dir / f"{FILE_PREFIX}-{idx}.npy"
 
     finally:
         # Save remainder so stopping from launcher doesn't lose data

versions/0.01/2-train_model.py

@@ -202,11 +202,23 @@ def _load_data(self):
         self.frames = np.array(all_frames)
         self.actions = np.array(all_actions, dtype=np.float32)
 
+        # Auto-detected width of the action vector. This is the single
+        # source of truth for sizing the model's output head. When mouse
+        # recording is enabled, collect_data appends 10 extra values
+        # (x, y, dx, dy, vx, vy, lmb, rmb, mmb, scroll) so the vector is
+        # 39 wide instead of 29 (keyboard=9 + gamepad=20). The trainer
+        # must size the head to match or BCEWithLogitsLoss raises
+        # "Target size ... must be the same as input size" (issue #64).
+        self.num_actions = (
+            self.actions.shape[1] if self.actions.ndim == 2 else len(self.actions[0])
+        )
+
         dt = time.time() - t0
         LOG.info(f"[Dataset] Loaded {len(self.frames)} sample(s) in {dt:.1f}s")
         LOG.info(f"[Dataset] Skipped files: {skipped_files}, skipped items: {skipped_items}")
         LOG.info(f"[Dataset] Frame shape: {self.frames[0].shape}")
         LOG.info(f"[Dataset] Action shape: {self.actions[0].shape}")
+        LOG.info(f"[Dataset] Auto-detected num_actions: {self.num_actions}")
 
     def __len__(self) -> int:
         return max(0, len(self.frames) - self.seq_len + 1)
@@ -430,7 +442,12 @@ def main(argv=None) -> int:
     )
     parser.add_argument("--lr", type=float, default=1e-4, help="Learning rate")
     parser.add_argument("--seq-len", type=int, default=4, help="Sequence length for temporal models")
-    parser.add_argument("--num-actions", type=int, default=29, help="Number of output actions")
+    parser.add_argument(
+        "--num-actions",
+        type=int,
+        default=0,
+        help="Number of output actions (0 = auto-detect from data; issue #64)",
+    )
     parser.add_argument("--val-split", type=float, default=0.1, help="Validation split ratio")
     parser.add_argument("--no-pretrained", action="store_true", help="Don't use pretrained weights")
     parser.add_argument("--cpu", action="store_true", help="Force CPU training")
@@ -537,11 +554,22 @@ def main(argv=None) -> int:
         if val_size > 0 else None
     )
 
+    # Resolve the output-head size. Prefer the value auto-detected from
+    # the dataset so the model head always matches the recorded action
+    # vector (29 keyboard+gamepad, or 39 when mouse recording is on).
+    # A positive --num-actions overrides it for advanced use. This fixes
+    # the "Target size ... must be the same as input size" crash (#64).
+    num_actions = args.num_actions if args.num_actions > 0 else dataset.num_actions
+    LOG.info(f"Action vector size: {num_actions}")
+    if num_actions > 29:
+        mouse_extra = num_actions - 29
+        LOG.info(f"  (keyboard=9 + gamepad=20 + mouse={mouse_extra})")
+
     # Create model
     LOG.info(f"Creating model: {args.model}")
     model = get_model(
         args.model,
-        num_actions=args.num_actions,
+        num_actions=num_actions,
         temporal_frames=seq_len,
         pretrained=not args.no_pretrained,
     )