1dmode

.gitignore

@@ -1,3 +1,13 @@
 **/*__pycache__
 **/*.pem
 **/*.sqlite3
+hudes_env/
+hudes/mnist_data/
+web/mnist_data/
+mnist_data/
+web/node_modules/
+web/dist/
+web/test-results/
+web/generated/
+web/bundle-visualization.html
+web/package-lock.json

.pre-commit-config.yaml

@@ -6,7 +6,6 @@ repos:
       - id: end-of-file-fixer
       - id: check-ast
       - id: check-yaml
-      - id: check-added-large-files
   - repo: https://github.com/astral-sh/ruff-pre-commit
     rev: v0.7.3
     hooks:

README.md

@@ -2,7 +2,8 @@
 
 ## Featured videos
 
-- [Human Descent speed run — October 20, 2025](https://youtu.be/PIFJnmvqMIk)
+- [3D speed run — November 18, 2025](https://www.youtube.com/watch?v=4exJirHbESM)
+- [1D speed run — November 18, 2025](https://www.youtube.com/watch?v=t4DESt8-PwQ)
 - [Human Descent — How to play](https://youtu.be/tspa15Ei3KI)
 
 **Live server:** The web version runs at [humandescent.net](https://humandescent.net) on an RTX 4090 GPU. Expect instability if many people connect simultaneously.
@@ -14,6 +15,11 @@ Human Descent opens the door for you to explore neural network optimization in h
 ### Older demo
 [Demo video](https://youtu.be/mqAmaBP3-Q4)
 
+## Game modes
+
+- **3D landscape (default):** renders three coupled loss surfaces that update in real time while you explore six random directions.
+- **1D slices (experimental):** renders six fast-updating loss curves, one per dimension. Launch with `?mode=1d` (e.g. `http://localhost:5173/?mode=1d`) or use the View toggle in the header to switch between modes.
+
 ## Batch-parameterized PyTorch layers
 
 The visualizations in this project rely on a custom set of layers found in `hudes/model_first/model_first_nn.py`. These helpers wrap familiar PyTorch modules (linear, convolution, pooling, flatten, sequential) so that every forward pass accepts not just a minibatch of data, but also a *batch of parameter tensors*. Key pieces include:
@@ -50,3 +56,15 @@ Iteratively selecting between new training batches and different random subspace
 For example its possible to train a 26,000 parameter MNIST model using a 6 dimensional keyboard input in about 10 minutes.
 
 ![Example snapshot](images/01_demo.jpg)
+
+## Testing
+
+- Run all frontend checks with `cd web && npm test` (wrap in the repo's virtualenv so the backend can start).
+- When iterating on the speed-run or highscores flows, capture deterministic Playwright logs with:
+
+```
+cd web
+(npm run -s test:e2e:chromium -- tests/speedrun.spec.mjs) 2>&1 3>&1 | tee /tmp/play-speedrun.log >/dev/null
+```
+
+That command mirrors what we use in CI so any browser/server output is preserved under `/tmp/play-speedrun.log` for later debugging.

cleanup_5s_runs.py

@@ -0,0 +1,36 @@
+import os
+import sqlite3
+import sys
+
+# Add current directory to sys.path so we can import hudes
+sys.path.append(os.getcwd())
+
+from hudes.high_scores import DB_DEFAULT_PATH
+
+
+def cleanup():
+    path = os.environ.get("HIGH_SCORES_PATH", DB_DEFAULT_PATH)
+    print(f"Opening DB at {path}")
+
+    if not os.path.exists(path):
+        print("Database file not found!")
+        return
+
+    try:
+        with sqlite3.connect(path) as conn:
+            cur = conn.execute("SELECT COUNT(*) FROM high_scores WHERE duration = 5")
+            count = cur.fetchone()[0]
+            print(f"Found {count} runs with duration 5s")
+
+            if count > 0:
+                conn.execute("DELETE FROM high_scores WHERE duration = 5")
+                conn.commit()
+                print(f"Successfully deleted {count} records.")
+            else:
+                print("No records found to delete.")
+    except Exception as e:
+        print(f"Error: {e}")
+
+
+if __name__ == "__main__":
+    cleanup()

hudes/high_scores.py

@@ -125,4 +125,10 @@ def get_rank(score: float, path: str | None = None) -> tuple[int, int]:
             (float(score),),
         )
         less = int(cur.fetchone()[0])
-        return less + 1 if total > 0 else 0, total
+        return less + 1, total
+
+
+def delete_high_score(row_id: int, path: str | None = None) -> bool:
+    with _conn(path) as conn:
+        cur = conn.execute("DELETE FROM high_scores WHERE id=?", (row_id,))
+        return cur.rowcount > 0

hudes/hudes.proto

@@ -15,6 +15,10 @@ message Config {
   required int32 mesh_grids = 5;
   required int32 batch_size = 6;
   required string dtype = 7;
+  optional bool mesh_enabled = 8 [default = true];
+  optional int32 loss_lines = 9;
+  optional bool resume_supported = 10 [default = true];
+  optional string client_session_token = 11;
 }
 
 message TrainLossAndPreds {
@@ -26,6 +30,8 @@ message TrainLossAndPreds {
 }
 message ValLoss {
   required double val_loss = 1;
+  optional int32 rank = 2;
+  optional int32 total_scores = 3;
 }
 
 message BatchExamples {
@@ -51,6 +57,7 @@ message Control {
     CONTROL_BATCH_EXAMPLES = 3;
     CONTROL_SPEED_RUN_START = 10;
     CONTROL_HIGH_SCORE_LOG = 11;
+    CONTROL_SPEED_RUN_CANCEL = 12;
     // Request the top 100 leaderboard entries (NAME/SCORE only) over WebSocket
     CONTROL_LEADERBOARD_REQUEST = 106;
     // Response containing the leaderboard arrays (parallel arrays)
@@ -60,6 +67,8 @@ message Control {
     CONTROL_CONFIG = 103;
     CONTROL_MESHGRID_RESULTS= 104;
     CONTROL_SGD_STEP = 105;
+    CONTROL_LOSS_LINE_RESULTS = 108;
+    CONTROL_RESUME = 109;
     CONTROL_QUIT = 901;
   }
   required Type type = 1;
@@ -82,6 +91,7 @@ message Control {
   optional int32 speed_run_seconds_remaining = 212;
   // True when a VAL message marks the end of the Speed Run window
   optional bool speed_run_finished = 213;
+  optional bool speed_run_active = 221;
   // Top leaderboard (NAME/SCORE only). Parallel arrays; same length.
   // Only set in CONTROL_LEADERBOARD_RESPONSE.
   repeated string leaderboard_names = 214;
@@ -94,5 +104,23 @@ message Control {
     optional int32 request_idx = 4;
   }
   optional HighScore high_score = 211; // used with CONTROL_HIGH_SCORE_LOG
+  repeated float loss_line_results = 216;
+  repeated int32 loss_line_shape = 217;
+  message Resume {
+    enum Status {
+      RESUME_OK = 0;
+      RESUME_NOT_FOUND = 1;
+      RESUME_EXPIRED = 2;
+      RESUME_REJECTED = 3;
+    }
+    optional Status status = 1;
+    optional string token = 2;
+    optional int32 last_request_idx = 3;
+    optional string client_session_token = 4;
+    optional string new_client_session_token = 5;
+  }
+  optional Resume resume = 218;
+  optional string resume_token = 219;
+  optional int32 total_eval_steps = 220;
 
 }

hudes/hudes_play.py

@@ -25,6 +25,11 @@ def main(argv=None):
         "--skip-help",
         action="store_true",
     )
+    parser.add_argument(
+        "--headless",
+        action="store_true",
+        help="Use non-OpenGL view when running keyboardGL (useful for CI/headless)",
+    )
     parser.add_argument("--max-loop-iterations", type=int, default=None)
     parser.add_argument("--max-seconds", type=float, default=None)
 
@@ -52,7 +57,10 @@ def main(argv=None):
             joystick_controller_key=args.controller,
             skip_help=args.skip_help,
         )
-        view = OpenGLView(grid_size=args.grid_size, grids=args.grids)
+        if args.headless:
+            view = View()
+        else:
+            view = OpenGLView(grid_size=args.grid_size, grids=args.grids)
     elif args.input == "xtouch":
         controller = XTouchClient(
             addr=args.addr,