From fadb4cf46ebfd804dd3185ac7198e4c99b05fd94 Mon Sep 17 00:00:00 2001
From: lukifer23 <56565060+lukifer23@users.noreply.github.com>
Date: Mon, 6 Oct 2025 14:08:59 -0500
Subject: [PATCH] Add regression test for enhanced SSL heads

---
 tests/test_ssl_heads.py | 90 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 90 insertions(+)
 create mode 100644 tests/test_ssl_heads.py

diff --git a/tests/test_ssl_heads.py b/tests/test_ssl_heads.py
new file mode 100644
index 0000000..867f330
--- /dev/null
+++ b/tests/test_ssl_heads.py
@@ -0,0 +1,90 @@
+import chess
+import torch
+
+from azchess.model.resnet import NetConfig, PolicyValueNet
+from azchess.ssl_algorithms import ChessSSLAlgorithms
+from experiments.grpo.utils.board_encoding import board_to_tensor
+
+
+def _make_batch(batch_size: int = 2) -> torch.Tensor:
+    sequences = [
+        [],
+        ["e4", "e5", "Nf3", "Nc6"],
+        ["d4", "d5", "c4", "dxc4", "Nf3", "Nf6"],
+    ]
+
+    boards = []
+    for seq in sequences:
+        board = chess.Board()
+        for san in seq:
+            board.push_san(san)
+        boards.append(board)
+
+    if batch_size > len(boards):
+        boards.extend(boards[: batch_size - len(boards)])
+
+    tensors = [board_to_tensor(b) for b in boards[:batch_size]]
+    return torch.cat(tensors, dim=0)
+
+
+def test_enhanced_ssl_heads_outputs_and_loss():
+    torch.manual_seed(0)
+    batch = _make_batch(batch_size=3)
+
+    cfg = NetConfig(
+        channels=32,
+        blocks=2,
+        attention=False,
+        chess_features=False,
+        self_supervised=True,
+        ssl_tasks=["piece", "threat", "pin", "fork", "control"],
+    )
+
+    model = PolicyValueNet(cfg)
+    model.eval()
+
+    # Forward pass to obtain SSL head outputs and shared features
+    _, _, ssl_outputs, feats = model.forward_with_features(batch, return_ssl=True)
+
+    expected_shapes = {
+        "piece": (batch.size(0), 13, 8, 8),
+        "threat": (batch.size(0), 1, 8, 8),
+        "pin": (batch.size(0), 1, 8, 8),
+        "fork": (batch.size(0), 1, 8, 8),
+        "control": (batch.size(0), 3, 8, 8),
+    }
+
+    assert set(ssl_outputs.keys()) == set(expected_shapes.keys())
+    for task, expected_shape in expected_shapes.items():
+        output = ssl_outputs[task]
+        assert output.shape == expected_shape
+        assert output.dtype == batch.dtype
+
+    ssl_alg = ChessSSLAlgorithms()
+    targets = ssl_alg.create_enhanced_ssl_targets(batch)
+
+    loss = model.get_enhanced_ssl_loss(batch, targets, feats=feats)
+    assert torch.isfinite(loss).item()
+    assert loss.dtype == torch.float32
+
+    stats_keys = {k for k in model._ssl_loss_stats if k.startswith("task:")}
+    assert stats_keys == {f"task:{name}" for name in expected_shapes.keys()}
+
+    # Regression: disabling tasks should immediately reflect in loss tracking
+    cfg_subset = NetConfig(
+        channels=32,
+        blocks=2,
+        attention=False,
+        chess_features=False,
+        self_supervised=True,
+        ssl_tasks=["piece", "threat"],
+    )
+
+    model_subset = PolicyValueNet(cfg_subset)
+    model_subset.eval()
+
+    _, _, _, feats_subset = model_subset.forward_with_features(batch, return_ssl=True)
+    model_subset.get_enhanced_ssl_loss(batch, targets, feats=feats_subset)
+
+    subset_stats_keys = {k for k in model_subset._ssl_loss_stats if k.startswith("task:")}
+    assert subset_stats_keys == {"task:piece", "task:threat"}