results: reject diagonal SDIL on PickupLoc

1 files changed, 262 insertions, 0 deletions

diff --git a/results/babyai_shared/pickup_p2/seed4103_clean_kp.json b/results/babyai_shared/pickup_p2/seed4103_clean_kp.json
new file mode 100644
index 0000000..5d862e9
--- /dev/null
+++ b/results/babyai_shared/pickup_p2/seed4103_clean_kp.json
@@ -0,0 +1,262 @@
+{
+  "condition": "clean_kp",
+  "config": {
+    "action_dim": 7,
+    "context_gain": 1.0,
+    "hidden_layers": 4,
+    "input_dim": 3964,
+    "learning_rate": 0.03,
+    "mission_dim": 23,
+    "momentum": 0.9,
+    "reciprocal_learning_rate": 0.03,
+    "weight_decay": 0.0001,
+    "width": 256
+  },
+  "data": {
+    "color_cardinality": 6,
+    "elapsed_seconds": 81.82925486564636,
+    "env_id": "BabyAI-PickupLoc-v0",
+    "minigrid_version": "3.1.0",
+    "object_cardinality": 11,
+    "protocol": "babyai_shared_feedback_b0",
+    "rollout_episodes": 500,
+    "rollout_seed_start": 200000,
+    "state_cardinality": 3,
+    "train_episodes": 20000,
+    "train_steps": 121467,
+    "validation_episodes": 2000,
+    "validation_steps": 12347,
+    "vocabulary": [
+      "<unk>",
+      "a",
+      "ball",
+      "behind",
+      "blue",
+      "box",
+      "front",
+      "green",
+      "grey",
+      "in",
+      "key",
+      "left",
+      "of",
+      "on",
+      "pick",
+      "purple",
+      "red",
+      "right",
+      "the",
+      "up",
+      "yellow",
+      "you",
+      "your"
+    ]
+  },
+  "epoch_history": [
+    {
+      "epoch": 1,
+      "train_loss": 0.6864033263608029
+    },
+    {
+      "epoch": 2,
+      "train_loss": 0.5136532720766569
+    },
+    {
+      "epoch": 3,
+      "train_loss": 0.4820757521453657
+    },
+    {
+      "epoch": 4,
+      "train_loss": 0.4678087641063489
+    },
+    {
+      "epoch": 5,
+      "train_loss": 0.4570458981237913
+    },
+    {
+      "epoch": 6,
+      "train_loss": 0.44718988864045395
+    },
+    {
+      "epoch": 7,
+      "train_loss": 0.43997702253492255
+    },
+    {
+      "epoch": 8,
+      "train_loss": 0.42566254320897556
+    },
+    {
+      "epoch": 9,
+      "train_loss": 0.4220644837931583
+    },
+    {
+      "epoch": 10,
+      "train_loss": 0.41289369489017286
+    },
+    {
+      "epoch": 11,
+      "train_loss": 0.405345781037682
+    },
+    {
+      "epoch": 12,
+      "train_loss": 0.40259038736945707
+    },
+    {
+      "epoch": 13,
+      "train_loss": 0.393158560112903
+    },
+    {
+      "epoch": 14,
+      "train_loss": 0.38850406364390727
+    },
+    {
+      "epoch": 15,
+      "train_loss": 0.3801098694926814
+    },
+    {
+      "epoch": 16,
+      "train_loss": 0.36695603276553906
+    },
+    {
+      "epoch": 17,
+      "train_loss": 0.35903412762441134
+    },
+    {
+      "epoch": 18,
+      "train_loss": 0.34711836858799583
+    },
+    {
+      "epoch": 19,
+      "train_loss": 0.3370369092727962
+    },
+    {
+      "epoch": 20,
+      "train_loss": 0.3200993647387153
+    },
+    {
+      "epoch": 21,
+      "train_loss": 0.3118796497897098
+    },
+    {
+      "epoch": 22,
+      "train_loss": 0.293176792420839
+    },
+    {
+      "epoch": 23,
+      "train_loss": 0.28278176423750423
+    },
+    {
+      "epoch": 24,
+      "train_loss": 0.2673768097789664
+    },
+    {
+      "epoch": 25,
+      "train_loss": 0.2630335626790398
+    },
+    {
+      "epoch": 26,
+      "train_loss": 0.2525263265559548
+    },
+    {
+      "epoch": 27,
+      "train_loss": 0.24664549178198764
+    },
+    {
+      "epoch": 28,
+      "train_loss": 0.23551284742982764
+    },
+    {
+      "epoch": 29,
+      "train_loss": 0.23284720298491027
+    },
+    {
+      "epoch": 30,
+      "train_loss": 0.22212432547619468
+    },
+    {
+      "epoch": 31,
+      "train_loss": 0.21526990802664506
+    },
+    {
+      "epoch": 32,
+      "train_loss": 0.20761467494462665
+    },
+    {
+      "epoch": 33,
+      "train_loss": 0.21185886748527225
+    },
+    {
+      "epoch": 34,
+      "train_loss": 0.20015101647690722
+    },
+    {
+      "epoch": 35,
+      "train_loss": 0.19496909729744258
+    },
+    {
+      "epoch": 36,
+      "train_loss": 0.1936807933763454
+    },
+    {
+      "epoch": 37,
+      "train_loss": 0.19314897574876483
+    },
+    {
+      "epoch": 38,
+      "train_loss": 0.1823754488794427
+    },
+    {
+      "epoch": 39,
+      "train_loss": 0.17861489517124077
+    },
+    {
+      "epoch": 40,
+      "train_loss": 0.17446756489967044
+    }
+  ],
+  "epochs_completed": 40,
+  "finite": true,
+  "first_nonfinite_epoch": null,
+  "mission_lesion_rollout": {
+    "episodes": 500,
+    "mean_length": 45.232,
+    "mean_return": 0.305125,
+    "success": 0.412
+  },
+  "mission_lesion_validation": {
+    "accuracy": 0.7619664695877542,
+    "loss": 0.7881599735844866
+  },
+  "predictor": [],
+  "provenance": {
+    "cuda_device_name": "NVIDIA GeForce GTX 1080",
+    "cuda_peak_allocated_bytes": 130533376,
+    "cuda_visible_devices": "1",
+    "device": "cuda:0",
+    "git_commit": "7b7d7a08a5e78c3379397f01f0c9af97d5d84fbd",
+    "git_dirty_tracked": false,
+    "minigrid_version": "3.1.0",
+    "torch_version": "2.3.1+cu118"
+  },
+  "rollout": {
+    "episodes": 500,
+    "mean_length": 28.356,
+    "mean_return": 0.56964375,
+    "success": 0.684
+  },
+  "stage": "babyai_pickup_p2",
+  "training": {
+    "batch_size": 256,
+    "history_steps": 4,
+    "model_seed": 4103,
+    "neutral_examples_per_epoch": 0,
+    "parameter_count_including_reciprocal_context_predictor": 1438215,
+    "shuffle_seed": 4103,
+    "total_wall_seconds_including_rollouts": 166.57137203216553,
+    "training_wall_seconds": 153.0613570213318
+  },
+  "validation": {
+    "accuracy": 0.8635296023325504,
+    "loss": 0.35984307350374345
+  }
+}