results: reject diagonal SDIL on PickupLoc

1 files changed, 262 insertions, 0 deletions

diff --git a/results/babyai_shared/pickup_p2/seed4102_bp.json b/results/babyai_shared/pickup_p2/seed4102_bp.json
new file mode 100644
index 0000000..294fb8c
--- /dev/null
+++ b/results/babyai_shared/pickup_p2/seed4102_bp.json
@@ -0,0 +1,262 @@
+{
+  "condition": "bp",
+  "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": 0.6358707290574124
+    },
+    {
+      "epoch": 2,
+      "train_loss": 0.4983976507814307
+    },
+    {
+      "epoch": 3,
+      "train_loss": 0.4675803022008193
+    },
+    {
+      "epoch": 4,
+      "train_loss": 0.4492099952697754
+    },
+    {
+      "epoch": 5,
+      "train_loss": 0.42707399907865023
+    },
+    {
+      "epoch": 6,
+      "train_loss": 0.39973109395880446
+    },
+    {
+      "epoch": 7,
+      "train_loss": 0.3692683135835748
+    },
+    {
+      "epoch": 8,
+      "train_loss": 0.34918095908666913
+    },
+    {
+      "epoch": 9,
+      "train_loss": 0.32295236201662764
+    },
+    {
+      "epoch": 10,
+      "train_loss": 0.3064782916558416
+    },
+    {
+      "epoch": 11,
+      "train_loss": 0.2867251298615807
+    },
+    {
+      "epoch": 12,
+      "train_loss": 0.2729152385498348
+    },
+    {
+      "epoch": 13,
+      "train_loss": 0.2690554935994901
+    },
+    {
+      "epoch": 14,
+      "train_loss": 0.2533571675576662
+    },
+    {
+      "epoch": 15,
+      "train_loss": 0.2478961335985284
+    },
+    {
+      "epoch": 16,
+      "train_loss": 0.24331268574062145
+    },
+    {
+      "epoch": 17,
+      "train_loss": 0.23489129813093887
+    },
+    {
+      "epoch": 18,
+      "train_loss": 0.22351061075925827
+    },
+    {
+      "epoch": 19,
+      "train_loss": 0.2218321235242643
+    },
+    {
+      "epoch": 20,
+      "train_loss": 0.21205236274945108
+    },
+    {
+      "epoch": 21,
+      "train_loss": 0.21070943317915264
+    },
+    {
+      "epoch": 22,
+      "train_loss": 0.2010072124474927
+    },
+    {
+      "epoch": 23,
+      "train_loss": 0.1968216976366545
+    },
+    {
+      "epoch": 24,
+      "train_loss": 0.19213551078972063
+    },
+    {
+      "epoch": 25,
+      "train_loss": 0.1923028072401097
+    },
+    {
+      "epoch": 26,
+      "train_loss": 0.18429007967835978
+    },
+    {
+      "epoch": 27,
+      "train_loss": 0.17929337776020954
+    },
+    {
+      "epoch": 28,
+      "train_loss": 0.1737828024908116
+    },
+    {
+      "epoch": 29,
+      "train_loss": 0.17116386283385127
+    },
+    {
+      "epoch": 30,
+      "train_loss": 0.16565450434622012
+    },
+    {
+      "epoch": 31,
+      "train_loss": 0.16759385447753103
+    },
+    {
+      "epoch": 32,
+      "train_loss": 0.1587610994514666
+    },
+    {
+      "epoch": 33,
+      "train_loss": 0.15462220474293356
+    },
+    {
+      "epoch": 34,
+      "train_loss": 0.15275993905569377
+    },
+    {
+      "epoch": 35,
+      "train_loss": 0.1518227168447093
+    },
+    {
+      "epoch": 36,
+      "train_loss": 0.14997457358397936
+    },
+    {
+      "epoch": 37,
+      "train_loss": 0.13869191097585779
+    },
+    {
+      "epoch": 38,
+      "train_loss": 0.14246609537225022
+    },
+    {
+      "epoch": 39,
+      "train_loss": 0.13483418814445797
+    },
+    {
+      "epoch": 40,
+      "train_loss": 0.1340645107156352
+    }
+  ],
+  "epochs_completed": 40,
+  "finite": true,
+  "first_nonfinite_epoch": null,
+  "mission_lesion_rollout": {
+    "episodes": 500,
+    "mean_length": 42.626,
+    "mean_return": 0.351171875,
+    "success": 0.506
+  },
+  "mission_lesion_validation": {
+    "accuracy": 0.743905402121973,
+    "loss": 1.0319232535939609
+  },
+  "predictor": [],
+  "provenance": {
+    "cuda_device_name": "NVIDIA GeForce GTX 1080",
+    "cuda_peak_allocated_bytes": 130533376,
+    "cuda_visible_devices": "4",
+    "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": 21.354,
+    "mean_return": 0.683309375,
+    "success": 0.836
+  },
+  "stage": "babyai_pickup_p2",
+  "training": {
+    "batch_size": 256,
+    "history_steps": 4,
+    "model_seed": 4102,
+    "neutral_examples_per_epoch": 0,
+    "parameter_count_including_reciprocal_context_predictor": 1438215,
+    "shuffle_seed": 4102,
+    "total_wall_seconds_including_rollouts": 331.3091230392456,
+    "training_wall_seconds": 311.8552315235138
+  },
+  "validation": {
+    "accuracy": 0.8787559731108772,
+    "loss": 0.35747272120397217
+  }
+}