results: reject diagonal SDIL on PickupLoc

1 files changed, 262 insertions, 0 deletions

diff --git a/results/babyai_shared/pickup_p2/seed4102_clean_kp.json b/results/babyai_shared/pickup_p2/seed4102_clean_kp.json
new file mode 100644
index 0000000..9cd9956
--- /dev/null
+++ b/results/babyai_shared/pickup_p2/seed4102_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": 0.681839566669966
+    },
+    {
+      "epoch": 2,
+      "train_loss": 0.5106817564838811
+    },
+    {
+      "epoch": 3,
+      "train_loss": 0.4866123088410026
+    },
+    {
+      "epoch": 4,
+      "train_loss": 0.46724821322842647
+    },
+    {
+      "epoch": 5,
+      "train_loss": 0.45275609010144285
+    },
+    {
+      "epoch": 6,
+      "train_loss": 0.444956036241431
+    },
+    {
+      "epoch": 7,
+      "train_loss": 0.4342330169050317
+    },
+    {
+      "epoch": 8,
+      "train_loss": 0.42846722878907856
+    },
+    {
+      "epoch": 9,
+      "train_loss": 0.41873464232996893
+    },
+    {
+      "epoch": 10,
+      "train_loss": 0.4157650209100623
+    },
+    {
+      "epoch": 11,
+      "train_loss": 0.40508943651851853
+    },
+    {
+      "epoch": 12,
+      "train_loss": 0.3999078395492152
+    },
+    {
+      "epoch": 13,
+      "train_loss": 0.3982298595654337
+    },
+    {
+      "epoch": 14,
+      "train_loss": 0.3850154632016232
+    },
+    {
+      "epoch": 15,
+      "train_loss": 0.3766156781347174
+    },
+    {
+      "epoch": 16,
+      "train_loss": 0.368439452836388
+    },
+    {
+      "epoch": 17,
+      "train_loss": 0.3616289800719211
+    },
+    {
+      "epoch": 18,
+      "train_loss": 0.3508304416505914
+    },
+    {
+      "epoch": 19,
+      "train_loss": 0.3401417228422667
+    },
+    {
+      "epoch": 20,
+      "train_loss": 0.3305343171797301
+    },
+    {
+      "epoch": 21,
+      "train_loss": 0.3165550116802517
+    },
+    {
+      "epoch": 22,
+      "train_loss": 0.3019398627469414
+    },
+    {
+      "epoch": 23,
+      "train_loss": 0.29349281496123264
+    },
+    {
+      "epoch": 24,
+      "train_loss": 0.2876167286697187
+    },
+    {
+      "epoch": 25,
+      "train_loss": 0.26772761492352737
+    },
+    {
+      "epoch": 26,
+      "train_loss": 0.26045800096110294
+    },
+    {
+      "epoch": 27,
+      "train_loss": 0.24848401973122045
+    },
+    {
+      "epoch": 28,
+      "train_loss": 0.2440636997787576
+    },
+    {
+      "epoch": 29,
+      "train_loss": 0.2351069336188467
+    },
+    {
+      "epoch": 30,
+      "train_loss": 0.23047932257777767
+    },
+    {
+      "epoch": 31,
+      "train_loss": 0.22535064524725865
+    },
+    {
+      "epoch": 32,
+      "train_loss": 0.22159839981480647
+    },
+    {
+      "epoch": 33,
+      "train_loss": 0.2125453116078126
+    },
+    {
+      "epoch": 34,
+      "train_loss": 0.20711057270828045
+    },
+    {
+      "epoch": 35,
+      "train_loss": 0.20062021220985213
+    },
+    {
+      "epoch": 36,
+      "train_loss": 0.19858844374355517
+    },
+    {
+      "epoch": 37,
+      "train_loss": 0.1931002575629636
+    },
+    {
+      "epoch": 38,
+      "train_loss": 0.1893660755063358
+    },
+    {
+      "epoch": 39,
+      "train_loss": 0.18631849710878573
+    },
+    {
+      "epoch": 40,
+      "train_loss": 0.18104866024694943
+    }
+  ],
+  "epochs_completed": 40,
+  "finite": true,
+  "first_nonfinite_epoch": null,
+  "mission_lesion_rollout": {
+    "episodes": 500,
+    "mean_length": 41.762,
+    "mean_return": 0.36832187499999997,
+    "success": 0.556
+  },
+  "mission_lesion_validation": {
+    "accuracy": 0.7485219081558273,
+    "loss": 0.7462387038617324
+  },
+  "predictor": [],
+  "provenance": {
+    "cuda_device_name": "NVIDIA GeForce GTX 1080",
+    "cuda_peak_allocated_bytes": 130533376,
+    "cuda_visible_devices": "5",
+    "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": 25.062,
+    "mean_return": 0.6245656249999999,
+    "success": 0.77
+  },
+  "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": 340.4907646179199,
+    "training_wall_seconds": 326.9207751750946
+  },
+  "validation": {
+    "accuracy": 0.8682270997003321,
+    "loss": 0.3341673115462806
+  }
+}