results: reject diagonal SDIL on PickupLoc

1 files changed, 262 insertions, 0 deletions

diff --git a/results/babyai_shared/pickup_p2/seed4103_raw_shared.json b/results/babyai_shared/pickup_p2/seed4103_raw_shared.json
new file mode 100644
index 0000000..2b57c87
--- /dev/null
+++ b/results/babyai_shared/pickup_p2/seed4103_raw_shared.json
@@ -0,0 +1,262 @@
+{
+  "condition": "raw_shared",
+  "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": 1.4279862976074218
+    },
+    {
+      "epoch": 2,
+      "train_loss": 1.6384362813046105
+    },
+    {
+      "epoch": 3,
+      "train_loss": 1.4120025303489283
+    },
+    {
+      "epoch": 4,
+      "train_loss": 1.5674286854894537
+    },
+    {
+      "epoch": 5,
+      "train_loss": 1.5401874258643702
+    },
+    {
+      "epoch": 6,
+      "train_loss": 1.4283890882291292
+    },
+    {
+      "epoch": 7,
+      "train_loss": 1.5478344112948368
+    },
+    {
+      "epoch": 8,
+      "train_loss": 1.6749039321196706
+    },
+    {
+      "epoch": 9,
+      "train_loss": 1.3940781457800615
+    },
+    {
+      "epoch": 10,
+      "train_loss": 1.3927083012932224
+    },
+    {
+      "epoch": 11,
+      "train_loss": 1.5125436822991623
+    },
+    {
+      "epoch": 12,
+      "train_loss": 1.6823845085344817
+    },
+    {
+      "epoch": 13,
+      "train_loss": 1.4184414868605764
+    },
+    {
+      "epoch": 14,
+      "train_loss": 1.4393992233276367
+    },
+    {
+      "epoch": 15,
+      "train_loss": 1.5265272190696315
+    },
+    {
+      "epoch": 16,
+      "train_loss": 1.6243808332242464
+    },
+    {
+      "epoch": 17,
+      "train_loss": 1.3540453097694798
+    },
+    {
+      "epoch": 18,
+      "train_loss": 1.499637666501497
+    },
+    {
+      "epoch": 19,
+      "train_loss": 1.3106657308026364
+    },
+    {
+      "epoch": 20,
+      "train_loss": 1.5018967904542622
+    },
+    {
+      "epoch": 21,
+      "train_loss": 1.3504434702270909
+    },
+    {
+      "epoch": 22,
+      "train_loss": 1.6013833044704637
+    },
+    {
+      "epoch": 23,
+      "train_loss": 1.6672487831115723
+    },
+    {
+      "epoch": 24,
+      "train_loss": 1.4507962395015517
+    },
+    {
+      "epoch": 25,
+      "train_loss": 1.3700003109480205
+    },
+    {
+      "epoch": 26,
+      "train_loss": 1.3767117627043473
+    },
+    {
+      "epoch": 27,
+      "train_loss": 1.4459412696487026
+    },
+    {
+      "epoch": 28,
+      "train_loss": 1.3630849577251234
+    },
+    {
+      "epoch": 29,
+      "train_loss": 1.6651583372919183
+    },
+    {
+      "epoch": 30,
+      "train_loss": 1.3656946495959634
+    },
+    {
+      "epoch": 31,
+      "train_loss": 1.410263954965692
+    },
+    {
+      "epoch": 32,
+      "train_loss": 1.3972627205597727
+    },
+    {
+      "epoch": 33,
+      "train_loss": 1.5788899810690629
+    },
+    {
+      "epoch": 34,
+      "train_loss": 1.5010775119379947
+    },
+    {
+      "epoch": 35,
+      "train_loss": 1.4368829202651978
+    },
+    {
+      "epoch": 36,
+      "train_loss": 1.3498843838039198
+    },
+    {
+      "epoch": 37,
+      "train_loss": 1.3490342864237335
+    },
+    {
+      "epoch": 38,
+      "train_loss": 1.3266521142658434
+    },
+    {
+      "epoch": 39,
+      "train_loss": 1.3577745432602732
+    },
+    {
+      "epoch": 40,
+      "train_loss": 1.2795040179553785
+    }
+  ],
+  "epochs_completed": 40,
+  "finite": true,
+  "first_nonfinite_epoch": null,
+  "mission_lesion_rollout": {
+    "episodes": 500,
+    "mean_length": 59.866,
+    "mean_return": 0.06493437499999999,
+    "success": 0.068
+  },
+  "mission_lesion_validation": {
+    "accuracy": 0.6101077184741233,
+    "loss": 1.1676633622739598
+  },
+  "predictor": [],
+  "provenance": {
+    "cuda_device_name": "NVIDIA GeForce GTX 1080",
+    "cuda_peak_allocated_bytes": 130029568,
+    "cuda_visible_devices": "2",
+    "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": 59.886,
+    "mean_return": 0.064653125,
+    "success": 0.068
+  },
+  "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": 171.7450816631317,
+    "training_wall_seconds": 154.85608053207397
+  },
+  "validation": {
+    "accuracy": 0.5960962177047056,
+    "loss": 1.1890508834481712
+  }
+}