results: reject diagonal SDIL on PickupLoc

1 files changed, 262 insertions, 0 deletions

diff --git a/results/babyai_shared/pickup_p2/seed4102_raw_shared.json b/results/babyai_shared/pickup_p2/seed4102_raw_shared.json
new file mode 100644
index 0000000..18eddd3
--- /dev/null
+++ b/results/babyai_shared/pickup_p2/seed4102_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": 1.426944848863702
+    },
+    {
+      "epoch": 2,
+      "train_loss": 1.4811008263889112
+    },
+    {
+      "epoch": 3,
+      "train_loss": 1.537802341109828
+    },
+    {
+      "epoch": 4,
+      "train_loss": 1.6175212340605887
+    },
+    {
+      "epoch": 5,
+      "train_loss": 1.4761648283506694
+    },
+    {
+      "epoch": 6,
+      "train_loss": 1.5078701266489531
+    },
+    {
+      "epoch": 7,
+      "train_loss": 1.228762877489391
+    },
+    {
+      "epoch": 8,
+      "train_loss": 1.3726487501044022
+    },
+    {
+      "epoch": 9,
+      "train_loss": 1.2172372989905509
+    },
+    {
+      "epoch": 10,
+      "train_loss": 1.1436676884952344
+    },
+    {
+      "epoch": 11,
+      "train_loss": 1.3721513145848325
+    },
+    {
+      "epoch": 12,
+      "train_loss": 1.3090164377814846
+    },
+    {
+      "epoch": 13,
+      "train_loss": 1.3085567035173116
+    },
+    {
+      "epoch": 14,
+      "train_loss": 1.24689010431892
+    },
+    {
+      "epoch": 15,
+      "train_loss": 1.438764830388521
+    },
+    {
+      "epoch": 16,
+      "train_loss": 1.4798997766093205
+    },
+    {
+      "epoch": 17,
+      "train_loss": 1.5367788784127487
+    },
+    {
+      "epoch": 18,
+      "train_loss": 1.4143389420760306
+    },
+    {
+      "epoch": 19,
+      "train_loss": 1.4158848059804816
+    },
+    {
+      "epoch": 20,
+      "train_loss": 1.2560336269830403
+    },
+    {
+      "epoch": 21,
+      "train_loss": 1.4611049637041593
+    },
+    {
+      "epoch": 22,
+      "train_loss": 1.3461816581926849
+    },
+    {
+      "epoch": 23,
+      "train_loss": 1.3200064413171066
+    },
+    {
+      "epoch": 24,
+      "train_loss": 1.209783155039737
+    },
+    {
+      "epoch": 25,
+      "train_loss": 1.2179505748497812
+    },
+    {
+      "epoch": 26,
+      "train_loss": 1.3550261970570212
+    },
+    {
+      "epoch": 27,
+      "train_loss": 1.0599337084669815
+    },
+    {
+      "epoch": 28,
+      "train_loss": 0.9845607308337563
+    },
+    {
+      "epoch": 29,
+      "train_loss": 0.9575978085869237
+    },
+    {
+      "epoch": 30,
+      "train_loss": 0.9305678352556731
+    },
+    {
+      "epoch": 31,
+      "train_loss": 0.9234064896483171
+    },
+    {
+      "epoch": 32,
+      "train_loss": 0.9524868351534793
+    },
+    {
+      "epoch": 33,
+      "train_loss": 0.9005411171913147
+    },
+    {
+      "epoch": 34,
+      "train_loss": 0.8903469843613474
+    },
+    {
+      "epoch": 35,
+      "train_loss": 0.9097682004225881
+    },
+    {
+      "epoch": 36,
+      "train_loss": 0.9102116966247559
+    },
+    {
+      "epoch": 37,
+      "train_loss": 0.8369047345613179
+    },
+    {
+      "epoch": 38,
+      "train_loss": 0.9056826622862565
+    },
+    {
+      "epoch": 39,
+      "train_loss": 1.0265569414590534
+    },
+    {
+      "epoch": 40,
+      "train_loss": 0.8987322279026634
+    }
+  ],
+  "epochs_completed": 40,
+  "finite": true,
+  "first_nonfinite_epoch": null,
+  "mission_lesion_rollout": {
+    "episodes": 500,
+    "mean_length": 52.798,
+    "mean_return": 0.176928125,
+    "success": 0.194
+  },
+  "mission_lesion_validation": {
+    "accuracy": 0.692556896412084,
+    "loss": 0.8453130396868214
+  },
+  "predictor": [],
+  "provenance": {
+    "cuda_device_name": "NVIDIA GeForce GTX 1080",
+    "cuda_peak_allocated_bytes": 130029568,
+    "cuda_visible_devices": "6",
+    "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": 50.642,
+    "mean_return": 0.211046875,
+    "success": 0.232
+  },
+  "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": 344.8074402809143,
+    "training_wall_seconds": 326.14055609703064
+  },
+  "validation": {
+    "accuracy": 0.6942577144245565,
+    "loss": 0.8413812175901195
+  }
+}