Upload dead feature stubs

tutorials/neuronpedia/generating_neuronpedia_outputs.ipynb

@@ -90,8 +90,7 @@
     "import os\n",
     "import requests\n",
     "\n",
-    "folder_path = \"../../neuronpedia_outputs/gpt2-small_blocks.0.hook_resid_pre_24576\" #runner.neuronpedia_folder\n",
-    "\n",
+    "FEATURE_OUTPUTS_FOLDER = runner.outputs_folder\n",
     "\n",
     "def nanToNeg999(obj: Any) -> Any:\n",
     "    if isinstance(obj, dict):\n",
@@ -110,13 +109,12 @@
     "\n",
     "# Server info\n",
     "host = \"http://localhost:3000\"\n",
-    "sourceName = str(LAYER) + \"-\" + SOURCE\n",
     "\n",
     "# Upload alive features\n",
-    "for file_name in os.listdir(folder_path):\n",
+    "for file_name in os.listdir(FEATURE_OUTPUTS_FOLDER):\n",
     "    if file_name.startswith(\"batch-\") and file_name.endswith(\".json\"):\n",
     "        print(\"Uploading file: \" + file_name)\n",
-    "        file_path = os.path.join(folder_path, file_name)\n",
+    "        file_path = os.path.join(FEATURE_OUTPUTS_FOLDER, file_name)\n",
     "        f = open(file_path, \"r\")\n",
     "        data = json.load(f)\n",
     "\n",
@@ -127,28 +125,18 @@
     "        url = host + \"/api/local/upload-features\"\n",
     "        resp = requests.post(\n",
     "            url,\n",
-    "            json={\n",
-    "                \"modelId\": MODEL,\n",
-    "                \"layer\": sourceName,\n",
-    "                \"features\": data,\n",
-    "            },\n",
+    "            json=data,\n",
     "        )\n",
     "\n",
-    "# Upload dead features (just makes blanks features)\n",
-    "# We want this for completeness\n",
-    "# skipped_path = os.path.join(folder_path, \"skipped_indexes.json\")\n",
-    "# f = open(skipped_path, \"r\")\n",
-    "# data = json.load(f)\n",
-    "# skipped_indexes = data[\"skipped_indexes\"]\n",
-    "# url = host + \"/api/internal/upload-dead-features\"\n",
-    "# resp = requests.post(\n",
-    "#     url,\n",
-    "#     json={\n",
-    "#         \"modelId\": MODEL,\n",
-    "#         \"layer\": sourceName,\n",
-    "#         \"deadIndexes\": skipped_indexes,\n",
-    "#     },\n",
-    "# )"
+    "# Upload dead feature stubs\n",
+    "skipped_path = os.path.join(FEATURE_OUTPUTS_FOLDER, \"skipped_indexes.json\")\n",
+    "f = open(skipped_path, \"r\")\n",
+    "data = json.load(f)\n",
+    "url = host + \"/api/local/upload-dead-features\"\n",
+    "resp = requests.post(\n",
+    "    url,\n",
+    "    json=data,\n",
+    ")"
    ]
   },
   {

tutorials/neuronpedia/upload_batch.py

@@ -45,19 +45,3 @@ def encode(self, o: Any, *args: Any, **kwargs: Any):
             url,
             json=data,
         )
-
-# Upload dead features (just makes blanks features)
-# We want this for completeness
-# skipped_path = os.path.join(folder_path, "skipped_indexes.json")
-# f = open(skipped_path, "r")
-# data = json.load(f)
-# skipped_indexes = data["skipped_indexes"]
-# url = host + "/api/internal/upload-dead-features"
-# resp = requests.post(
-#     url,
-#     json={
-#         "modelId": MODEL,
-#         "layer": sourceName,
-#         "deadIndexes": skipped_indexes,
-#     },
-# )

tutorials/neuronpedia/upload_dead_feature_stubs.py

@@ -0,0 +1,18 @@
+import json
+import os
+import requests
+import sys
+
+FEATURE_OUTPUTS_FOLDER = sys.argv[1]
+
+# Server info
+host = "http://localhost:3000"
+
+skipped_path = os.path.join(FEATURE_OUTPUTS_FOLDER, "skipped_indexes.json")
+f = open(skipped_path, "r")
+data = json.load(f)
+url = host + "/api/local/upload-dead-features"
+resp = requests.post(
+    url,
+    json=data,
+)

tutorials/neuronpedia/upload_dead_feature_stubs.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+
+echo "===== This will create stubs for dead features using skipped_indexes.json."
+echo "===== You'll need Neuronpedia running at localhost:3000 for this to work."
+
+echo ""
+echo "(Step 1 of 1)"
+echo "What is the absolute, full local DIRECTORY PATH to your Neuronpedia batch outputs?"
+read outputfilesdir
+
+echo ""
+read -p "===== Hit ENTER to start uploading!" start
+
+echo "RUNNING: python upload_dead_feature_stubs.py $outputfilesdir"
+python upload_dead_feature_stubs.py $outputfilesdir
+
+echo ""
+echo "===== ALL DONE."
+echo "===== Go to http://localhost:3000 to browse your features"

tutorials/neuronpedia/upload_features.sh

@@ -1,8 +1,5 @@
 #!/bin/bash
 
-# we use a script around python to work around OOM issues - this ensures every batch gets the whole system available memory
-# better fix is to investigate and fix the memory issues
-
 echo "===== This will start upload the feature batch files to Neuronpedia."
 echo "===== You'll need Neuronpedia running at localhost:3000 for this to work."