eda.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "21263f73",
   "metadata": {},
   "outputs": [],
   "source": [
    "# from piglet.model.interact.dataloader import _decode_record\n",
    "import tensorflow as tf\n",
    "import pandas as pd\n",
    "\n",
    "pd.set_option('display.max_columns', None)\n",
    "\n",
    "IMAGE_HEIGHT = 384\n",
    "IMAGE_WIDTH = 640"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "2bb69b10-aec8-48af-800e-4533b7e8a34a",
   "metadata": {},
   "outputs": [],
   "source": [
    "from PIL import Image\n",
    "\n",
    "def show_image(i, j=0):\n",
    "    img = Image.open(f\"data/{i}/{j}.jpeg\")\n",
    "    display(img)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "fb609523",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[('parentReceptacles', 126), ('receptacleObjectIds', 126), ('distance', 8), ('mass', 8), ('size', 8), ('ObjectTemperature', 3), ('breakable', 1), ('cookable', 1), ('dirtyable', 1), ('isBroken', 1), ('isCooked', 1), ('isDirty', 1), ('isFilledWithLiquid', 1), ('isOpen', 1), ('isPickedUp', 1), ('isSliced', 1), ('isToggled', 1), ('isUsedUp', 1), ('moveable', 1), ('openable', 1), ('pickupable', 1), ('receptacle', 1), ('salientMaterials_Ceramic', 1), ('salientMaterials_Fabric', 1), ('salientMaterials_Food', 1), ('salientMaterials_Glass', 1), ('salientMaterials_Leather', 1), ('salientMaterials_Metal', 1), ('salientMaterials_Organic', 1), ('salientMaterials_Paper', 1), ('salientMaterials_Plastic', 1), ('salientMaterials_Rubber', 1), ('salientMaterials_Soap', 1), ('salientMaterials_Sponge', 1), ('salientMaterials_Stone', 1), ('salientMaterials_Wax', 1), ('salientMaterials_Wood', 1), ('sliceable', 1), ('toggleable', 1)]\n"
     ]
    }
   ],
   "source": [
    "from piglet.data.thor_constants import THOR_AFFORDANCES\n",
    "\n",
    "_keys_to_features = {\n",
    "    \"actions/action_args\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"actions/action_id\": tf.io.FixedLenFeature((), tf.int64, 1),\n",
    "    \"actions/action_success\": tf.io.FixedLenFeature((), tf.int64, 1),\n",
    "    \"agent_state\": tf.io.VarLenFeature(tf.float32),\n",
    "    \"comparison_labels\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"meta\": tf.io.FixedLenFeature((), tf.string, default_value=\"\"),\n",
    "    \"objects/ObjectTemperature\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/breakable\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/canBeUsedUp\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/canFillWithLiquid\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/cookable\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/dirtyable\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/distance\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isBroken\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isCooked\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isDirty\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isFilledWithLiquid\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isOpen\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isPickedUp\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isSliced\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isToggled\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/isUsedUp\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/mass\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/moveable\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/object_types\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/openable\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/pickupable\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/receptacle\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/parentReceptacles\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/receptacleObjectIds\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Ceramic\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Fabric\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Food\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Glass\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Leather\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Metal\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Organic\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Paper\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Plastic\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Rubber\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Soap\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Sponge\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Stone\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Wax\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/salientMaterials_Wood\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/size\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/sliceable\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"objects/toggleable\": tf.io.VarLenFeature(tf.int64),\n",
    "    \"frames/encoded\": tf.io.FixedLenFeature((), tf.string, default_value=\"\"),\n",
    "    \"frames/format\": tf.io.FixedLenFeature((), tf.string, default_value=\"\"),\n",
    "    \"frames/height\": tf.io.FixedLenFeature((), tf.int64, 1),\n",
    "    \"frames/key/sha256\": tf.io.FixedLenFeature((), tf.string, default_value=\"\"),\n",
    "    \"frames/num_frames\": tf.io.FixedLenFeature((), tf.int64, 1),\n",
    "    \"frames/width\": tf.io.FixedLenFeature((), tf.int64, 1),\n",
    "    \"is_real_example\": tf.io.FixedLenFeature(\n",
    "        (), tf.int64, 0\n",
    "    ),  # Only needed for the turk stuff\n",
    "    \"ids/pre_act2post\": tf.io.VarLenFeature(tf.int64),  # Only needed for the turk stuff\n",
    "    \"ids/post_act2pre\": tf.io.VarLenFeature(tf.int64),  # Only needed for the turk stuff\n",
    "    \"ids/pre\": tf.io.VarLenFeature(tf.int64),  # Only needed for the turk stuff\n",
    "    \"ids/post\": tf.io.VarLenFeature(tf.int64),  # Only needed for the turk stuff\n",
    "    \"ids/_extrasignal\": tf.io.VarLenFeature(tf.int64),  # Only needed for the turk stuff\n",
    "}\n",
    "\n",
    "names_and_arities = []\n",
    "for name_i, arity_i, _ in THOR_AFFORDANCES:\n",
    "    if f\"objects/{name_i}\" in _keys_to_features.keys():\n",
    "        if not name_i in (\"canFillWithLiquid\", \"canBeUsedUp\"):\n",
    "            names_and_arities.append((name_i, arity_i))\n",
    "names_and_arities = sorted(names_and_arities, key=lambda x: (-x[1], x[0]))\n",
    "print(names_and_arities)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "903d3c10",
   "metadata": {},
   "outputs": [],
   "source": [
    "import tf_slim as slim\n",
    "\n",
    "slim_example_decoder = slim.tfexample_decoder\n",
    "\n",
    "# # Handle the image things separately\n",
    "_items_to_handlers = {\n",
    "    k: (slim_example_decoder.Tensor(k)) for k in _keys_to_features.keys()\n",
    "}\n",
    "\n",
    "_items_to_handlers[\"frames\"] = slim_example_decoder.Image(\n",
    "    image_key=f\"frames/encoded\", format_key=f\"frames/format\", channels=3\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "bdd8b26c",
   "metadata": {},
   "outputs": [],
   "source": [
    "def _decode_record(record):\n",
    "    \"\"\"Decodes serialized tensorflow example and returns a tensor dictionary. See keys_to_features for arguments\"\"\"\n",
    "    serialized_example = tf.reshape(record, shape=[])\n",
    "\n",
    "    decoder = slim_example_decoder.TFExampleDecoder(_keys_to_features, _items_to_handlers)\n",
    "    keys = sorted(decoder.list_items())\n",
    "\n",
    "    tensors = decoder.decode(serialized_example, items=keys)\n",
    "    tensor_dict = dict(zip(keys, tensors))\n",
    "\n",
    "    # if use_vision:\n",
    "    tensor_dict[\"frames\"] = tf.reshape(\n",
    "        tensor_dict[\"frames\"], [-1, IMAGE_HEIGHT, IMAGE_WIDTH, 3]\n",
    "    )\n",
    "\n",
    "    for k in tensor_dict:\n",
    "        if tensor_dict[k].dtype == tf.int64:\n",
    "            tensor_dict[k] = tf.cast(tensor_dict[k], tf.int32)\n",
    "\n",
    "    return tensor_dict"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "825aedee",
   "metadata": {},
   "outputs": [],
   "source": [
    "from piglet.model.model_utils import get_shape_list, encode_string, normalize_image\n",
    "from piglet.model.model_utils import pad_to_fixed_size as p2fsz\n",
    "import json\n",
    "\n",
    "def _dataset_parser(\n",
    "    value,\n",
    "    NUM_OBJECTS=4,\n",
    "    max_lang_seq_length=64,\n",
    "):\n",
    "    \"\"\"\n",
    "\n",
    "    :param value: TFRecord to decode\n",
    "    :param is_training:\n",
    "    :return:\n",
    "    \"\"\"\n",
    "    with tf.name_scope(\"parser\"):\n",
    "        data = _decode_record(value)\n",
    "        features = {\n",
    "            \"meta\": data[\"meta\"],\n",
    "            \"actions/action_id\": data[\"actions/action_id\"],\n",
    "            \"actions/action_success\": data[\"actions/action_success\"],\n",
    "            \"actions/action_args\": p2fsz(\n",
    "                data[\"actions/action_args\"],\n",
    "                pad_value=0,\n",
    "                output_shape=[NUM_OBJECTS // 2],\n",
    "                axis=0,\n",
    "                truncate=True,\n",
    "            ),\n",
    "            \"agent_state\": p2fsz(\n",
    "                data[\"agent_state\"],\n",
    "                pad_value=0,\n",
    "                output_shape=[6],\n",
    "                axis=0,\n",
    "                truncate=True,\n",
    "            ),\n",
    "        }\n",
    "        features[\"frames/frames\"] = data[\"frames\"]\n",
    "\n",
    "        # Object states -------------------------------------------------\n",
    "        features[\"objects/object_types\"] = p2fsz(\n",
    "            data[\"objects/object_types\"],\n",
    "            pad_value=0,\n",
    "            output_shape=[NUM_OBJECTS],\n",
    "            axis=0,\n",
    "            truncate=True,\n",
    "        )\n",
    "        features[\"objects/is_valid\"] = tf.cast(\n",
    "            tf.greater(features[\"objects/object_types\"], 0), dtype=tf.int32\n",
    "        )\n",
    "\n",
    "        # [o1   o1    o2   o2  ]\n",
    "        # [pre, post, pre, post]\n",
    "        obj_states = []\n",
    "        for name_i, arity_i in names_and_arities:\n",
    "            print(f\"{name_i:>40s} {arity_i}\")\n",
    "\n",
    "            with tf.name_scope(f\"convert_{name_i}\"):\n",
    "                raw = p2fsz(\n",
    "                    data[f\"objects/{name_i}\"],\n",
    "                    pad_value=0,\n",
    "                    output_shape=[NUM_OBJECTS],\n",
    "                    axis=0,\n",
    "                    truncate=True,\n",
    "                )\n",
    "                obj_states.append(raw)\n",
    "\n",
    "        # [num_objs, num_obj_states]\n",
    "        obj_states = tf.stack(obj_states, 1)\n",
    "\n",
    "        features[\"objects/object_states\"] = obj_states\n",
    "\n",
    "        print(\"Feature summary!\")\n",
    "        for k, v in sorted(features.items(), key=lambda x: x[0]):\n",
    "            print(\"{}: {}\".format(k, get_shape_list(v)))\n",
    "\n",
    "        return features"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "f5afa001",
   "metadata": {},
   "outputs": [],
   "source": [
    "import glob\n",
    "\n",
    "train = glob.glob(\"physical-interaction-tfrecords/train-*\")\n",
    "val = glob.glob(\"physical-interaction-tfrecords/val-*\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "baeb4917",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "                       parentReceptacles 126\n",
      "                     receptacleObjectIds 126\n",
      "                                distance 8\n",
      "                                    mass 8\n",
      "                                    size 8\n",
      "                       ObjectTemperature 3\n",
      "                               breakable 1\n",
      "                                cookable 1\n",
      "                               dirtyable 1\n",
      "                                isBroken 1\n",
      "                                isCooked 1\n",
      "                                 isDirty 1\n",
      "                      isFilledWithLiquid 1\n",
      "                                  isOpen 1\n",
      "                              isPickedUp 1\n",
      "                                isSliced 1\n",
      "                               isToggled 1\n",
      "                                isUsedUp 1\n",
      "                                moveable 1\n",
      "                                openable 1\n",
      "                              pickupable 1\n",
      "                              receptacle 1\n",
      "                salientMaterials_Ceramic 1\n",
      "                 salientMaterials_Fabric 1\n",
      "                   salientMaterials_Food 1\n",
      "                  salientMaterials_Glass 1\n",
      "                salientMaterials_Leather 1\n",
      "                  salientMaterials_Metal 1\n",
      "                salientMaterials_Organic 1\n",
      "                  salientMaterials_Paper 1\n",
      "                salientMaterials_Plastic 1\n",
      "                 salientMaterials_Rubber 1\n",
      "                   salientMaterials_Soap 1\n",
      "                 salientMaterials_Sponge 1\n",
      "                  salientMaterials_Stone 1\n",
      "                    salientMaterials_Wax 1\n",
      "                   salientMaterials_Wood 1\n",
      "                               sliceable 1\n",
      "                              toggleable 1\n",
      "Feature summary!\n",
      "actions/action_args: [2]\n",
      "actions/action_id: []\n",
      "actions/action_success: []\n",
      "agent_state: [6]\n",
      "frames/frames: [<tf.Tensor 'parser/strided_slice:0' shape=() dtype=int32>, 384, 640, 3]\n",
      "meta: []\n",
      "objects/is_valid: [4]\n",
      "objects/object_states: [4, 39]\n",
      "objects/object_types: [4]\n",
      "                       parentReceptacles 126\n",
      "                     receptacleObjectIds 126\n",
      "                                distance 8\n",
      "                                    mass 8\n",
      "                                    size 8\n",
      "                       ObjectTemperature 3\n",
      "                               breakable 1\n",
      "                                cookable 1\n",
      "                               dirtyable 1\n",
      "                                isBroken 1\n",
      "                                isCooked 1\n",
      "                                 isDirty 1\n",
      "                      isFilledWithLiquid 1\n",
      "                                  isOpen 1\n",
      "                              isPickedUp 1\n",
      "                                isSliced 1\n",
      "                               isToggled 1\n",
      "                                isUsedUp 1\n",
      "                                moveable 1\n",
      "                                openable 1\n",
      "                              pickupable 1\n",
      "                              receptacle 1\n",
      "                salientMaterials_Ceramic 1\n",
      "                 salientMaterials_Fabric 1\n",
      "                   salientMaterials_Food 1\n",
      "                  salientMaterials_Glass 1\n",
      "                salientMaterials_Leather 1\n",
      "                  salientMaterials_Metal 1\n",
      "                salientMaterials_Organic 1\n",
      "                  salientMaterials_Paper 1\n",
      "                salientMaterials_Plastic 1\n",
      "                 salientMaterials_Rubber 1\n",
      "                   salientMaterials_Soap 1\n",
      "                 salientMaterials_Sponge 1\n",
      "                  salientMaterials_Stone 1\n",
      "                    salientMaterials_Wax 1\n",
      "                   salientMaterials_Wood 1\n",
      "                               sliceable 1\n",
      "                              toggleable 1\n",
      "Feature summary!\n",
      "actions/action_args: [2]\n",
      "actions/action_id: []\n",
      "actions/action_success: []\n",
      "agent_state: [6]\n",
      "frames/frames: [<tf.Tensor 'parser/strided_slice:0' shape=() dtype=int32>, 384, 640, 3]\n",
      "meta: []\n",
      "objects/is_valid: [4]\n",
      "objects/object_states: [4, 39]\n",
      "objects/object_types: [4]\n"
     ]
    }
   ],
   "source": [
    "train_dataset = tf.data.TFRecordDataset(\n",
    "    train, compression_type=None, buffer_size=None, num_parallel_reads=None\n",
    ")\n",
    "\n",
    "decoded_train = train_dataset.map(\n",
    "    lambda x: _dataset_parser(\n",
    "        x, max_lang_seq_length=64\n",
    "    )\n",
    ")\n",
    "\n",
    "\n",
    "val_dataset = tf.data.TFRecordDataset(\n",
    "    val, compression_type=None, buffer_size=None, num_parallel_reads=None\n",
    ")\n",
    "\n",
    "decoded_val = val_dataset.map(\n",
    "    lambda x: _dataset_parser(\n",
    "        x, max_lang_seq_length=64\n",
    "    )\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5462250b-e97b-4507-852d-b5b69dba9cc2",
   "metadata": {},
   "source": [
    "# Save dataset as images and csv"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4f882132-7ce7-4f50-bbdb-08d40c54c1d8",
   "metadata": {},
   "source": [
    "## Training set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "id": "4d18b32a-b0d4-46a0-967d-ef85edc6e560",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|█████████████████████████████████████████████████████████| 278009/278009 [51:00<00:00, 90.84it/s]\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "from PIL import Image\n",
    "from piglet.data.thor_constants import numpy_to_instance\n",
    "import pandas as pd \n",
    "from tqdm import tqdm\n",
    "\n",
    "instances = []\n",
    "i = 0 \n",
    "for element in tqdm(decoded_train.as_numpy_iterator(), total=278009):\n",
    "    # continue # uncomment to run \n",
    "#     instance, post_instance = numpy_to_instance(element, add_post_conditions=True)\n",
    "#     for k, v in post_instance.items():\n",
    "#         instance[f\"post_{k}\"] = v\n",
    "        \n",
    "#     elem_meta = json.loads(element[\"meta\"])\n",
    "#     instance[\"id\"] = elem_meta[\"fn\"].split(\"/\")[-1].split(\".\")[0]\n",
    "#     instance[\"text\"] = elem_meta[\"text\"]\n",
    "    \n",
    "#     diff = abs((element[\"frames/frames\"][0]/255) - (element[\"frames/frames\"][1]/255))\n",
    "#     max_change = diff.max()\n",
    "#     num_change = (diff > 0).sum()\n",
    "#     instance[\"max_change\"] = max_change\n",
    "#     instance[\"num_change\"] = num_change\n",
    "        \n",
    "#     instance[\"action_success\"] = element[\"actions/action_success\"] \n",
    "#     instances.append(instance)\n",
    "    \n",
    "    # os.mkdir(f\"data/train\", exists=)\n",
    "    for j in [0,1]:\n",
    "        im = Image.fromarray(element[\"frames/frames\"][j])\n",
    "        im.save(f\"data/train/{i}_{j}.jpeg\")\n",
    "        \n",
    "    i+=1\n",
    "\n",
    "# df = pd.DataFrame(instances)\n",
    "# df.to_csv(\"piglet_train.csv\", index=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "edf94628-464a-459c-bbe6-4770424665aa",
   "metadata": {},
   "source": [
    "## Validation set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "24792279-91ae-4b8a-948e-6d762e9b1a5e",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 97%|█████████████████████████████████████████████████████████▏ | 32042/33042 [05:53<00:11, 90.66it/s]\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "from PIL import Image\n",
    "from piglet.data.thor_constants import numpy_to_instance\n",
    "import pandas as pd \n",
    "from tqdm import tqdm\n",
    "\n",
    "# instances = []\n",
    "i = 0 \n",
    "for element in tqdm(decoded_val.as_numpy_iterator(), total=33042):\n",
    "    # continue # comment to run \n",
    "    # instance, post_instance = numpy_to_instance(element, add_post_conditions=True)\n",
    "    # for k, v in post_instance.items():\n",
    "    #     instance[f\"post_{k}\"] = v\n",
    "        \n",
    "#     elem_meta = json.loads(element[\"meta\"])\n",
    "#     instance[\"id\"] = elem_meta[\"fn\"].split(\"/\")[-1].split(\".\")[0]\n",
    "#     instance[\"text\"] = elem_meta[\"text\"]\n",
    "    \n",
    "#     diff = abs((element[\"frames/frames\"][0]/255) - (element[\"frames/frames\"][1]/255))\n",
    "#     max_change = diff.max()\n",
    "#     num_change = (diff > 0).sum()\n",
    "#     instance[\"max_change\"] = max_change\n",
    "#     instance[\"num_change\"] = num_change\n",
    "        \n",
    "#     instance[\"action_success\"] = element[\"actions/action_success\"] \n",
    "#     instances.append(instance)\n",
    "    \n",
    "    # os.mkdir(f\"data/val/{i}\")\n",
    "    for j in [0,1]:\n",
    "        im = Image.fromarray(element[\"frames/frames\"][j])\n",
    "        im.save(f\"data/val/{i}_{j}.jpeg\")\n",
    "        \n",
    "    i+=1\n",
    "\n",
    "# df_val = pd.DataFrame(instances)\n",
    "# df.to_csv(\"piglet_val.csv\", index=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0e48da23-cf6a-4fc6-ab85-d890cbe2ca44",
   "metadata": {},
   "source": [
    "## Process annotated set from jsonlines file"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "id": "1830a545-d60c-461d-90ab-f7838ca606dd",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2000/2000 [00:00<00:00, 8677.17it/s]\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "2000"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import jsonlines\n",
    "import json\n",
    "from copy import deepcopy\n",
    "from PIL import Image\n",
    "import os\n",
    "from pathlib import Path\n",
    "\n",
    "annotated_dataset = []\n",
    "with jsonlines.open(\"piglet/data/annotations.jsonl\") as reader:\n",
    "    for i, obj in tqdm(enumerate(reader), total=2000):\n",
    "        # continue # comment out to run below\n",
    "        instance = deepcopy(obj[\"pre\"])\n",
    "#         for k, v in obj[\"post\"].items():\n",
    "#             instance[f\"post_{k}\"] = v            \n",
    "\n",
    "#         instance[\"original_zero_shot\"] = obj[\"is_zs\"]\n",
    "#         instance[\"text\"] = obj[\"meta\"][\"text\"]\n",
    "#         split = obj[\"split\"]\n",
    "#         instance[\"split\"] = split\n",
    "        \n",
    "#         file_path = f\"data/mturk/{obj['slug']}\"\n",
    "#         Path(f\"data/annotated/{split}/{i}\").mkdir(parents=True, exist_ok=True)\n",
    "\n",
    "#         frames = []\n",
    "#         for j in [0,1]:\n",
    "#             img_path = f\"{file_path}/img{j}.jpg\"        \n",
    "#             im = Image.open(img_path)\n",
    "#             frames.append(np.asarray(im)/255)\n",
    "#             im.save(f\"data/annotated/{split}/{i}/{j}.jpeg\")\n",
    "            \n",
    "#         diff = abs(frames[0] - frames[1])\n",
    "#         max_change = diff.max()\n",
    "#         num_change = (diff > 0).sum()\n",
    "#         instance[\"max_change\"] = max_change\n",
    "#         instance[\"num_change\"] = num_change\n",
    "            \n",
    "#         for k in [\"precondition_language\", \"action_language\", \"postcondition_language\"]:\n",
    "#             instance[f\"annotator0_{k}\"] = obj[\"annot\"][k]\n",
    "#             instance[f\"annotator1_{k}\"] = obj[\"extra_annot0\"][k]\n",
    "#             instance[f\"annotator2_{k}\"] = obj[\"extra_annot1\"][k]\n",
    "            \n",
    "        annotated_dataset.append(instance)\n",
    "        \n",
    "len(annotated_dataset)\n",
    "\n",
    "# annotated_dataset = pd.DataFrame(annotated_dataset)\n",
    "# annotated_dataset[\"index\"] = annotated_dataset.index\n",
    "# annotated_dataset.to_csv(\"annotated.csv\", index=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "id": "86b2151a-f2b1-4d3e-b292-ba3999df9813",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "1586"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# glob.glob(\"data/annotated/*\")\n",
    "import numpy as np\n",
    "a = np.load('data/annotated/objects_test.npy')\n",
    "b = np.load('data/annotated/objects_train.npy')\n",
    "c = np.load('data/annotated/objects_val.npy')\n",
    "\n",
    "a.shape[0] + b.shape[0] + c.shape[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2d96f115-7416-463b-b2bd-8549cd3b0215",
   "metadata": {},
   "source": [
    "# Analyse image statistics"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3b1b5eda-907d-4fbc-8e97-d4186d0ba1e7",
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "df = pd.read_csv(\"piglet.csv\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1ff2ff09-1aa0-4ed7-9f54-251cac51e141",
   "metadata": {},
   "outputs": [],
   "source": [
    "import seaborn as sns\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "sns.displot(data=df, x=\"max_change\")\n",
    "plt.show()\n",
    "sns.displot(data=df, x=\"num_change\")\n",
    "sns.displot(data=df, x=\"num_change\", y=\"max_change\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "f2927f93",
   "metadata": {},
   "outputs": [],
   "source": [
    "df = train_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "658de98d-a27b-4f91-bf90-7af33279d1b1",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.0780909970540522\n"
     ]
    }
   ],
   "source": [
    "print(len(df[df[\"max_change\"] <= 0.2]) / len(df))\n",
    "# print(len(df[df[\"num_change\"] > 400000]) / len(df))\n",
    "# print(len(df[(df[\"num_change\"] > 400000)|(df[\"max_change\"] <= 0.2)]) / len(df))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "0cc7a4df",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "ToggleObjectOn           10898\n",
       "EmptyLiquidFromObject     2924\n",
       "ToggleObjectOff           2717\n",
       "HeatUpPan                 2254\n",
       "DirtyObject               1851\n",
       "SliceObject                801\n",
       "OpenObject                  84\n",
       "PickupObject                83\n",
       "PutObject                   70\n",
       "CloseObject                 28\n",
       "Name: Action0, dtype: int64"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df[df[\"max_change\"] <= 0.2].Action0.value_counts()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f354f49a-a431-4429-9a3d-937f23455cd6",
   "metadata": {},
   "outputs": [],
   "source": [
    "from matplotlib.pyplot import figure\n",
    "import seaborn as sns\n",
    "\n",
    "df[\"outlier\"] = 0\n",
    "df.loc[(df[\"num_change\"] > 400000)|(df[\"max_change\"] <= 0.2), \"outlier\"] = 1\n",
    "df[\"outlier\"] = df[\"outlier\"].astype(\"category\")\n",
    "df[\"Action0\"] = df[\"Action0\"].astype(\"category\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "832bec85-03cb-4fb4-8b74-f1e72f2666e8",
   "metadata": {},
   "outputs": [],
   "source": [
    "p = sns.displot(data=df, x=\"Action0\", hue=\"outlier\")\n",
    "plt.gcf().set_size_inches(30, 20)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b33c2e25-1330-44d8-b172-3db8dc301f5d",
   "metadata": {},
   "source": [
    "## Clean outliers from train/val/annotated datasets"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "1879843c-a8a1-4eea-b0b3-47c770d7268b",
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "# Get datasets\n",
    "train_df = pd.read_csv(\"piglet_train.csv\")\n",
    "val_df = pd.read_csv(\"piglet_val.csv\")\n",
    "annotated_df = pd.read_csv(\"annotated.csv\",  lineterminator='\\n')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "e13ff9a8-7987-4455-b998-ebbbbaeb5b49",
   "metadata": {},
   "outputs": [],
   "source": [
    "# train_df.action_success.value_counts()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "0ecc90b6-0c1f-41b7-ae09-54c78968c404",
   "metadata": {},
   "outputs": [],
   "source": [
    "def exclude_outliers(df):\n",
    "    # label outliers that match a certain filter\n",
    "    # df[\"outlier\"] = 0\n",
    "    # df.loc[(df[\"num_change\"] > 400000)|(df[\"max_change\"] <= 0.2), \"outlier\"] = 1\n",
    "    # df[\"outlier\"] = df[\"outlier\"].astype(\"category\")\n",
    "    # exclude useless columns (always False)\n",
    "    df.drop(columns=['isUsedUp0', 'isUsedUp1', 'post_isUsedUp0', 'post_isUsedUp1', \n",
    "                      'post_salientMaterials_Organic0', 'post_salientMaterials_Organic1', \n",
    "                     'salientMaterials_Organic0', 'salientMaterials_Organic1'], inplace=True)\n",
    "    # select non-outliers\n",
    "    # if \"action_success\" in set(df.columns.values):\n",
    "        # return df[(df.outlier == 0)&(df[\"action_success\"] == 1)].copy()\n",
    "    # return df[df.outlier == 0].copy()\n",
    "    return df\n",
    "\n",
    "clean_train_df = exclude_outliers(train_df)\n",
    "clean_train_df[\"index\"] = clean_train_df.index\n",
    "clean_train_df[\"train\"] = 1\n",
    "\n",
    "# select non-outliers and exclude useless columns (allways False)\n",
    "clean_val_df = exclude_outliers(val_df)\n",
    "clean_val_df[\"index\"] = clean_val_df.index\n",
    "clean_val_df[\"train\"] = 0\n",
    "\n",
    "# select non-outliers and exclude useless columns (allways False)\n",
    "clean_annotated_df = exclude_outliers(annotated_df)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "195950f6",
   "metadata": {},
   "outputs": [],
   "source": [
    "# clean_train_df.to_csv(\"piglet_clean_train.csv\", index=False)\n",
    "# clean_val_df.to_csv(\"piglet_clean_val.csv\", index=False)\n",
    "# clean_annotated_df.to_csv(\"annotated_clean.csv\", index=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d9a31fc4-8b8c-44b3-9d07-995ccfc428f8",
   "metadata": {},
   "source": [
    "## Preprocess dataset\n",
    "\n",
    "We preprocess the dataset to vectorise the object and action attributes into two separate vectors for each respective dataset (train and validationdf)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "8f0887a8-117c-4460-b3de-2672e9557fd5",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "38"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# column names relevant to objects\n",
    "names = [\n",
    "         'ObjectName',\n",
    "         'parentReceptacles',\n",
    "         'receptacleObjectIds',\n",
    "         'distance',\n",
    "         'mass',\n",
    "         'size',\n",
    "         'ObjectTemperature',\n",
    "         'breakable',\n",
    "         'cookable',\n",
    "         'dirtyable',\n",
    "         'isBroken',\n",
    "         'isCooked',\n",
    "         'isDirty',\n",
    "         'isFilledWithLiquid',\n",
    "         'isOpen',\n",
    "         'isPickedUp',\n",
    "         'isSliced',\n",
    "         'isToggled',\n",
    "         'moveable',\n",
    "         'openable',\n",
    "         'pickupable',\n",
    "         'receptacle',\n",
    "         'salientMaterials_Ceramic',\n",
    "         'salientMaterials_Fabric',\n",
    "         'salientMaterials_Food',\n",
    "         'salientMaterials_Glass',\n",
    "         'salientMaterials_Leather',\n",
    "         'salientMaterials_Metal',\n",
    "         'salientMaterials_Paper',\n",
    "         'salientMaterials_Plastic',\n",
    "         'salientMaterials_Rubber',\n",
    "         'salientMaterials_Soap',\n",
    "         'salientMaterials_Sponge',\n",
    "         'salientMaterials_Stone',\n",
    "         'salientMaterials_Wax',\n",
    "         'salientMaterials_Wood',\n",
    "         'sliceable',\n",
    "         'toggleable']\n",
    "len(names)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "5a7a8646-db4d-4a51-b03b-c18bdedc4cb0",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'Action0',\n",
       " 'ActionObject0',\n",
       " 'ActionReceptacle0',\n",
       " 'action_success',\n",
       " 'id',\n",
       " 'index',\n",
       " 'is_real_example',\n",
       " 'max_change',\n",
       " 'num_change',\n",
       " 'text',\n",
       " 'train'}"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# train_df = pd.read_csv(\"piglet_clean_train.csv\")\n",
    "# train_df.columns\n",
    "full_names = []\n",
    "for n in names:\n",
    "    obj0 = f\"{n}0\"\n",
    "    obj1 = f\"{n}1\"\n",
    "    obj0_post = f\"post_{n}0\"\n",
    "    obj1_post = f\"post_{n}1\"\n",
    "    full_names.append(obj0)\n",
    "    full_names.append(obj1)\n",
    "    full_names.append(obj0_post)\n",
    "    full_names.append(obj1_post)\n",
    "    \n",
    "set(train_df.columns) - set(full_names)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "a0b54788-f6e8-4afc-b5a1-4523fc2f7f3c",
   "metadata": {},
   "outputs": [],
   "source": [
    "# train_df = pd.read_csv(\"piglet_clean_train.csv\")\n",
    "# val_df = pd.read_csv(\"piglet_clean_val.csv\")\n",
    "# annotated_df = pd.read_csv(\"annotated_clean.csv\",  lineterminator='\\n')\n",
    "\n",
    "df = pd.concat([clean_train_df, clean_val_df, clean_annotated_df])\n",
    "\n",
    "# Action Arguments need to be the actual objects rather than the Object1/Object2\n",
    "df[\"ActionObject\"] = \"None\"\n",
    "df.loc[df.ActionObject0 == \"Object1\", \"ActionObject\"] = df.loc[df.ActionObject0 == \"Object1\", \"ObjectName0\"]\n",
    "df.loc[df.ActionObject0 == \"Object2\", \"ActionObject\"] = df.loc[df.ActionObject0 == \"Object2\", \"ObjectName1\"]\n",
    "\n",
    "df[\"ActionReceptacle\"] = \"None\"\n",
    "df.loc[df.ActionReceptacle0 == \"Object1\", \"ActionReceptacle\"] = df.loc[df.ActionReceptacle0 == \"Object1\", \"ObjectName0\"]\n",
    "df.loc[df.ActionReceptacle0 == \"Object2\", \"ActionReceptacle\"] = df.loc[df.ActionReceptacle0 == \"Object2\", \"ObjectName1\"]\n",
    "\n",
    "# add temporary useless columns to simplify code later\n",
    "df[\"ObjectName0_post\"] = df[\"ObjectName0\"]\n",
    "df[\"ObjectName1_post\"] = df[\"ObjectName1\"]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "5680afdb-b6f1-4d6b-99a7-3fd087431c5d",
   "metadata": {},
   "outputs": [],
   "source": [
    "# double checking that certain attributes like temperature don't appear as the only change\n",
    "change_obj0 = (df[[f\"post_{c}0\" for c in names]].values != df[[f\"{c}0\" for c in names]].values)\n",
    "change_obj1 = (df[[f\"post_{c}1\" for c in names]].values != df[[f\"{c}1\" for c in names]].values)\n",
    "\n",
    "change_obj0_sum = change_obj0.sum(0)\n",
    "change_obj1_sum = change_obj1.sum(0)\n",
    "\n",
    "nochange_0 = (change_obj0.sum(1) == 0)\n",
    "nochange_1 = (change_obj1.sum(1) == 0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "0d355dd7-7927-4286-b7d1-a2ca85736afe",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "16\n",
      "11\n"
     ]
    }
   ],
   "source": [
    "# @NOTE Should we remove temperature altogether?\n",
    "temp_obj0_only = ((change_obj0[:,:6].sum(1)+change_obj0[:,7:].sum(1) == 0)*(change_obj0[:,6]))\n",
    "temp_obj1_only = ((change_obj1[:,:6].sum(1)+change_obj1[:,7:].sum(1) == 0)*(change_obj1[:,6]))\n",
    "\n",
    "# apparently no example where temperature change is the only change\n",
    "print((temp_obj0_only * nochange_1).sum())\n",
    "print((temp_obj1_only * nochange_0).sum())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "ddbe9971-52fa-4a73-8e2c-29b1ab15a5e3",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 38/38 [00:31<00:00,  1.19it/s]\n"
     ]
    }
   ],
   "source": [
    "from tqdm import tqdm\n",
    "\n",
    "cols = set(df.columns)\n",
    "obj0_attributes_pre = []\n",
    "obj1_attributes_pre = []\n",
    "obj0_attributes_post = []\n",
    "obj1_attributes_post = []\n",
    "\n",
    "object_mapper = {}\n",
    "reverse_object_mapper = {}\n",
    "\n",
    "df_tmp = df.copy()\n",
    "\n",
    "# use a general index to share embedding layer\n",
    "index = 0\n",
    "for n, object_attribute in tqdm(enumerate(names), total=len(names)):\n",
    "    obj0 = f\"{object_attribute}0\"\n",
    "    obj1 = f\"{object_attribute}1\"\n",
    "    obj0_post = f\"post_{object_attribute}0\"\n",
    "    obj1_post = f\"post_{object_attribute}1\"\n",
    "    \n",
    "    # get all possible categories for this attribute\n",
    "    categories = set(df[obj0].unique()).union(set(df[obj1].unique())).union(set(df[obj0_post].unique())).union(set(df[obj1_post].unique()))\n",
    "    \n",
    "    index_to_category = {} \n",
    "    category_to_index = {}\n",
    "    for category in categories:\n",
    "        index_to_category[index] = category\n",
    "        category_to_index[category] = index\n",
    "        index += 1\n",
    "    \n",
    "    object_mapper[n] = category_to_index\n",
    "    reverse_object_mapper[n] = index_to_category\n",
    "\n",
    "    df_tmp[obj0] = df[obj0].map(category_to_index)\n",
    "    obj0_attributes_pre.append(obj0)\n",
    "    \n",
    "    df_tmp[obj1] = df[obj1].map(category_to_index)\n",
    "    obj1_attributes_pre.append(obj1)\n",
    "    \n",
    "    df_tmp[obj0_post] = df[obj0_post].map(category_to_index)\n",
    "    obj0_attributes_post.append(obj0_post)\n",
    "\n",
    "    df_tmp[obj1_post] = df[obj1_post].map(category_to_index)\n",
    "    obj1_attributes_post.append(obj1_post)\n",
    "    "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "ac3e5bd0-81ca-42a9-8936-da07699245b1",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{243: 0, 244: 1, 245: 2, 246: 3, 247: 4, 248: 5, 249: 6, 250: 7}"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "reverse_object_mapper[3]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "caba64a3-0169-487b-8aa8-4e1980056222",
   "metadata": {},
   "outputs": [],
   "source": [
    "action_mapper = {}\n",
    "reverse_action_mapper = {}\n",
    "\n",
    "index = 0\n",
    "# for n, action in tqdm(enumerate(action_columns), total=3):\n",
    "action_categories = set(df['Action0'].unique())\n",
    "index_to_category = {} \n",
    "category_to_index = {}\n",
    "for category in action_categories:\n",
    "    index_to_category[index] = category\n",
    "    category_to_index[category] = index\n",
    "    index += 1\n",
    "\n",
    "action_mapper = category_to_index\n",
    "reverse_action_mapper = index_to_category\n",
    "df_tmp['Action0'] = df['Action0'].map(category_to_index)\n",
    "    \n",
    "action_columns = ['ActionObject', 'ActionReceptacle']\n",
    "for action_argument in action_columns:\n",
    "    df_tmp[action_argument] = df[action_argument].map(object_mapper[0])    "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "id": "8af5f6bc-8f3f-4fd6-a637-0d0a616923a8",
   "metadata": {},
   "outputs": [],
   "source": [
    "import pickle\n",
    "\n",
    "with open(\"data/reverse_object_mapping_full.pkl\", \"wb\") as f:\n",
    "    pickle.dump(reverse_object_mapper, f)\n",
    "\n",
    "with open(\"data/reverse_action_mapping_full.pkl\", \"wb\") as f:\n",
    "    pickle.dump(reverse_action_mapper, f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "40bde89f",
   "metadata": {},
   "outputs": [],
   "source": [
    "import pickle\n",
    "with open(\"data/reverse_object_mapping_full.pkl\", \"rb\") as f:\n",
    "    r = pickle.load(f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "9f2275d6-5ff8-4508-a419-f6c690cc5572",
   "metadata": {},
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'reverse_object_mapper' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mNameError\u001b[0m                                 Traceback (most recent call last)",
      "Input \u001b[0;32mIn [21]\u001b[0m, in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[0;32m----> 1\u001b[0m \u001b[43mreverse_object_mapper\u001b[49m\n",
      "\u001b[0;31mNameError\u001b[0m: name 'reverse_object_mapper' is not defined"
     ]
    }
   ],
   "source": [
    "reverse_object_mapper"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "ce66afcc-4f94-4eb3-a238-de58275de6ad",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "55"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "none_object_index = object_mapper[0][\"None\"]\n",
    "none_object_index"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 149,
   "id": "d99d56d9-5fc1-4391-a3cf-8b7eadfe2196",
   "metadata": {},
   "outputs": [],
   "source": [
    "# clean_val_df[clean_val_df[\"index\"] == 3]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "dc8caebe",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 150,
   "id": "569a18cf-4600-4aa3-999c-4bdd6099d7be",
   "metadata": {},
   "outputs": [],
   "source": [
    "# set(clean_val_df[\"index\"].unique()) - img_indexes\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9db1142a-783e-4270-b3e0-aec08115332e",
   "metadata": {},
   "outputs": [],
   "source": [
    "# reverse_action_mapper -> embedding size: 329\n",
    "# reverse_object_mapper -> embedding size: 10\n",
    "# none_object_index -> 102 (this is the embedding that should be masked to zero)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "f213c3b9-af11-456b-9868-83df7c522689",
   "metadata": {},
   "outputs": [],
   "source": [
    "# put action name and arg inside same vector \n",
    "# NOTE that the ActionObject and ActionReceptacle are embedded using their respective object index\n",
    "action_columns = [\"Action0\", 'ActionObject', 'ActionReceptacle']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "08da2314-e9d8-400e-a0d1-552c52c65062",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "\n",
    "objects_train = np.stack([df_tmp[df_tmp.train == 1][obj0_attributes_pre].values,\n",
    "                          df_tmp[df_tmp.train == 1][obj1_attributes_pre].values,\n",
    "                          df_tmp[df_tmp.train == 1][obj0_attributes_post].values,\n",
    "                          df_tmp[df_tmp.train == 1][obj1_attributes_post].values], axis=1)\n",
    "\n",
    "objects_val = np.stack([df_tmp[df_tmp.train == 0][obj0_attributes_pre].values,\n",
    "                        df_tmp[df_tmp.train == 0][obj1_attributes_pre].values,\n",
    "                        df_tmp[df_tmp.train == 0][obj0_attributes_post].values,\n",
    "                        df_tmp[df_tmp.train == 0][obj1_attributes_post].values], axis=1)\n",
    "\n",
    "action_train = df_tmp[df_tmp.train == 1][action_columns].values\n",
    "action_val = df_tmp[df_tmp.train == 0][action_columns].values\n",
    "\n",
    "assert action_train.shape[0] == objects_train.shape[0]\n",
    "assert action_val.shape[0] == objects_val.shape[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "bc1681ba-0220-4e27-bc7e-d8bccc156090",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(278009, 4, 38)"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# objects_train.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "16affb8b-716e-4271-a96d-84c93f9a5cfc",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Save vectors to numpy arrays\n",
    "np.save(\"data/img_indices_train_full.npy\", train_df[\"index\"].values)\n",
    "np.save(\"data/img_indices_val_full.npy\", val_df[\"index\"].values)\n",
    "np.save(\"data/objects_train_full.npy\", objects_train)\n",
    "np.save(\"data/objects_val_full.npy\", objects_val)\n",
    "np.save(\"data/actions_train_full.npy\", action_train)\n",
    "np.save(\"data/actions_val_full.npy\", action_val)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 114,
   "id": "c296a804-591c-4f95-aaf1-2207a537327d",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "312051"
      ]
     },
     "execution_count": 114,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(df_tmp)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "id": "d77b1583-8a79-45d2-befe-9da723682df9",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 2, 33, 55],\n",
       "       [ 5, 65, 55],\n",
       "       [ 0, 65, 55],\n",
       "       ...,\n",
       "       [ 3, 83, 55],\n",
       "       [12, 55, 55],\n",
       "       [ 6, 88, 55]])"
      ]
     },
     "execution_count": 50,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.load(\"data/annotated/actions_train_full.npy\", allow_pickle=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2ff9d1ce-9d56-425b-9527-56a8f73cf8a9",
   "metadata": {},
   "source": [
    "#### Save annotated set vectors"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "id": "c729def7-564e-4930-818c-61fb0b680ed4",
   "metadata": {},
   "outputs": [],
   "source": [
    "annotated_objects_train = np.stack([df_tmp[df_tmp[\"split\"] == \"train\"][obj0_attributes_pre].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"train\"][obj1_attributes_pre].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"train\"][obj0_attributes_post].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"train\"][obj1_attributes_post].values], axis=1)\n",
    "\n",
    "annotated_objects_val = np.stack([df_tmp[df_tmp[\"split\"] == \"val\"][obj0_attributes_pre].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"val\"][obj1_attributes_pre].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"val\"][obj0_attributes_post].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"val\"][obj1_attributes_post].values], axis=1)\n",
    "\n",
    "annotated_objects_test = np.stack([df_tmp[df_tmp[\"split\"] == \"test\"][obj0_attributes_pre].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"test\"][obj1_attributes_pre].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"test\"][obj0_attributes_post].values,\n",
    "                                        df_tmp[df_tmp[\"split\"] == \"test\"][obj1_attributes_post].values], axis=1)\n",
    "\n",
    "annotated_action_train = df_tmp[df_tmp[\"split\"] == \"train\"][action_columns].values\n",
    "annotated_action_val = df_tmp[df_tmp[\"split\"] == \"val\"][action_columns].values\n",
    "annotated_action_test = df_tmp[df_tmp[\"split\"] == \"test\"][action_columns].values"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "id": "38a0fd3c-79ac-4b0e-8dc2-0ace978eff39",
   "metadata": {},
   "outputs": [],
   "source": [
    "# save precondition text for all three annotators for train/val/test\n",
    "np.save(\"data/annotated/precondition_language_train_full.npy\", df_tmp[df_tmp[\"split\"] == \"train\"][[f\"annotator{i}_precondition_language\" for i in range(3)]].values)\n",
    "np.save(\"data/annotated/precondition_language_val_full.npy\", df_tmp[df_tmp[\"split\"] == \"val\"][[f\"annotator{i}_precondition_language\" for i in range(3)]].values)\n",
    "np.save(\"data/annotated/precondition_language_test_full.npy\", df_tmp[df_tmp[\"split\"] == \"test\"][[f\"annotator{i}_precondition_language\" for i in range(3)]].values)\n",
    "\n",
    "# save action text for all three annotators for train/val/test\n",
    "np.save(\"data/annotated/action_language_train_full.npy\", df_tmp[df_tmp[\"split\"] == \"train\"][[f\"annotator{i}_action_language\" for i in range(3)]].values)\n",
    "np.save(\"data/annotated/action_language_val_full.npy\", df_tmp[df_tmp[\"split\"] == \"val\"][[f\"annotator{i}_action_language\" for i in range(3)]].values)\n",
    "np.save(\"data/annotated/action_language_test_full.npy\", df_tmp[df_tmp[\"split\"] == \"test\"][[f\"annotator{i}_action_language\" for i in range(3)]].values)\n",
    "\n",
    "# save postcondition text for all three annotators for train/val/test\n",
    "np.save(\"data/annotated/postcondition_language_train_full.npy\", df_tmp[df_tmp[\"split\"] == \"train\"][[f\"annotator{i}_postcondition_language\" for i in range(3)]].values)\n",
    "np.save(\"data/annotated/postcondition_language_val_full.npy\", df_tmp[df_tmp[\"split\"] == \"val\"][[f\"annotator{i}_postcondition_language\" for i in range(3)]].values)\n",
    "np.save(\"data/annotated/postcondition_language_test_full.npy\", df_tmp[df_tmp[\"split\"] == \"test\"][[f\"annotator{i}_postcondition_language\" for i in range(3)]].values)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "id": "824c2b4e-1b4f-4d08-92aa-8e88a94f1e41",
   "metadata": {},
   "outputs": [],
   "source": [
    "# save image indices\n",
    "np.save(\"data/annotated/img_indices_train_full.npy\", df_tmp[df_tmp[\"split\"] == \"train\"][\"index\"].values)\n",
    "np.save(\"data/annotated/img_indices_val_full.npy\", df_tmp[df_tmp[\"split\"] == \"val\"][\"index\"].values)\n",
    "np.save(\"data/annotated/img_indices_test_full.npy\", df_tmp[df_tmp[\"split\"] == \"test\"][\"index\"].values)\n",
    "\n",
    "# save vectors for objects and actions\n",
    "np.save(\"data/annotated/objects_train_full.npy\", annotated_objects_train)\n",
    "np.save(\"data/annotated/objects_val_full.npy\", annotated_objects_val)\n",
    "np.save(\"data/annotated/objects_test_full.npy\", annotated_objects_test)\n",
    "np.save(\"data/annotated/actions_train_full.npy\", annotated_action_train)\n",
    "np.save(\"data/annotated/actions_val_full.npy\", annotated_action_val)\n",
    "np.save(\"data/annotated/actions_test_full.npy\", annotated_action_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 122,
   "id": "f5b3a258-8086-49b7-88d3-8474ef2a81d8",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(500, 2)"
      ]
     },
     "execution_count": 122,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "annotated_action_val.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fe4b7db9-aaba-49e1-9fb7-53f055261145",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Cleanup unused images\n",
    "# train_indices = set(train_df[\"index\"].values)\n",
    "# files = glob.glob(\"data/train/*\")\n",
    "# for file in tqdm(files):\n",
    "#     if int(file.split(\"/\")[-1]) not in train_indices:\n",
    "#         shutil.rmtree(file)\n",
    "\n",
    "# val_indices = set(val_df[\"index\"].values)\n",
    "# files = glob.glob(\"data/val/*\")\n",
    "# for file in tqdm(files):\n",
    "#     if int(file.split(\"/\")[-1]) not in val_indices:\n",
    "#         shutil.rmtree(file)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3db00471-965c-400d-9b24-e45be96ba70d",
   "metadata": {},
   "source": [
    "# Calculate Image statistics"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d57aaa7e-36c9-4aee-aa26-cba8f6f2c389",
   "metadata": {},
   "outputs": [],
   "source": [
    "all_images = glob.glob(\"data/train/*/*.jpeg\") + glob.glob(\"data/annotated/*/*/*.jpeg\") + glob.glob(\"data/val/*/*.jpeg\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8d49eb6f",
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import torch\n",
    "\n",
    "MEAN = torch.tensor([0.49458119, 0.43375753, 0.34601003])\n",
    "STD = torch.tensor([0.19409628, 0.19771467, 0.19638838])\n",
    "\n",
    "\n",
    "def normalize_image(image):\n",
    "    normed = torch.from_numpy(image).clone()\n",
    "    normed = normed.permute(-1, 0, 1)\n",
    "    normed = (normed - MEAN[:, None, None]) / STD[:, None, None]\n",
    "    return normed\n",
    "\n",
    "\n",
    "def denormalize_image(image):\n",
    "    denormed = image * STD[:, None, None] + MEAN[:, None, None]\n",
    "    denormed = denormed.permute(1, 2, 0)\n",
    "    return denormed\n",
    "\n",
    "\n",
    "unnorm_1 = (np.asarray(Image.open(\"data/train/1/0.jpeg\"))/255)\n",
    "plt.imshow(unnorm_1)\n",
    "plt.show()\n",
    "\n",
    "norm_1 = normalize_image(unnorm_1)\n",
    "norm_1_channel_last = norm_1.permute(1, 2, 0)\n",
    "plt.imshow(norm_1_channel_last)\n",
    "plt.show()\n",
    "\n",
    "norm_1_channel_last.min()\n",
    "unnorm_1 = denormalize_image(norm_1)\n",
    "plt.imshow(unnorm_1)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ae29bbb5",
   "metadata": {},
   "outputs": [],
   "source": [
    "from PIL import Image\n",
    "\n",
    "channels_psum = np.zeros(3)\n",
    "channels_psum_sq = np.zeros(3)\n",
    "for image in tqdm(all_images):\n",
    "    arr = np.asarray(Image.open(image)) / 255\n",
    "    channels_psum += arr.sum((0,1))\n",
    "    channels_psum_sq += (arr ** 2).sum((0,1))\n",
    "    \n",
    "count = 384*640*len(all_images)\n",
    "total_mean = channels_psum / count\n",
    "total_std = np.sqrt(channels_psum_sq / count - total_mean ** 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "id": "27954877-5a42-43cd-9af2-61e5c772667c",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "239392"
      ]
     },
     "execution_count": 51,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.load(\"data/objects_train.npy\").shape[0] #+ np.load(\"data/objects_val.npy\").shape[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b42f5705-660a-42fd-9a1c-cd9e3f39ee00",
   "metadata": {},
   "source": [
    "# Zero Shot"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 482,
   "id": "6968248a-c804-47f9-9755-857ad10917b9",
   "metadata": {},
   "outputs": [],
   "source": [
    "# actions_train = np.concatenate((np.load(\"data/actions_train.npy\"), np.load(\"data/annotated/actions_train.npy\")), axis=0)\n",
    "# actions_train = np.load(\"data/actions_train.npy\")\n",
    "# actions_test = np.load(\"data/annotated/actions_test.npy\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "e722c845-8903-4e45-be00-3103daf9505e",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 2, 33, 55],\n",
       "       [ 5, 65, 55],\n",
       "       [ 0, 65, 55],\n",
       "       ...,\n",
       "       [ 3, 83, 55],\n",
       "       [12, 55, 55],\n",
       "       [ 6, 88, 55]])"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import numpy as np\n",
    "np.load(\"data/annotated/actions_train_full.npy\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 575,
   "id": "6d34a4bb-030a-4f6a-9b33-5c21347fc83f",
   "metadata": {},
   "outputs": [],
   "source": [
    "from collections import Counter, defaultdict\n",
    "import pickle \n",
    "\n",
    "# action_mapping = pickle.load(open(\"data/reverse_action_mapping.pkl\", \"rb\"))\n",
    "# object_mapping = pickle.load(open(\"data/reverse_object_mapping.pkl\", \"rb\"))\n",
    "\n",
    "# def action_triplet_to_string(action_triplet):\n",
    "#     return action_mapping[action_triplet[0]], object_mapping[0][action_triplet[1]], object_mapping[0][action_triplet[2]]\n",
    "\n",
    "# action_triplets_train = []\n",
    "# action_triplets_test = []\n",
    "\n",
    "# for action_triplet in actions_train:\n",
    "#     act = action_triplet_to_string(action_triplet)\n",
    "#     action_triplets_train.append({\"action\": act[0], \"object_0\": act[1], \"object_1\": act[2], \"arr\": action_triplet})\n",
    "    \n",
    "# for action_triplet in actions_test:\n",
    "#     act = action_triplet_to_string(action_triplet)\n",
    "#     action_triplets_test.append({\"action\": act[0], \"object_0\": act[1], \"object_1\": act[2], \"arr\": action_triplet})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 647,
   "id": "6e439b80-06d9-41d1-97be-af96829b128c",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "27"
      ]
     },
     "execution_count": 647,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# train_df = pd.DataFrame(action_triplets_train)\n",
    "# test_df = pd.DataFrame(action_triplets_test)\n",
    "len(hold_out_action_object_pairs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 577,
   "id": "014e70d7-e2f3-4d82-9129-a905806091ba",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 3401 examples\n",
    "hold_out_objects = ['HandTowel', 'Towel', 'Plunger', 'Watch', 'CD', 'SoapBottle',\n",
    "                   'Pen', 'RemoteControl', 'SoapBar', 'Box', 'Bottle', 'CreditCard',\n",
    "                   'Statue', 'KeyChain']\n",
    "\n",
    "# 3278 examples\n",
    "hold_out_action_object_pairs = [('CloseObject', 'Toilet'),\n",
    "                                 ('DirtyObject', 'Pan'),\n",
    "                                 ('DirtyObject', 'Pot'),\n",
    "                                 ('EmptyLiquidFromObject', 'Bottle'),\n",
    "                                 ('EmptyLiquidFromObject', 'Pot'),\n",
    "                                 ('OpenObject', 'Toilet'),\n",
    "                                 ('PickupObject', 'Box'),\n",
    "                                 ('PickupObject', 'CellPhone'),\n",
    "                                 ('PickupObject', 'CreditCard'),\n",
    "                                 ('PickupObject', 'KeyChain'),\n",
    "                                 ('PutObject', 'CD'),\n",
    "                                 ('PutObject', 'CreditCard'),\n",
    "                                 ('PutObject', 'HandTowel'),\n",
    "                                 ('PutObject', 'Laptop'),\n",
    "                                 ('PutObject', 'Lettuce'),\n",
    "                                 ('PutObject', 'Pen'),\n",
    "                                 ('PutObject', 'Plunger'),\n",
    "                                 ('PutObject', 'Pot'),\n",
    "                                 ('PutObject', 'RemoteControl'),\n",
    "                                 ('PutObject', 'SoapBar'),\n",
    "                                 ('PutObject', 'SoapBottle'),\n",
    "                                 ('PutObject', 'Statue'),\n",
    "                                 ('PutObject', 'ToiletPaper'),\n",
    "                                 ('PutObject', 'Towel'),\n",
    "                                 ('PutObject', 'Watch'),\n",
    "                                 ('ToggleObjectOff', 'CellPhone'),\n",
    "                                 ('ToggleObjectOff', 'Television')]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 533,
   "id": "261ee555-083b-4fb6-bb0e-7ea9a8c620ac",
   "metadata": {},
   "outputs": [],
   "source": [
    "# set(hold_out_action_object_pairs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 578,
   "id": "7079c184-e7f9-4f75-bc13-9800c6857daa",
   "metadata": {},
   "outputs": [],
   "source": [
    "# c = 0\n",
    "# for o in hold_out_objects:\n",
    "#     c += len(train_df[(train_df[\"object_0\"] == o)])\n",
    "# print(c)\n",
    "# indices = []\n",
    "# for pair in hold_out_action_object_pairs:\n",
    "#     if pair[1] in hold_out_objects:\n",
    "#         continue\n",
    "#     c += len(train_df[(train_df[\"action\"] == pair[0])&(train_df[\"object_0\"] == pair[1])])\n",
    "#     indices += list(train_df[(train_df[\"action\"] == pair[0])&(train_df[\"object_0\"] == pair[1])].index.values)\n",
    "# print(c)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 579,
   "id": "96260f11-4883-4f2d-948c-5313637a3e81",
   "metadata": {},
   "outputs": [],
   "source": [
    "# len(indices)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 585,
   "id": "d68228c0-4417-4f96-a2f2-6dd229a7eb67",
   "metadata": {},
   "outputs": [],
   "source": [
    "# excluded_triplets\n",
    "obj_name_to_index = dict(zip(object_mapping[0].values(), object_mapping[0].keys()))\n",
    "act_name_to_index = dict(zip(action_mapping.values(), action_mapping.keys()))\n",
    "\n",
    "holdout_objects_indexes = np.array([obj_name_to_index[obj] for obj in hold_out_objects])\n",
    "holdout_objects_pairs_indexes = np.array([[act_name_to_index[a], obj_name_to_index[o]] for a, o in hold_out_action_object_pairs])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 565,
   "id": "c69340d6-4143-44b3-b740-ea19e7751da9",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Excluding 739 examples for data/val\n",
      "Excluding 6767 examples for data/train\n",
      "Excluding 22 examples for data/annotated/val\n",
      "Excluding 49 examples for data/annotated/train\n"
     ]
    }
   ],
   "source": [
    "for annotated in [False, True]:\n",
    "    for split in [\"val\", \"train\"]:\n",
    "        base_path = f\"data\"\n",
    "        if annotated:\n",
    "            base_path += \"/annotated\"\n",
    "        \n",
    "        img_indices_path = f\"{base_path}/img_indices_{split}.npy\"\n",
    "        actions_path = f\"{base_path}/actions_{split}.npy\"\n",
    "        objects_path = f\"{base_path}/objects_{split}.npy\"\n",
    "\n",
    "        img_indices = np.load(img_indices_path)\n",
    "        actions = np.load(actions_path)\n",
    "        objects = np.load(objects_path)\n",
    "\n",
    "        holdout_objects = (actions[:, 1].reshape(-1, 1) == holdout_objects_indexes.reshape(1, -1)).any(1)\n",
    "        holdout_pairs = (actions[:, :2].reshape(-1, 1,  2) == holdout_objects_pairs_indexes.reshape(1, -1, 2)).all(axis=2).any(axis=1)\n",
    "        keep = ~(holdout_objects|holdout_pairs)\n",
    "        print(f\"Excluding {sum(keep == False)} examples for {base_path}/{split}\", )\n",
    "\n",
    "        img_indices = img_indices[keep]\n",
    "        actions = actions[keep]\n",
    "        objects = objects[keep]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 584,
   "id": "11f0df47-141c-4c81-8012-2a77b9d66c56",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "data/annotated/actions_test.npy\n",
      "42\n",
      "0.10552763819095477\n"
     ]
    }
   ],
   "source": [
    "split = \"test\"\n",
    "actions_path = f\"{base_path}/actions_{split}.npy\"\n",
    "\n",
    "print(actions_path)\n",
    "actions = np.load(actions_path)\n",
    "\n",
    "holdout_objects = (actions[:, 1].reshape(-1, 1) == np.array(holdout_objects_indexes).reshape(1, -1)).any(1)\n",
    "holdout_pairs = (actions[:, :2].reshape(-1, 1,  2) == np.array(holdout_objects_pairs_indexes).reshape(1, -1, 2)).all(axis=2).any(axis=1)\n",
    "zero_shot = holdout_objects|holdout_pairs\n",
    "\n",
    "print(zero_shot.sum())\n",
    "print(zero_shot.sum()/actions.shape[0])\n",
    "\n",
    "np.save(f\"{base_path}/zero_shot_test.npy\", zero_shot)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "id": "beffa886-8a17-4768-a828-617818e70e8e",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "240"
      ]
     },
     "execution_count": 51,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# zero_shot.sum\n",
    "import numpy as np\n",
    "import pickle\n",
    "\n",
    "\n",
    "# self.object_attributes_idx_to_mapper = get_objects_mapper(data_dir_path)\n",
    "\n",
    "with open(f\"data/reverse_object_mapping.pkl\", \"rb\") as f:\n",
    "    # select first index since the first index is the objet name encoded as an int\n",
    "    object_mapper =  pickle.load(f)\n",
    "    \n",
    "# object_mapper\n",
    "# list(object_mapper[idx].values())\n",
    "\n",
    "# min_idx = list(object_mapper[idx].keys())[0]\n",
    "# class_names = list(object_mapper[idx].values())\n",
    "# print(f\"object_attribute_name: {object_attribute_name}\")\n",
    "# print(f\"idx: {idx}\")\n",
    "# print(f\"min_idx: {min_idx}\")\n",
    "# print(f\"class_names: {class_names}\")\n",
    "# print(f\"preds: {preds[:, idx].numpy()}\")\n",
    "# print(f\"y_true: {labels[:, idx].numpy()}\")\n",
    "\n",
    "list(object_mapper[3].keys())[0]\n",
    "\n",
    "# object_attribute_name: distance\n",
    "# idx: 3\n",
    "# min_idx: 240\n",
    "# class_names: [0, 1, 2, 3, 4, 5, 6, 7]\n",
    "# preds: [243 245 243 ... 246 245 243]\n",
    "# y_true: [248 247 244 ... 247 247 247]\n",
    "\n",
    "# np.load(\"data/annotated/seen_test.npy\").shape\n",
    "\n",
    "# np.load(\"data/annotated/actions_test.npy\").shape[0]\n",
    "# np.ones(398).astype(bool)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "id": "e0b06334-7e3c-40ed-b835-2062c8a68e82",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "243"
      ]
     },
     "execution_count": 53,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.load(\"data/objects_train_full.npy\")[:,:,3].min()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 586,
   "id": "de5fde44-bbe2-4b3f-8f61-993cbeb273f8",
   "metadata": {},
   "outputs": [],
   "source": [
    "obj_name_to_index = dict(zip(object_mapping[0].values(), object_mapping[0].keys()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 616,
   "id": "b443d541-bb34-45bd-ae95-08c5209d0500",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Actual</th>\n",
       "      <th>Predicted</th>\n",
       "      <th>_defaultColorIndex</th>\n",
       "      <th>id</th>\n",
       "      <th>nPredictions</th>\n",
       "      <th>name</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>0</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>0</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>0</td>\n",
       "      <td>2</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>0</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>0</td>\n",
       "      <td>3</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>0</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>0</td>\n",
       "      <td>4</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>0</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>59</th>\n",
       "      <td>7</td>\n",
       "      <td>3</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>1</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>60</th>\n",
       "      <td>7</td>\n",
       "      <td>4</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>1</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>61</th>\n",
       "      <td>7</td>\n",
       "      <td>5</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>2</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>62</th>\n",
       "      <td>7</td>\n",
       "      <td>6</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>0</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>63</th>\n",
       "      <td>7</td>\n",
       "      <td>7</td>\n",
       "      <td>8</td>\n",
       "      <td>42_2xl68fp3</td>\n",
       "      <td>4</td>\n",
       "      <td>images_fuse_h64_l3_42_nlu</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>64 rows × 6 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "    Actual  Predicted  _defaultColorIndex           id  nPredictions  \\\n",
       "0        0          0                   8  42_2xl68fp3             0   \n",
       "1        0          1                   8  42_2xl68fp3             0   \n",
       "2        0          2                   8  42_2xl68fp3             0   \n",
       "3        0          3                   8  42_2xl68fp3             0   \n",
       "4        0          4                   8  42_2xl68fp3             0   \n",
       "..     ...        ...                 ...          ...           ...   \n",
       "59       7          3                   8  42_2xl68fp3             1   \n",
       "60       7          4                   8  42_2xl68fp3             1   \n",
       "61       7          5                   8  42_2xl68fp3             2   \n",
       "62       7          6                   8  42_2xl68fp3             0   \n",
       "63       7          7                   8  42_2xl68fp3             4   \n",
       "\n",
       "                         name  \n",
       "0   images_fuse_h64_l3_42_nlu  \n",
       "1   images_fuse_h64_l3_42_nlu  \n",
       "2   images_fuse_h64_l3_42_nlu  \n",
       "3   images_fuse_h64_l3_42_nlu  \n",
       "4   images_fuse_h64_l3_42_nlu  \n",
       "..                        ...  \n",
       "59  images_fuse_h64_l3_42_nlu  \n",
       "60  images_fuse_h64_l3_42_nlu  \n",
       "61  images_fuse_h64_l3_42_nlu  \n",
       "62  images_fuse_h64_l3_42_nlu  \n",
       "63  images_fuse_h64_l3_42_nlu  \n",
       "\n",
       "[64 rows x 6 columns]"
      ]
     },
     "execution_count": 616,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import seaborn as sns\n",
    "\n",
    "df = pd.read_csv(\"~/Downloads/aaa.csv\")\n",
    "df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 633,
   "id": "4ca4cb00-1e3e-4ac3-8c6d-f190cfe47317",
   "metadata": {},
   "outputs": [],
   "source": [
    "cf_matrix = np.zeros((8,8))\n",
    "for i, row in df.iterrows():\n",
    "    cf_matrix[row[\"Actual\"],row[\"Predicted\"]] = row[\"nPredictions\"]\n",
    "    \n",
    "cf_matrix = cf_matrix[1:,1:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 645,
   "id": "50f6732c-db41-4fc3-8b92-bdcafac8ff3d",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 432x432 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "import seaborn as sns\n",
    "\n",
    "fig = plt.gcf()\n",
    "fig.set_size_inches(6, 6)\n",
    "sns.set_context(\"paper\")\n",
    "\n",
    "group_counts = [\"{0:0.0f}\".format(value) for value in cf_matrix.flatten()]\n",
    "group_percentages = [\"{0:.2%}\".format(value) for value in cf_matrix.flatten()/np.sum(cf_matrix)]\n",
    "\n",
    "labels = [f\"{v1}\\n{v2}\" for v1, v2 in\n",
    "      zip(group_counts,group_percentages)]\n",
    "\n",
    "labels = np.asarray(labels).reshape(cf_matrix.shape)\n",
    "\n",
    "sns.heatmap(cf_matrix, annot=labels, fmt='', cmap='Blues', cbar=False, annot_kws={\"size\": 14})\n",
    "plt.title(\"Confusion Matrix for distance\", fontsize=16)\n",
    "plt.xlabel(\"Predicted label\", fontsize=16)\n",
    "plt.ylabel(\"True label\", fontsize=16)\n",
    "plt.tick_params(labelsize=16)\n",
    "plt.tight_layout()\n",
    "plt.savefig(\"distance.png\", dpi=300)\n",
    "# plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e66890ad-a031-4f7d-bb2e-19c3c6f581c5",
   "metadata": {},
   "outputs": [],
   "source": [
    "object_mapping"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 649,
   "id": "0f8a94ab-7ae6-4121-86e1-3a0d1301899d",
   "metadata": {},
   "outputs": [],
   "source": [
    "df = pd.read_csv(\"piglet_train.csv\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 654,
   "id": "de68264f-75a7-42e2-a21b-649aabb92795",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "4    94461\n",
       "3    61080\n",
       "2    46786\n",
       "5    36595\n",
       "1    31566\n",
       "7     3911\n",
       "6     3587\n",
       "0       23\n",
       "Name: distance0, dtype: int64"
      ]
     },
     "execution_count": 654,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df[\"distance0\"].value_counts()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "63f1edd3-3e6a-4062-a6cb-0e05b5c457bc",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "interpreter": {
   "hash": "b10fccac95882d786ab3b514f693eb0985db25808b52f4156dc2bb4315fa1cbb"
  },
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}