download_android_control.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!git clone https://github.com/google-research/google-research.git\n",
    "!pip install tensorflow"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Download images"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "CgoFvmqCtps3",
    "outputId": "64381a1a-5e3d-4ce4-92dd-6a11e43f61ed"
   },
   "outputs": [],
   "source": "# -*- coding: utf-8 -*-\nimport io\nimport json\nfrom pathlib import Path\nfrom typing import Any, Dict, List, Optional, Tuple\n\nimport tensorflow as tf\nfrom PIL import Image\n\n# =========================\n# TF.Feature -> Python value conversion\n# =========================\ndef feature_to_list_bytes(f: tf.train.Feature) -> List[bytes]:\n    return list(f.bytes_list.value)\n\ndef feature_to_list_int(f: tf.train.Feature) -> List[int]:\n    return list(f.int64_list.value)\n\ndef feature_to_list_float(f: tf.train.Feature) -> List[float]:\n    return list(f.float_list.value)\n\ndef get_feature(example: tf.train.Example, key: str) -> Optional[tf.train.Feature]:\n    fmap = example.features.feature\n    return fmap[key] if key in fmap else None\n\n# =========================\n# Extract coordinates from action (supports both normalized/pixel)\n# =========================\ndef extract_points_from_action(\n    action: Dict[str, Any],\n    img_w: int,\n    img_h: int\n) -> List[Tuple[int, int]]:\n    \"\"\"\n    Returns: [(x_px, y_px), ...]\n    - Single point (tap) returns length 1\n    - Drag/swipe returns 2 points (start/end)\n    - Empty list if no points\n    \"\"\"\n    pts: List[Tuple[int, int]] = []\n\n    def to_px(x: float, y: float, normalized: Optional[bool]=None) -> Tuple[int,int]:\n        if normalized is None:\n            normalized = (0.0 <= x <= 1.0 and 0.0 <= y <= 1.0)\n        if normalized:\n            return (int(round(x * img_w)), int(round(y * img_h)))\n        else:\n            return (int(round(x)), int(round(y)))\n\n    # 1) Top-level x, y\n    if \"x\" in action and \"y\" in action:\n        pts.append(to_px(float(action[\"x\"]), float(action[\"y\"]), None))\n\n    # 2) point / click / tap / press / long_press / long_tap\n    for k in [\"point\", \"click\", \"tap\", \"press\", \"long_press\", \"long_tap\"]:\n        if k in action and isinstance(action[k], dict):\n            px = action[k]\n            if \"x\" in px and \"y\" in px:\n                pts.append(to_px(float(px[\"x\"]), float(px[\"y\"]), None))\n        if k in action and isinstance(action[k], list):\n            for px in action[k]:\n                if isinstance(px, dict) and \"x\" in px and \"y\" in px:\n                    pts.append(to_px(float(px[\"x\"]), float(px[\"y\"]), None))\n\n    # 3) from/to, start/end\n    for a, b in [(\"from\", \"to\"), (\"start\", \"end\")]:\n        if a in action and b in action and isinstance(action[a], dict) and isinstance(action[b], dict):\n            ax, ay = action[a].get(\"x\"), action[a].get(\"y\")\n            bx, by = action[b].get(\"x\"), action[b].get(\"y\")\n            if ax is not None and ay is not None and bx is not None and by is not None:\n                pts.append(to_px(float(ax), float(ay), None))\n                pts.append(to_px(float(bx), float(by), None))\n\n    # 4) start_x/start_y/end_x/end_y\n    cand = {\"start_x\": None, \"start_y\": None, \"end_x\": None, \"end_y\": None}\n    found = False\n    for ck in cand.keys():\n        if ck in action:\n            cand[ck] = float(action[ck])\n            found = True\n    if found and cand[\"start_x\"] is not None and cand[\"start_y\"] is not None:\n        pts.append(to_px(cand[\"start_x\"], cand[\"start_y\"], None))\n        if cand[\"end_x\"] is not None and cand[\"end_y\"] is not None:\n            pts.append(to_px(cand[\"end_x\"], cand[\"end_y\"], None))\n\n    # Remove duplicates\n    uniq: List[Tuple[int,int]] = []\n    seen = set()\n    for p in pts:\n        if p not in seen:\n            uniq.append(p)\n            seen.add(p)\n    return uniq\n\n# =========================\n# Parse episode from TF Example\n# =========================\ndef load_episode_from_example(ex: tf.train.Example) -> Dict[str, Any]:\n    f = ex.features.feature\n\n    screenshots_bytes = feature_to_list_bytes(f[\"screenshots\"])\n    a11y_bytes_list = feature_to_list_bytes(f[\"accessibility_trees\"])\n    widths = feature_to_list_int(f[\"screenshot_widths\"])\n    heights = feature_to_list_int(f[\"screenshot_heights\"])\n\n    actions_json_list = [b.decode(\"utf-8\") for b in feature_to_list_bytes(f[\"actions\"])]\n    step_insts = [b.decode(\"utf-8\") for b in feature_to_list_bytes(f[\"step_instructions\"])]\n    actions = [json.loads(s) for s in actions_json_list]\n\n    goal = f[\"goal\"].bytes_list.value[0].decode(\"utf-8\")\n    episode_id = int(f[\"episode_id\"].int64_list.value[0]) if f[\"episode_id\"].int64_list.value else int(\n        f[\"episode_id\"].bytes_list.value[0].decode(\"utf-8\")\n    )\n\n    assert len(screenshots_bytes) == len(widths) == len(heights), \"screenshot/width/height length mismatch\"\n    assert len(actions) == len(step_insts) == (len(screenshots_bytes) - 1), \\\n        \"actions/step_instructions must equal screenshots-1\"\n\n    return {\n        \"episode_id\": episode_id,\n        \"goal\": goal,\n        \"screenshots\": screenshots_bytes,\n        \"a11y\": a11y_bytes_list,\n        \"widths\": widths,\n        \"heights\": heights,\n        \"actions\": actions,\n        \"step_instructions\": step_insts,\n    }\n\n# =========================\n# Action mapping & utilities\n# =========================\ndef _center_xy(w: int, h: int) -> Tuple[int,int]:\n    return (int(round(w/2)), int(round(h/2)))\n\ndef _norm_dir(d: Optional[str]) -> str:\n    if not d: return \"down\"\n    d = str(d).lower()\n    if d in [\"up\",\"down\",\"left\",\"right\"]:\n        return d\n    if d in [\"u\",\"top\"]: return \"up\"\n    if d in [\"d\",\"bottom\"]: return \"down\"\n    if d in [\"l\"]: return \"left\"\n    if d in [\"r\"]: return \"right\"\n    return \"down\"\n\ndef map_action(\n    action: Dict[str, Any],\n    w: int,\n    h: int,\n    pts: List[Tuple[int,int]],\n) -> Optional[Dict[str, Any]]:\n    \"\"\"\n    Allowed mappings:\n    click -> {\"type\": \"touch\", \"x\": <x>, \"y\": <y>}\n    long_press -> {\"type\": \"long_touch\", \"x\": <x>, \"y\": <y>}\n    input_text -> {\"type\": \"set_text\", \"text\": \"...\", \"x\": <x>, \"y\": <y>}\n    scroll -> {\"type\": \"scroll\", \"direction\": \"up|down|left|right\", \"x\": <center_x>, \"y\": <center_y>}\n    navigate_home -> {\"type\": \"press\", \"key\": \"home\"}\n    navigate_back -> {\"type\": \"press\", \"key\": \"back\"}\n    \"\"\"\n    atype = (action.get(\"action_type\") or action.get(\"type\") or action.get(\"action\") or \"\").lower()\n    x, y = (pts[0] if pts else _center_xy(w, h))\n\n    if atype in [\"click\", \"tap\", \"press\", \"click_view\"]:\n        return {\"type\": \"touch\", \"x\": x, \"y\": y}\n\n    if atype in [\"long_press\", \"long_tap\", \"long_click\"]:\n        return {\"type\": \"long_touch\", \"x\": x, \"y\": y}\n\n    if atype in [\"input_text\", \"set_text\", \"type_text\", \"enter_text\", \"text\"]:\n        text = action.get(\"text\") or action.get(\"input_text\") or action.get(\"value\") or \"\"\n        return {\"type\": \"set_text\", \"text\": str(text), \"x\": x, \"y\": y}\n\n    if atype in [\"scroll\", \"swipe\"]:\n        if len(pts) >= 2:\n            cx = (pts[0][0] + pts[1][0]) // 2\n            cy = (pts[0][1] + pts[1][1]) // 2\n        else:\n            cx, cy = _center_xy(w, h)\n        return {\"type\": \"scroll\", \"direction\": _norm_dir(action.get(\"direction\")), \"x\": cx, \"y\": cy}\n\n    if atype in [\"navigate_home\", \"home\", \"press_home\"]:\n        return {\"type\": \"press\", \"key\": \"home\"}\n\n    if atype in [\"navigate_back\", \"back\", \"press_back\"]:\n        return {\"type\": \"press\", \"key\": \"back\"}\n\n    # Others (open_app, wait, etc.) → skip saving\n    return None\n\ndef save_clean_image(img_bytes: bytes, episode_id: int, step_idx: int, base_dir: str = \"and_ctrl\") -> str:\n    \"\"\"\n    Save image as: out_episode_{EP}_step_{STEP:03d}.png (without overlay)\n    \"\"\"\n    Path(base_dir).mkdir(parents=True, exist_ok=True)\n    fname = f\"out_episode_{episode_id}_step_{step_idx:03d}.png\"\n    fpath = Path(base_dir) / fname\n    Image.open(io.BytesIO(img_bytes)).convert(\"RGB\").save(fpath)\n    # Return just the relative path from base_dir\n    return f\"{base_dir}/{fname}\"\n\n# =========================\n# Export messages to JSON\n# =========================\ndef export_messages(ds, limit_episodes: int = 5, out_json: str = \"and_ctrl.json\", image_dir: str = \"and_ctrl\"):\n    \"\"\"\n    Extract allowed actions from first N episodes in TFRecordDataset\n    and save them in request format to and_ctrl.json.\n    \"\"\"\n    all_items: List[Dict[str, Any]] = []\n    ep_cnt = 0\n\n    for raw in ds:\n        ex = tf.train.Example()\n        ex.ParseFromString(raw.numpy())\n        ep = load_episode_from_example(ex)\n\n        ep_id = ep[\"episode_id\"]\n        for i, (action, inst) in enumerate(zip(ep[\"actions\"], ep[\"step_instructions\"])):\n            w, h = ep[\"widths\"][i], ep[\"heights\"][i]\n            img_bytes = ep[\"screenshots\"][i]\n            pts = extract_points_from_action(action, w, h)\n            mapped = map_action(action, w, h, pts)\n            if not mapped:\n                continue  # Skip\n\n            img_path = save_clean_image(img_bytes, ep_id, i, base_dir=image_dir)\n\n            all_items.append({\n                \"messages\": [\n                    {\"role\": \"user\", \"content\": f\"<image>\\n{inst}\"},\n                    # Save as Python dict string (single quotes), not JSON\n                    {\"role\": \"assistant\", \"content\": str(mapped)}\n                ],\n                \"images\": [img_path]\n            })\n\n        ep_cnt += 1\n        if ep_cnt >= limit_episodes:\n            break\n\n    with open(out_json, \"w\", encoding=\"utf-8\") as f:\n        json.dump(all_items, f, ensure_ascii=False, indent=2)\n\n    print(f\"[DONE] episodes processed: {ep_cnt}, items saved: {len(all_items)} → {out_json}\")\n\n# =========================\n# Main entry point\n# =========================\ndef main():\n    # Adjust path pattern if needed\n    filenames = tf.io.gfile.glob('gs://gresearch/android_control/android_control*')\n    ds = tf.data.TFRecordDataset(filenames, compression_type='GZIP')\n    export_messages(ds, limit_episodes=50, out_json=\"and_ctrl.json\", image_dir=\"and_ctrl\")\n\nif __name__ == \"__main__\":\n    main()"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "9wgmlG4CxlkL",
    "outputId": "8dfce9bd-7003-4637-98ad-6843f2aeb9f9"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "46.5 GiB     gs://gresearch/android_control/android_control*\n"
     ]
    }
   ],
   "source": [
    "!gsutil du -sh gs://gresearch/android_control/android_control*\n"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "gpuType": "T4",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "and_ctrl",
   "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": "3.12.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}