{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ca6b600d-980f-4e4b-87c5-4f0b693d26ca",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2\n",
    "%config InlineBackend.figure_format = \"retina\"\n",
    "\n",
    "import my_nb_path  # isort: split\n",
    "import a2rl\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "import torch\n",
    "\n",
    "from flight_sales.flight_sales_gym import flight_sales_gym\n",
    "\n",
    "model = torch.load(\"results/model-profit/model.pt\")\n",
    "tokenizer = a2rl.utils.pickle_load(\"results/model-profit/tokenizer.pt\")\n",
    "simulator = a2rl.Simulator(tokenizer, model, max_steps=365)\n",
    "env = flight_sales_gym()\n",
    "\n",
    "\n",
    "def history_2_context(df):\n",
    "    custom_context = df.values.ravel()[: -len(df.actions) - len(df.rewards)]\n",
    "    return custom_context"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1255191d-d3b0-47d1-b032-a6e44c6ec8c6",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "env.reset()\n",
    "env.step(0.5)\n",
    "\n",
    "ctx = np.tile(history_2_context(env.context(tail=tokenizer.block_size_row, fillna=True)), (4, 1))\n",
    "ctx.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ec40b863-0a6c-4ca7-91ee-f8229086d155",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "simulator.sample(ctx, max_size=2, as_token=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e7f490fd-ada6-4a99-abf2-24c506e805ac",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "display(\n",
    "    # These are equivalent\n",
    "    env.history.iloc[: env.day],\n",
    "    env.context(),\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "08bba524-f7c1-4d67-9ae6-786bd755a1f0",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "env.reset()\n",
    "env.step(0.5)\n",
    "ctx = env.context(fillna=True)\n",
    "display(\n",
    "    (tokenizer.block_size_row, tokenizer.block_size),\n",
    "    ctx,\n",
    "    tokenizer.field_tokenizer.transform(ctx),\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f269ddc4-cd9b-4c72-b5c3-4547218815e2",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "display(\n",
    "    tokenizer.df.tail(tokenizer.block_size_row),\n",
    "    tokenizer.df.iloc[-tokenizer.block_size_row :],\n",
    "    history_2_context(tokenizer.df.iloc[-tokenizer.block_size_row :]),\n",
    "    history_2_context(tokenizer.df.tail(tokenizer.block_size_row)),\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9baa1891-d7b9-4670-8db3-5744d2059905",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "# ctx = (s, a, r, ..., s), where len([first_s, ..., last_s]) == block_size\n",
    "ctx = history_2_context(tokenizer.df.tail(tokenizer.block_size_row))\n",
    "batch_ctx = np.asarray([ctx, ctx, ctx])  # A batch of 3 trajectories\n",
    "display(\n",
    "    simulator.sample(ctx, max_size=5),\n",
    "    simulator.sample(batch_ctx, max_size=5),\n",
    "    simulator.sample(batch_ctx, max_size=500).shape,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "546bd50f-4592-4ef2-a97c-ce406cac151f",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "batch_ctx = np.asarray([ctx, ctx, ctx])  # A batch of 3 trajectories\n",
    "trajectories_cnt = batch_ctx.shape[0]\n",
    "batch_results = simulator.sample(batch_ctx, max_size=500)\n",
    "display(batch_ctx.shape, batch_results.shape)\n",
    "\n",
    "# Let's try a few different ways to cut the batch dataframe into per-trajectory objects.\n",
    "\n",
    "#### 01: numpy manipulation ####\n",
    "trajectory_results_a = batch_results.values.reshape(\n",
    "    trajectories_cnt,\n",
    "    -1,\n",
    "    len(batch_results.actions) + len(batch_results.rewards),\n",
    ")\n",
    "display(\n",
    "    trajectory_results_a.shape,\n",
    "    # Verify first two rows and last two rows are the same.\n",
    "    pd.concat([batch_results.head(2), batch_results.tail(2)]),\n",
    "    [\n",
    "        trajectory_results_a[0, 0:2, :],\n",
    "        trajectory_results_a[-1, -2:, :],\n",
    "    ],\n",
    ")\n",
    "\n",
    "#### 02: pandas manipulation ####\n",
    "trajectory_results_df: list[a2rl.WiDataFrame] = np.array_split(batch_results, trajectories_cnt)\n",
    "display(\n",
    "    [tdf.shape for tdf in trajectory_results_df],\n",
    "    # Verify first two rows and last two rows are the same.\n",
    "    pd.concat([batch_results.head(2), batch_results.tail(2)]),\n",
    "    pd.concat([trajectory_results_df[0].head(2), trajectory_results_df[-1].tail(2)]),\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9ce40a38-51ea-4e4c-847c-f5af4f44aa1f",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "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": "3.10.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}