nbs/16k-fit-one-cycle.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "16k-fit-one-cycle.ipynb",
      "provenance": []
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "LmtceCFGGajO",
        "colab_type": "text"
      },
      "source": [
        "หลังจากที่ใน ep ก่อน เราได้ใช้ LR_Find Callback หา Learning Rate ที่ดีที่สุดได้แล้ว แล้วเราจะนำมาใช้อย่างไร Learning Rate ถือว่าเป็นหนึ่งใน Hyperparameter ที่สำคัญที่สุดในการเทรน Machine Learning \n",
        "\n",
        "มีแนวคิดจากหลากหลาย Paper ที่ว่า ในแต่ละ State ของการเทรน Deep Neural Network นั้นต้องการ Hyperparameter ต่างกันไป เช่น ตอนเริ่มเทรน เรายังไม่รู้ว่า Global Minima อยู่ไหน โมเดลจะไปทางไหนดี ไม่ควรรีบร้อนใช้ Learning Rate สูง จะทำให้กระโดดไปผิดทางได้, ช่วงท้ายของการเทรน เราก็ควรลด Learning Rate ลงเพื่อ Fine Tune ไม่ให้โมเดลกระโดดออกมาจากหลุม Global Minima, etc.\n",
        "\n",
        "ดังนั้นเราควรจะ Schedule ทุก Hyperparamter ใน ep นี้เราเริ่มต้นด้วยการ มาดูตัวอย่างการ Schedule Learning Rate ด้วย Callback กัน เริ่มต้นที่[หัวข้อ 6.1 ParamScheduler](#6.1-ParamScheduler)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "3drCJqK97wZ5",
        "colab_type": "text"
      },
      "source": [
        "# 0. Magic"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "As6QTNhIo7LC",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "%load_ext autoreload\n",
        "%autoreload 2\n",
        "%matplotlib inline"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "aJdvUE9d7y7L",
        "colab_type": "text"
      },
      "source": [
        "# 1. Import"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "M0k1vCN_7zFe",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "import torch\n",
        "from torch import tensor\n",
        "from torch.nn import *\n",
        "import torch.nn.functional as F\n",
        "from torch.utils.data import *\n",
        "from fastai import datasets\n",
        "from fastai.metrics import accuracy\n",
        "from fastai.basic_data import *\n",
        "from fastai.basic_train import *\n",
        "import pickle, gzip, math, torch, re\n",
        "from IPython.core.debugger import set_trace\n",
        "import matplotlib.pyplot as plt\n",
        "from functools import partial"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "lYLQCdBq7zNn",
        "colab_type": "text"
      },
      "source": [
        "# 2. Data"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "MqtAa1OttrhQ",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "class Dataset(Dataset):\n",
        "    def __init__(self, x, y):\n",
        "        self.x, self.y = x, y\n",
        "    def __len__(self):\n",
        "        return len(self.x)\n",
        "    def __getitem__(self, i):\n",
        "        return self.x[i], self.y[i]"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "fiPq9nOu8V-y",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "MNIST_URL='http://deeplearning.net/data/mnist/mnist.pkl'"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "j8Jl_Ekx7zVi",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def get_data():\n",
        "    path = datasets.download_data(MNIST_URL, ext='.gz')\n",
        "    with gzip.open(path, 'rb') as f:\n",
        "        ((x_train, y_train), (x_valid, y_valid), _) = pickle.load(f, encoding='latin-1')\n",
        "    return map(tensor, (x_train, y_train, x_valid, y_valid))"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "LzugVJ09eCrk",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "x_train, y_train, x_valid, y_valid = get_data()"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "kWKmfCFVLOmn",
        "colab_type": "text"
      },
      "source": [
        "# 3. Data Preprocessing"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "1d9ixSpfgUeS",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def normalize(x, m, s): \n",
        "    return (x-m)/s"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lrGYLaZU7age",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "from typing import *\n",
        "\n",
        "def listify(o):\n",
        "    if o is None: return []\n",
        "    if isinstance(o, list): return o\n",
        "    if isinstance(o, str): return [o]\n",
        "    if isinstance(o, Iterable): return list(o)\n",
        "    return [o]"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ykdiURV28_se",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "_camel_re1 = re.compile('(.)([A-Z][a-z]+)')\n",
        "_camel_re2 = re.compile('([a-z0-9])([A-Z])')\n",
        "\n",
        "def camel2snake(name):\n",
        "    s1 = re.sub(_camel_re1, r'\\1_\\2', name)\n",
        "    return re.sub(_camel_re2, r'\\1_\\2', s1).lower()   "
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "IEtfU8TjgO-t",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "train_mean, train_std = x_train.mean(), x_train.std()\n",
        "x_train = normalize(x_train, train_mean, train_std)\n",
        "x_valid = normalize(x_valid, train_mean, train_std)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lwr_S24teD1-",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "nh, bs = 100, 256\n",
        "n, m = x_train.shape\n",
        "c = (y_train.max()+1).numpy()\n",
        "loss_func = F.cross_entropy"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lxDyNxD07zeH",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "train_ds, valid_ds = Dataset(x_train, y_train), Dataset(x_valid, y_valid)\n",
        "train_dl, valid_dl = DataLoader(train_ds, bs), DataLoader(valid_ds, bs)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Z3OzWeXWBzLl",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "data = DataBunch(train_dl, valid_dl)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "o9EbSM4m8lef",
        "colab_type": "text"
      },
      "source": [
        "# 4. Model"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "CtoT7A2DggGz",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "lr = 0.001\n",
        "epoch = 10\n",
        "nh = 32"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "NBp4jzQoeWLe",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def get_model():\n",
        "    # loss function\n",
        "    loss_func = F.cross_entropy\n",
        "    model = Sequential(Linear(m, nh), ReLU(), Linear(nh,c))\n",
        "    return model, loss_func"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5p_H1kWNd9ab",
        "colab_type": "text"
      },
      "source": [
        "# 5. Training Loop"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5yFdOKNfgbVB",
        "colab_type": "text"
      },
      "source": [
        "Training Loop ที่รองรับ Callback"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "tp3wybPiGylV",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "class Runner():\n",
        "    def __init__(self, cbs=None, cb_funcs=None):\n",
        "        cbs = listify(cbs)\n",
        "        for cbf in listify(cb_funcs):\n",
        "            cb = cbf()\n",
        "            setattr(self, cb.name, cb)\n",
        "            cbs.append(cb)\n",
        "        self.stop, self.cbs = False, [TrainEvalCallback()]+cbs\n",
        "\n",
        "    @property\n",
        "    def opt_func(self):     return self.learn.opt_func\n",
        "    @property\n",
        "    def model(self):        return self.learn.model\n",
        "    @property\n",
        "    def loss_func(self):    return self.learn.loss_func\n",
        "    @property\n",
        "    def data(self):         return self.learn.data\n",
        "\n",
        "    def one_batch(self, xb, yb):\n",
        "        try: \n",
        "            self.xb, self.yb = xb, yb\n",
        "            self('begin_batch')\n",
        "            self.pred = self.model(xb)\n",
        "            self('after_pred')\n",
        "            self.loss = self.loss_func(self.pred, yb)\n",
        "            self('after_loss')\n",
        "            if not self.in_train: return\n",
        "            self.loss.backward()\n",
        "            self('after_backward')\n",
        "            self.opt_func.step()\n",
        "            self('after_step')\n",
        "            self.opt_func.zero_grad()\n",
        "        except CancelBatchException: self('after_cancel_batch')\n",
        "        finally: self('after_batch')\n",
        "    \n",
        "    def all_batches(self, dl):\n",
        "        self.iters = len(dl)\n",
        "        try:\n",
        "            for xb, yb in dl:\n",
        "                self.one_batch(xb, yb)\n",
        "        except CancelEpochException: self('after_cancel_epoch')\n",
        "    \n",
        "    def fit(self, epochs, learn):\n",
        "        self.epochs, self.learn, self.loss = epochs, learn, tensor(0.)\n",
        "\n",
        "        try:\n",
        "            for cb in self.cbs: cb.set_runner(self)\n",
        "            self('begin_fit')\n",
        "            for epoch in range(epochs):\n",
        "                self.epoch = epoch\n",
        "                if not self('begin_epoch'): self.all_batches(self.data.train_dl)\n",
        "\n",
        "                with torch.no_grad():\n",
        "                    if not self('begin_validate'): self.all_batches(self.data.valid_dl)\n",
        "                self('after_epoch')\n",
        "        except CancelTrainException: self('after_cancel_train')\n",
        "        finally: \n",
        "            self('after_fit')\n",
        "            self.train = None\n",
        "\n",
        "    def __call__(self, cb_name):\n",
        "        # return True = Cancel, return False = Continue (Default)\n",
        "        res = False\n",
        "        # check if at least one True return True\n",
        "        for cb in sorted(self.cbs, key=lambda x: x._order): res = res or cb(cb_name)\n",
        "        return res        "
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "U-kacUKlJk6j",
        "colab_type": "text"
      },
      "source": [
        "# 6. Callbacks"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "G0O9bxvAG5BW",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "class Callback():\n",
        "    _order = 0\n",
        "    def set_runner(self, run): self.run = run\n",
        "    def __getattr__(self, k): return getattr(self.run, k)\n",
        "\n",
        "    @property\n",
        "    def name(self):\n",
        "        name = re.sub(r'Callback$', '', self.__class__.__name__)\n",
        "        return camel2snake(name or 'callback')\n",
        "    \n",
        "    def __call__(self, cb_name):\n",
        "        f = getattr(self, cb_name, None)\n",
        "        if f and f(): return True\n",
        "        return False\n",
        "\n",
        "class TrainEvalCallback(Callback):\n",
        "    def begin_fit(self):\n",
        "        self.run.n_epochs = 0.\n",
        "        self.run.n_iter = 0\n",
        "    \n",
        "    def begin_epoch(self):\n",
        "        self.run.n_epochs = self.epoch  \n",
        "        self.model.train()\n",
        "        self.run.in_train=True\n",
        "\n",
        "    def after_batch(self):\n",
        "        if not self.in_train: return\n",
        "        self.run.n_epochs += 1./self.iters\n",
        "        self.run.n_iter += 1\n",
        "\n",
        "    def begin_validate(self):\n",
        "        self.model.eval()\n",
        "        self.run.in_train=False    \n",
        "           \n",
        "class CancelTrainException(Exception): pass\n",
        "class CancelEpochException(Exception): pass\n",
        "class CancelBatchException(Exception): pass"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "u5zJVLbeqK8W",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "class AvgStatsCallback(Callback):\n",
        "    def __init__(self, metrics):\n",
        "        self.train_stats, self.valid_stats = AvgStats(metrics, True), AvgStats(metrics, False)\n",
        "    def begin_epoch(self):\n",
        "        self.train_stats.reset()\n",
        "        self.valid_stats.reset()\n",
        "    \n",
        "    def after_loss(self):\n",
        "        stats = self.train_stats if self.in_train else self.valid_stats\n",
        "        with torch.no_grad(): stats.accumulate(self.run)\n",
        "\n",
        "    def after_epoch(self):\n",
        "        print(self.train_stats)\n",
        "        print(self.valid_stats)\n",
        "\n",
        "class AvgStats():\n",
        "    def __init__(self, metrics, in_train): \n",
        "        self.metrics, self.in_train = listify(metrics), in_train\n",
        "    \n",
        "    def reset(self):\n",
        "        self.tot_loss, self.count = 0., 0\n",
        "        self.tot_mets = [0.] * len(self.metrics)\n",
        "\n",
        "    @property\n",
        "    def all_stats(self): return [self.tot_loss.item()] + self.tot_mets\n",
        "    @property\n",
        "    def avg_stats(self): return [o/self.count for o in self.all_stats]\n",
        "\n",
        "    def __repr__(self):\n",
        "        if not self.count: return \"\"\n",
        "        return f\"{'train' if self.in_train else 'valid'}: {self.avg_stats}\"\n",
        "\n",
        "    def accumulate(self, run):\n",
        "        bn = run.xb.shape[0]\n",
        "        self.tot_loss += run.loss * bn\n",
        "        self.count += bn\n",
        "        \n",
        "        for i, m in enumerate(self.metrics):\n",
        "            self.tot_mets[i] += m(run.pred, run.yb) * bn\n",
        "\n",
        "class Recorder(Callback):\n",
        "    def begin_fit(self): \n",
        "        self.lrs = [[] for _ in self.opt_func.param_groups]\n",
        "        self.losses = []\n",
        "\n",
        "    def after_batch(self):\n",
        "        if not self.in_train: return\n",
        "        for pg, lr in zip(self.opt_func.param_groups, self.lrs): lr.append(pg['lr'])\n",
        "        self.losses.append(self.loss.detach().cpu())\n",
        "    \n",
        "    def plot_lr(self, pgid=-1): plt.plot(self.lrs[pgid])\n",
        "    def plot_loss(self, skip_last=0): plt.plot(self.losses[:len(self.losses)-skip_last])\n",
        "    def plot(self, skip_last=0, pgid=-1):\n",
        "        losses = [o.item() for o in self.losses]\n",
        "        lrs = self.lrs[pgid]\n",
        "        n = len(losses)-skip_last\n",
        "        plt.xscale('log')\n",
        "        plt.plot(lrs[:n], losses[:n])"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ArRWTAm6tWl0",
        "colab_type": "text"
      },
      "source": [
        "## 6.1 ParamScheduler"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "dKDri3fGJlx2",
        "colab_type": "text"
      },
      "source": [
        "เราจะสร้าง Callback ที่จะ Schedule Hyperparameter โดยรับชื่อ Hyperparameter และ Schdule Function จากภายนอก\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "dvtme4ONJmMm",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "class ParamScheduler(Callback):\n",
        "    _order = 1\n",
        "    def __init__(self, pname, sched_funcs): self.pname, self.sched_funcs = pname, sched_funcs\n",
        "\n",
        "    def begin_fit(self): \n",
        "        if not isinstance(self.sched_funcs, (list, tuple)): \n",
        "            self.sched_funcs = [self.sched_funcs] * len(self.opt_func.param_groups)\n",
        "\n",
        "    def set_param(self):\n",
        "        assert len(self.opt_func.param_groups) == len(self.sched_funcs)\n",
        "        for pg, f in zip(self.opt_func.param_groups, self.sched_funcs):\n",
        "            pg[self.pname] = f(self.n_epochs/self.epochs)\n",
        "\n",
        "    def begin_batch(self):\n",
        "        if self.in_train: self.set_param()"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "I36khLdmEIm0",
        "colab_type": "text"
      },
      "source": [
        "เราจะใช้ Decorator มาอำนวยความสะดวก ในการสร้าง Partial Function ของ Schedule Function ใด ๆ ที่รับ start, end, pos ให้เหลือเป็น pos ตัวเดียว พร้อมใช้ส่งให้กับ ParamScheduler.set_param ที่จะเรียก f(self.n_epochs/self.epochs)"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "fv1iNbOp_K5I",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def annealer(f):\n",
        "    def _inner(start, end): return partial(f, start, end)\n",
        "    return _inner "
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "d6e8P4_qFPvs",
        "colab_type": "text"
      },
      "source": [
        "ประกาศ Linear Schedule Function ฟังก์ชันกำหนดการแบบเส้นตรง"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "rutKoz1aFObU",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "@annealer\n",
        "def sched_lin(start, end, pos): return start + pos * (end - start)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "WURQ4ipPDgLX",
        "colab_type": "text"
      },
      "source": [
        "ลองเทส sched_lin ช่วงระหว่าง 0 ถึง 5 เทสเรียกด้วย State ต่าง ๆ ของการเทรน"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "eUrnI24L_iJt",
        "colab_type": "code",
        "outputId": "5d594b72-7084-43d7-c6f1-dc7f463dbe2e",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        }
      },
      "source": [
        "f = sched_lin(0, 5)\n",
        "f(0.1), f(0.3), f(0.6), f(0.9)"
      ],
      "execution_count": 22,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(0.5, 1.5, 3.0, 4.5)"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 22
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "aZ9aw08ZD4rP",
        "colab_type": "text"
      },
      "source": [
        "ประกาศ Schedule Function แบบต่าง ๆ เช่น Cosine, No Schedule, Exponential"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "OScLdPMbFVGQ",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "@annealer\n",
        "def sched_cos(start, end, pos): return start + (1 + math.cos(math.pi*(1-pos))) * (end-start) / 2\n",
        "\n",
        "@annealer\n",
        "def sched_no(start, end, pos): return start\n",
        "\n",
        "@annealer\n",
        "def sched_exp(start, end, pos): return start * (end/start) ** pos"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "gLdT8BrrEIJG",
        "colab_type": "text"
      },
      "source": [
        "กำหนด Property Tensor.ndim เพื่อใช้ในการพล็อตกราฟด้วย Matplotlib"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "jlx2hdeTFVWc",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "torch.Tensor.ndim = property(lambda x: len(x.shape))"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "HR7AvLFWEfYn",
        "colab_type": "text"
      },
      "source": [
        "ลองพล็อตกราฟ ฟังก์ชัน Schedule ทั้ง 4 แบบ"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Dg_cIlfNFVee",
        "colab_type": "code",
        "outputId": "3dc7b04e-78f9-456b-9b52-1801ca5bff78",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 269
        }
      },
      "source": [
        "annealings = \"NO LINEAR COS EXP\".split()\n",
        "\n",
        "a = torch.arange(0, 100)\n",
        "p = torch.linspace(0.01, 1, 100)\n",
        "\n",
        "fns = [sched_no, sched_lin, sched_cos, sched_exp]\n",
        "for fn, t in zip(fns, annealings):\n",
        "    f = fn(1e-2, 3)\n",
        "    plt.plot(a, [f(o) for o in p], label=t)\n",
        "plt.legend();"
      ],
      "execution_count": 25,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXcAAAD8CAYAAACMwORRAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9u4mg2bYEoy5",
        "colab_type": "text"
      },
      "source": [
        "ในการใช้งานจริงของ Schedule เรามักจะรวม Schedule แบบต่าง ๆ ในช่วงต่าง ๆ เข้าด้วยกัน เพื่อให้ได้ Curve ที่เราต้องการ"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "yun3oVfCKXHj",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def combine_scheds(pcts, scheds):\n",
        "    assert sum(pcts) == 1.\n",
        "    pcts = tensor([0] + listify(pcts))\n",
        "    assert torch.all(pcts >= 0)\n",
        "    pcts = torch.cumsum(pcts, 0)\n",
        "\n",
        "    def _inner(pos):\n",
        "        idx = (pos >= pcts).nonzero().max()\n",
        "        actual_pos = (pos-pcts[idx]) / (pcts[idx+1]-pcts[idx])\n",
        "        return scheds[idx](actual_pos)\n",
        "    return _inner    "
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "G8vAvZ--Enwp",
        "colab_type": "text"
      },
      "source": [
        "เราจะรวม Cosine Schedule Function 2 ช่วงเข้าด้วยกัน ด้วยสัดส่วน 30/70 ให้ได้กราฟของ Learning Rate เริ่มที่ 0.03 แล้วค่อย ๆ เพิ่มขึ้นจนถึงค่า Max = 0.1 แล้วค่อย ๆ ลดลงมาที่ 0.02 แบบ Smooth "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "miyUFwsJMU0_",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "sched = combine_scheds([0.3, 0.7], [sched_cos(0.03, 0.3), sched_cos(0.3, 0.02)])"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "OxpE0eeZMUqc",
        "colab_type": "code",
        "outputId": "069fad6d-d3d6-4212-ffb4-3aed3719c367",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 286
        }
      },
      "source": [
        "plt.plot(a, [sched(o) for o in p])"
      ],
      "execution_count": 28,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[<matplotlib.lines.Line2D at 0x7fc40b3c8860>]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 28
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAX0AAAD8CAYAAACb4nSYAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nf66KwZXrVesoAAAAAElFTkSuQmCC\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "KwWPi6H0JaKK",
        "colab_type": "text"
      },
      "source": [
        "เราจะใช้ 2 Callback ในการเทรนครั้งนี้ คือ Record Callback ไว้บันทึก Learning Rate, Loss และ ParamScheduler Callback ไว้ Schedule lr (Learning Rate) ดังกราฟด้านบน"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "eb_lm_zyQKac",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "cbfs = [Recorder, partial(ParamScheduler, 'lr', sched)]"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "oCgO5uqZ7TLa",
        "colab_type": "text"
      },
      "source": [
        "# 7. Train"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "Qm6H1b9V8Jbu"
      },
      "source": [
        "เริ่มต้นเทรน ด้วย SGD ไปแค่ 1 Epoch ก็พอ"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "W1tXBrGr8Jbv",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 357
        },
        "outputId": "9eae22c7-6108-4478-cb49-0dcd6b8db309"
      },
      "source": [
        "model, loss_func = get_model()\n",
        "opt = torch.optim.SGD(model.parameters(), lr=lr)\n",
        "learn = Learner(data, model, opt, loss_func=loss_func)\n",
        "\n",
        "run = Runner(cb_funcs=cbfs, cbs=AvgStatsCallback([accuracy]))\n",
        "run.fit(10, learn)"
      ],
      "execution_count": 30,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "train: [0.729900859375, tensor(0.8040)]\n",
            "valid: [0.3770200439453125, tensor(0.8845)]\n",
            "train: [0.3286458203125, tensor(0.9011)]\n",
            "valid: [0.2499768798828125, tensor(0.9273)]\n",
            "train: [0.23506021484375, tensor(0.9287)]\n",
            "valid: [0.1764240478515625, tensor(0.9486)]\n",
            "train: [0.162588486328125, tensor(0.9517)]\n",
            "valid: [0.14234881591796875, tensor(0.9598)]\n",
            "train: [0.125389228515625, tensor(0.9632)]\n",
            "valid: [0.128112841796875, tensor(0.9617)]\n",
            "train: [0.1038401171875, tensor(0.9695)]\n",
            "valid: [0.11910045166015625, tensor(0.9643)]\n",
            "train: [0.089089912109375, tensor(0.9742)]\n",
            "valid: [0.1110241943359375, tensor(0.9651)]\n",
            "train: [0.0791958349609375, tensor(0.9771)]\n",
            "valid: [0.1064181396484375, tensor(0.9659)]\n",
            "train: [0.0729940234375, tensor(0.9791)]\n",
            "valid: [0.1030518310546875, tensor(0.9676)]\n",
            "train: [0.0695858642578125, tensor(0.9806)]\n",
            "valid: [0.101595654296875, tensor(0.9681)]\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "6Wtv7NLWLHjv",
        "colab_type": "text"
      },
      "source": [
        "เทรนเสร็จเรียบร้อย ได้ Accuracy ประมาณ 97%"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "FTxndie97Zuj",
        "colab_type": "text"
      },
      "source": [
        "# 8. Interpret"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "Pa1Pfp0L8Jb1"
      },
      "source": [
        "เราลองพล็อตกราฟ Learning Rate"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "outputId": "e368c3e6-bb48-4f8b-e791-460446294a2e",
        "id": "rajIwpjF8Jb2",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 269
        }
      },
      "source": [
        "run.recorder.plot_lr()"
      ],
      "execution_count": 31,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAX4AAAD8CAYAAABw1c+bAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "M6kFsM668Jb8"
      },
      "source": [
        "พล็อตกราฟ Loss"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "outputId": "cc6a38b7-1cd5-4ab5-8c13-ab97074a4687",
        "id": "Lyaaoryg8Jb9",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 269
        }
      },
      "source": [
        "run.recorder.plot_loss()"
      ],
      "execution_count": 32,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXcAAAD8CAYAAACMwORRAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "kXKZ8_gY8JcE"
      },
      "source": [
        "# Credit"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "-1sztQ5R8JcF"
      },
      "source": [
        "* https://course.fast.ai/videos/?lesson=9\n",
        "* https://arxiv.org/pdf/1803.09820.pdf\n",
        "* http://yann.lecun.com/exdb/mnist/"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "BqGffrmU2juj",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        ""
      ],
      "execution_count": 0,
      "outputs": []
    }
  ]
}