diff --git a/dev/01a_callbacks.ipynb b/dev/01a_callbacks.ipynb
index 9b8e7c4..f2dc0a3 100644
--- a/dev/01a_callbacks.ipynb
+++ b/dev/01a_callbacks.ipynb
@@ -34,7 +34,7 @@
     "    Loss = 10\n",
     "    Metric = 100\n",
     "    Schedule = 200\n",
-    "    History = 300\n",
+    "    Tracker = 300\n",
     "    Logging = 1000\n",
     "    \n",
     "    def __call__(self, index=0):\n",
@@ -116,7 +116,7 @@
     "class History(Callback):\n",
     "    \"\"\"A callback that collects model's metrics during its training.\"\"\"\n",
     "    \n",
-    "    order = Order.History()\n",
+    "    order = Order.Tracker()\n",
     "    \n",
     "    def training_started(self, **kwargs):\n",
     "        self.recorded = None\n",
@@ -153,9 +153,9 @@
     "    \"\"\"\n",
     "    order = Order.Metric()\n",
     "    \n",
-    "    def __init__(self, metric_fn: 'callable', name: str=None):\n",
+    "    def __init__(self, metric_fn: 'callable', alias: str=None):\n",
     "        self.metric_fn = metric_fn\n",
-    "        self.name = name or self.metric_fn.__name__\n",
+    "        self.name = alias or self.metric_fn.__name__\n",
     "    \n",
     "    def epoch_started(self, **kwargs):\n",
     "        self.values = defaultdict(int)\n",
@@ -197,9 +197,14 @@
     "        metrics = merge_dicts([p.last_metrics for p in phases])\n",
     "        values = [f'{k}={autoformat(v)}' for k, v in metrics.items()]\n",
     "        values_string = ', '.join(values)\n",
-    "        string = f'Epoch: {epoch:4d} | {values_string}\\n'\n",
+    "        self.write(f'Epoch: {epoch:4d} | {values_string}\\n')\n",
+    "        \n",
+    "    def interrupted(self, exc, **kwargs):\n",
+    "        self.write(exc)\n",
+    "    \n",
+    "    def write(self, msg):\n",
     "        for stream in self.streams:\n",
-    "            stream.write(string)\n",
+    "            stream.write(msg)\n",
     "            stream.flush()"
    ]
   },
@@ -216,9 +221,9 @@
     "    Each observer has a backward reference to its group via 'group' attribute. The group\n",
     "    keeps a reference to the model which can be used by the \n",
     "    \"\"\"\n",
-    "    def __init__(self, cbs):\n",
+    "    def __init__(self, cbs, model=None):\n",
     "        self._init(cbs)\n",
-    "        self._model = None\n",
+    "        self.model = model\n",
     "        \n",
     "    def _init(self, cbs):\n",
     "        if not cbs:\n",
@@ -234,9 +239,6 @@
     "    def add(self, cb, *cbs):\n",
     "        cbs = [cb] + list(cbs)\n",
     "        self._init(cbs)\n",
-    "        \n",
-    "    def set_model(self, model):\n",
-    "        self._model = model\n",
     "\n",
     "    def training_started(self, **kwargs): self('training_started', **kwargs)    \n",
     "    def training_ended(self, **kwargs): self('training_ended', **kwargs)\n",
diff --git a/dev/02b_phase.ipynb b/dev/02b_phase.ipynb
index 99ba76c..a227fb9 100644
--- a/dev/02b_phase.ipynb
+++ b/dev/02b_phase.ipynb
@@ -105,6 +105,9 @@
     "        for name, values in self.metrics.items():\n",
     "            metrics[f'{self.name}_{name}'] = values\n",
     "        return metrics\n",
+    "    \n",
+    "    def get_last_value(self, metric):\n",
+    "        return self.last_metrics[f'{self.name}_{metric}']\n",
     "\n",
     "    def update(self, loss: float):\n",
     "        self.losses.append(loss)\n",
diff --git a/dev/02c_training.ipynb b/dev/02c_training.ipynb
index bad0a04..5780521 100644
--- a/dev/02c_training.ipynb
+++ b/dev/02c_training.ipynb
@@ -74,7 +74,7 @@
     "        model.to(device)\n",
     "        opt = opt_fn(model, **(opt_params or {}))\n",
     "        cb = create_callbacks(cbs, default_cb)\n",
-    "        cb.set_model(model)\n",
+    "        cb.model = model\n",
     "        \n",
     "        self.model = model\n",
     "        self.opt = opt\n",
@@ -94,7 +94,7 @@
     "                self.train_one_epoch(phases, epoch)\n",
     "            self.cb.training_ended(phases=phases)\n",
     "        except TrainingInterrupted as e:\n",
-    "            self.cb.interrupted(reason=e)\n",
+    "            self.cb.interrupted(exc=e)\n",
     "    \n",
     "    def train_one_epoch(self, phases: list, curr_epoch: int=1):\n",
     "        cb, model, opt = self.cb, self.model, self.opt\n",
@@ -141,7 +141,9 @@
     "#export\n",
     "class TrainingInterrupted(Exception):\n",
     "    def __init__(self, context=None):\n",
-    "        self.context = context"
+    "        self.context = context\n",
+    "    def __str__(self):\n",
+    "        return str(self.context)"
    ]
   },
   {
diff --git a/dev/03a_schedule.ipynb b/dev/03a_schedule.ipynb
index 0cd19fb..212286d 100644
--- a/dev/03a_schedule.ipynb
+++ b/dev/03a_schedule.ipynb
@@ -2,7 +2,7 @@
  "cells": [
   {
    "cell_type": "code",
-   "execution_count": 99,
+   "execution_count": 1,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -13,7 +13,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 100,
+   "execution_count": 2,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -23,13 +23,13 @@
     "import numpy as np\n",
     "import pandas as pd\n",
     "\n",
-    "from loop.callbacks import Callback\n",
+    "from loop.callbacks import Callback, Order\n",
     "from loop.utils import calculate_layout"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 101,
+   "execution_count": 3,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -39,7 +39,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 102,
+   "execution_count": 4,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -48,7 +48,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 103,
+   "execution_count": 5,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -62,7 +62,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 104,
+   "execution_count": 6,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -89,7 +89,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 105,
+   "execution_count": 7,
    "metadata": {},
    "outputs": [
     {
@@ -111,7 +111,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 106,
+   "execution_count": 8,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -144,7 +144,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 107,
+   "execution_count": 9,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -159,7 +159,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 108,
+   "execution_count": 10,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -169,7 +169,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 109,
+   "execution_count": 11,
    "metadata": {},
    "outputs": [
     {
@@ -191,7 +191,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 110,
+   "execution_count": 12,
    "metadata": {},
    "outputs": [
     {
@@ -213,7 +213,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 116,
+   "execution_count": 13,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -221,6 +221,8 @@
     "class Scheduler(Callback):\n",
     "    \"\"\"Updates optimizer's learning rates using provided scheduling function.\"\"\"\n",
     "    \n",
+    "    order = Order.Schedule()\n",
+    "    \n",
     "    def __init__(self, opt, schedule: 'callable', params: list=None, mode: str='batch'):\n",
     "        assert mode in {'batch', 'epoch'}\n",
     "        params = _make_sched_params(params)\n",
@@ -229,14 +231,14 @@
     "        self.params = params\n",
     "        self.n_steps = 0\n",
     "        \n",
-    "    def training_started(self, **params):\n",
+    "    def training_started(self, **kwargs):\n",
     "        self.history = defaultdict(list)\n",
     "        \n",
-    "    def epoch_started(self, epoch, **params):\n",
+    "    def epoch_started(self, epoch, **kwargs):\n",
     "        if self.mode == 'epoch': \n",
     "            self.step(epoch)\n",
     "    \n",
-    "    def batch_started(self, phase, **params):\n",
+    "    def batch_started(self, phase, **kwargs):\n",
     "        if self.mode == 'batch': \n",
     "            if phase.grad:\n",
     "                self.step(phase.batch_index)\n",
@@ -276,7 +278,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 117,
+   "execution_count": 14,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -297,7 +299,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 118,
+   "execution_count": 15,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -312,7 +314,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 119,
+   "execution_count": 16,
    "metadata": {},
    "outputs": [
     {
@@ -347,39 +349,6 @@
     "    \n",
     "sched.plot(['lr', 'weight_decay']);"
    ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 98,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Exported: /home/ck/code/loop/dev/00a_annotations.ipynb -> loop/annotations.py\r\n",
-      "Exported: /home/ck/code/loop/dev/00b_config.ipynb -> loop/config.py\r\n",
-      "Exported: /home/ck/code/loop/dev/00c_utils.ipynb -> loop/utils.py\r\n",
-      "Exported: /home/ck/code/loop/dev/01a_callbacks.ipynb -> loop/callbacks.py\r\n",
-      "Exported: /home/ck/code/loop/dev/01b_modules.ipynb -> loop/modules.py\r\n",
-      "Exported: /home/ck/code/loop/dev/02a_metrics.ipynb -> loop/metrics.py\r\n",
-      "Exported: /home/ck/code/loop/dev/02b_phase.ipynb -> loop/phase.py\r\n",
-      "Exported: /home/ck/code/loop/dev/02c_training.ipynb -> loop/training.py\r\n",
-      "Exported: /home/ck/code/loop/dev/03a_schedule.ipynb -> loop/schedule.py\r\n",
-      "9 notebook(s) exported into folder: loop\r\n"
-     ]
-    }
-   ],
-   "source": [
-    "!python export.py -o loop"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": []
   }
  ],
  "metadata": {
diff --git a/dev/03b_early_stopping.ipynb b/dev/03b_early_stopping.ipynb
new file mode 100644
index 0000000..258f371
--- /dev/null
+++ b/dev/03b_early_stopping.ipynb
@@ -0,0 +1,288 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%matplotlib inline\n",
+    "%reload_ext autoreload\n",
+    "%autoreload 2"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#export\n",
+    "from pathlib import Path\n",
+    "\n",
+    "import torch\n",
+    "\n",
+    "from loop.callbacks import Callback, Order\n",
+    "from loop.utils import autoformat"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#export\n",
+    "class BestMetric(Callback):\n",
+    "    \"\"\"A callback that memorizes the best value of metric.\n",
+    "    \n",
+    "    The class is intended to be a base class for other types of metric trackers that\n",
+    "    perform some action when metric stops to improve.\n",
+    "    \"\"\"\n",
+    "\n",
+    "    def __init__(self, phase: str='valid', metric: str='loss', better: 'callable'=min):\n",
+    "        self.phase = phase\n",
+    "        self.metric = metric\n",
+    "        self.better = better\n",
+    "        \n",
+    "    @property\n",
+    "    def formatted_best(self):\n",
+    "        return f'{self.phase}_{self.metric}={autoformat(self.best_value)}'\n",
+    "        \n",
+    "    def training_started(self, **kwargs):\n",
+    "        self.best_value = None\n",
+    "        \n",
+    "    def epoch_started(self, **kwargs):\n",
+    "        self.updated = False\n",
+    "        \n",
+    "    def phase_ended(self, phase, **kwargs):\n",
+    "        ignore = phase.name != self.phase\n",
+    "        if not ignore:\n",
+    "            self.update_best(phase, **kwargs)\n",
+    "        return ignore\n",
+    "        \n",
+    "    def update_best(self, phase, **kwargs):\n",
+    "        new_value = phase.get_last_value(self.metric)\n",
+    "        if self.best_value is None:\n",
+    "            self.best_value = new_value\n",
+    "        else:\n",
+    "            self.best_value = self.better(self.best_value, new_value)\n",
+    "        self.updated = self.best_value == new_value\n",
+    "        return self.updated"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#export\n",
+    "class EarlyStopping(BestMetric):\n",
+    "    \n",
+    "    order = Order.Tracker(1)\n",
+    "    \n",
+    "    def __init__(self, patience: int=1, **kwargs):\n",
+    "        super().__init__(**kwargs)\n",
+    "        self.patience = patience\n",
+    "    \n",
+    "    def training_started(self, **kwargs):\n",
+    "        super().training_started(**kwargs)\n",
+    "        self.trials = 0\n",
+    "        self.running = True\n",
+    "        \n",
+    "    def phase_ended(self, phase, **kwargs):\n",
+    "        ignore = super().phase_ended(phase=phase, **kwargs)\n",
+    "        if ignore: return\n",
+    "        if self.updated: \n",
+    "            self.trials = 0\n",
+    "        else:\n",
+    "            self.trials += 1\n",
+    "            if self.trials >= self.patience:\n",
+    "                breakpoint()\n",
+    "                self.running = False\n",
+    "        \n",
+    "    def epoch_ended(self, phases, epoch, **kwargs):\n",
+    "        super().epoch_ended(phases=phases, epoch=epoch, **kwargs)\n",
+    "        if self.running: return\n",
+    "        from loop.training import TrainingInterrupted\n",
+    "        msg = f'Early stopping at epoch {epoch} with {self.formatted_best}'\n",
+    "        raise TrainingInterrupted(msg)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#export\n",
+    "class ModelSaver(BestMetric):\n",
+    "    \n",
+    "    order = Order.Tracker(2)\n",
+    "    \n",
+    "    def __init__(self, mode: str='every', root: Path=Path.cwd(), **kwargs):\n",
+    "        super().__init__(**kwargs)\n",
+    "        assert mode in {'every', 'best'}\n",
+    "        self.root = Path(root)\n",
+    "        self.mode = mode\n",
+    "        \n",
+    "    def training_started(self, **kwargs):\n",
+    "        super().training_started(**kwargs)\n",
+    "        if not self.root.exists():\n",
+    "            self.root.mkdir(parents=True)\n",
+    "        self.last_saved = None\n",
+    "        \n",
+    "    def epoch_ended(self, phases, epoch, **kwargs):\n",
+    "        super().epoch_ended(phases=phases, epoch=epoch, **kwargs)\n",
+    "        fname = f'model__{self.formatted_best}__epoch={epoch}.pth'\n",
+    "        if self.mode == 'every' or self.updated:\n",
+    "            path = self.root/fname\n",
+    "            torch.save(self.group.model, path)\n",
+    "            self.last_saved = path\n",
+    "            \n",
+    "    def load_last_saved_state(self, model=None):\n",
+    "        if self.last_saved is None:\n",
+    "            raise ValueError('nothing was saved during training')\n",
+    "        model = model or self.group.model\n",
+    "        if model is None:\n",
+    "            raise ValueError('no model provided to restore the saved state')\n",
+    "        model.load_state_dict(torch.load(self.last_saved))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 20,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Exception ignored in: <function _DataLoaderIter.__del__ at 0x7eff21d1d950>\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/home/ck/anaconda3/envs/fastai_10/lib/python3.7/site-packages/torch/utils/data/dataloader.py\", line 717, in __del__\n",
+      "    self._shutdown_workers()\n",
+      "  File \"/home/ck/anaconda3/envs/fastai_10/lib/python3.7/site-packages/torch/utils/data/dataloader.py\", line 713, in _shutdown_workers\n",
+      "    w.join()\n",
+      "  File \"/home/ck/anaconda3/envs/fastai_10/lib/python3.7/multiprocessing/process.py\", line 138, in join\n",
+      "    assert self._parent_pid == os.getpid(), 'can only join a child process'\n",
+      "AssertionError: can only join a child process\n",
+      "Exception ignored in: <function _DataLoaderIter.__del__ at 0x7eff21d1d950>\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/home/ck/anaconda3/envs/fastai_10/lib/python3.7/site-packages/torch/utils/data/dataloader.py\", line 717, in __del__\n",
+      "    self._shutdown_workers()\n",
+      "  File \"/home/ck/anaconda3/envs/fastai_10/lib/python3.7/site-packages/torch/utils/data/dataloader.py\", line 713, in _shutdown_workers\n",
+      "    w.join()\n",
+      "  File \"/home/ck/anaconda3/envs/fastai_10/lib/python3.7/multiprocessing/process.py\", line 138, in join\n",
+      "    assert self._parent_pid == os.getpid(), 'can only join a child process'\n",
+      "AssertionError: can only join a child process\n"
+     ]
+    },
+    {
+     "ename": "KeyError",
+     "evalue": "'valid_acc'",
+     "output_type": "error",
+     "traceback": [
+      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
+      "\u001b[0;31mKeyError\u001b[0m                                  Traceback (most recent call last)",
+      "\u001b[0;32m<ipython-input-20-81e2813f91e1>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m      6\u001b[0m        ModelSaver(mode='best', root=Path.home()/'models')]\n\u001b[1;32m      7\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 8\u001b[0;31m \u001b[0mtrain_classifier_with_callbacks\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mTinyNet\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m1\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mcbs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mcbs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mn\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m10000\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mbs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m100\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m      9\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     10\u001b[0m \u001b[0mloop\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcb\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m'history'\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mplot\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m~/code/loop/dev/loop/testing.py\u001b[0m in \u001b[0;36mtrain_classifier_with_callbacks\u001b[0;34m(model, cbs, n, bs)\u001b[0m\n\u001b[1;32m     38\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mtrain_classifier_with_callbacks\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mcbs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mn\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mbs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m1024\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     39\u001b[0m     \u001b[0mloop\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mLoop\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mcbs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mcbs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mloss_fn\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mF\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcross_entropy\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 40\u001b[0;31m     \u001b[0mloop\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mfit_datasets\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0mget_mnist\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepochs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mn\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mbatch_size\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mbs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     41\u001b[0m     \u001b[0;32mreturn\u001b[0m \u001b[0mloop\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m~/code/loop/dev/loop/training.py\u001b[0m in \u001b[0;36mfit_datasets\u001b[0;34m(self, trn_ds, val_ds, epochs, batch_size)\u001b[0m\n\u001b[1;32m     50\u001b[0m         phases = Phase.make_train_valid(\n\u001b[1;32m     51\u001b[0m             trn_ds, val_ds, bs=batch_size, num_workers=defaults.n_jobs)\n\u001b[0;32m---> 52\u001b[0;31m         \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtrain\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphases\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepochs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     53\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     54\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mtrain\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mphases\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mlist\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepochs\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mint\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m1\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m~/code/loop/dev/loop/training.py\u001b[0m in \u001b[0;36mtrain\u001b[0;34m(self, phases, epochs)\u001b[0m\n\u001b[1;32m     56\u001b[0m             \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcb\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtraining_started\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphases\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mphases\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     57\u001b[0m             \u001b[0;32mfor\u001b[0m \u001b[0mepoch\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mrange\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m1\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepochs\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0;36m1\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 58\u001b[0;31m                 \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtrain_one_epoch\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphases\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepoch\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     59\u001b[0m             \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcb\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtraining_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphases\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mphases\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     60\u001b[0m         \u001b[0;32mexcept\u001b[0m \u001b[0mTrainingInterrupted\u001b[0m \u001b[0;32mas\u001b[0m \u001b[0me\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m~/code/loop/dev/loop/training.py\u001b[0m in \u001b[0;36mtrain_one_epoch\u001b[0;34m(self, phases, curr_epoch)\u001b[0m\n\u001b[1;32m     92\u001b[0m                 \u001b[0mcb\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbatch_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphase\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mphase\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0moutput\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mout\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtarget\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0my\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     93\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 94\u001b[0;31m             \u001b[0mcb\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mphase_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphase\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mphase\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     95\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     96\u001b[0m         \u001b[0mcb\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mepoch_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphases\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mphases\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepoch\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mcurr_epoch\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m~/code/loop/dev/loop/callbacks.py\u001b[0m in \u001b[0;36mphase_ended\u001b[0;34m(self, **kwargs)\u001b[0m\n\u001b[1;32m    193\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mepoch_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'epoch_ended'\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    194\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mphase_started\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'phase_started'\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 195\u001b[0;31m     \u001b[0;32mdef\u001b[0m \u001b[0mphase_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'phase_ended'\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    196\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mbatch_started\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'batch_started'\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    197\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mbatch_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'batch_ended'\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m~/code/loop/dev/loop/callbacks.py\u001b[0m in \u001b[0;36m__call__\u001b[0;34m(self, name, **kwargs)\u001b[0m\n\u001b[1;32m    222\u001b[0m             \u001b[0;32mif\u001b[0m \u001b[0mmethod\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    223\u001b[0m                 \u001b[0;32mcontinue\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 224\u001b[0;31m             \u001b[0mmethod\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
+      "\u001b[0;32m<ipython-input-12-a648c3b38f8c>\u001b[0m in \u001b[0;36mphase_ended\u001b[0;34m(self, phase, **kwargs)\u001b[0m\n\u001b[1;32m     14\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     15\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mphase_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mphase\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 16\u001b[0;31m         \u001b[0mignore\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0msuper\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mphase_ended\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphase\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mphase\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     17\u001b[0m         \u001b[0;32mif\u001b[0m \u001b[0mignore\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0;32mreturn\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     18\u001b[0m         \u001b[0;32mif\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mupdated\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m<ipython-input-11-b03ebf0bf7a8>\u001b[0m in \u001b[0;36mphase_ended\u001b[0;34m(self, phase, **kwargs)\u001b[0m\n\u001b[1;32m     25\u001b[0m         \u001b[0mignore\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mphase\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mname\u001b[0m \u001b[0;34m!=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mphase\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     26\u001b[0m         \u001b[0;32mif\u001b[0m \u001b[0;32mnot\u001b[0m \u001b[0mignore\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 27\u001b[0;31m             \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mupdate_best\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mphase\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     28\u001b[0m         \u001b[0;32mreturn\u001b[0m \u001b[0mignore\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     29\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m<ipython-input-11-b03ebf0bf7a8>\u001b[0m in \u001b[0;36mupdate_best\u001b[0;34m(self, phase, **kwargs)\u001b[0m\n\u001b[1;32m     29\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     30\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mupdate_best\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mphase\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 31\u001b[0;31m         \u001b[0mnew_value\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mphase\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mget_last_value\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmetric\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     32\u001b[0m         \u001b[0;32mif\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbest_value\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     33\u001b[0m             \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbest_value\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mnew_value\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m~/code/loop/dev/loop/phase.py\u001b[0m in \u001b[0;36mget_last_value\u001b[0;34m(self, metric)\u001b[0m\n\u001b[1;32m     80\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     81\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mget_last_value\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmetric\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 82\u001b[0;31m         \u001b[0;32mreturn\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mlast_metrics\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34mf'{self.name}_{metric}'\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     83\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     84\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mupdate\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mloss\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mfloat\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;31mKeyError\u001b[0m: 'valid_acc'"
+     ]
+    }
+   ],
+   "source": [
+    "from loop.testing import train_classifier_with_callbacks\n",
+    "from loop.modules import TinyNet\n",
+    "\n",
+    "cbs = [C.Average(accuracy, alias='acc'), \n",
+    "       EarlyStopping(metric='acc', patience=1),\n",
+    "       ModelSaver(mode='best', metric='acc', root=Path.home()/'models')]\n",
+    "\n",
+    "train_classifier_with_callbacks(TinyNet(1), cbs=cbs, n=10000, bs=100)\n",
+    "\n",
+    "loop.cb['history'].plot()\n",
+    "cbs['early_stopping'].load_last_saved_state()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Exported: /home/ck/code/loop/dev/00a_annotations.ipynb -> loop/annotations.py\r\n",
+      "Exported: /home/ck/code/loop/dev/00b_config.ipynb -> loop/config.py\r\n",
+      "Exported: /home/ck/code/loop/dev/00c_utils.ipynb -> loop/utils.py\r\n",
+      "Exported: /home/ck/code/loop/dev/01a_callbacks.ipynb -> loop/callbacks.py\r\n",
+      "Exported: /home/ck/code/loop/dev/01b_modules.ipynb -> loop/modules.py\r\n",
+      "Exported: /home/ck/code/loop/dev/02a_metrics.ipynb -> loop/metrics.py\r\n",
+      "Exported: /home/ck/code/loop/dev/02b_phase.ipynb -> loop/phase.py\r\n",
+      "Exported: /home/ck/code/loop/dev/02c_training.ipynb -> loop/training.py\r\n",
+      "Exported: /home/ck/code/loop/dev/03a_schedule.ipynb -> loop/schedule.py\r\n",
+      "Exported: /home/ck/code/loop/dev/03b_early_stopping.ipynb -> loop/early_stopping.py\r\n",
+      "Exported: /home/ck/code/loop/dev/99_testing.ipynb -> loop/testing.py\r\n",
+      "11 notebook(s) exported into folder: loop\r\n"
+     ]
+    }
+   ],
+   "source": [
+    "!python export.py -o loop"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "fastai (cuda 10)",
+   "language": "python",
+   "name": "fastai"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.7.1"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
diff --git a/dev/99_testing.ipynb b/dev/99_testing.ipynb
new file mode 100644
index 0000000..5c0ed17
--- /dev/null
+++ b/dev/99_testing.ipynb
@@ -0,0 +1,97 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#export\n",
+    "from torch.nn import functional as F\n",
+    "\n",
+    "from loop import callbacks as C\n",
+    "from loop.config import defaults\n",
+    "from loop.metrics import accuracy\n",
+    "from loop.modules import TinyNet\n",
+    "from loop.training import Loop"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#export\n",
+    "def get_mnist():\n",
+    "    from torchvision.datasets import MNIST\n",
+    "    from torchvision import transforms as T\n",
+    "    \n",
+    "    root = defaults.datasets/'mnist'\n",
+    "\n",
+    "    mnist_stats = ([0.15]*1, [0.15]*1)\n",
+    "\n",
+    "    trn_ds = MNIST(root, train=True, transform=T.Compose([\n",
+    "        T.Resize(32),\n",
+    "        T.RandomAffine(5, translate=(0.05, 0.05), scale=(0.9, 1.1)),\n",
+    "        T.ToTensor(),\n",
+    "        T.Normalize(*mnist_stats)\n",
+    "    ]))\n",
+    "    val_ds = MNIST(root, train=False, transform=T.Compose([\n",
+    "        T.Resize(32),\n",
+    "        T.ToTensor(),\n",
+    "        T.Normalize(*mnist_stats)\n",
+    "    ]))\n",
+    "    \n",
+    "    return trn_ds, val_ds"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#export\n",
+    "def train_classifier_with_callbacks(model, cbs, n, bs=1024):\n",
+    "    loop = Loop(model, cbs=cbs, loss_fn=F.cross_entropy)\n",
+    "    loop.fit_datasets(*get_mnist(), epochs=n, batch_size=bs)\n",
+    "    return loop"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "fastai (cuda 10)",
+   "language": "python",
+   "name": "fastai"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.7.1"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
diff --git a/dev/loop/callbacks.py b/dev/loop/callbacks.py
index 964dc2f..d5c59cc 100644
--- a/dev/loop/callbacks.py
+++ b/dev/loop/callbacks.py
@@ -23,7 +23,7 @@ class Order(IntFlag):
     Loss = 10
     Metric = 100
     Schedule = 200
-    History = 300
+    Tracker = 300
     Logging = 1000
 
     def __call__(self, index=0):
@@ -84,7 +84,7 @@ def epoch_ended(self, phases, **kwargs):
 class History(Callback):
     """A callback that collects model's metrics during its training."""
 
-    order = Order.History()
+    order = Order.Tracker()
 
     def training_started(self, **kwargs):
         self.recorded = None
@@ -114,9 +114,9 @@ class Average(Callback):
     """
     order = Order.Metric()
 
-    def __init__(self, metric_fn: 'callable', name: str=None):
+    def __init__(self, metric_fn: 'callable', alias: str=None):
         self.metric_fn = metric_fn
-        self.name = name or self.metric_fn.__name__
+        self.name = alias or self.metric_fn.__name__
 
     def epoch_started(self, **kwargs):
         self.values = defaultdict(int)
@@ -151,9 +151,14 @@ def epoch_ended(self, phases, epoch, **kwargs):
         metrics = merge_dicts([p.last_metrics for p in phases])
         values = [f'{k}={autoformat(v)}' for k, v in metrics.items()]
         values_string = ', '.join(values)
-        string = f'Epoch: {epoch:4d} | {values_string}\n'
+        self.write(f'Epoch: {epoch:4d} | {values_string}\n')
+
+    def interrupted(self, exc, **kwargs):
+        self.write(exc)
+
+    def write(self, msg):
         for stream in self.streams:
-            stream.write(string)
+            stream.write(msg)
             stream.flush()
 
 
@@ -163,9 +168,9 @@ class Group(Callback):
     Each observer has a backward reference to its group via 'group' attribute. The group
     keeps a reference to the model which can be used by the
     """
-    def __init__(self, cbs):
+    def __init__(self, cbs, model=None):
         self._init(cbs)
-        self._model = None
+        self.model = model
 
     def _init(self, cbs):
         if not cbs:
@@ -182,9 +187,6 @@ def add(self, cb, *cbs):
         cbs = [cb] + list(cbs)
         self._init(cbs)
 
-    def set_model(self, model):
-        self._model = model
-
     def training_started(self, **kwargs): self('training_started', **kwargs)
     def training_ended(self, **kwargs): self('training_ended', **kwargs)
     def epoch_started(self, **kwargs): self('epoch_started', **kwargs)
diff --git a/dev/loop/early_stopping.py b/dev/loop/early_stopping.py
new file mode 100644
index 0000000..297b3aa
--- /dev/null
+++ b/dev/loop/early_stopping.py
@@ -0,0 +1,114 @@
+# -----------------------------------------
+# THIS FILE WAS AUTOGENERATED! DO NOT EDIT!
+# -----------------------------------------
+# file to edit: 03b_early_stopping.ipynb
+
+from pathlib import Path
+
+import torch
+
+from loop.callbacks import Callback, Order
+from loop.utils import autoformat
+
+
+class BestMetric(Callback):
+    """A callback that memorizes the best value of metric.
+
+    The class is intended to be a base class for other types of metric trackers that
+    perform some action when metric stops to improve.
+    """
+
+    def __init__(self, phase: str='valid', metric: str='loss', better: 'callable'=min):
+        self.phase = phase
+        self.metric = metric
+        self.better = better
+
+    @property
+    def formatted_best(self):
+        return f'{self.phase}_{self.metric}={autoformat(self.best_value)}'
+
+    def training_started(self, **kwargs):
+        self.best_value = None
+
+    def epoch_started(self, **kwargs):
+        self.updated = False
+
+    def phase_ended(self, phase, **kwargs):
+        ignore = phase.name != self.phase
+        if not ignore:
+            self.update_best(phase, **kwargs)
+        return ignore
+
+    def update_best(self, phase, **kwargs):
+        new_value = phase.get_last_value(self.metric)
+        if self.best_value is None:
+            self.best_value = new_value
+        else:
+            self.best_value = self.better(self.best_value, new_value)
+        self.updated = self.best_value == new_value
+        return self.updated
+
+
+class EarlyStopping(BestMetric):
+
+    order = Order.Tracker(1)
+
+    def __init__(self, patience: int=1, **kwargs):
+        super().__init__(**kwargs)
+        self.patience = patience
+
+    def training_started(self, **kwargs):
+        super().training_started(**kwargs)
+        self.trials = 0
+        self.running = True
+
+    def phase_ended(self, phase, **kwargs):
+        ignore = super().phase_ended(phase=phase, **kwargs)
+        if ignore: return
+        if self.updated:
+            self.trials = 0
+        else:
+            self.trials += 1
+            if self.trials >= self.patience:
+                breakpoint()
+                self.running = False
+
+    def epoch_ended(self, phases, epoch, **kwargs):
+        super().epoch_ended(phases=phases, epoch=epoch, **kwargs)
+        if self.running: return
+        from loop.training import TrainingInterrupted
+        msg = f'Early stopping at epoch {epoch} with {self.formatted_best}'
+        raise TrainingInterrupted(msg)
+
+
+class ModelSaver(BestMetric):
+
+    order = Order.Tracker(2)
+
+    def __init__(self, mode: str='every', root: Path=Path.cwd(), **kwargs):
+        super().__init__(**kwargs)
+        assert mode in {'every', 'best'}
+        self.root = Path(root)
+        self.mode = mode
+
+    def training_started(self, **kwargs):
+        super().training_started(**kwargs)
+        if not self.root.exists():
+            self.root.mkdir(parents=True)
+        self.last_saved = None
+
+    def epoch_ended(self, phases, epoch, **kwargs):
+        super().epoch_ended(phases=phases, epoch=epoch, **kwargs)
+        fname = f'model__{self.formatted_best}__epoch={epoch}.pth'
+        if self.mode == 'every' or self.updated:
+            path = self.root/fname
+            torch.save(self.group.model, path)
+            self.last_saved = path
+
+    def load_last_saved_state(self, model=None):
+        if self.last_saved is None:
+            raise ValueError('nothing was saved during training')
+        model = model or self.group.model
+        if model is None:
+            raise ValueError('no model provided to restore the saved state')
+        model.load_state_dict(torch.load(self.last_saved))
diff --git a/dev/loop/phase.py b/dev/loop/phase.py
index 71b3855..e2afab9 100644
--- a/dev/loop/phase.py
+++ b/dev/loop/phase.py
@@ -78,6 +78,9 @@ def metrics_history(self):
             metrics[f'{self.name}_{name}'] = values
         return metrics
 
+    def get_last_value(self, metric):
+        return self.last_metrics[f'{self.name}_{metric}']
+
     def update(self, loss: float):
         self.losses.append(loss)
 
diff --git a/dev/loop/schedule.py b/dev/loop/schedule.py
index 7839599..42bc560 100644
--- a/dev/loop/schedule.py
+++ b/dev/loop/schedule.py
@@ -8,7 +8,7 @@
 import numpy as np
 import pandas as pd
 
-from loop.callbacks import Callback
+from loop.callbacks import Callback, Order
 from loop.utils import calculate_layout
 
 
@@ -72,6 +72,8 @@ def __call__(self, t):
 class Scheduler(Callback):
     """Updates optimizer's learning rates using provided scheduling function."""
 
+    order = Order.Schedule()
+
     def __init__(self, opt, schedule: 'callable', params: list=None, mode: str='batch'):
         assert mode in {'batch', 'epoch'}
         params = _make_sched_params(params)
@@ -80,14 +82,14 @@ def __init__(self, opt, schedule: 'callable', params: list=None, mode: str='batc
         self.params = params
         self.n_steps = 0
 
-    def training_started(self, **params):
+    def training_started(self, **kwargs):
         self.history = defaultdict(list)
 
-    def epoch_started(self, epoch, **params):
+    def epoch_started(self, epoch, **kwargs):
         if self.mode == 'epoch':
             self.step(epoch)
 
-    def batch_started(self, phase, **params):
+    def batch_started(self, phase, **kwargs):
         if self.mode == 'batch':
             if phase.grad:
                 self.step(phase.batch_index)
@@ -121,7 +123,8 @@ def plot(self, params=None, axes=None):
             f, axes = plt.subplots(*calculate_layout(n))
         for i, param in enumerate(params):
             df.plot(x='iteration', y=param, ax=axes.flat[i])
-        return f
+        f.tight_layout()
+        return axes
 
 
 def _make_sched_params(params):
diff --git a/dev/loop/testing.py b/dev/loop/testing.py
new file mode 100644
index 0000000..24e5025
--- /dev/null
+++ b/dev/loop/testing.py
@@ -0,0 +1,41 @@
+# -----------------------------------------
+# THIS FILE WAS AUTOGENERATED! DO NOT EDIT!
+# -----------------------------------------
+# file to edit: 99_testing.ipynb
+
+from torch.nn import functional as F
+
+from loop import callbacks as C
+from loop.config import defaults
+from loop.metrics import accuracy
+from loop.modules import TinyNet
+from loop.training import Loop
+
+
+def get_mnist():
+    from torchvision.datasets import MNIST
+    from torchvision import transforms as T
+
+    root = defaults.datasets/'mnist'
+
+    mnist_stats = ([0.15]*1, [0.15]*1)
+
+    trn_ds = MNIST(root, train=True, transform=T.Compose([
+        T.Resize(32),
+        T.RandomAffine(5, translate=(0.05, 0.05), scale=(0.9, 1.1)),
+        T.ToTensor(),
+        T.Normalize(*mnist_stats)
+    ]))
+    val_ds = MNIST(root, train=False, transform=T.Compose([
+        T.Resize(32),
+        T.ToTensor(),
+        T.Normalize(*mnist_stats)
+    ]))
+
+    return trn_ds, val_ds
+
+
+def train_classifier_with_callbacks(model, cbs, n, bs=1024):
+    loop = Loop(model, cbs=cbs, loss_fn=F.cross_entropy)
+    loop.fit_datasets(*get_mnist(), epochs=n, batch_size=bs)
+    return loop
diff --git a/dev/loop/training.py b/dev/loop/training.py
index 2cdff9f..f79de62 100644
--- a/dev/loop/training.py
+++ b/dev/loop/training.py
@@ -38,7 +38,7 @@ def __init__(self, model: nn.Module, cbs: list=None,
         model.to(device)
         opt = opt_fn(model, **(opt_params or {}))
         cb = create_callbacks(cbs, default_cb)
-        cb.set_model(model)
+        cb.model = model
 
         self.model = model
         self.opt = opt
@@ -58,7 +58,7 @@ def train(self, phases: list, epochs: int=1):
                 self.train_one_epoch(phases, epoch)
             self.cb.training_ended(phases=phases)
         except TrainingInterrupted as e:
-            self.cb.interrupted(reason=e)
+            self.cb.interrupted(exc=e)
 
     def train_one_epoch(self, phases: list, curr_epoch: int=1):
         cb, model, opt = self.cb, self.model, self.opt
@@ -99,6 +99,8 @@ def train_one_epoch(self, phases: list, curr_epoch: int=1):
 class TrainingInterrupted(Exception):
     def __init__(self, context=None):
         self.context = context
+    def __str__(self):
+        return str(self.context)
 
 
 def place_and_unwrap(batch, device):