{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%install '.package(path: \"$cwd/FastaiNotebook_04_callbacks\")' FastaiNotebook_04_callbacks"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Demonstrating a simple MNIST model with an array of layers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import FastaiNotebook_04_callbacks\n",
    "var data = mnistDataBunch(flat: true)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import TensorFlow\n",
    "\n",
    "struct DynamicModel: Layer {\n",
    "    public var hiddens: [Dense<Float>]\n",
    "    public var last: Dense<Float>\n",
    "    \n",
    "    init(nIn: Int, nHiddens: [Int], nOut: Int) {\n",
    "        var nLast = nIn\n",
    "        self.hiddens = []\n",
    "        for nHidden in nHiddens {\n",
    "            self.hiddens.append(Dense.init(inputSize: nLast, outputSize: nHidden, activation: relu))\n",
    "            nLast = nHidden\n",
    "        }\n",
    "        self.last = Dense.init(inputSize: nLast, outputSize: nOut, activation: identity)\n",
    "    }\n",
    "    \n",
    "    @differentiable\n",
    "    func applied(to input: Tensor<Float>, in context: Context) -> Tensor<Float> {\n",
    "        return input.sequenced(in: context, through: hiddens, last)\n",
    "    }\n",
    "}\n",
    "\n",
    "/// Model with a fixed number of layers, for comparison.\n",
    "struct FixedModel: Layer {\n",
    "    public var layer1: Dense<Float>\n",
    "    public var layer2: Dense<Float>\n",
    "    public var layer3: Dense<Float>\n",
    "    \n",
    "    public init(nIn: Int, nHid: Int, nOut: Int){\n",
    "        layer1 = Dense(inputSize: nIn, outputSize: nHid, activation: relu)\n",
    "        layer2 = Dense(inputSize: nHid, outputSize: nHid)\n",
    "        layer3 = Dense(inputSize: nHid, outputSize: nOut)\n",
    "    }\n",
    "    \n",
    "    @differentiable\n",
    "    public func applied(to input: Tensor<Float>, in context: Context) -> Tensor<Float> {\n",
    "        return input.sequenced(in: context, through: layer1, layer2, layer3)\n",
    "    }\n",
    "}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "let nIn = 784\n",
    "let nHid = 50\n",
    "let nOut = 10\n",
    "\n",
    "let modelDynamic = DynamicModel(nIn: nIn, nHiddens: [nHid, nHid], nOut: nOut)\n",
    "var modelFixed = FixedModel(nIn: nIn, nHid: nHid, nOut: nOut)\n",
    "\n",
    "// Turn off shuffling, and make the model weights the same, to verify that both models do the same thing.\n",
    "data.shuffleTrain = false\n",
    "modelFixed.layer1 = modelDynamic.hiddens[0]\n",
    "modelFixed.layer2 = modelDynamic.hiddens[1]\n",
    "modelFixed.layer3 = modelDynamic.last\n",
    "\n",
    "let optDynamic = SimpleSGD<DynamicModel, Float>(learningRate: 1e-2)\n",
    "func modelInitDynamic() -> DynamicModel { return modelDynamic }\n",
    "let learnerDynamic = Learner(data: data, lossFunction: softmaxCrossEntropy, optimizer: optDynamic, initializingWith: modelInitDynamic)\n",
    "learnerDynamic.delegates = [learnerDynamic.makeTrainEvalDelegate(), learnerDynamic.makeAvgMetric(metrics: [accuracy])]\n",
    "\n",
    "let optFixed = SimpleSGD<FixedModel, Float>(learningRate: 1e-2)\n",
    "func modelInitFixed() -> FixedModel { return modelFixed }\n",
    "let learnerFixed = Learner(data: data, lossFunction: softmaxCrossEntropy, optimizer: optFixed, initializingWith: modelInitFixed)\n",
    "learnerFixed.delegates = [learnerFixed.makeTrainEvalDelegate(), learnerFixed.makeAvgMetric(metrics: [accuracy])]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "learnerDynamic.fit(3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "learnerFixed.fit(3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Swift",
   "language": "swift",
   "name": "swift"
  },
  "language_info": {
   "file_extension": ".swift",
   "mimetype": "text/x-swift",
   "name": "swift",
   "version": ""
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}