ICTP_2022.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "collapsed_sections": [],
      "toc_visible": true,
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/emiliopaolini/ICTP_2022/blob/main/ICTP_2022.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Quantization in Neural Networks: Advantages and Limitations\n"
      ],
      "metadata": {
        "id": "D4wurFQQVmxk"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Library that we will use:\n",
        "\n",
        "\n",
        "*   TensorFlow + Keras\n",
        "*   Larq\n",
        "\n"
      ],
      "metadata": {
        "id": "9y1xWn1gXfKC"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!pip install larq"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "0-nPtDFLXywX",
        "outputId": "837a865c-e661-4089-de0e-3f13a19926c3"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/\n",
            "Requirement already satisfied: larq in /usr/local/lib/python3.7/dist-packages (0.13.0)\n",
            "Requirement already satisfied: numpy<2.0,>=1.15.4 in /usr/local/lib/python3.7/dist-packages (from larq) (1.21.6)\n",
            "Requirement already satisfied: terminaltables>=3.1.0 in /usr/local/lib/python3.7/dist-packages (from larq) (3.1.10)\n",
            "Requirement already satisfied: importlib-metadata<4.0,>=2.0 in /usr/local/lib/python3.7/dist-packages (from larq) (3.10.1)\n",
            "Requirement already satisfied: packaging>=19.2 in /usr/local/lib/python3.7/dist-packages (from larq) (21.3)\n",
            "Requirement already satisfied: zipp>=0.5 in /usr/local/lib/python3.7/dist-packages (from importlib-metadata<4.0,>=2.0->larq) (3.10.0)\n",
            "Requirement already satisfied: typing-extensions>=3.6.4 in /usr/local/lib/python3.7/dist-packages (from importlib-metadata<4.0,>=2.0->larq) (4.1.1)\n",
            "Requirement already satisfied: pyparsing!=3.0.5,>=2.0.2 in /usr/local/lib/python3.7/dist-packages (from packaging>=19.2->larq) (3.0.9)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "Lvdlc2cZVkuY"
      },
      "outputs": [],
      "source": [
        "#import packages\n",
        "import numpy as np\n",
        "import tensorflow as tf\n",
        "from keras.datasets import mnist\n",
        "import larq as lq\n",
        "import matplotlib.pyplot as plt\n",
        "%matplotlib inline\n",
        "\n",
        "\n",
        "# Set seed\n",
        "tf.random.set_seed(0)\n",
        "tf.keras.utils.set_random_seed(0)   \n",
        "tf.config.experimental.enable_op_determinism()"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Data preprocessing"
      ],
      "metadata": {
        "id": "YFwiCftjZfR8"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "#Load dataset as train and test sets\n",
        "(x_train, y_train), (x_test, y_test) = mnist.load_data()\n",
        "#Plot images\n",
        "num = 10\n",
        "images = x_train[:num]\n",
        "labels = y_train[:num]\n",
        "num_row = 2\n",
        "num_col = 5\n",
        "fig, axes = plt.subplots(num_row, num_col, figsize=(1.5*num_col,2*num_row))\n",
        "for i in range(num):\n",
        "    ax = axes[i//num_col, i%num_col]\n",
        "    ax.imshow(images[i], cmap='gray')\n",
        "    ax.set_title('Label: {}'.format(labels[i]))\n",
        "plt.tight_layout()\n",
        "plt.show()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 278
        },
        "id": "Y2QPtaNyYLsc",
        "outputId": "ce7177b3-5bd5-4709-b574-e620a8f8d452"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 540x288 with 10 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "print(x_train.shape)\n",
        "# Flatten the image: from (60000,28,28) to (60000,28x28)\n",
        "x_train = x_train.reshape(60000, 784,1)\n",
        "x_test = x_test.reshape(10000, 784,1)\n",
        "# interpret as float32\n",
        "x_train = x_train.astype('float32')\n",
        "x_test = x_test.astype('float32')\n",
        "# normalize between [0,1]\n",
        "x_train /= 255\n",
        "x_test /= 255"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "7o5eKARoY0y2",
        "outputId": "c63eb651-547a-4cdb-aa98-3a5aaedcb785"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(60000, 28, 28)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Model definition"
      ],
      "metadata": {
        "id": "cgRIw-mbZyVt"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# Define the type of the model: Sequential\n",
        "model = tf.keras.models.Sequential()\n",
        "\n",
        "# First layer: Fully-Connected, 512 Neurons, ReLU activation function\n",
        "model.add(tf.keras.layers.Dense(512, input_shape=(784, ), activation=\"relu\"))\n",
        "model.add(tf.keras.layers.BatchNormalization())\n",
        "# Second layer: Fully-Connected, 64 Neurons, ReLU activation function\n",
        "model.add(tf.keras.layers.Dense(64, activation=\"relu\"))\n",
        "model.add(tf.keras.layers.BatchNormalization())\n",
        "# Output layer: Fully-Connected, 10 Neurons, Softmax to perform classification \n",
        "model.add(tf.keras.layers.Dense(10, activation=\"softmax\"))\n",
        "\n",
        "model.summary()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "D_rrPVVUY0u_",
        "outputId": "8af5e9e3-d5ec-4f62-f4fc-3ff3b33a5df0"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Model: \"sequential_15\"\n",
            "_________________________________________________________________\n",
            " Layer (type)                Output Shape              Param #   \n",
            "=================================================================\n",
            " dense_9 (Dense)             (None, 512)               401920    \n",
            "                                                                 \n",
            " batch_normalization (BatchN  (None, 512)              2048      \n",
            " ormalization)                                                   \n",
            "                                                                 \n",
            " dense_10 (Dense)            (None, 64)                32832     \n",
            "                                                                 \n",
            " batch_normalization_1 (Batc  (None, 64)               256       \n",
            " hNormalization)                                                 \n",
            "                                                                 \n",
            " dense_11 (Dense)            (None, 10)                650       \n",
            "                                                                 \n",
            "=================================================================\n",
            "Total params: 437,706\n",
            "Trainable params: 436,554\n",
            "Non-trainable params: 1,152\n",
            "_________________________________________________________________\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Float32 Model Training"
      ],
      "metadata": {
        "id": "aHrvHUFUbx7F"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "model.compile(loss=tf.keras.losses.sparse_categorical_crossentropy, optimizer='adam', metrics=['accuracy'])\n",
        "\n",
        "# Fitting the model with the dataset\n",
        "hist = model.fit(x_train, y_train, batch_size=128, epochs=20, verbose=1, validation_data=(x_test,y_test))\n",
        "\n",
        "# Evaluate the model on the test set\n",
        "score = model.evaluate(x_test, y_test, verbose=1)\n",
        "float_accuracy = score[1]\n",
        "\n",
        "print(\"Test loss {:.4f}, accuracy {:.2f}%\".format(score[0], score[1] * 100))\n",
        "\n",
        "#Save model weights to use them later\n",
        "model.save('./float_model')"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "K-VuE_pKb1UH",
        "outputId": "e7d1be90-4061-46d1-b7f7-e9b1e34a98b5"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch 1/20\n",
            "469/469 [==============================] - 8s 15ms/step - loss: 0.1955 - accuracy: 0.9440 - val_loss: 0.1103 - val_accuracy: 0.9662\n",
            "Epoch 2/20\n",
            "469/469 [==============================] - 8s 16ms/step - loss: 0.0737 - accuracy: 0.9779 - val_loss: 0.0897 - val_accuracy: 0.9714\n",
            "Epoch 3/20\n",
            "469/469 [==============================] - 8s 16ms/step - loss: 0.0499 - accuracy: 0.9845 - val_loss: 0.0819 - val_accuracy: 0.9744\n",
            "Epoch 4/20\n",
            "469/469 [==============================] - 7s 14ms/step - loss: 0.0352 - accuracy: 0.9893 - val_loss: 0.0724 - val_accuracy: 0.9774\n",
            "Epoch 5/20\n",
            "469/469 [==============================] - 7s 14ms/step - loss: 0.0294 - accuracy: 0.9907 - val_loss: 0.0756 - val_accuracy: 0.9765\n",
            "Epoch 6/20\n",
            "469/469 [==============================] - 6s 13ms/step - loss: 0.0230 - accuracy: 0.9928 - val_loss: 0.0765 - val_accuracy: 0.9768\n",
            "Epoch 7/20\n",
            "469/469 [==============================] - 6s 12ms/step - loss: 0.0197 - accuracy: 0.9937 - val_loss: 0.0746 - val_accuracy: 0.9794\n",
            "Epoch 8/20\n",
            "469/469 [==============================] - 6s 12ms/step - loss: 0.0189 - accuracy: 0.9941 - val_loss: 0.0774 - val_accuracy: 0.9785\n",
            "Epoch 9/20\n",
            "469/469 [==============================] - 6s 12ms/step - loss: 0.0160 - accuracy: 0.9950 - val_loss: 0.0778 - val_accuracy: 0.9774\n",
            "Epoch 10/20\n",
            "469/469 [==============================] - 6s 12ms/step - loss: 0.0131 - accuracy: 0.9959 - val_loss: 0.0848 - val_accuracy: 0.9788\n",
            "Epoch 11/20\n",
            "469/469 [==============================] - 6s 12ms/step - loss: 0.0129 - accuracy: 0.9959 - val_loss: 0.0742 - val_accuracy: 0.9790\n",
            "Epoch 12/20\n",
            "469/469 [==============================] - 8s 16ms/step - loss: 0.0127 - accuracy: 0.9955 - val_loss: 0.0736 - val_accuracy: 0.9807\n",
            "Epoch 13/20\n",
            "469/469 [==============================] - 7s 14ms/step - loss: 0.0126 - accuracy: 0.9957 - val_loss: 0.0819 - val_accuracy: 0.9796\n",
            "Epoch 14/20\n",
            "469/469 [==============================] - 6s 14ms/step - loss: 0.0100 - accuracy: 0.9966 - val_loss: 0.0857 - val_accuracy: 0.9785\n",
            "Epoch 15/20\n",
            "469/469 [==============================] - 6s 13ms/step - loss: 0.0103 - accuracy: 0.9968 - val_loss: 0.0928 - val_accuracy: 0.9772\n",
            "Epoch 16/20\n",
            "469/469 [==============================] - 6s 13ms/step - loss: 0.0086 - accuracy: 0.9970 - val_loss: 0.0913 - val_accuracy: 0.9776\n",
            "Epoch 17/20\n",
            "469/469 [==============================] - 6s 13ms/step - loss: 0.0074 - accuracy: 0.9975 - val_loss: 0.0834 - val_accuracy: 0.9806\n",
            "Epoch 18/20\n",
            "469/469 [==============================] - 6s 13ms/step - loss: 0.0089 - accuracy: 0.9971 - val_loss: 0.0800 - val_accuracy: 0.9810\n",
            "Epoch 19/20\n",
            "469/469 [==============================] - 6s 13ms/step - loss: 0.0093 - accuracy: 0.9970 - val_loss: 0.0855 - val_accuracy: 0.9786\n",
            "Epoch 20/20\n",
            "469/469 [==============================] - 7s 14ms/step - loss: 0.0084 - accuracy: 0.9973 - val_loss: 0.0932 - val_accuracy: 0.9776\n",
            "313/313 [==============================] - 2s 6ms/step - loss: 0.0932 - accuracy: 0.9776\n",
            "Test loss 0.0932, accuracy 97.76%\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Direct Quantization"
      ],
      "metadata": {
        "id": "lzWSDAdcoGLs"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "bitwidth = 1\n",
        "\n",
        "# Larq definition of input and kernel quantizers\n",
        "kwargs = dict(input_quantizer=lq.quantizers.DoReFa(k_bit=bitwidth, mode=\"activations\"),\n",
        "          kernel_quantizer=lq.quantizers.DoReFa(k_bit=bitwidth, mode=\"weights\"),\n",
        "          kernel_constraint=\"weight_clip\"\n",
        "          )\n",
        "\n",
        "model = tf.keras.models.Sequential()\n",
        "\n",
        "model.add(lq.layers.QuantDense(512, activation=\"relu\", **kwargs,input_shape=(784, )))\n",
        "model.add(tf.keras.layers.BatchNormalization())\n",
        "\n",
        "model.add(lq.layers.QuantDense(64, activation=\"relu\", **kwargs))\n",
        "model.add(tf.keras.layers.BatchNormalization())\n",
        "\n",
        "model.add(lq.layers.QuantDense(10, activation=\"softmax\", **kwargs))\n",
        "\n",
        "model.compile(loss=tf.keras.losses.sparse_categorical_crossentropy, optimizer='adam', metrics=['accuracy'])\n",
        "\n",
        "# Load trained model (float32) \n",
        "model.load_weights('./float_model')\n",
        "\n",
        "\n",
        "\n",
        "# Evaluate the direct quantization\n",
        "score = model.evaluate(x_test, y_test, verbose=1)\n",
        "print(\"Test loss {:.4f}, accuracy {:.2f}%\".format(score[0], score[1] * 100))\n",
        "direct_quantization_accuracy = score[1]"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "hp4I14XLoH_7",
        "outputId": "bfd48dbe-16ca-4254-d800-4f74749a9b07"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "313/313 [==============================] - 3s 6ms/step - loss: 2.8426 - accuracy: 0.5178\n",
            "Test loss 2.8426, accuracy 51.78%\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Quantization-Aware Training"
      ],
      "metadata": {
        "id": "dgAnJyyoqefx"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "bitwidth = 1\n",
        "\n",
        "# Larq definition of input and kernel quantizers\n",
        "kwargs = dict(input_quantizer=lq.quantizers.DoReFa(k_bit=bitwidth, mode=\"activations\"),\n",
        "          kernel_quantizer=lq.quantizers.DoReFa(k_bit=bitwidth, mode=\"weights\"),\n",
        "          kernel_constraint=\"weight_clip\"\n",
        "          )\n",
        "\n",
        "model = tf.keras.models.Sequential()\n",
        "\n",
        "model.add(lq.layers.QuantDense(512, activation=\"relu\", **kwargs,input_shape=(784, )))\n",
        "model.add(tf.keras.layers.BatchNormalization())\n",
        "\n",
        "model.add(lq.layers.QuantDense(64, activation=\"relu\", **kwargs))\n",
        "model.add(tf.keras.layers.BatchNormalization())\n",
        "\n",
        "model.add(lq.layers.QuantDense(10, activation=\"softmax\", **kwargs))\n",
        "\n",
        "model.compile(loss=tf.keras.losses.sparse_categorical_crossentropy, optimizer='adam', metrics=['accuracy'])\n",
        "\n",
        "lq.models.summary(model)\n",
        "\n",
        "# Quantization Aware Training\n",
        "hist = model.fit(x_train, y_train, batch_size=128, epochs=20, verbose=1, validation_data=(x_test,y_test))\n",
        "\n",
        "# Evaluate the QAT\n",
        "score = model.evaluate(x_test, y_test, verbose=1)\n",
        "print(\"Test loss {:.4f}, accuracy {:.2f}%\".format(score[0], score[1] * 100))\n",
        "QAT_accuracy = score[1]\n",
        "\n"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "f9SaDfb4qgjG",
        "outputId": "88342aef-f851-433a-fb00-a2edcde39251"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "+sequential_19 stats-------------------------------------------------------------------+\n",
            "| Layer                  Input prec.    Outputs  # 1-bit  # 32-bit  Memory  1-bit MACs |\n",
            "|                              (bit)                 x 1       x 1    (kB)             |\n",
            "+--------------------------------------------------------------------------------------+\n",
            "| quant_dense_36                   1  (-1, 512)   401408       512   51.00      401408 |\n",
            "| batch_normalization_6            -  (-1, 512)        0      1024    4.00           0 |\n",
            "| quant_dense_37                   1   (-1, 64)    32768        64    4.25       32768 |\n",
            "| batch_normalization_7            -   (-1, 64)        0       128    0.50           0 |\n",
            "| quant_dense_38                   1   (-1, 10)      640        10    0.12         640 |\n",
            "+--------------------------------------------------------------------------------------+\n",
            "| Total                                           434816      1738   59.87      434816 |\n",
            "+--------------------------------------------------------------------------------------+\n",
            "+sequential_19 summary------------------------+\n",
            "| Total params                      437 k     |\n",
            "| Trainable params                  435 k     |\n",
            "| Non-trainable params              1.15 k    |\n",
            "| Model size                        59.87 KiB |\n",
            "| Model size (8-bit FP weights)     54.78 KiB |\n",
            "| Float-32 Equivalent               1.67 MiB  |\n",
            "| Compression Ratio of Memory       0.04      |\n",
            "| Number of MACs                    435 k     |\n",
            "| Ratio of MACs that are binarized  1.0000    |\n",
            "+---------------------------------------------+\n",
            "Epoch 1/20\n",
            "469/469 [==============================] - 11s 19ms/step - loss: 0.7812 - accuracy: 0.7988 - val_loss: 0.4538 - val_accuracy: 0.8749\n",
            "Epoch 2/20\n",
            "469/469 [==============================] - 9s 18ms/step - loss: 0.3806 - accuracy: 0.8943 - val_loss: 0.3478 - val_accuracy: 0.9053\n",
            "Epoch 3/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.2947 - accuracy: 0.9156 - val_loss: 0.3069 - val_accuracy: 0.9127\n",
            "Epoch 4/20\n",
            "469/469 [==============================] - 9s 19ms/step - loss: 0.2438 - accuracy: 0.9296 - val_loss: 0.3019 - val_accuracy: 0.9162\n",
            "Epoch 5/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.2119 - accuracy: 0.9378 - val_loss: 0.2542 - val_accuracy: 0.9293\n",
            "Epoch 6/20\n",
            "469/469 [==============================] - 9s 19ms/step - loss: 0.1910 - accuracy: 0.9423 - val_loss: 0.2367 - val_accuracy: 0.9346\n",
            "Epoch 7/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.1698 - accuracy: 0.9494 - val_loss: 0.2264 - val_accuracy: 0.9397\n",
            "Epoch 8/20\n",
            "469/469 [==============================] - 10s 20ms/step - loss: 0.1546 - accuracy: 0.9538 - val_loss: 0.2015 - val_accuracy: 0.9471\n",
            "Epoch 9/20\n",
            "469/469 [==============================] - 9s 19ms/step - loss: 0.1463 - accuracy: 0.9560 - val_loss: 0.2021 - val_accuracy: 0.9440\n",
            "Epoch 10/20\n",
            "469/469 [==============================] - 9s 18ms/step - loss: 0.1343 - accuracy: 0.9601 - val_loss: 0.1996 - val_accuracy: 0.9503\n",
            "Epoch 11/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.1270 - accuracy: 0.9614 - val_loss: 0.1982 - val_accuracy: 0.9422\n",
            "Epoch 12/20\n",
            "469/469 [==============================] - 9s 18ms/step - loss: 0.1239 - accuracy: 0.9619 - val_loss: 0.2114 - val_accuracy: 0.9422\n",
            "Epoch 13/20\n",
            "469/469 [==============================] - 9s 18ms/step - loss: 0.1196 - accuracy: 0.9636 - val_loss: 0.2633 - val_accuracy: 0.9364\n",
            "Epoch 14/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.1133 - accuracy: 0.9658 - val_loss: 0.2099 - val_accuracy: 0.9451\n",
            "Epoch 15/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.1120 - accuracy: 0.9668 - val_loss: 0.2591 - val_accuracy: 0.9313\n",
            "Epoch 16/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.1041 - accuracy: 0.9689 - val_loss: 0.2046 - val_accuracy: 0.9466\n",
            "Epoch 17/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.0992 - accuracy: 0.9701 - val_loss: 0.2005 - val_accuracy: 0.9487\n",
            "Epoch 18/20\n",
            "469/469 [==============================] - 8s 18ms/step - loss: 0.0967 - accuracy: 0.9711 - val_loss: 0.1912 - val_accuracy: 0.9509\n",
            "Epoch 19/20\n",
            "469/469 [==============================] - 9s 18ms/step - loss: 0.0914 - accuracy: 0.9723 - val_loss: 0.1955 - val_accuracy: 0.9543\n",
            "Epoch 20/20\n",
            "469/469 [==============================] - 9s 18ms/step - loss: 0.0919 - accuracy: 0.9730 - val_loss: 0.2033 - val_accuracy: 0.9536\n",
            "313/313 [==============================] - 2s 6ms/step - loss: 0.2033 - accuracy: 0.9536\n",
            "Test loss 0.2033, accuracy 95.36%\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "print(\"Accuracy results:\")\n",
        "print(\"Float32 {:.2f}%, Direct quantization {:.2f}%, QAT {:.2f}%\".format(float_accuracy*100,direct_quantization_accuracy*100,QAT_accuracy*100))"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "Ybf3J-bLuwm1",
        "outputId": "595a1bd7-308d-4da5-95a1-ac2cb7ea2a6a"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Accuracy results:\n",
            "Float32 97.76%, Direct quantization 51.78%, QAT 95.22%\n"
          ]
        }
      ]
    }
  ]
}