tflite-part-2.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "tf-lite-part-2.ipynb",
      "provenance": [],
      "private_outputs": true,
      "collapsed_sections": []
    },
    "kernelspec": {
      "display_name": "Python 3",
      "name": "python3"
    }
  },
  "cells": [
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "dzLKpmZICaWN",
        "colab": {}
      },
      "source": [
        "import os\n",
        "import tensorflow as tf\n",
        "import h5py\n",
        "from tensorflow import keras\n",
        "import numpy as np\n",
        "import matplotlib.pyplot as plt\n",
        "from tensorflow.keras.layers import Flatten\n",
        "from tensorflow.keras.layers import Dense\n",
        "from tensorflow.keras.losses import SparseCategoricalCrossentropy\n",
        "from sklearn.metrics import accuracy_score\n",
        "from sys import getsizeof"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "NZKNhx8x_1jq",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "print(tf.__version__)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ymcbpqPdLJxW",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def get_file_size(file_path):\n",
        "    size = os.path.getsize(file_path)\n",
        "    return size"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "_vfRLdF_LKUK",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def convert_bytes(size, unit=None):\n",
        "    if unit == \"KB\":\n",
        "        return print('File size: ' + str(round(size / 1024, 3)) + ' Kilobytes')\n",
        "    elif unit == \"MB\":\n",
        "        return print('File size: ' + str(round(size / (1024 * 1024), 3)) + ' Megabytes')\n",
        "    else:\n",
        "        return print('File size: ' + str(size) + ' bytes')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "yR0EdgrLCaWR"
      },
      "source": [
        "## Import the Fashion MNIST dataset"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "DLdCchMdCaWQ"
      },
      "source": [
        "This guide uses the [Fashion MNIST](https://github.com/zalandoresearch/fashion-mnist) dataset which contains 70,000 grayscale images in 10 categories. The images show individual articles of clothing at low resolution (28 by 28 pixels), as seen here:\n",
        "\n",
        "<table>\n",
        "  <tr><td>\n",
        "    <img src=\"https://tensorflow.org/images/fashion-mnist-sprite.png\"\n",
        "         alt=\"Fashion MNIST sprite\"  width=\"600\">\n",
        "  </td></tr>\n",
        "  <tr><td align=\"center\">\n",
        "    <b>Figure 1.</b> <a href=\"https://github.com/zalandoresearch/fashion-mnist\">Fashion-MNIST samples</a> (by Zalando, MIT License).<br/>&nbsp;\n",
        "  </td></tr>\n",
        "</table>\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "7MqDQO0KCaWS",
        "colab": {}
      },
      "source": [
        "fashion_mnist = keras.datasets.fashion_mnist\n",
        "(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "IjnLH5S2CaWx",
        "colab": {}
      },
      "source": [
        "class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',\n",
        "               'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "Brm0b_KACaWX"
      },
      "source": [
        "## Explore the data"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "zW5k_xz1CaWX",
        "colab": {}
      },
      "source": [
        "train_images.shape"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "TRFYHB2mCaWb",
        "colab": {}
      },
      "source": [
        "len(train_labels)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "XKnCTHz4CaWg",
        "colab": {}
      },
      "source": [
        "np.unique(train_labels)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "sqTxmbVBASQF",
        "colab_type": "text"
      },
      "source": [
        "# Test Dataset"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "2KFnYlcwCaWl",
        "colab": {}
      },
      "source": [
        "test_images.shape"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "iJmPr5-ACaWn",
        "colab": {}
      },
      "source": [
        "len(test_labels)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "ES6uQoLKCaWr"
      },
      "source": [
        "## Preprocessing"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "m4VEw8Ud9Quh",
        "colab": {}
      },
      "source": [
        "plt.figure()\n",
        "plt.imshow(train_images[88])\n",
        "plt.colorbar()\n",
        "plt.grid(False)\n",
        "plt.show()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "bW5WzIPlCaWv",
        "colab": {}
      },
      "source": [
        "train_images = train_images / 255.0\n",
        "test_images = test_images / 255.0"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "colab_type": "text",
        "id": "59veuiEZCaW4"
      },
      "source": [
        "## Build & Compile the model\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "9ODch-OFCaW4",
        "colab": {}
      },
      "source": [
        "model = keras.Sequential([\n",
        "    Flatten(input_shape=(28, 28)),\n",
        "    Dense(128, activation='relu'),\n",
        "    Dense(10)\n",
        "])"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "Lhan11blCaW7",
        "colab": {}
      },
      "source": [
        "model.compile(optimizer='adam',\n",
        "              loss= SparseCategoricalCrossentropy(from_logits=True),\n",
        "              metrics=['accuracy'])"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab_type": "code",
        "id": "xvwvpA64CaW_",
        "colab": {}
      },
      "source": [
        "model.fit(train_images, train_labels, epochs=10)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "4n4_BRNKLRwo",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "KERAS_MODEL_NAME = \"tf_model_fashion_mnist.h5\""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "A_3Y3F4GBWud",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "model.save(KERAS_MODEL_NAME)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "19sd1FJkCHID",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "convert_bytes(get_file_size(KERAS_MODEL_NAME), \"MB\")"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Y_deKtkhCHMu",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "test_loss, test_acc = model.evaluate(test_images,  test_labels, verbose=2)\n",
        "print('\\nTest accuracy:', test_acc)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "1BFE9lvwLfgv",
        "colab_type": "text"
      },
      "source": [
        "# TF Lite Model"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ORAx4Yc7LjwM",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "TF_LITE_MODEL_FILE_NAME = \"tf_lite_model.tflite\""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "NY57t7EwCW8P",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "tf_lite_converter = tf.lite.TFLiteConverter.from_keras_model(model)\n",
        "tf_lite_converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]\n",
        "# tf_lite_converter.optimizations = [tf.lite.Optimize.DEFAULT]\n",
        "# tf_lite_converter.target_spec.supported_types = [tf.float16]\n",
        "tflite_model = tf_lite_converter.convert()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "CZxnt1IsCXBb",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "tflite_model_name = TF_LITE_MODEL_FILE_NAME\n",
        "open(tflite_model_name, \"wb\").write(tflite_model)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "XtIdP296CXFl",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "convert_bytes(get_file_size(TF_LITE_MODEL_FILE_NAME), \"KB\")"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "i_gttI3-M45M",
        "colab_type": "text"
      },
      "source": [
        "# Check Input Tensor Shape"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "IJXJuGLFGfjB",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "interpreter = tf.lite.Interpreter(model_path = TF_LITE_MODEL_FILE_NAME)\n",
        "input_details = interpreter.get_input_details()\n",
        "output_details = interpreter.get_output_details()\n",
        "print(\"Input Shape:\", input_details[0]['shape'])\n",
        "print(\"Input Type:\", input_details[0]['dtype'])\n",
        "print(\"Output Shape:\", output_details[0]['shape'])\n",
        "print(\"Output Type:\", output_details[0]['dtype'])"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "3_X-k46vMzAt",
        "colab_type": "text"
      },
      "source": [
        "# Resize Tensor Shape"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "M6wo8RNFGpXw",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "interpreter.resize_tensor_input(input_details[0]['index'], (10000, 28, 28))\n",
        "interpreter.resize_tensor_input(output_details[0]['index'], (10000, 10))\n",
        "interpreter.allocate_tensors()\n",
        "input_details = interpreter.get_input_details()\n",
        "output_details = interpreter.get_output_details()\n",
        "print(\"Input Shape:\", input_details[0]['shape'])\n",
        "print(\"Input Type:\", input_details[0]['dtype'])\n",
        "print(\"Output Shape:\", output_details[0]['shape'])\n",
        "print(\"Output Type:\", output_details[0]['dtype'])"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "zsv7ROq8Gf0f",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "test_images.dtype"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "SM1alowAHy2d",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "test_imgs_numpy = np.array(test_images, dtype=np.float32)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "wm4bxTIwHErB",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "interpreter.set_tensor(input_details[0]['index'], test_imgs_numpy)\n",
        "interpreter.invoke()\n",
        "tflite_model_predictions = interpreter.get_tensor(output_details[0]['index'])\n",
        "print(\"Prediction results shape:\", tflite_model_predictions.shape)\n",
        "prediction_classes = np.argmax(tflite_model_predictions, axis=1)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "yRKoF30HIZP2",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "acc = accuracy_score(prediction_classes, test_labels)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "FZIF6OenI_5i",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "print('Test accuracy TFLITE model :', acc)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Vw7r5WyUV307",
        "colab_type": "text"
      },
      "source": [
        "# Get Weights of Keras Model"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Lzqy6REGPxIA",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "print(model.get_weights()[0][0][0], type(model.get_weights()[0][0][0]))"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "drZEEbY5RXJk",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "keras_weight_var = np.array([model.get_weights()[0][0][0]], dtype=\"float64\")"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "oIwkH2EYWKom",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "getsizeof(keras_weight_var[0])"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "SuSUIDA_WZ9V",
        "colab_type": "text"
      },
      "source": [
        "# Access Quantized Weights of TFLite"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "fbzyyvetWhCs",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "TF_LITE_WEIGHTS_TEMP_FILE = \"temp_weights_from_tflite.h5\""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "hV83qjttPz5E",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "interpreter = tf.lite.Interpreter(model_path=TF_LITE_MODEL_FILE_NAME)\n",
        "interpreter.allocate_tensors()\n",
        "input_details = interpreter.get_input_details()\n",
        "output_details = interpreter.get_output_details()\n",
        "all_layers_details = interpreter.get_tensor_details() \n",
        "f = h5py.File(TF_LITE_WEIGHTS_TEMP_FILE, \"w\")   \n",
        "for layer in all_layers_details:\n",
        "     grp = f.create_group(str(layer['index']))\n",
        "     grp.attrs[\"name\"] = layer['name']\n",
        "     grp.attrs[\"shape\"] = layer['shape']\n",
        "     grp.attrs[\"quantization\"] = layer['quantization']\n",
        "     grp.create_dataset(\"weights\", data=interpreter.get_tensor(layer['index']))\n",
        "f.close()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "oTi8OzHmUZF1",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "temp_file = h5py.File(TF_LITE_WEIGHTS_TEMP_FILE, 'r')\n",
        "print(temp_file[\"5\"][\"weights\"][0][6], type(temp_file[\"5\"][\"weights\"][0][6]))"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "dcdUtejoW73F",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "quantized_weight_var = np.array([temp_file[\"5\"][\"weights\"][0][6]])"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Udfbq6fTXCxa",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "getsizeof(quantized_weight_var[0])"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "zUF9xC2iZVsb",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "temp_file.close()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Vv4y9DS5ZX71",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    }
  ]
}