{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "afdd07a1",
   "metadata": {},
   "source": [
    "# Convolutional networks in Keras 3 and TensorFlow\n",
    "\n",
    "The same ideas the TensorFlow way: Keras 3 layers, `compile`/`fit`, callbacks, and `tf.GradientTape` when you want your own training loop. TensorFlow 2.21 is stable and mostly receives fixes now; Google recommends Keras 3, JAX or PyTorch for new generative-AI work, and Keras 3 code like this also runs on JAX or PyTorch by changing one line.\n",
    "\n",
    "[Open in Google Colab](https://colab.research.google.com/github/mahmoudyosrimahmoud13/inpace-library/blob/main/courses/dl-4-cnns/files/dl4_convolutional_networks_keras.ipynb) · part of the inPace course library. Slides © DeepLearning.AI (CC BY-SA 2.0); this notebook: inPace (CC BY-SA 4.0).\n",
    "\n",
    "**Run it top to bottom** (Runtime → Run all in Colab). Each part says what to look for."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7a22c818",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Works in Google Colab as it is (TensorFlow and Keras are already installed there).\n",
    "# On your own computer:  pip install tensorflow keras scikit-learn scipy matplotlib\n",
    "# One framework per notebook: PyTorch and TensorFlow can clash when loaded into the same Python process.\n",
    "import os\n",
    "os.environ[\"KERAS_BACKEND\"] = \"tensorflow\"   # Keras 3 also runs on \"jax\" or \"torch\": change it here\n",
    "os.environ.setdefault(\"TF_CPP_MIN_LOG_LEVEL\", \"3\")\n",
    "import numpy as np, matplotlib.pyplot as plt\n",
    "import tensorflow as tf, keras\n",
    "keras.utils.set_random_seed(0)\n",
    "print(\"TensorFlow\", tf.__version__, \"| Keras\", keras.__version__, \"| backend:\", keras.backend.backend(),\n",
    "      \"| GPU:\", bool(tf.config.list_physical_devices(\"GPU\")))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "db0850bf",
   "metadata": {},
   "source": [
    "## 1. A CNN on digits\n",
    "Keras images are channels-last: `(batch, height, width, channels)`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7ea77fe4",
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.datasets import load_digits\n",
    "from sklearn.model_selection import train_test_split\n",
    "d = load_digits()\n",
    "Xa, Xb, ya, yb = train_test_split((d.images / 16.0).astype(\"float32\")[..., None], d.target, test_size=0.3, random_state=0)\n",
    "cnn = keras.Sequential([keras.Input((8, 8, 1)),\n",
    "    keras.layers.Conv2D(16, 3, padding=\"same\", activation=\"relu\"), keras.layers.MaxPooling2D(),\n",
    "    keras.layers.Conv2D(32, 3, padding=\"same\", activation=\"relu\"), keras.layers.MaxPooling2D(),\n",
    "    keras.layers.Flatten(), keras.layers.Dense(10)])\n",
    "cnn.compile(\"adamw\", keras.losses.SparseCategoricalCrossentropy(from_logits=True), metrics=[\"accuracy\"])\n",
    "cnn.fit(Xa, ya, epochs=60, batch_size=64, verbose=0)\n",
    "print(f\"parameters {cnn.count_params():,} | test accuracy {cnn.evaluate(Xb, yb, verbose=0)[1]:.3f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "62ac53b8",
   "metadata": {},
   "source": [
    "## 2. A residual block with the functional API"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "aef843a6",
   "metadata": {},
   "outputs": [],
   "source": [
    "def residual_block(x, filters):\n",
    "    shortcut = x\n",
    "    x = keras.layers.Conv2D(filters, 3, padding=\"same\", use_bias=False)(x)\n",
    "    x = keras.layers.ReLU()(keras.layers.BatchNormalization()(x))\n",
    "    x = keras.layers.BatchNormalization()(keras.layers.Conv2D(filters, 3, padding=\"same\", use_bias=False)(x))\n",
    "    return keras.layers.ReLU()(keras.layers.Add()([x, shortcut]))     # the skip connection\n",
    "inp = keras.Input((8, 8, 16)); print(\"shape kept:\", keras.Model(inp, residual_block(inp, 16)).output_shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8bc431a3",
   "metadata": {},
   "source": [
    "## 3. Depthwise separable convolution"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9213c57e",
   "metadata": {},
   "outputs": [],
   "source": [
    "x = keras.Input((32, 32, 64))\n",
    "normal = keras.Model(x, keras.layers.Conv2D(128, 3, padding=\"same\")(x))\n",
    "separable = keras.Model(x, keras.layers.SeparableConv2D(128, 3, padding=\"same\")(x))\n",
    "print(f\"weights: normal {normal.count_params():,} vs separable {separable.count_params():,}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "790688f9",
   "metadata": {},
   "source": [
    "## 4. Augmentation layers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "70c481f1",
   "metadata": {},
   "outputs": [],
   "source": [
    "aug = keras.Sequential([keras.layers.RandomRotation(0.05), keras.layers.RandomZoom(0.15), keras.layers.RandomContrast(0.3)])\n",
    "fig, ax = plt.subplots(1, 8, figsize=(10, 1.5))\n",
    "for i, a in enumerate(ax):\n",
    "    a.imshow((Xa[0] if i == 0 else aug(Xa[:1], training=True)[0])[..., 0], cmap=\"gray_r\"); a.axis(\"off\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8df42c1f",
   "metadata": {},
   "source": [
    "## 5. Non-max suppression with TensorFlow\n",
    "TensorFlow's boxes are `[y1, x1, y2, x2]`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8a2b9d2b",
   "metadata": {},
   "outputs": [],
   "source": [
    "boxes = tf.constant([[30, 30, 120, 130], [38, 40, 128, 140], [25, 35, 125, 125], [60, 200, 160, 290], [70, 210, 165, 300]], tf.float32)\n",
    "scores = tf.constant([0.9, 0.75, 0.6, 0.85, 0.7])\n",
    "print(\"kept:\", tf.image.non_max_suppression(boxes, scores, max_output_size=10, iou_threshold=0.5).numpy().tolist())"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d18e4ed6",
   "metadata": {},
   "source": [
    "## 6. Transfer learning with keras.applications\n",
    "Needs internet to download the ImageNet weights (runs in Colab)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "363ca96f",
   "metadata": {},
   "outputs": [],
   "source": [
    "try:\n",
    "    base = keras.applications.MobileNetV3Small(input_shape=(224, 224, 3), include_top=False, weights=\"imagenet\")\n",
    "    base.trainable = False\n",
    "    inp = keras.Input((224, 224, 3))\n",
    "    out = keras.layers.Dense(3)(keras.layers.GlobalAveragePooling2D()(base(inp, training=False)))\n",
    "    model = keras.Model(inp, out)\n",
    "    print(\"trainable weights:\", sum(int(np.prod(w.shape)) for w in model.trainable_weights))\n",
    "except Exception as e:\n",
    "    print(\"Needs internet to download the weights (run this in Colab):\", type(e).__name__)"
   ]
  }
 ],
 "metadata": {
  "colab": {
   "provenance": [],
   "toc_visible": true
  },
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}