{"metadata":{"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"name":"python","version":"3.12.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceType":"competition","sourceId":3362,"databundleVersionId":31148,"isSourceIdPinned":false},{"sourceType":"competition","sourceId":5441,"databundleVersionId":38425,"isSourceIdPinned":false},{"sourceType":"datasetVersion","sourceId":1003830,"datasetId":550917,"databundleVersionId":1032543},{"sourceType":"datasetVersion","sourceId":791828,"datasetId":119698,"databundleVersionId":813580}],"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"id":"1121e723","cell_type":"markdown","source":"
\n\n

Deep Learning — Practical Assignment

\n\n\n\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
Team Members
01Abdelaziz Omar AbdelazizID: 221849
02Abdallah Mohamed AbdelazizID: 220090
03Emad Alaa Eldin ZakiID: 222689
04Mustafa Somaa MahmoudID: 221751
CourseDeep Learning
FrameworkTensorFlow 2.x / Keras
HardwareKaggle — GPU Tesla T4 ×2
\n\n---\n\n
\n\n## Table of Contents\n\n| # | Section | Description |\n|:---:|:---|:---|\n| 1 | Setup & Libraries | Imports, seeds, GPU configuration |\n| **A** | **Part A — Cats vs Dogs** | Binary image classification |\n| A.1 | Data Extraction | Unpack ZIP, build labeled DataFrame |\n| A.2 | Data Exploration | Class balance, sample images |\n| A.3 | Train / Val / Test Split | Stratified 80 / 10 / 10 |\n| A.4 | Augmentation & Generators | ImageDataGenerator pipeline |\n| A.5 | CNN Architecture | 4-block deep CNN with BatchNorm |\n| A.6 | Training | Callbacks & learning rate schedule |\n| A.7 | Evaluation | Accuracy, loss curves, confusion matrix |\n| A.8 | Prediction Visualization | Sample predictions |\n| **B** | **Part B — MNIST** | 10-class handwritten digit recognition |\n| B.1 | Load & Explore | Dataset overview |\n| B.2 | Preprocessing | Normalize, reshape, one-hot encode |\n| B.3 | CNN Architecture | 2-block lightweight CNN |\n| B.4 | Training & Evaluation | Full evaluation pipeline |\n| **C** | **Part C — Fashion MNIST** | 10-class clothing classification |\n| C.1 | Load & Explore | Dataset overview |\n| C.2 | CNN Architecture | 3-block CNN |\n| C.3 | Training & Evaluation | Full evaluation pipeline |\n| **D** | **Conclusion & Comparison** | Cross-dataset summary |\n\n---\n\n> **Datasets used:**\n> - `/kaggle/input/competitions/dogs-vs-cats/train.zip` — Kaggle Dogs vs Cats (25,000 images)\n> - `/kaggle/input/competitions/dogs-vs-cats/test1.zip` — Kaggle test set\n> - `keras.datasets.mnist` — Built-in MNIST (70,000 images)\n> - `keras.datasets.fashion_mnist` — Built-in Fashion MNIST (70,000 images)\n","metadata":{}},{"id":"003f1a69","cell_type":"markdown","source":"---\n## 1. Setup & Libraries\n\nAll libraries required across the three parts of this notebook are imported in a single cell.\nCentralizing imports makes the notebook easier to audit and avoids hidden dependencies.\n\n**Reproducibility** is ensured by fixing the random seed across Python, NumPy, and TensorFlow.\nWith the same seed, every split, weight initialization, and augmentation sequence is identical across runs.\n\n| Library Group | Libraries | Purpose |\n|:---|:---|:---|\n| Standard | `os`, `zipfile`, `random`, `warnings` | File I/O, ZIP extraction, randomness control |\n| Numerical | `numpy`, `pandas` | Array operations, DataFrame for label management |\n| Visualization | `matplotlib`, `seaborn` | Training curves, confusion matrices, image grids |\n| Image | `PIL` (Pillow) | Opening and resizing image files |\n| Deep Learning | `tensorflow`, `keras` | CNN construction, training, callbacks, evaluation |\n| Evaluation | `sklearn` | Stratified split, confusion matrix, classification report |\n\n![Deep Learning Pipeline](https://i.imgur.com/T82z202.jpeg)\n\n*Figure 1: A typical CNN pipeline — input image passes through convolutional layers, pooling layers, and a fully connected classifier.*\n","metadata":{}},{"id":"1e7123d6","cell_type":"code","source":"# ── Standard ──────────────────────────────────────────────────────────────────\nimport os\nimport zipfile\nimport random\nimport warnings\nwarnings.filterwarnings('ignore')\n\n# ── Numerical & Data ──────────────────────────────────────────────────────────\nimport numpy as np\nimport pandas as pd\nfrom PIL import Image\nfrom matplotlib.image import imread\n\n# ── Visualization ─────────────────────────────────────────────────────────────\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nsns.set_style('darkgrid')\n\n# ── Deep Learning ─────────────────────────────────────────────────────────────\nimport tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import (\n Conv2D, MaxPooling2D, Dense, Flatten,\n Dropout, BatchNormalization\n)\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\nfrom tensorflow.keras.callbacks import ReduceLROnPlateau, EarlyStopping\n\n# ── Evaluation ────────────────────────────────────────────────────────────────\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import classification_report, confusion_matrix\n\n# ── Reproducibility ───────────────────────────────────────────────────────────\nSEED = 42\nrandom.seed(SEED)\nnp.random.seed(SEED)\ntf.random.set_seed(SEED)\n\n# ── Hardware Info ─────────────────────────────────────────────────────────────\nprint(\"TensorFlow version :\", tf.__version__)\ngpus = tf.config.list_physical_devices('GPU')\nprint(\"GPU available :\", len(gpus) > 0)\nfor g in gpus:\n print(\" ->\", g)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:17:52.140189Z","iopub.execute_input":"2026-04-26T22:17:52.141033Z","iopub.status.idle":"2026-04-26T22:18:33.565539Z","shell.execute_reply.started":"2026-04-26T22:17:52.140990Z","shell.execute_reply":"2026-04-26T22:18:33.564685Z"}},"outputs":[],"execution_count":null},{"id":"070165de","cell_type":"markdown","source":"\n\n---\n**Part A — Cats vs Dogs**\n---\n\n\n\n**Objective:** Train a CNN to distinguish cat photographs from dog photographs\nwith accuracy exceeding 90% on unseen test data.\n\n| Property | Value |\n|:---|:---|\n| Dataset | Kaggle Dogs vs Cats Competition — 25,000 labeled JPEG images |\n| Task | Binary Classification: Cat = 0   /   Dog = 1 |\n| Input resolution | 128 x 128 pixels, RGB (3 channels) |\n| Target accuracy | > 90% on held-out test set |\n\n**Why this task is challenging:**\nUnlike MNIST or Fashion MNIST, these are real-world photographs with significant\nvariation in lighting, background, pose, scale, and image quality.\nThe model must learn robust visual representations that generalize across all these variations.\n\n![CNN Layers](https://i.imgur.com/2JlmaU3.jpeg)\n\n*Figure A.0: A neural network learns hierarchical representations.\nEarly layers detect edges and gradients; deeper layers recognize semantic structures\nsuch as ears, snouts, and body shapes that distinguish cats from dogs.*\n","metadata":{}},{"id":"2a648d62","cell_type":"markdown","source":"### A.1 Data Extraction from ZIP\n\nThe Kaggle Dogs vs Cats dataset is distributed as a compressed ZIP archive.\nWe extract it to the Kaggle working directory and construct a labeled DataFrame\nby parsing the filename prefix of each image.\n\n**Why filename-based labeling?**\nThe dataset follows a strict naming convention: `cat.N.jpg` and `dog.N.jpg`.\nParsing the prefix is deterministic and error-free, without requiring any\nfolder structure or external metadata file.\n","metadata":{}},{"id":"129248d4","cell_type":"code","source":"WORKING = \"/kaggle/working/\"\nTRAIN_ZIP = \"/kaggle/input/competitions/dogs-vs-cats/train.zip\"\nTEST_ZIP = \"/kaggle/input/competitions/dogs-vs-cats/test1.zip\"\n\nprint(\"Extracting training archive...\")\nwith zipfile.ZipFile(TRAIN_ZIP, 'r') as z:\n z.extractall(WORKING)\nprint(\" train/ extracted successfully.\")\n\nprint(\"Extracting test archive...\")\nwith zipfile.ZipFile(TEST_ZIP, 'r') as z:\n z.extractall(WORKING)\nprint(\" test1/ extracted successfully.\")\n\n# ── Build labeled DataFrame ────────────────────────────────────────\nIMAGE_DIR = os.path.join(WORKING, \"train\")\nfilenames = os.listdir(IMAGE_DIR)\nlabels = [f.split(\".\")[0] for f in filenames] # \"cat\" or \"dog\"\n\ndata = pd.DataFrame({\"filename\": filenames, \"label\": labels})\n\nprint(f\"\\nTotal images : {len(data):,}\")\nprint(\"\\nClass distribution:\")\nprint(data['label'].value_counts().to_string())\nprint()\ndata.head(8)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:18:33.567037Z","iopub.execute_input":"2026-04-26T22:18:33.567602Z","iopub.status.idle":"2026-04-26T22:18:44.231973Z","shell.execute_reply.started":"2026-04-26T22:18:33.567576Z","shell.execute_reply":"2026-04-26T22:18:44.230960Z"}},"outputs":[],"execution_count":null},{"id":"d48406dd","cell_type":"markdown","source":"### A.2 Data Exploration\n\nBefore any model is built, we examine the dataset to understand:\n\n1. **Class balance** — are Cat and Dog counts equal?\n2. **Visual diversity** — do images vary in size, angle, and background?\n\n**Key observations:**\n- The dataset is perfectly balanced: 12,500 cats and 12,500 dogs.\n- Images vary significantly in resolution, aspect ratio, background, and pose.\n- Some images contain multiple animals, partial views, or cluttered backgrounds.\n- These factors motivate the use of data augmentation during training.\n","metadata":{}},{"id":"9ab0c8d6","cell_type":"code","source":"# ── Sample Dog images ─────────────────────────────────────────────\nplt.figure(figsize=(20, 3))\nplt.suptitle(\"Sample Dog Images from Training Set\",\n fontsize=13, fontweight='bold', y=1.05)\nfor i in range(10):\n plt.subplot(1, 10, i + 1)\n img = imread(os.path.join(IMAGE_DIR, f\"dog.{i}.jpg\"))\n plt.imshow(img)\n plt.title(\"Dog\", fontsize=9, color='#e74c3c')\n plt.axis('off')\nplt.tight_layout()\nplt.show()\n\n# ── Sample Cat images ─────────────────────────────────────────────\nplt.figure(figsize=(20, 3))\nplt.suptitle(\"Sample Cat Images from Training Set\",\n fontsize=13, fontweight='bold', y=1.05)\nfor i in range(10):\n plt.subplot(1, 10, i + 1)\n img = imread(os.path.join(IMAGE_DIR, f\"cat.{i}.jpg\"))\n plt.imshow(img)\n plt.title(\"Cat\", fontsize=9, color='#3498db')\n plt.axis('off')\nplt.tight_layout()\nplt.show()\n\n# ── Class distribution chart ──────────────────────────────────────\nfig, ax = plt.subplots(figsize=(7, 4), dpi=130)\ncounts = data['label'].value_counts()\nsns.barplot(x=counts.index, y=counts.values,\n palette=['#3498db', '#e74c3c'], width=0.35, ax=ax)\nax.set_title(\"Class Distribution\", fontsize=12, fontweight='bold')\nax.set_xlabel(\"Class\"); ax.set_ylabel(\"Number of Images\")\nfor bar in ax.patches:\n ax.text(bar.get_x() + bar.get_width() / 2,\n bar.get_height() + 80,\n f\"{int(bar.get_height()):,}\",\n ha='center', fontsize=11, fontweight='bold')\nax.set_ylim(0, 14500)\nplt.tight_layout()\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:18:44.233047Z","iopub.execute_input":"2026-04-26T22:18:44.233367Z","iopub.status.idle":"2026-04-26T22:18:45.871682Z","shell.execute_reply.started":"2026-04-26T22:18:44.233335Z","shell.execute_reply":"2026-04-26T22:18:45.870989Z"}},"outputs":[],"execution_count":null},{"id":"1d95eba6","cell_type":"markdown","source":"### A.3 Train / Validation / Test Split\n\nThe dataset is divided into three non-overlapping subsets using **stratified sampling**.\nStratification preserves the 50/50 class balance in every subset.\n\n| Subset | Ratio | Images | Purpose |\n|:---|:---:|:---:|:---|\n| Training | 80% | 20,000 | Update model weights via backpropagation |\n| Validation | 10% | 2,500 | Monitor generalization during training |\n| Test | 10% | 2,500 | Final unbiased evaluation — never seen during training |\n\n**Why a separate test set?**\nThe validation set is used to make decisions during training (early stopping, LR reduction),\nso the model indirectly adapts to it. The test set is the only truly unbiased performance measure.\n","metadata":{}},{"id":"c488f8a8","cell_type":"code","source":"labels_col = data['label']\n\n# Step 1: Training vs (Validation + Test)\nX_train, X_temp = train_test_split(\n data, test_size=0.2, stratify=labels_col, random_state=SEED\n)\n\n# Step 2: Validation vs Test\nX_val, X_test = train_test_split(\n X_temp, test_size=0.5,\n stratify=X_temp['label'], random_state=SEED\n)\n\nprint(f\"Training set : {len(X_train):>6,} images ({len(X_train)/len(data)*100:.0f}%)\")\nprint(f\"Validation set : {len(X_val):>6,} images ({len(X_val)/len(data)*100:.0f}%)\")\nprint(f\"Test set : {len(X_test):>6,} images ({len(X_test)/len(data)*100:.0f}%)\")\n\n# ── Class balance per split ────────────────────────────────────────\nfig, axes = plt.subplots(1, 3, figsize=(16, 5), dpi=130)\nfig.suptitle(\"Class Distribution per Split\", fontsize=13, fontweight='bold')\n\nfor ax, (df, title) in zip(axes, [\n (X_train, \"Training (80%)\"),\n (X_val, \"Validation (10%)\"),\n (X_test, \"Test (10%)\")\n]):\n counts = df['label'].value_counts()\n sns.barplot(x=counts.index, y=counts.values,\n palette=['#3498db', '#e74c3c'], width=0.35, ax=ax)\n ax.set_title(title, fontsize=12, fontweight='bold')\n ax.set_xlabel(\"Class\"); ax.set_ylabel(\"Images\")\n for bar in ax.patches:\n ax.text(bar.get_x() + bar.get_width() / 2,\n bar.get_height() + 30,\n f\"{int(bar.get_height()):,}\",\n ha='center', fontsize=10, fontweight='bold')\n\nplt.tight_layout()\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:18:45.873396Z","iopub.execute_input":"2026-04-26T22:18:45.873837Z","iopub.status.idle":"2026-04-26T22:18:46.276600Z","shell.execute_reply.started":"2026-04-26T22:18:45.873814Z","shell.execute_reply":"2026-04-26T22:18:46.275988Z"}},"outputs":[],"execution_count":null},{"id":"b66de2e2","cell_type":"markdown","source":"### A.4 Data Augmentation & Generators\n\n**What is Data Augmentation?**\nAugmentation applies random transformations to training images before each epoch,\npreventing memorization and teaching the model to be invariant to irrelevant variations.\n\n| Transform | Value | Justification |\n|:---|:---:|:---|\n| `rotation_range` | 15 deg | Cameras are not always perfectly level |\n| `horizontal_flip` | True | Cats and dogs are bilaterally symmetric |\n| `zoom_range` | 0.2 | Subjects photographed at different distances |\n| `width_shift_range` | 0.1 | Subject not always centered horizontally |\n| `height_shift_range` | 0.1 | Subject not always centered vertically |\n| `shear_range` | 0.1 | Perspective distortion from camera angle |\n| `fill_mode` | reflect | Fills border pixels naturally |\n\nValidation and test generators use **normalization only** — no augmentation.\n\n**Why `flow_from_dataframe`?**\nIt accepts an explicit filename-to-label mapping, giving precise control over splits\nand working correctly on Kaggle's read-only filesystem.\n\n![Data Augmentation Example](https://i.imgur.com/Kw0GhMi.jpeg)\n\n*Figure A.4: Common data augmentation operations applied to training images.\nEach epoch the model sees a different random transformation of the same image.*\n","metadata":{}},{"id":"7577b1b4","cell_type":"code","source":"IMAGE_SIZE = 128\nBATCH_SIZE = 32\nIMAGE_CHANNEL = 3\n\n# ── Training generator: normalize + augment ────────────────────────\ntrain_datagen = ImageDataGenerator(\n rescale = 1.0 / 255,\n rotation_range = 15,\n horizontal_flip = True,\n zoom_range = 0.2,\n shear_range = 0.1,\n fill_mode = 'reflect',\n width_shift_range = 0.1,\n height_shift_range = 0.1\n)\n\n# ── Evaluation generator: normalize only ──────────────────────────\neval_datagen = ImageDataGenerator(rescale=1.0 / 255)\n\n# ── Flow from DataFrame ───────────────────────────────────────────\ntrain_generator = train_datagen.flow_from_dataframe(\n X_train,\n directory = IMAGE_DIR,\n x_col = 'filename',\n y_col = 'label',\n batch_size = BATCH_SIZE,\n target_size = (IMAGE_SIZE, IMAGE_SIZE),\n class_mode = 'binary'\n)\n\nval_generator = eval_datagen.flow_from_dataframe(\n X_val,\n directory = IMAGE_DIR,\n x_col = 'filename',\n y_col = 'label',\n batch_size = BATCH_SIZE,\n target_size = (IMAGE_SIZE, IMAGE_SIZE),\n class_mode = 'binary',\n shuffle = False\n)\n\ntest_generator = eval_datagen.flow_from_dataframe(\n X_test,\n directory = IMAGE_DIR,\n x_col = 'filename',\n y_col = 'label',\n batch_size = BATCH_SIZE,\n target_size = (IMAGE_SIZE, IMAGE_SIZE),\n class_mode = 'binary',\n shuffle = False\n)\n\nprint(f\"Class indices : {train_generator.class_indices}\")\nprint(f\"Training batches : {len(train_generator)}\")\nprint(f\"Validation batches : {len(val_generator)}\")\nprint(f\"Test batches : {len(test_generator)}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:18:46.277379Z","iopub.execute_input":"2026-04-26T22:18:46.277765Z","iopub.status.idle":"2026-04-26T22:18:46.475519Z","shell.execute_reply.started":"2026-04-26T22:18:46.277742Z","shell.execute_reply":"2026-04-26T22:18:46.474935Z"}},"outputs":[],"execution_count":null},{"id":"ceebe85a","cell_type":"code","source":"# ── Visualize augmented samples ────────────────────────────────────\nsample_imgs, sample_labels = next(train_generator)\n\nfig, axes = plt.subplots(2, 5, figsize=(16, 7))\nfig.suptitle(\"Augmented Training Samples (as seen by the model each epoch)\",\n fontsize=12, fontweight='bold')\n\nfor idx in range(10):\n ax = axes[idx // 5][idx % 5]\n label = \"Dog\" if int(sample_labels[idx]) == 1 else \"Cat\"\n color = '#e74c3c' if label == \"Dog\" else '#3498db'\n ax.imshow(sample_imgs[idx])\n ax.set_title(label, color=color, fontsize=10, fontweight='bold')\n ax.axis('off')\n\nplt.tight_layout()\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:18:46.476397Z","iopub.execute_input":"2026-04-26T22:18:46.476703Z","iopub.status.idle":"2026-04-26T22:18:47.543903Z","shell.execute_reply.started":"2026-04-26T22:18:46.476680Z","shell.execute_reply":"2026-04-26T22:18:47.543016Z"}},"outputs":[],"execution_count":null},{"id":"ded7a84a","cell_type":"markdown","source":"### A.5 CNN Architecture\n\nThe model uses four convolutional blocks followed by a fully connected classification head.\n\n```\nInput: 128 x 128 x 3 (RGB image)\n +---------------------------------------------------------------+\n | Block 1 -- Conv2D(32) + BatchNorm + MaxPool(2) + Drop(0.2) | Output: 63 x 63 x 32\n | Block 2 -- Conv2D(64) + BatchNorm + MaxPool(2) + Drop(0.2) | Output: 30 x 30 x 64\n | Block 3 -- Conv2D(128) + BatchNorm + MaxPool(2) + Drop(0.2) | Output: 14 x 14 x 128\n | Block 4 -- Conv2D(256) + BatchNorm + MaxPool(2) + Drop(0.2) | Output: 6 x 6 x 256\n | Head -- Flatten + Dense(512) + BatchNorm + Dropout(0.2) |\n | Output -- Dense(1) + Sigmoid | Probability in [0, 1]\n +---------------------------------------------------------------+\n```\n\n| Component | Choice | Reason |\n|:---|:---|:---|\n| Filter progression | 32 to 64 to 128 to 256 | Deeper layers need more filters for complex patterns |\n| BatchNormalization | After every Conv layer | Stabilizes activations, speeds convergence |\n| MaxPooling2D | After every block | Reduces spatial dimensions, builds translation invariance |\n| Dropout(0.2) | After every block | Prevents co-adaptation of feature detectors |\n| Dense(1) + Sigmoid | Output layer | Binary output: probability < 0.5 = Cat, >= 0.5 = Dog |\n| binary_crossentropy | Loss function | Correct loss for single-unit sigmoid binary classification |\n\n![CNN Feature Maps](https://i.imgur.com/66cr55x.jpeg)\n\n*Figure A.5: Feature maps in a CNN become smaller spatially but richer in learned representations as depth increases.*\n","metadata":{}},{"id":"48c4e093","cell_type":"code","source":"model = Sequential([\n # Block 1 -- low-level features: edges, color boundaries\n Conv2D(32, (3, 3), activation='relu',\n input_shape=(IMAGE_SIZE, IMAGE_SIZE, IMAGE_CHANNEL)),\n BatchNormalization(),\n MaxPooling2D(pool_size=(2, 2)),\n Dropout(0.2),\n\n # Block 2 -- mid-level features: textures, fur patterns\n Conv2D(64, (3, 3), activation='relu'),\n BatchNormalization(),\n MaxPooling2D(pool_size=(2, 2)),\n Dropout(0.2),\n\n # Block 3 -- high-level features: shapes, object parts\n Conv2D(128, (3, 3), activation='relu'),\n BatchNormalization(),\n MaxPooling2D(pool_size=(2, 2)),\n Dropout(0.2),\n\n # Block 4 -- semantic features: ears, snouts, body silhouettes\n Conv2D(256, (3, 3), activation='relu'),\n BatchNormalization(),\n MaxPooling2D(pool_size=(2, 2)),\n Dropout(0.2),\n\n # Classification head\n Flatten(),\n Dense(512, activation='relu'),\n BatchNormalization(),\n Dropout(0.2),\n\n # Binary output -- Sigmoid maps to [0, 1]\n Dense(1, activation='sigmoid')\n], name=\"CatsVsDogs_CNN\")\n\nmodel.compile(\n optimizer = 'adam',\n loss = 'binary_crossentropy',\n metrics = ['accuracy']\n)\n\nmodel.summary()\nprint(f\"\\nTotal parameters: {model.count_params():,}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:18:47.545077Z","iopub.execute_input":"2026-04-26T22:18:47.545410Z","iopub.status.idle":"2026-04-26T22:18:49.972576Z","shell.execute_reply.started":"2026-04-26T22:18:47.545385Z","shell.execute_reply":"2026-04-26T22:18:49.971951Z"}},"outputs":[],"execution_count":null},{"id":"e60810f8","cell_type":"markdown","source":"### A.6 Training\n\n| Callback | Monitor | Trigger | Action |\n|:---|:---|:---|:---|\n| ReduceLROnPlateau | val_accuracy | No improvement for 2 epochs | Multiply learning rate by 0.5 |\n| EarlyStopping | val_loss | No improvement for 3 epochs | Stop; restore best weights |\n\n**Why reduce the learning rate?**\nWhen the model stops improving, smaller updates allow finer descent into the loss landscape.\n\n**Why early stopping?**\nHalts training automatically when validation loss plateaus, protecting against overfitting.\n`restore_best_weights=True` ensures the final model corresponds to the best validation epoch.\n","metadata":{}},{"id":"64cd90d9","cell_type":"code","source":"lr_reduction = ReduceLROnPlateau(\n monitor = 'val_accuracy',\n patience = 2,\n factor = 0.5,\n min_lr = 0.00001,\n verbose = 1\n)\n\nearly_stop = EarlyStopping(\n monitor = 'val_loss',\n patience = 3,\n restore_best_weights = True,\n verbose = 1\n)\n\nprint(\"Starting training -- Cats vs Dogs CNN...\")\nprint(f\" Epochs (max) : 30 | Batch size: {BATCH_SIZE} | Image size: {IMAGE_SIZE}x{IMAGE_SIZE}\")\nprint()\n\nhistory_cvd = model.fit(\n train_generator,\n validation_data = val_generator,\n callbacks = [early_stop, lr_reduction],\n epochs = 30,\n verbose = 1\n)\nprint(\"\\nTraining complete.\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:18:49.973373Z","iopub.execute_input":"2026-04-26T22:18:49.973656Z","iopub.status.idle":"2026-04-26T22:58:59.013101Z","shell.execute_reply.started":"2026-04-26T22:18:49.973635Z","shell.execute_reply":"2026-04-26T22:58:59.012487Z"}},"outputs":[],"execution_count":null},{"id":"f83d8861","cell_type":"markdown","source":"### A.7 Evaluation & Confusion Matrix\n\n**Training curves** show how loss and accuracy evolve across epochs.\nA healthy run shows train and validation curves close together without a large gap.\n\n**Confusion matrix** decomposes accuracy into per-class counts:\n- Diagonal cells: correct predictions\n- Off-diagonal cells: misclassifications\n\nThe **Classification Report** adds Precision, Recall, and F1-score per class.\n","metadata":{}},{"id":"a1eab274","cell_type":"code","source":"# ── Training curves ───────────────────────────────────────────────\nerror = pd.DataFrame(history_cvd.history)\n\nfig, axes = plt.subplots(1, 2, figsize=(16, 5), dpi=150)\nfig.suptitle(\"Cats vs Dogs -- Training History\", fontsize=13, fontweight='bold')\n\naxes[0].plot(error['loss'], lw=2.5, label='Train Loss', color='#3498db')\naxes[0].plot(error['val_loss'], lw=2.5, label='Validation Loss', color='#e74c3c', ls='--')\naxes[0].set_title(\"Loss per Epoch\"); axes[0].set_xlabel(\"Epoch\"); axes[0].set_ylabel(\"Loss\")\naxes[0].legend(); axes[0].grid(alpha=0.3)\n\naxes[1].plot(error['accuracy'], lw=2.5, label='Train Accuracy', color='#3498db')\naxes[1].plot(error['val_accuracy'], lw=2.5, label='Validation Accuracy', color='#e74c3c', ls='--')\naxes[1].set_title(\"Accuracy per Epoch\"); axes[1].set_xlabel(\"Epoch\"); axes[1].set_ylabel(\"Accuracy\")\naxes[1].legend(); axes[1].grid(alpha=0.3)\n\nplt.tight_layout()\nplt.show()\n\nprint(\"\\nFinal Evaluation:\")\nprint(\"-\" * 50)\nfor gen, name in [\n (train_generator, \"Training\"),\n (val_generator, \"Validation\"),\n (test_generator, \"Test\")\n]:\n loss, acc = model.evaluate(gen, batch_size=BATCH_SIZE, verbose=0)\n print(f\" {name:<14} Accuracy: {acc*100:.2f}% Loss: {loss:.4f}\")\nprint(\"-\" * 50)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T22:58:59.014129Z","iopub.execute_input":"2026-04-26T22:58:59.014500Z","iopub.status.idle":"2026-04-26T23:00:36.460776Z","shell.execute_reply.started":"2026-04-26T22:58:59.014466Z","shell.execute_reply":"2026-04-26T23:00:36.460068Z"}},"outputs":[],"execution_count":null},{"id":"e02fc9a2","cell_type":"code","source":"# ── Confusion matrix ──────────────────────────────────────────────\nresult = model.predict(test_generator, batch_size=BATCH_SIZE, verbose=0)\ny_pred = (result.flatten() > 0.5).astype(int)\ny_true = test_generator.labels\nclass_labels = ['Cat', 'Dog']\n\ncm = confusion_matrix(y_true, y_pred)\n\nfig, axes = plt.subplots(1, 2, figsize=(14, 5), dpi=140)\nfig.suptitle(\"Test Set Evaluation -- Cats vs Dogs\", fontsize=13, fontweight='bold')\n\nsns.heatmap(cm, annot=True, fmt='d', cmap='Blues',\n xticklabels=class_labels, yticklabels=class_labels,\n linewidths=0.5, ax=axes[0], cbar=False, annot_kws={\"size\": 14})\naxes[0].set_title(\"Confusion Matrix (Counts)\"); axes[0].set_xlabel(\"Predicted\"); axes[0].set_ylabel(\"True\")\n\ncm_pct = cm.astype(float) / cm.sum(axis=1, keepdims=True) * 100\nsns.heatmap(cm_pct, annot=True, fmt='.1f', cmap='Blues',\n xticklabels=class_labels, yticklabels=class_labels,\n linewidths=0.5, ax=axes[1], cbar=False, annot_kws={\"size\": 14})\naxes[1].set_title(\"Confusion Matrix (Row %)\"); axes[1].set_xlabel(\"Predicted\"); axes[1].set_ylabel(\"True\")\n\nplt.tight_layout()\nplt.show()\n\nprint(\"\\nClassification Report:\")\nprint(classification_report(y_true, y_pred, target_names=class_labels))\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:00:36.463053Z","iopub.execute_input":"2026-04-26T23:00:36.463376Z","iopub.status.idle":"2026-04-26T23:00:41.003470Z","shell.execute_reply.started":"2026-04-26T23:00:36.463352Z","shell.execute_reply":"2026-04-26T23:00:41.002526Z"}},"outputs":[],"execution_count":null},{"id":"884cb945","cell_type":"markdown","source":"### A.8 Prediction Visualization\n\nSample test images displayed with true and predicted labels.\nGreen = correct prediction. Red = incorrect prediction.\n","metadata":{}},{"id":"b2e982e3","cell_type":"code","source":"label_map = {0: 'Cat', 1: 'Dog'}\nX_test_reset = X_test.reset_index(drop=True)\n\nfig, axes = plt.subplots(2, 5, figsize=(18, 8))\nfig.suptitle(\"Sample Test Predictions -- green = correct | red = incorrect\",\n fontsize=12, fontweight='bold')\n\nfor idx in range(10):\n ax = axes[idx // 5][idx % 5]\n fname = X_test_reset.iloc[idx]['filename']\n true = X_test_reset.iloc[idx]['label'].capitalize()\n pred = label_map[y_pred[idx]]\n color = '#2ecc71' if pred.lower() == true.lower() else '#e74c3c'\n\n try:\n img = Image.open(os.path.join(IMAGE_DIR, fname)).resize((200, 200))\n ax.imshow(img)\n except Exception:\n ax.set_facecolor('#eeeeee')\n\n ax.set_title(f\"True: {true} | Pred: {pred}\", color=color, fontsize=9, fontweight='bold')\n ax.axis('off')\n\nplt.tight_layout()\nplt.show()\n\nmodel.save(\"cats_dogs_model.h5\")\nprint(\"Model saved: cats_dogs_model.h5\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:00:41.004521Z","iopub.execute_input":"2026-04-26T23:00:41.004831Z","iopub.status.idle":"2026-04-26T23:00:42.311479Z","shell.execute_reply.started":"2026-04-26T23:00:41.004808Z","shell.execute_reply":"2026-04-26T23:00:42.310695Z"}},"outputs":[],"execution_count":null},{"id":"6a6a35ce","cell_type":"markdown","source":"\n---\n**Part B — MNIST**\n---\n\n| Property | Value |\n|:---|:---|\n| Dataset | MNIST -- 70,000 grayscale images (60,000 train / 10,000 test) |\n| Source | Built into Keras -- no external download required |\n| Image size | 28 x 28 pixels, 1 channel (grayscale) |\n| Task | 10-class classification |\n\n**Why is MNIST simpler than Cats vs Dogs?**\nMNIST images are small (28x28), centered, uniform-background, and free of noise.\nA lightweight 2-block CNN can achieve above 99% accuracy in minutes.\n\n","metadata":{}},{"id":"7e2579ca","cell_type":"markdown","source":"### B.1 Load & Explore","metadata":{}},{"id":"e574528d","cell_type":"code","source":"(X_train_m, y_train_m), (X_test_m, y_test_m) = keras.datasets.mnist.load_data()\n\nprint(\"MNIST dataset loaded from keras.datasets.\")\nprint(f\" Training images : {X_train_m.shape}\")\nprint(f\" Test images : {X_test_m.shape}\")\nprint(f\" Pixel range : [{X_train_m.min()}, {X_train_m.max()}]\")\nprint(f\" Classes : {np.unique(y_train_m)}\")\n\nfig, axes = plt.subplots(2, 10, figsize=(16, 4))\nfig.suptitle(\"MNIST -- Two Samples per Digit Class\", fontsize=12, fontweight='bold')\nfor digit in range(10):\n for row, offset in enumerate([0, 5]):\n idx = np.where(y_train_m == digit)[0][offset]\n axes[row][digit].imshow(X_train_m[idx], cmap='Blues')\n axes[row][digit].set_title(str(digit), fontsize=10, fontweight='bold')\n axes[row][digit].axis('off')\nplt.tight_layout()\nplt.show()\n\nfig, ax = plt.subplots(figsize=(11, 4), dpi=130)\nunique, counts = np.unique(y_train_m, return_counts=True)\nsns.barplot(x=unique, y=counts, palette='icefire', ax=ax, width=0.5)\nax.set_title(\"MNIST -- Class Distribution (Training Set)\", fontsize=12, fontweight='bold')\nax.set_xlabel(\"Digit Class\"); ax.set_ylabel(\"Number of Images\")\nfor bar, count in zip(ax.patches, counts):\n ax.text(bar.get_x() + bar.get_width() / 2,\n bar.get_height() + 30, f\"{count:,}\", ha='center', fontsize=9)\nax.set_ylim(0, 7500)\nplt.tight_layout()\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:00:42.312543Z","iopub.execute_input":"2026-04-26T23:00:42.313233Z","iopub.status.idle":"2026-04-26T23:00:43.636305Z","shell.execute_reply.started":"2026-04-26T23:00:42.313208Z","shell.execute_reply":"2026-04-26T23:00:43.635502Z"}},"outputs":[],"execution_count":null},{"id":"360816ee","cell_type":"markdown","source":"### B.2 Preprocessing\n\n| Step | Operation | Reason |\n|:---|:---|:---|\n| Normalization | Divide by 255 | Scale pixels to [0, 1] for stable training |\n| Reshape | Add channel dim | Conv2D expects (H, W, C); grayscale has no channel |\n| One-hot encode | to_categorical | categorical_crossentropy needs probability vectors |\n\nExample: label `3` becomes `[0, 0, 0, 1, 0, 0, 0, 0, 0, 0]`\n","metadata":{}},{"id":"471fdc07","cell_type":"code","source":"X_train_m = X_train_m.astype('float32') / 255.0\nX_test_m = X_test_m.astype('float32') / 255.0\n\nX_train_m = X_train_m[..., np.newaxis]\nX_test_m = X_test_m[..., np.newaxis]\n\ny_train_m_oh = keras.utils.to_categorical(y_train_m, 10)\ny_test_m_oh = keras.utils.to_categorical(y_test_m, 10)\n\nprint(\"Preprocessing complete.\")\nprint(f\" X_train shape : {X_train_m.shape} dtype: {X_train_m.dtype}\")\nprint(f\" y_train[0] = digit {y_train_m[0]} -> {y_train_m_oh[0]}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:00:43.637332Z","iopub.execute_input":"2026-04-26T23:00:43.637588Z","iopub.status.idle":"2026-04-26T23:00:43.776667Z","shell.execute_reply.started":"2026-04-26T23:00:43.637567Z","shell.execute_reply":"2026-04-26T23:00:43.775834Z"}},"outputs":[],"execution_count":null},{"id":"5ac61cc1","cell_type":"markdown","source":"### B.3 CNN Architecture\n\n```\nInput: 28 x 28 x 1\n +------------------------------------------------------------------+\n | Block 1 -- Conv2D(32) + BatchNorm + MaxPool(2) + Dropout(0.25) | 14 x 14 x 32\n | Block 2 -- Conv2D(64) + BatchNorm + MaxPool(2) + Dropout(0.25) | 7 x 7 x 64\n | Head -- Flatten + Dense(128) + Dropout(0.5) |\n | Output -- Dense(10) + Softmax | 10 probabilities\n +------------------------------------------------------------------+\n```\n\n**Why simpler than Cats vs Dogs?**\nMNIST images are 28x28 grayscale and highly standardized.\nTwo convolutional blocks are sufficient for state-of-the-art accuracy.\n","metadata":{}},{"id":"7d5dc26c","cell_type":"code","source":"model_mnist = Sequential([\n Conv2D(32, (3, 3), activation='relu', padding='same', input_shape=(28, 28, 1)),\n BatchNormalization(),\n MaxPooling2D(2, 2),\n Dropout(0.25),\n\n Conv2D(64, (3, 3), activation='relu', padding='same'),\n BatchNormalization(),\n MaxPooling2D(2, 2),\n Dropout(0.25),\n\n Flatten(),\n Dense(128, activation='relu'),\n Dropout(0.5),\n Dense(10, activation='softmax')\n], name=\"MNIST_CNN\")\n\nmodel_mnist.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\nmodel_mnist.summary()\nprint(f\"\\nTotal parameters: {model_mnist.count_params():,}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:00:43.777728Z","iopub.execute_input":"2026-04-26T23:00:43.778045Z","iopub.status.idle":"2026-04-26T23:00:43.857019Z","shell.execute_reply.started":"2026-04-26T23:00:43.778011Z","shell.execute_reply":"2026-04-26T23:00:43.856472Z"}},"outputs":[],"execution_count":null},{"id":"639c408d","cell_type":"markdown","source":"### B.4 Training & Evaluation","metadata":{}},{"id":"7ed6f797","cell_type":"code","source":"callbacks_mnist = [\n EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True, verbose=1),\n ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=3, verbose=1)\n]\n\nprint(\"Training MNIST CNN...\")\nhistory_mnist = model_mnist.fit(\n X_train_m, y_train_m_oh,\n epochs=20, batch_size=128, validation_split=0.1,\n callbacks=callbacks_mnist, verbose=1\n)\n\ntest_loss_m, test_acc_m = model_mnist.evaluate(X_test_m, y_test_m_oh, verbose=0)\nprint(f\"\\nTest Accuracy : {test_acc_m*100:.2f}% | Test Loss: {test_loss_m:.4f}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:00:43.857859Z","iopub.execute_input":"2026-04-26T23:00:43.858157Z","iopub.status.idle":"2026-04-26T23:01:30.415377Z","shell.execute_reply.started":"2026-04-26T23:00:43.858137Z","shell.execute_reply":"2026-04-26T23:01:30.414490Z"}},"outputs":[],"execution_count":null},{"id":"c8e1ffd5","cell_type":"code","source":"error_m = pd.DataFrame(history_mnist.history)\nfig, axes = plt.subplots(1, 2, figsize=(16, 5), dpi=150)\nfig.suptitle(\"MNIST CNN -- Training History\", fontsize=13, fontweight='bold')\n\naxes[0].plot(error_m['loss'], lw=2.5, label='Train', color='#f39c12')\naxes[0].plot(error_m['val_loss'], lw=2.5, label='Validation', color='#8e44ad', ls='--')\naxes[0].set_title(\"Loss\"); axes[0].set_xlabel(\"Epoch\"); axes[0].legend(); axes[0].grid(alpha=0.3)\n\naxes[1].plot(error_m['accuracy'], lw=2.5, label='Train', color='#f39c12')\naxes[1].plot(error_m['val_accuracy'], lw=2.5, label='Validation', color='#8e44ad', ls='--')\naxes[1].set_title(\"Accuracy\"); axes[1].set_xlabel(\"Epoch\"); axes[1].legend(); axes[1].grid(alpha=0.3)\nplt.tight_layout(); plt.show()\n\ny_pred_m = np.argmax(model_mnist.predict(X_test_m, verbose=0), axis=1)\ncm_m = confusion_matrix(y_test_m, y_pred_m)\n\nfig, ax = plt.subplots(figsize=(9, 7), dpi=130)\nsns.heatmap(cm_m, annot=True, fmt='d', cmap='Blues', ax=ax,\n linewidths=0.3, cbar=False, annot_kws={\"size\": 10})\nax.set_title(\"Confusion Matrix -- MNIST Test Set\", fontsize=12, fontweight='bold', pad=12)\nax.set_xlabel(\"Predicted Digit\"); ax.set_ylabel(\"True Digit\")\nplt.tight_layout(); plt.show()\n\nprint(\"\\nClassification Report:\")\nprint(classification_report(y_test_m, y_pred_m))\n\nfig, axes = plt.subplots(2, 10, figsize=(16, 4))\nfig.suptitle(\"MNIST Predictions -- green = correct | red = incorrect\", fontsize=12, fontweight='bold')\nfor i in range(20):\n ax = axes[i // 10][i % 10]\n ax.imshow(X_test_m[i].squeeze(), cmap='Blues')\n pred = y_pred_m[i]; true = y_test_m[i]\n ax.set_title(f\"P:{pred}\", fontsize=9, color='#2ecc71' if pred==true else '#e74c3c', fontweight='bold')\n ax.axis('off')\nplt.tight_layout(); plt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:01:30.416558Z","iopub.execute_input":"2026-04-26T23:01:30.416878Z","iopub.status.idle":"2026-04-26T23:01:33.438188Z","shell.execute_reply.started":"2026-04-26T23:01:30.416843Z","shell.execute_reply":"2026-04-26T23:01:33.437505Z"}},"outputs":[],"execution_count":null},{"id":"4da9fd45","cell_type":"markdown","source":"---\n**Part C — Fashion MNIST**\n---\n\n| Property | Value |\n|:---|:---|\n| Dataset | Fashion MNIST -- 70,000 grayscale images (60,000 train / 10,000 test) |\n| Source | Built into Keras -- no external download required |\n| Image size | 28 x 28 pixels |\n| Task | 10-class classification |\n\n| Label | Class | Label | Class |\n|:---:|:---|:---:|:---|\n| 0 | T-shirt / top | 5 | Sandal |\n| 1 | Trouser | 6 | Shirt |\n| 2 | Pullover | 7 | Sneaker |\n| 3 | Dress | 8 | Bag |\n| 4 | Coat | 9 | Ankle Boot |\n\n**Why harder than MNIST?**\nClothing items share visual features: Shirt vs T-shirt, Coat vs Pullover.\nThese fine-grained distinctions require a deeper model.\n\n![Fashion MNIST](https://i.imgur.com/xSg5QJS.jpeg)\n\n*Figure C.0: All 10 classes in the Fashion MNIST dataset.*\n","metadata":{}},{"id":"f28c9c3c","cell_type":"markdown","source":"### C.1 Load & Explore","metadata":{}},{"id":"708a23cb","cell_type":"code","source":"FASHION_CLASSES = [\n 'T-shirt', 'Trouser', 'Pullover', 'Dress', 'Coat',\n 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle Boot'\n]\n\n(X_train_f, y_train_f), (X_test_f, y_test_f) = keras.datasets.fashion_mnist.load_data()\n\nX_train_f = X_train_f.astype('float32') / 255.0\nX_test_f = X_test_f.astype('float32') / 255.0\nX_train_f = X_train_f[..., np.newaxis]\nX_test_f = X_test_f[..., np.newaxis]\ny_train_f_oh = keras.utils.to_categorical(y_train_f, 10)\ny_test_f_oh = keras.utils.to_categorical(y_test_f, 10)\n\nprint(f\"Training : {X_train_f.shape} | Test: {X_test_f.shape}\")\n\npalette = ['#3498db','#e74c3c','#2ecc71','#9b59b6','#e67e22',\n '#1abc9c','#f39c12','#c0392b','#16a085','#8e44ad']\n\nfig, axes = plt.subplots(2, 5, figsize=(15, 7))\nfig.suptitle(\"Fashion MNIST -- One Sample per Class\", fontsize=12, fontweight='bold')\nfor i, cls in enumerate(FASHION_CLASSES):\n ax = axes[i // 5][i % 5]\n idx = np.where(y_train_f == i)[0][2]\n ax.imshow(X_train_f[idx].squeeze(), cmap='plasma')\n ax.set_title(cls, color=palette[i], fontsize=10, fontweight='bold')\n ax.axis('off')\nplt.tight_layout(); plt.show()\n\nfig, ax = plt.subplots(figsize=(13, 4), dpi=130)\nunique_f, counts_f = np.unique(y_train_f, return_counts=True)\nsns.barplot(x=[FASHION_CLASSES[i] for i in unique_f], y=counts_f, palette=palette, ax=ax, width=0.6)\nax.set_title(\"Fashion MNIST -- Class Distribution\", fontsize=12, fontweight='bold')\nax.set_xlabel(\"Class\"); ax.set_ylabel(\"Images\")\nplt.xticks(rotation=30, ha='right')\nfor bar, count in zip(ax.patches, counts_f):\n ax.text(bar.get_x() + bar.get_width() / 2,\n bar.get_height() + 30, f\"{count:,}\", ha='center', fontsize=8)\nplt.tight_layout(); plt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:01:33.439154Z","iopub.execute_input":"2026-04-26T23:01:33.439469Z","iopub.status.idle":"2026-04-26T23:01:35.118922Z","shell.execute_reply.started":"2026-04-26T23:01:33.439434Z","shell.execute_reply":"2026-04-26T23:01:35.118081Z"}},"outputs":[],"execution_count":null},{"id":"7cb8df4e","cell_type":"markdown","source":"### C.2 CNN Architecture\n\n```\nInput: 28 x 28 x 1\n +------------------------------------------------------------------------------+\n | Block 1 -- Conv2D(32)x2 + BN + MaxPool(2) + Dropout(0.3) | 14 x 14 x 32 |\n | Block 2 -- Conv2D(64)x2 + BN + MaxPool(2) + Dropout(0.3) | 7 x 7 x 64 |\n | Block 3 -- Conv2D(128) + BN + MaxPool(2) + Dropout(0.3) | 3 x 3 x 128 |\n | Head -- Flatten + Dense(256) + BN + Dropout(0.5) | |\n | Output -- Dense(10) + Softmax | 10 probs |\n +------------------------------------------------------------------------------+\n```\n\n3 blocks (vs 2 for MNIST) to capture finer garment-level distinctions.\nDouble convolution per block increases receptive field without immediate pooling.\n","metadata":{}},{"id":"5fc79ff9","cell_type":"code","source":"model_fm = Sequential([\n Conv2D(32, (3,3), activation='relu', padding='same', input_shape=(28,28,1)),\n BatchNormalization(),\n Conv2D(32, (3,3), activation='relu', padding='same'),\n BatchNormalization(),\n MaxPooling2D(2,2), Dropout(0.3),\n\n Conv2D(64, (3,3), activation='relu', padding='same'),\n BatchNormalization(),\n Conv2D(64, (3,3), activation='relu', padding='same'),\n BatchNormalization(),\n MaxPooling2D(2,2), Dropout(0.3),\n\n Conv2D(128, (3,3), activation='relu', padding='same'),\n BatchNormalization(),\n MaxPooling2D(2,2), Dropout(0.3),\n\n Flatten(),\n Dense(256, activation='relu'),\n BatchNormalization(),\n Dropout(0.5),\n Dense(10, activation='softmax')\n], name=\"FashionMNIST_CNN\")\n\nmodel_fm.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\nmodel_fm.summary()\nprint(f\"\\nTotal parameters: {model_fm.count_params():,}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:01:35.119933Z","iopub.execute_input":"2026-04-26T23:01:35.120246Z","iopub.status.idle":"2026-04-26T23:01:35.250693Z","shell.execute_reply.started":"2026-04-26T23:01:35.120225Z","shell.execute_reply":"2026-04-26T23:01:35.250088Z"}},"outputs":[],"execution_count":null},{"id":"80df0be3","cell_type":"markdown","source":"### C.3 Training & Evaluation","metadata":{}},{"id":"39a0e730","cell_type":"code","source":"callbacks_fm = [\n EarlyStopping(monitor='val_loss', patience=7, restore_best_weights=True, verbose=1),\n ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=4, verbose=1)\n]\n\nprint(\"Training Fashion MNIST CNN...\")\nhistory_fm = model_fm.fit(\n X_train_f, y_train_f_oh,\n epochs=25, batch_size=128, validation_split=0.1,\n callbacks=callbacks_fm, verbose=1\n)\n\ntest_loss_f, test_acc_f = model_fm.evaluate(X_test_f, y_test_f_oh, verbose=0)\nprint(f\"\\nTest Accuracy : {test_acc_f*100:.2f}% | Test Loss: {test_loss_f:.4f}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:01:35.251548Z","iopub.execute_input":"2026-04-26T23:01:35.251847Z","iopub.status.idle":"2026-04-26T23:03:53.324625Z","shell.execute_reply.started":"2026-04-26T23:01:35.251827Z","shell.execute_reply":"2026-04-26T23:03:53.323822Z"}},"outputs":[],"execution_count":null},{"id":"2aa61d2e","cell_type":"code","source":"error_f = pd.DataFrame(history_fm.history)\nfig, axes = plt.subplots(1, 2, figsize=(16, 5), dpi=150)\nfig.suptitle(\"Fashion MNIST CNN -- Training History\", fontsize=13, fontweight='bold')\n\naxes[0].plot(error_f['loss'], lw=2.5, label='Train', color='#2ecc71')\naxes[0].plot(error_f['val_loss'], lw=2.5, label='Validation', color='#8e44ad', ls='--')\naxes[0].set_title(\"Loss\"); axes[0].set_xlabel(\"Epoch\"); axes[0].legend(); axes[0].grid(alpha=0.3)\n\naxes[1].plot(error_f['accuracy'], lw=2.5, label='Train', color='#2ecc71')\naxes[1].plot(error_f['val_accuracy'], lw=2.5, label='Validation', color='#8e44ad', ls='--')\naxes[1].set_title(\"Accuracy\"); axes[1].set_xlabel(\"Epoch\"); axes[1].legend(); axes[1].grid(alpha=0.3)\nplt.tight_layout(); plt.show()\n\ny_pred_f = np.argmax(model_fm.predict(X_test_f, verbose=0), axis=1)\ncm_f = confusion_matrix(y_test_f, y_pred_f)\n\nfig, axes = plt.subplots(1, 2, figsize=(20, 7), dpi=130)\nfig.suptitle(\"Fashion MNIST -- Confusion Matrix\", fontsize=13, fontweight='bold')\n\nsns.heatmap(cm_f, annot=True, fmt='d', cmap='Blues',\n xticklabels=FASHION_CLASSES, yticklabels=FASHION_CLASSES,\n ax=axes[0], linewidths=0.3, cbar=False, annot_kws={\"size\": 9})\naxes[0].set_title(\"Counts\"); axes[0].set_xlabel(\"Predicted\"); axes[0].set_ylabel(\"True\")\nplt.setp(axes[0].get_xticklabels(), rotation=40, ha='right')\n\ncm_f_pct = cm_f.astype(float) / cm_f.sum(axis=1, keepdims=True) * 100\nsns.heatmap(cm_f_pct, annot=True, fmt='.1f', cmap='Blues',\n xticklabels=FASHION_CLASSES, yticklabels=FASHION_CLASSES,\n ax=axes[1], linewidths=0.3, cbar=False, annot_kws={\"size\": 9})\naxes[1].set_title(\"Row-Normalized (%)\"); axes[1].set_xlabel(\"Predicted\"); axes[1].set_ylabel(\"True\")\nplt.setp(axes[1].get_xticklabels(), rotation=40, ha='right')\n\nplt.tight_layout(); plt.show()\n\nprint(\"\\nClassification Report:\")\nprint(classification_report(y_test_f, y_pred_f, target_names=FASHION_CLASSES))\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:03:53.325492Z","iopub.execute_input":"2026-04-26T23:03:53.325967Z","iopub.status.idle":"2026-04-26T23:03:56.226634Z","shell.execute_reply.started":"2026-04-26T23:03:53.325941Z","shell.execute_reply":"2026-04-26T23:03:56.225932Z"}},"outputs":[],"execution_count":null},{"id":"e882b147","cell_type":"markdown","source":"---\n
\n\n## D. Conclusion & Comparison\n\n
\n\n---\n","metadata":{}},{"id":"8b2951b5","cell_type":"code","source":"cvd_loss, cvd_acc = model.evaluate(test_generator, batch_size=BATCH_SIZE, verbose=0)\n\nresults = {\n 'Cats vs Dogs' : cvd_acc * 100,\n 'MNIST' : test_acc_m * 100,\n 'Fashion MNIST': test_acc_f * 100,\n}\n\nprint(\"=\" * 65)\nprint(f\" {'Dataset':<18} {'Task':<22} {'Test Accuracy':>13}\")\nprint(\"=\" * 65)\nprint(f\" {'Cats vs Dogs':<18} {'Binary (2 classes)':<22} {results['Cats vs Dogs']:>12.2f}%\")\nprint(f\" {'MNIST':<18} {'Multi (10 classes)':<22} {results['MNIST']:>12.2f}%\")\nprint(f\" {'Fashion MNIST':<18} {'Multi (10 classes)':<22} {results['Fashion MNIST']:>12.2f}%\")\nprint(\"=\" * 65)\n\nfig, ax = plt.subplots(figsize=(11, 6), dpi=150)\nbars = ax.bar(results.keys(), results.values(),\n color=['#3498db', '#f39c12', '#2ecc71'],\n edgecolor='black', linewidth=0.8, width=0.45)\nax.set_ylabel('Test Accuracy (%)', fontsize=12)\nax.set_ylim(0, 112)\nax.set_title('Test Accuracy Comparison -- All Three Datasets', fontsize=13, fontweight='bold')\nax.axhline(y=90, color='gray', ls='--', alpha=0.6, lw=1.5, label='90% reference')\nax.legend(fontsize=10); ax.grid(axis='y', alpha=0.3)\n\nfor bar, acc in zip(bars, results.values()):\n ax.text(bar.get_x() + bar.get_width() / 2, bar.get_height() + 0.8,\n f'{acc:.2f}%', ha='center', va='bottom', fontsize=13, fontweight='bold')\n\nplt.tight_layout(); plt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-26T23:03:56.227436Z","iopub.execute_input":"2026-04-26T23:03:56.227823Z","iopub.status.idle":"2026-04-26T23:03:59.728130Z","shell.execute_reply.started":"2026-04-26T23:03:56.227801Z","shell.execute_reply":"2026-04-26T23:03:59.727480Z"}},"outputs":[],"execution_count":null},{"id":"3dbb4d14","cell_type":"markdown","source":"### Key Findings\n\n**1. Architecture depth must match task complexity.**\nCats vs Dogs required 4 convolutional blocks (32 to 64 to 128 to 256 filters).\nFashion MNIST needed 3 blocks. MNIST required only 2.\n\n**2. Data augmentation is essential for real-world photographic data.**\nFor MNIST and Fashion MNIST, standardized images require no augmentation.\nFor Cats vs Dogs, augmentation (flip, rotation, zoom) was critical to prevent overfitting.\n\n**3. `flow_from_dataframe` is the correct approach on Kaggle.**\nIt accepts an explicit filename-to-label mapping and works on read-only filesystems,\nunlike `flow_from_directory` which requires images pre-sorted into class folders.\n\n**4. BatchNormalization and Dropout are complementary.**\nBatchNorm stabilizes activations; Dropout prevents neuron co-adaptation.\nTogether they enable deeper networks without explicit weight decay.\n\n**5. EarlyStopping with `restore_best_weights=True` is essential.**\nIt halts training when validation loss plateaus and saves the best model automatically.\n\n**6. Confusion matrices expose class-level failure modes.**\nFor Fashion MNIST: Shirt vs T-shirt and Coat vs Pullover are the hardest pairs.\nFor MNIST: digit pairs 4/9 and 3/8 produce the most errors.\n\n---\n\n### Acknowledgements\n\n**Dr. Sherif Al-Itreby** — for his expert instruction in Deep Learning theory and practice,\nand for designing the project framework that motivated this investigation.\n\n**Eng. Omar Amir** — for his technical mentorship and invaluable assistance\nduring the implementation phase.\n","metadata":{}}]}