Deep Learning Project: CNN (Fashion-MNIST) and BiLSTM (SST-2)

This repository contains two end-to-end notebook workflows:

• A Convolutional Neural Network (CNN) for Fashion-MNIST image classification
• A Bidirectional LSTM (BiLSTM) for SST-2 sentiment analysis

---

Dataset details

Fashion-MNIST (for CNN)

• 70,000 grayscale images of clothing items (60,000 train / 10,000 test)
• Image size: 28×28 pixels, 10 classes:
  • T-shirt/top, Trouser, Pullover, Dress, Coat, Sandal, Shirt, Sneaker, Bag, Ankle boot
• Loading: tensorflow.keras.datasets.fashion_mnist.load_data()
• Preprocessing:
  • Normalize pixel values to [0, 1]
  • Expand channel dimension to shape (28, 28, 1) for CNN input
  • One-hot encode labels to 10 classes

GLUE SST-2 (for BiLSTM)

• Stanford Sentiment Treebank binary classification (positive/negative)
• Loaded via tensorflow_datasets as glue/sst2
• Splits used in the notebook:
  • Train: split='train'
  • Validation/Test: split='validation'
• Preprocessing:
  • Keras Tokenizer with num_words=10000 and OOV token
  • Convert texts to integer sequences
  • Pad/truncate sequences to max_length=50 (post-padding)
  • Labels cast to float32 for training

---

Model architectures

CNN (Fashion-MNIST)

A compact architecture suitable for 28×28 grayscale images:

• Input: (28, 28, 1)
• Conv2D(32, 3×3, ReLU)
• MaxPooling2D(2×2)
• Conv2D(64, 3×3, ReLU)
• MaxPooling2D(2×2)
• Flatten
• Dense(64, ReLU)
• Dropout(0.5)
• Dense(10, Softmax)

Compile and training settings:

• Optimizer: adam
• Loss: categorical_crossentropy
• Metric: accuracy
• Epochs: 10
• Batch size: 32
• Validation: uses the provided test split

BiLSTM (SST-2 Sentiment)

A text classifier with embeddings and stacked bidirectional LSTMs:

• Input: integer sequences of length 50, vocabulary size 10,000
• Embedding(input_dim=10000, output_dim=256, input_length=50)
• SpatialDropout1D(0.2)
• Bidirectional(LSTM(128, return_sequences=True))
• Dropout(0.3)
• Bidirectional(LSTM(64))
• Dropout(0.3)
• Dense(64, ReLU)
• Dropout(0.5)
• Dense(1, Sigmoid)

Compile and training settings:

• Optimizer: adam
• Loss: binary_crossentropy
• Metrics: accuracy
• Epochs: 5
• Batch size: 32
• Validation split: 0.2 from training set
• Callbacks:
  • EarlyStopping(monitor=val_loss, patience=5, restore_best_weights=True)
  • ReduceLROnPlateau(monitor=val_loss, factor=0.5, patience=3, min_lr=1e-7)

---

Evaluation, Metrics, and Outputs

CNN on Fashion-MNIST

Dataset: Fashion-MNIST (60,000 training images, 10,000 test images)
Model Type: Convolutional Neural Network (2 Conv2D + 2 MaxPool + Dense layers)
Loss Function: Categorical Cross-Entropy
Optimizer: Adam

Metrics Summary

Metric	Training	Validation	Test
Accuracy	91.3%	90.6%	90.64%
Loss	0.238	0.265	0.2654

Classification Report

Class	Precision	Recall	F1-Score	Support
T-shirt/Top	0.87	0.84	0.85	1000
Trouser	0.99	0.98	0.99	1000
Pullover	0.82	0.87	0.85	1000
Dress	0.90	0.92	0.91	1000
Coat	0.84	0.88	0.86	1000
Sandal	0.98	0.98	0.98	1000
Shirt	0.75	0.69	0.72	1000
Sneaker	0.95	0.97	0.96	1000
Bag	0.99	0.97	0.98	1000
Ankle Boot	0.98	0.96	0.97	1000

Macro Avg: 0.91 Weighted Avg: 0.91 Overall Accuracy: 0.9064

Visual Outputs

• Training Curves:
  • Accuracy improved steadily to ~91% across 10 epochs.
  • Validation loss decreased and stabilized around 0.27.
• Confusion Matrix: Highlights strong classification in classes like Trouser, Sandal, and Ankle Boot, with minor confusion among garment categories such as Shirt and Coat.

BiLSTM for Sentiment Analysis

Dataset: GLUE SST-2 (Stanford Sentiment Treebank)
Model Type: Bidirectional LSTM with Embedding and Dropout
Loss Function: Binary Cross-Entropy
Optimizer: Adam
Callbacks: EarlyStopping, ReduceLROnPlateau

Metrics Summary

Metric	Training	Validation	Test
Accuracy	96.8%	92.0%	82.91%
Loss	0.074	0.338	0.436

Confusion Matrix

	Predicted Negative	Predicted Positive
Actual Negative	340	88
Actual Positive	61	383

Precision: 0.81 Recall: 0.86 F1 Score: 0.84

Visual Outputs

• Training Curves:
  • Training accuracy increased to 96% by epoch 5, with validation convergence around 92%.
  • Overfitting mitigated by dropout and early stopping.
• Confusion Matrix Plot:
  Shows good discrimination between positive and negative sentiment classes, minor false positives.