Deep Learning Lab-BAI701-Program 5

 

Program 5

Design and implement a deep learning network for classification of textual documents.

Step 1:

 

·        TensorFlow/Keras → to build and train the deep learning model.

·        IMDB dataset → built-in dataset of 50,000 movie reviews (positive/negative).

·        pad_sequences → ensures all reviews are the same length.

·        Matplotlib → for plotting graphs of training/validation performance.

Step 2:

 

·        VOCAB_SIZE = Only keep the 10,000 most common words.

·        MAX_LEN = Each review will be cut/padded to 256 words.

·        EMBEDDING_DIM = Each word becomes a 128-dimensional vector.

·        BATCH_SIZE = Train with 64 reviews at a time.

·        EPOCHS = Train the model for up to 12 passes through the dataset.

·        SEED = Fixes randomness for reproducibility.

Step 3

·        Loads training and test data. Each review is already converted into word indices.

·        pad_sequences → ensures all reviews are the same length (256).

·        Short reviews → padded with zeros.

·        Long reviews → cut at 256.

 

Step 4:

 

·        Splits training data into:

·        Train set (80%)

·        Validation set (20%) → checks performance during training.

·        Test set is used only at the end.

 

Step 5

·        Embedding → converts word indices into dense word vectors.

·        SpatialDropout1D → randomly drops word vectors during training (prevents overfitting).

·        Bidirectional LSTM → reads the review both forward and backward, capturing meaning from context.

·        Pooling layers → reduce sequence into a fixed vector:

·        GlobalMaxPooling1D → picks strongest features.

·        GlobalAveragePooling1D → averages features.

·        Dense(64, relu) → hidden layer to learn patterns.

·        Dropout(0.4) → prevents overfitting by turning off neurons randomly.

·        Dense(1, sigmoid) → outputs probability (positive review vs negative review).

 

Step 6:

·        Adam optimizer → adjusts learning rate automatically.

·         Binary crossentropy → best suited for 2-class problems.

·        Metrics:

·        Accuracy → % of correct predictions.

·        AUC → measures how well the model separates the two classes.

 

Step 7:

 

·        EarlyStopping → stops training if validation doesn’t improve (avoids overfitting).

·        ModelCheckpoint → saves the best model automatically.

·        ReduceLROnPlateau → lowers learning rate if training stalls.

 

Step 8:

• Model learns patterns from training reviews.
• Validation set checks progress each epoch.

Step 9: