Deep Learning Lab Assignments

Complete implementations of all 10 Deep Learning lab assignments.

Setup

pip install -r requirements.txt

Labs

#	Topic	File
1	PCA – Dimensionality Reduction	Lab1_PCA/lab1_pca.py
2	Simple & Multiple Linear Regression	Lab2_LinearRegression/lab2_linear_regression.py
3	K-Nearest Neighbours (Iris)	Lab3_KNN/lab3_knn.py
4	Naive Bayes Classifier	Lab4_NaiveBayes/lab4_naive_bayes.py
5	Decision Tree + Random Forest + AdaBoost	Lab5_DecisionTree_RF_AdaBoost/lab5_tree_rf_adaboost.py
6	CNN – MNIST Digit Classification	Lab6_CNN_MNIST/lab6_cnn_mnist.py
7	CNN – Plant Disease Detection	Lab7_CNN_PlantDisease/lab7_plant_disease.py
8	CNN – Fashion MNIST Classification	Lab8_CNN_FashionMNIST/lab8_fashion_mnist.py
9	RNN – Stock Price Prediction	Lab9_RNN_StockPrice/lab9_rnn_stock.py
10	LSTM – Weather Prediction	Lab10_LSTM_Weather/lab10_lstm_weather.py

How to Run

# Example
cd Lab1_PCA
python lab1_pca.py

Labs 1–6, 8–10 use built-in datasets and run without any external data files.
Lab 7 requires a plant disease dataset folder (e.g., PlantVillage from Kaggle).

Notes

• Labs 9 & 10 include synthetic data generators so they run immediately even without a real CSV. Swap in your actual CSV when available (instructions in each file).
• All plots are saved as .png in the same folder as the script.

lab 1 Viva questions (click to reveal answer)

What is PCA? › Principal Component Analysis. It reduces dimensions by finding new axes (principal components) that capture maximum variance in data. Why standardize before PCA? › PCA is variance-based. If features have different scales (e.g., height in cm vs weight in kg), larger-scale features dominate. Standardization fixes this. What is a scree plot? › A line graph of explained variance ratio vs. number of components. The 'elbow' point tells us how many components to keep. What are principal components? › New orthogonal axes created by PCA. PC1 has max variance, PC2 is orthogonal to PC1 with next max variance, etc. What is explained_variance_ratio_? › Fraction of total variance captured by each principal component. Sum = 1.0 for all components. Difference between fit() and fit_transform()? › fit() learns parameters (mean, std). transform() applies them. fit_transform() does both in one step.

lab 2 Viva questions (click to reveal answer)

What is Simple vs Multiple Linear Regression? › SLR: one input feature predicts one output. MLR: multiple input features predict one output. What does R² score mean? › Coefficient of determination. R²=0.85 means model explains 85% of variance in target. Closer to 1 is better. What is the difference between MSE and RMSE? › MSE is in squared units. RMSE is sqrt(MSE), in same unit as target — easier to interpret as average error. Why do we split into train and test? › To evaluate how well model generalizes to unseen data. Training on all data gives overly optimistic metrics. What assumption does Linear Regression make? › Linear relationship between features and target, no multicollinearity, homoscedasticity (constant variance of errors), errors are normally distributed. What is overfitting? › Model learns training data too well including noise, performs poorly on test data. MLR with many features risks this.

lab 3

Viva questions (click to reveal answer)

How does KNN work? › KNN stores training data. For a new point, it finds K nearest neighbors using distance (Euclidean by default) and assigns the majority class among those K neighbors. What is the effect of K value? › Small K (e.g., 1) = overfits, sensitive to noise. Large K = smoother boundary, might underfit. Optimal K found using cross-validation. What distance metric does KNN use? › Default is Euclidean distance: sqrt(sum of (xi-yi)²). Can also use Manhattan, Minkowski. Is KNN a lazy learner? › Yes. It does no training computation — just stores data. All computation happens at prediction time. Why scale features for KNN? › KNN uses distance. Features with large values dominate. StandardScaler ensures equal contribution from all features. What is Iris dataset? › Famous dataset with 150 samples, 3 classes (Setosa, Versicolor, Virginica), 4 features (sepal length/width, petal length/width).

lab 4

What is Bayes theorem? › P(A|B) = P(B|A) × P(A) / P(B). In classification: P(class|features) ∝ P(features|class) × P(class). What is the 'Naive' assumption? › Features are conditionally independent given the class. In reality this is rarely true, but the classifier still works well in practice. When to use Gaussian NB vs Multinomial NB? › Gaussian NB: continuous features (e.g., height, weight). Multinomial NB: discrete counts (e.g., word frequencies in text). What is precision vs recall? › Precision = TP/(TP+FP) — how many predicted positives are actually positive. Recall = TP/(TP+FN) — how many actual positives were found. What is F1-score? › Harmonic mean of precision and recall: 2×(P×R)/(P+R). Good when class imbalance exists. Advantage of Naive Bayes? › Fast, works well with small data, handles high dimensions well, good baseline for text classification.

lab 5

How does a Decision Tree split? › Finds the feature and threshold that maximizes information gain (or minimizes Gini impurity) at each node. What is overfitting in decision trees? › A deep tree memorizes training data including noise, gives high train accuracy but low test accuracy. What is cost complexity pruning? › After full tree is built, branches are removed if they don't significantly improve accuracy. ccp_alpha controls aggressiveness. How does Random Forest reduce overfitting? › Each tree is trained on a random bootstrap sample of data with random feature subset. Averaging many uncorrelated trees reduces variance. What is bagging vs boosting? › Bagging (Random Forest): trains trees in parallel on random subsets, averages results. Boosting (AdaBoost): trains sequentially, each model corrects errors of previous. What is a decision stump? › A decision tree with depth=1. Makes one split. Used as weak learner in AdaBoost. What is Gini impurity? › Measure of how often a randomly chosen element would be misclassified. 0 = pure node. Used to decide splits.

lab 6

mnist.load_data() Downloads MNIST: 60k train + 10k test images of handwritten digits 0–9. X_tr.reshape(-1, 28, 28, 1) Add channel dimension. CNN expects (batch, height, width, channels). Grayscale = 1 channel. / 255.0 Normalize pixel values from 0-255 to 0-1. Helps neural network train faster. Conv2D(32, (3,3), activation='relu') 32 filters of 3×3 size. Each filter detects a different pattern (edge, curve, etc.). MaxPooling2D((2,2)) Reduces spatial size by half by taking max in 2×2 window. Reduces params, adds translation invariance. Flatten() Converts 2D feature maps to 1D vector to feed into Dense layers. Dense(64, activation='relu') Fully connected layer with 64 neurons. Learns combinations of features. Dense(10, activation='softmax') 10 neurons for 10 digit classes. Softmax gives probability distribution summing to 1. sparse_categorical_crossentropy Loss for multi-class when labels are integers (not one-hot encoded). np.argmax(m.predict(X_te), axis=1) predict() gives probabilities. argmax picks the class with highest probability. confusion_matrix + sns.heatmap Shows which classes are confused with each other. Diagonal = correct predictions.

What is a convolutional layer? › Applies learnable filters to detect spatial patterns (edges, textures). Each filter produces a feature map. What is MaxPooling? › Downsamples feature maps by taking maximum value in a window. Reduces computation and adds translation invariance. What is ReLU activation? › f(x) = max(0,x). Removes negative values. Solves vanishing gradient. Most common hidden layer activation. What is Softmax? › Converts raw scores to probabilities that sum to 1. Used in output layer for multi-class classification. What is Adam optimizer? › Adaptive learning rate optimizer. Combines momentum and RMSProp. Default choice for most deep learning. What does a confusion matrix show? › Row = actual class, Column = predicted. Diagonal = correct. Off-diagonal = misclassifications. Shows which classes get confused. What is sparse_categorical_crossentropy vs categorical_crossentropy? › sparse: labels are integers (0,1,2...). categorical: labels are one-hot encoded ([0,1,0,...]).

lab 7

ImageDataGenerator(rescale=1./255, rotation_range=20, horizontal_flip=True) Data augmentation: rescales pixels to 0-1, randomly rotates ±20°, flips horizontally. Creates more training variety. flow_from_directory('plant_data', target_size=(64,64)) Reads images from folder structure (one subfolder per class). Resizes to 64×64. class_mode='categorical' One-hot encodes labels. For multi-class with categorical_crossentropy. subset='training' / subset='validation' Splits data using validation_split=0.2. 80% train, 20% validation. n_cls = len(train_data.class_indices) Automatically gets number of plant disease classes from folder names. Dropout(0.5) Randomly disables 50% of neurons during training. Prevents overfitting. Dense(128, activation='relu') Larger dense layer to learn complex disease patterns. Dense(n_cls, activation='softmax') Output layer with one neuron per disease class. h = m.fit(...) h (history) stores accuracy/loss per epoch. Used for plotting. h.history['accuracy'] vs h.history['val_accuracy'] Compare train vs validation accuracy per epoch. Large gap = overfitting.

What is data augmentation? › Artificially expanding training data by applying transformations (rotation, flip, zoom). Reduces overfitting, especially with small datasets. What is Dropout? › Regularization technique: randomly sets neuron outputs to 0 during training. Forces network to learn redundant representations. What is overfitting and how to detect it? › Train accuracy >> validation accuracy. Model memorizes training data instead of generalizing. Why use color images (3 channels) for plant disease? › Disease symptoms (yellowing, spots, discoloration) are color-dependent. 3 channels (RGB) preserve this info. What datasets are used for plant disease? › PlantVillage dataset is most common: 50k+ images of 38 plant disease classes. What is transfer learning? (bonus) › Using a pre-trained model (like VGG, ResNet) as base and fine-tuning for new task. Much better for small datasets.

lab 8

names = [...] 10 clothing category names corresponding to labels 0–9. X_tr[..., np.newaxis] Adds channel dimension. ... means 'all existing dims'. Same as reshape(-1,28,28,1). / 255.0 Normalizes pixels to 0–1 range. BatchNormalization() Normalizes layer outputs during training. Speeds up training, acts as regularizer. Dropout(0.4) 40% dropout. Fashion MNIST is harder than MNIST, more regularization needed. batch_size=64 Process 64 images at once per gradient update. Balances speed and stability. m.evaluate(X_te, y_te) Returns final test loss and accuracy. m.predict(X_te[:5]) Get probability predictions for first 5 test images. names[np.argmax(pred[i])] Convert probability array to class name using argmax.

What is Fashion MNIST? › Drop-in replacement for MNIST. 70k grayscale 28×28 images of 10 clothing categories. Harder than digit MNIST. What is Batch Normalization? › Normalizes the input of each layer (zero mean, unit variance) during training. Reduces internal covariate shift, allows higher learning rates. What is batch size? › Number of training samples processed before updating model weights. Larger batch = faster but needs more memory and may converge to worse solution. Why is Fashion MNIST harder than MNIST? › Digits have distinct shapes; clothing items share more visual similarity (e.g., shirt vs. coat vs. pullover). What is sparse_categorical_crossentropy? › Cross-entropy loss for multi-class when labels are integers. If labels were one-hot vectors, use categorical_crossentropy. What is an epoch? › One complete pass through the entire training dataset. Multiple epochs allow the model to refine weights.

lab 9

df['Close'].values.reshape(-1,1) Get closing prices as column vector. reshape needed for MinMaxScaler. MinMaxScaler() Scale values to 0-1 range. RNNs train poorly on large raw stock prices. def mk_seq(data, steps=60) Create sliding window sequences. Each X is 60 past days, y is next day price. X.append(data[i-steps:i, 0]) Take 60 consecutive values as input sequence. SimpleRNN(64, return_sequences=True) 64 RNN units. return_sequences=True passes output at each step to next RNN layer. SimpleRNN(32) Second RNN layer, returns only final output (return_sequences=False by default). Dense(1) Output: single predicted price (regression, not classification). loss='mse' Mean Squared Error for regression. No activation on output layer. sc.inverse_transform(pred) Convert normalized predictions back to actual dollar values.

What is an RNN? › Recurrent Neural Network. Has internal state (memory) that persists across time steps. Processes sequences by passing hidden state from one step to the next. Why use RNN for stock prediction? › Stock prices are time series — future prices depend on past prices. RNNs capture temporal dependencies. What is the vanishing gradient problem? › In RNNs, gradients shrink exponentially through many time steps, making it hard to learn long-term dependencies. Solved by LSTM/GRU. Why MinMaxScaler for RNN (not StandardScaler)? › RNNs are sensitive to input scale. MinMaxScaler bounds values to [0,1] which suits sigmoid/tanh activations. What is return_sequences=True? › When stacking RNN layers, intermediate layers must pass output at every time step (not just last). return_sequences=True enables this. Why 60 time steps? › Common heuristic for stock data (~3 trading months). Model sees 60 previous days to predict day 61.

lab 10

df['Temperature'].values.reshape(-1,1) Extract temperature column as 2D array for scaler. MinMaxScaler() + fit_transform Normalize temperatures to [0,1]. Helps LSTM converge faster. mk_seq(d, steps=30) 30-day sliding windows. Input = 30 past temperatures, output = next day. LSTM(64, return_sequences=True) 64 LSTM units. return_sequences=True for stacking another LSTM layer below. LSTM(32) Second LSTM reads full sequence from above, outputs one vector at final step. Dropout(0.2) 20% dropout after each LSTM to prevent overfitting. Dense(1) Predict one value (tomorrow's temperature). loss='mse' Regression task. MSE penalizes large prediction errors more. sc.inverse_transform(pred) De-normalize back to real temperatures (°C or °F).

What is LSTM and how is it better than SimpleRNN? › Long Short-Term Memory. Has gates (input, forget, output) that control what information to keep or discard. Solves vanishing gradient — can learn long-term dependencies. What are LSTM gates? › Forget gate: what to erase from cell state. Input gate: what new info to add. Output gate: what to output. Cell state carries long-term memory. Difference between RNN and LSTM? › RNN: simple hidden state, suffers vanishing gradient, short memory. LSTM: cell state + hidden state, gated mechanism, can remember across 100s of steps. What is GRU? › Gated Recurrent Unit. Simplified LSTM with 2 gates instead of 3. Faster to train, similar performance. Why is weather prediction a time series problem? › Temperature depends on recent history (trends, seasons). Past values predict future values. What is the difference between regression and classification in output layer? › Regression: Dense(1) with no activation, loss=MSE. Classification: Dense(n_classes) with softmax, loss=cross-entropy.