evaluation.py

from itertools import product
import numpy as np
import matplotlib.pyplot as plt


def confusion_matrix(y_pred, y_true, num_classes):
    """
    Create a confusion matrix for label encodings in PyTorch.

    Parameters:
    y_pred (torch.Tensor): Predicted labels tensor.
    y_true (torch.Tensor): True labels tensor.
    num_classes (int): Number of classes.

    Returns:
    numpy.ndarray: Confusion matrix.
    """ 
    # if len(y_pred) != len(y_true):
    #     raise ValueError("Shapes of predictions and true labels must match. y_pred shape: {} y_true shape: {}".format(y_pred.shape, y_true.shape))

    conf_matrix = np.zeros((num_classes, num_classes), dtype=np.int64)

    y_pred_np = y_pred.argmax(dim=1).cpu().numpy()
    y_true_np = y_true.cpu().numpy()

    for pred, true in zip(y_pred_np, y_true_np):
        conf_matrix[pred, true] += 1

    return conf_matrix

def calculate_confusion_matrix(test_emb, test_labels, model):
    model.eval()
    output = model(test_emb)
    return confusion_matrix(output, test_labels, 6)

def class_accuracy(conf_matrix):
    """
    Calculate accuracy for each class based on a confusion matrix.

    Parameters:
    conf_matrix (numpy.ndarray): Confusion matrix.

    Returns:
    list: List of accuracies for each class.
    """
    # diagonal = np.diag(conf_matrix)

    # row_sums = conf_matrix.sum(axis=1)

    # accuracies = diagonal / row_sums.astype(float)

    # return accuracies
    
    diagonal = np.diag(conf_matrix)
    row_sums = conf_matrix.sum(axis=1)

    with np.errstate(divide='ignore', invalid='ignore'):
        accuracies = np.where(row_sums != 0, diagonal / row_sums.astype(float), 0.0)

    return accuracies

def class_f1_score(conf_matrix, epsilon=1e-7):
    """
    Calculate F1 score for each class based on a confusion matrix.

    Parameters:
    conf_matrix (numpy.ndarray): Confusion matrix.
    epsilon (float): Smoothing term to avoid division by zero.

    Returns:
    list: List of F1 scores for each class.
    """
    
    tp = np.diag(conf_matrix)
    fp = conf_matrix.sum(axis=0) - tp
    fn = conf_matrix.sum(axis=1) - tp

    precision = tp / (tp + fp + epsilon)
    recall = tp / (tp + fn + epsilon)

    f1_scores = 2 * (precision * recall) / (precision + recall + epsilon)

    return f1_scores

def class_wise_precision_recall(conf_matrix: np.ndarray):
    num_classes = conf_matrix.shape[0]
    precision = np.zeros(num_classes)
    recall = np.zeros(num_classes)

    for i in range(num_classes):
        true_positives = conf_matrix[i, i]
        false_positives = np.sum(conf_matrix[:, i]) - true_positives
        false_negatives = np.sum(conf_matrix[i, :]) - true_positives

        precision[i] = true_positives / (true_positives + false_positives)
        recall[i] = true_positives / (true_positives + false_negatives)

    return precision, recall

def plot_confusion_matrix(conf_matrix, modelname, plot= False):
    labels = {0: "SADNESS", 1: "JOY", 2: "LOVE", 3: "ANGER", 4: "FEAR", 5: "SURPRISE"}
    plt.style.use("seaborn")
    plt.figure(figsize=(7,7))
    norm_conf_mat = np.divide(conf_matrix, np.sum(conf_matrix, axis= 0))
    plt.imshow(norm_conf_mat, cmap='coolwarm', interpolation='nearest')
    plt.colorbar()
    
    plt.xticks(list(labels.keys()), [labels[i] for i in range(len(labels))])
    plt.yticks(list(labels.keys()), [labels[i] for i in range(len(labels))])
    
    fmt = '.2f'
    for i, j in product(range(norm_conf_mat.shape[0]), range(norm_conf_mat.shape[1])):
        plt.text(j, i, format(norm_conf_mat[i, j], fmt),
                 horizontalalignment="center",
                 color="black")
    
    plt.xlabel('Predicted label',fontdict={"fontweight":"bold"})
    plt.ylabel('True label',fontdict={"fontweight":"bold"})
    plt.title('Confusion Matrix',fontdict={"fontsize":15,"fontweight":"bold"})
    plt.grid(False)
    if plot:
        plt.show()
    else:
        if modelname == 'bilstm':
            plt.savefig(fname= "./images/bilstm_confmat.png", dpi=300)
        elif modelname == 'cnn':
            plt.savefig(fname= "./images/cnn_confmat.png", dpi=300)