Shingome/Image_Processing

1import numpy as np2import matplotlib.pyplot as plt3 4 5# Hyper arguments6INPUT_DIM = 647OUT_DIM = 58H1_DIM = 169H2_DIM = 1010 11ALPHA = 0.00512NUM_EPOCHS = 10013 14# Global15img_map = []16loss_arr = []17 18 19def relu(t):20    return np.maximum(t, 0)21 22 23def softmax(t):24    out = np.exp(t)25    return out / np.sum(out)26 27 28def sparse_cross_entropy(z, y):29    return -np.log(z[0, y])30 31 32def to_full(y, num_classes):33    y_full = np.zeros((1, num_classes))34    y_full[0, y] = 135    return y_full36 37 38def relu_deriv(t):39    return (t >= 0).astype(float)40 41 42def predict(x):43    t1 = x @ W1 + b144    h1 = relu(t1)45    t2 = h1 @ W2 + b246    h2 = relu(t2)47    t3 = h2 @ W3 + b348    z = softmax(t3)49    return z50 51 52def calc_accuracy():53    correct = 054    for y, x in dataset:55        z = predict(x)56        y_pred = np.argmax(z)57        if y_pred == y:58            correct += 159    acc = correct / len(dataset)60    return acc61 62 63def create_map():64    for x in image:65        z = predict(x)66        y_pred = np.argmax(z)67        img_map.append(y_pred)68 69 70if __name__ == "__main__":71    # Load infomation72    dataset = np.load('./../dataset.npy', allow_pickle=True)73 74    image = np.load('./../image.npy')75 76    # Random synapses77    W1 = np.random.rand(INPUT_DIM, H1_DIM)78    b1 = np.random.rand(1, H1_DIM)79    W2 = np.random.rand(H1_DIM, H2_DIM)80    b2 = np.random.rand(1, H2_DIM)81    W3 = np.random.rand(H2_DIM, OUT_DIM)82    b3 = np.random.rand(1, OUT_DIM)83 84    W1 = (W1 - 0.5) * 2 * np.sqrt(1/INPUT_DIM)85    b1 = (b1 - 0.5) * 2 * np.sqrt(1/INPUT_DIM)86    W2 = (W2 - 0.5) * 2 * np.sqrt(1/H1_DIM)87    b2 = (b2 - 0.5) * 2 * np.sqrt(1/H1_DIM)88    W3 = (W3 - 0.5) * 2 * np.sqrt(1/H2_DIM)89    b3 = (b3 - 0.5) * 2 * np.sqrt(1/H2_DIM)90 91    loss = 092 93    # Backpropagation94    for ep in range(NUM_EPOCHS):95        print(ep)96        np.random.shuffle(dataset)97        for i in range(len(dataset)):98 99            x = dataset[i][1]100            y = dataset[i][0]101 102            # Forward103            t1 = x @ W1 + b1104            h1 = relu(t1)105            t2 = h1 @ W2 + b2106            h2 = relu(t2)107            t3 = h2 @ W3 + b3108            z = softmax(t3)109            E = sparse_cross_entropy(z, y)110 111            # Backward112            y_full = to_full(y, OUT_DIM)113            dE_dt3 = z - y_full114            dE_dW3 = h2.T @ dE_dt3115            dE_db3 = np.sum(dE_dt3, axis=0, keepdims=True)116            dE_dh2 = dE_dt3 @ W3.T117            dE_dt2 = dE_dh2 * relu_deriv(t2)118            dE_dW2 = h1.T @ dE_dt2119            dE_db2 = np.sum(dE_dt2, axis=0, keepdims=True)120            dE_dh1 = dE_dt2 @ W2.T121            dE_dt1 = dE_dh1 * relu_deriv(t1)122            dE_dW1 = x.T @ dE_dt1123            dE_db1 = np.sum(dE_dt1, axis=0, keepdims=True)124 125            # Update126            W1 = W1 - ALPHA * dE_dW1127            b1 = b1 - ALPHA * dE_db1128            W2 = W2 - ALPHA * dE_dW2129            b2 = b2 - ALPHA * dE_db2130            W3 = W3 - ALPHA * dE_dW3131            b3 = b3 - ALPHA * dE_db3132 133            loss += E134 135        loss_arr.append(loss)136        loss = 0137 138    # Accuracy139    accuracy = calc_accuracy()140    print("Accuracy:", accuracy)141 142    # Map143    create_map()144    np.save('map', np.asarray(img_map))145 146    # Plot147    plt.plot(loss_arr)148    plt.show()149 150    # Save synapses151    np.savez('./../synapses', W1, b1, W2, b2, W3, b3)152