Function visualize

examples/SimilarityLearning/mnist-embeddings.py:362–410

(model_path, model, algo_name)

def visualize(model_path, model, algo_name):
    if not MATPLOTLIB_AVAIBLABLE:
        logger.error("visualize requires matplotlib package ...")
        return
    pred = OfflinePredictor(PredictConfig(
        session_init=SmartInit(model_path),
        model=model(),
        input_names=['input'],
        output_names=['emb']))

    NUM_BATCHES = 6
    BATCH_SIZE = 128
    images = np.zeros((BATCH_SIZE * NUM_BATCHES, 28, 28))  # the used digits
    embed = np.zeros((BATCH_SIZE * NUM_BATCHES, 2))  # the actual embeddings in 2-d

    # get only the embedding model data (MNIST test)
    ds = get_test_data()
    ds.reset_state()

    for offset, dp in enumerate(ds):
        digit, label = dp
        prediction = pred(digit)[0]
        embed[offset * BATCH_SIZE:offset * BATCH_SIZE + BATCH_SIZE, ...] = prediction
        images[offset * BATCH_SIZE:offset * BATCH_SIZE + BATCH_SIZE, ...] = digit
        offset += 1
        if offset == NUM_BATCHES:
            break

    plt.figure()
    ax = plt.subplot(111)
    ax_min = np.min(embed, 0)
    ax_max = np.max(embed, 0)

    ax_dist_sq = np.sum((ax_max - ax_min)**2)
    ax.axis('off')
    shown_images = np.array([[1., 1.]])
    for i in range(embed.shape[0]):
        dist = np.sum((embed[i] - shown_images)**2, 1)
        if np.min(dist) < 3e-4 * ax_dist_sq:     # don't show points that are too close
            continue
        shown_images = np.r_[shown_images, [embed[i]]]
        imagebox = offsetbox.AnnotationBbox(offsetbox.OffsetImage(np.reshape(images[i, ...], [28, 28]),
                                            zoom=0.6, cmap=plt.cm.gray_r), xy=embed[i], frameon=False)
        ax.add_artist(imagebox)

    plt.axis([ax_min[0], ax_max[0], ax_min[1], ax_max[1]])
    plt.xticks([]), plt.yticks([])
    plt.title('Embedding using %s-loss' % algo_name)
    plt.savefig('%s.jpg' % algo_name)


if __name__ == '__main__':