Function plot_distribution

src/utils.py:323–394

Visualize the feature X in the 2-D space Params: - X: a feature matrix has dims (num_samples, hidden_dims) - Y: a label list has dims (num_samples) - label_list: a list has dims (num_classes) and it represents the name of each class

(X, Y, label_list, class_center_matrix=None, sample_ratio=1.0, select_labels=None)

                size=15)

def plot_distribution(X, Y, label_list, class_center_matrix=None, sample_ratio=1.0, select_labels=None):
    '''
        Visualize the feature X in the 2-D space

        Params:
            - X: a feature matrix has dims (num_samples, hidden_dims)
            - Y: a label list has dims (num_samples)
            - label_list: a list has dims (num_classes) 
            and it represents the name of each class
            - class_center_matrix: if not None, plot the class center of each class;
            it has dims (num_classes, hidden_dims)
            - sample_ratio: the ratio of the samples used for visualization
            - select_labels: a list represents the selected labels for visualization
    '''
    # clone and convert to tensor
    if isinstance(X, list):
        _X = torch.tensor(X)
    else:
        _X = X.clone().detach().cpu()
    if isinstance(Y, list):
        _Y = torch.tensor(Y)
    else:
        _Y = Y.clone().detach().cpu()
    num_samples = _Y.shape[0]
    print('Total %d samples for visualization'%num_samples)

    # random sampling
    if sample_ratio<1.0:
        assert sample_ratio>0.0, "Invalid sample ratio!!!"
        
        sample_lst = list(range(num_samples))
        random.shuffle(sample_lst)
        sample_lst = sample_lst[:int(num_samples*sample_ratio)]
        _X = _X[sample_lst]
        _Y = _Y[sample_lst]
        print('Select %d samples for visualization'%_Y.shape[0])

    if select_labels!=None and len(select_labels)>0:
        for i,l in enumerate(select_labels):
            if i==0:
                class_mask = np.equal(_Y,l)
            else:
                class_mask = np.logical_or(class_mask,np.equal(_Y,l))
        _Y = _Y[class_mask]
        _X = _X[class_mask]
    
    # t-SNE for visualization
    tsne = TSNE(n_components=2)
    if not class_center_matrix is None:
        assert len(label_list)==class_center_matrix.shape[0], "Number of classes is not consistent!!!"
        num_class = class_center_matrix.shape[0]
        concat_X = torch.cat((_X, class_center_matrix),dim=0)
        concat_low_repre = torch.tensor(tsne.fit_transform(concat_X))

        # scale to 0-1
        x_min, x_max = torch.min(concat_low_repre, 0)[0], torch.max(concat_low_repre, 0)[0]
        concat_low_repre = (concat_low_repre - x_min) / (x_max - x_min)