About the proposals feature cluster visualization

[megvii-research]

About the proposals feature cluster visualization

[Chauncy-Cai]

You can use the code below to visualize the features.

def plot_embedding(data, label, title,show=None):
    # param data:data
    # param label:label
    # param title:title of output
    # param show:(int) if you have too much proposals to draw, you can draw part of them
    # return: tsne-image
    
    if show is not None:
        temp = [i for i in range(len(data))]
        random.shuffle(temp)
        data = data[temp]
        data = data[:show]
        label = torch.tensor(label)[temp]
        label = label[:show]
        label.numpy().tolist()

    x_min, x_max = np.min(data, 0), np.max(data, 0)
    data = (data - x_min) / (x_max - x_min) # norm data
    fig = plt.figure() 

    # go through all the samples
    data = data.tolist()
    label = label.squeeze().tolist()
    
    for i in range(len(data)):
        plt.text(data[i][0], data[i][1], ".",fontsize=18, color=plt.cm.tab20(label[i] / 20))
    plt.title(title, fontsize=14)
    return fig

# weight:(n proposals * 1024-D) input of the classifier
# label: the label of the proposals/ground truth 
# we only select foreground proposals to visualize
# you can try to visualize the weight of different classes by extracting weight during training or testing stage

ts = TSNE(n_components=2,init='pca', random_state=0)
weight = ts.fit_transform(weight)
fig = plot_embedding(weight, label, 't-SNE feature child')
plt.show()

[Chauncy-Cai]

(1) Weight is the feature extract from the input of the classifier (1024-dimension). (I think it should be named feature)
(2) Weight and label are torch.tensor(), if you use another type, remember to convert them.
(3) TSNE is imported from sklearn.manifold, and here is the guild. https://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html. Moreover, perplexity may need to be changed to get a better result.

[chengyu0910]

(1) Weight is the feature extract from the input of the classifier (1024-dimension). (I think it should be named feature)
(2) Weight and label are torch.tensor(), if you use another type, remember to convert them.
(3) TSNE is imported from sklearn.manifold, and here is the guild. https://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html. Moreover, perplexity may need to be changed to get a better result.

Are the perplexitys same for features learned with CPEloss and without CPEloss ?

[Chauncy-Cai]

(1) Weight is the feature extract from the input of the classifier (1024-dimension). (I think it should be named feature)
(2) Weight and label are torch.tensor(), if you use another type, remember to convert them.
(3) TSNE is imported from sklearn.manifold, and here is the guild. https://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html. Moreover, perplexity may need to be changed to get a better result.

Are the perplexitys same for features learned with CPEloss and without CPEloss ?

Yes.
We set perplexity=30, which is the default number recommended by the guild. You can just ignore this parameter.

[chengyu0910]

(1) Weight is the feature extract from the input of the classifier (1024-dimension). (I think it should be named feature)
(2) Weight and label are torch.tensor(), if you use another type, remember to convert them.
(3) TSNE is imported from sklearn.manifold, and here is the guild. https://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html. Moreover, perplexity may need to be changed to get a better result.

Are the perplexitys same for features learned with CPEloss and without CPEloss ?

Yes

Got it. Thanks a lot!

[MrCrightH]

hello, where should this code be added?

[hjfdsssdg]

@chengyu0910
Please ask you a question, I will not have the true value label when I test the model, so how do I pass the label parameter when I draw T-SNE？

[zzzjoey]

@chengyu0910 Please ask you a question, I will not have the true value label when I test the model, so how do I pass the label parameter when I draw T-SNE？

You can use the labels predicted by the model, as the model's predictions are relatively confident in test stage.
Or you can extract features and draw the T-SNE figure during training, where the sample's true label can be determined by the iou-matching with the ground truth bounding box annotations.