Code about adavqa

[guoyang9]

The repo is too old, and even I cannot read the code...
But I can offer you some insights if you're interested.

• Two important tools: Matplotlib and T-SNE (dimension reduction to 2D).

• The data points are from the validation set, where each point corresponds to the outputted feature of the last layer of the network (then use T-SNE to reduce dimension).

• The scatter plot is quite simple, I guess you are more interested in the one lying on a circle. In fact, this one is also easy to obtain. After you have the 2D feature, simply normalize it to a unit vector (so that x2 + y2 = 1, and the point is mapped to a circle).

[darwann]

The repo is too old, and even I cannot read the code... But I can offer you some insights if you're interested.

• Two important tools: Matplotlib and T-SNE (dimension reduction to 2D).
• The data points are from the validation set, where each point corresponds to the outputted feature of the last layer of the network (then use T-SNE to reduce dimension).
• The scatter plot is quite simple, I guess you are more interested in the one lying on a circle. In fact, this one is also easy to obtain. After you have the 2D feature, simply normalize it to a unit vector (so that x_2 + y_2 = 1, and the point is mapped to a circle).

I got it. Thanks for your reply.