Calculate Inception Score (IS) using PyTorch
In this video, I tried to explain how 𝐈𝐧𝐜𝐞𝐩𝐭𝐢𝐨𝐧 𝐒𝐜𝐨𝐫𝐞 (𝐈𝐒) can be calculated using PyTorch. Inception score is a metric that is often used for evaluating the quality of synthetic images provided by generative models.
Inception score estimates the quality of a collection of synthetic images based on how well the pretrained 𝙄𝙣𝙘𝙚𝙥𝙩𝙞𝙤𝙣𝙑3 model classifies them as one of 1000 known objects. Inception score combines the confidence of the conditional class distribution (𝐪𝐮𝐚𝐥𝐢𝐭𝐲) and the integral of the marginal class distribution (𝐝𝐢𝐯𝐞𝐫𝐬𝐢𝐭𝐲). In other words, conditional class distribution should have a low entropy and marginal class distribution should have a high entropy.
However, the inception score is limited by what the 𝙄𝙣𝙘𝙚𝙥𝙩𝙞𝙤𝙣𝙑3 model can detect. If the generated images are not present in the model's training data, the score will be low despite generating high-quality images since the image does not get recognized as a distinct class.
𝑮𝒊𝒕𝑯𝒖𝒃 𝒂𝒅𝒅𝒓𝒆𝒔𝒔: https://github.com/randomaccess2023/M...
𝑰𝒎𝒑𝒐𝒓𝒕𝒂𝒏𝒕 𝒕𝒊𝒎𝒆𝒔𝒕𝒂𝒎𝒑𝒔:
00:55 - Import required libraries
05:04 - Define 𝗜𝗻𝗰𝗲𝗽𝘁𝗶𝗼𝗻 function
15:56 - Define 𝗜𝗺𝗮𝗴𝗲𝗣𝗮𝘁𝗵𝗗𝗮𝘁𝗮𝘀𝗲𝘁 class
18:47 - Calculate 𝗜𝗻𝗰𝗲𝗽𝘁𝗶𝗼𝗻 𝗦𝗰𝗼𝗿𝗲
#inceptionscore #generativemodelevaluation #datascience