3 Questions on transfer learning

Hi everyone,

I’ve just watched the transfer learning class, and 3 question arose in my head:

  1. Once we start a project on image recognition, where can we start searching for reliable pre-trained weights?

  2. How do we implement it in code? I’ve understand the logic of how pre-train works, but we haven’t been told how to use those weights that we can download, on our data.

  3. Besides the pre-trained items being of the same format (images, audio, etc.), do they need to be the same size (in case of images), too? Or we can use pre-trained weights for images of sizes 100x100 pixels in our 50x50 images? (I’m concerned with the first layer of the NN, as it will be of different size!)

Can someone help me figuring out these topics?


Hello @AlvaroViudez,

For your Q1 and Q2, please check out this post which has the link to a tutorial for transfer learning. The tutorial has a notebook which initiate a pre-trained model for learning and prediction.

For your Q3, the image size has to follow the requirement by the pre-trained network. if the network is designed for 100 x 100 images, then you need to resize your 50 x 50 images to become 100 x 100.



That’s great! Thanks @rmwkwok ! :slight_smile:

1 Like

You’re welcome @AlvaroViudez :slight_smile: