Build a Convolutional AutoEncoder (CAE) using PyTorch - Example with USPS dataset
In this tutorial, I used USPS dataset that consists of digit images of very low resolution (16 x 16 spatial size) to train a Convolutional AutoEncoder (CAE) model. During this process, I tried to reconstruct the original USPS images using a CAE. I used PyTorch as the deep learning framework for making this video and Jupyter Notebook for writing code in Python programming language.
AutoEncoders are basically encoder-decoder modules. In CAE, we need to use convolutional layers in the encoder and transposed convolutional layers in the decoder. We don't necessarily have to flatten the input before training CAE like we have to do in AE. Traditionally, these models are trained with Mean Squared Error (MSE) for calculating the reconstruction loss. A lower reconstruction loss value ensures better prediction that can be made by the model.
00:32 - 𝙄𝙢𝙥𝙤𝙧𝙩 𝙩𝙝𝙚 𝙡𝙞𝙗𝙧𝙖𝙧𝙞𝙚𝙨
02:01 - 𝘿𝙚𝙛𝙞𝙣𝙚 𝙩𝙝𝙚 𝙙𝙚𝙫𝙞𝙘𝙚
04:06 - 𝙇𝙤𝙖𝙙 𝙐𝙎𝙋𝙎 𝙙𝙖𝙩𝙖𝙨𝙚𝙩
07:43 - 𝘿𝙚𝙛𝙞𝙣𝙚 𝙙𝙖𝙩𝙖𝙡𝙤𝙖𝙙𝙚𝙧𝙨
09:09 - 𝘿𝙚𝙛𝙞𝙣𝙚 𝙩𝙝𝙚 𝙣𝙚𝙩𝙬𝙤𝙧𝙠 𝙖𝙧𝙘𝙝𝙞𝙩𝙚𝙘𝙩𝙪𝙧𝙚
15:22 - 𝙑𝙞𝙨𝙪𝙖𝙡𝙞𝙯𝙚 𝙩𝙝𝙚 𝙢𝙤𝙙𝙚𝙡 𝙖𝙧𝙘𝙝𝙞𝙩𝙚𝙘𝙩𝙪𝙧𝙚
17:06 - 𝙏𝙧𝙖𝙞𝙣 𝙗𝙖𝙩𝙘𝙝
18:34 - 𝙏𝙚𝙨𝙩 𝙗𝙖𝙩𝙘𝙝
20:24 - 𝙎𝙚𝙩 𝙪𝙥 𝙢𝙤𝙙𝙚𝙡 (𝘾𝘼𝙀), 𝙘𝙧𝙞𝙩𝙚𝙧𝙞𝙤𝙣 (𝙡𝙤𝙨𝙨 𝙛𝙪𝙣𝙘𝙩𝙞𝙤𝙣) 𝙖𝙣𝙙 𝙤𝙥𝙩𝙞𝙢𝙞𝙯𝙚𝙧
21:36 - 𝙏𝙧𝙖𝙞𝙣 𝙩𝙝𝙚 𝙢𝙤𝙙𝙚𝙡
28:47 - 𝙋𝙡𝙤𝙩 𝙢𝙤𝙙𝙚𝙡 𝙡𝙤𝙨𝙨
30:50 - 𝙈𝙤𝙙𝙚𝙡 𝙥𝙧𝙚𝙙𝙞𝙘𝙩𝙞𝙤𝙣
GitHub address: https://github.com/randomaccess2023/M...
An example with color images: https://analyticsindiamag.com/how-to-...
I made another video of this type using AutoEncoder (AE). Check it out if you are interested: • Build an AutoEncoder (AE) using PyTorch - ...
#deep_learning #unsupervised_learning #pytorch #jupyter_notebook #data_science #usps_dataset #autoencoder_models