Lets Build Our Own PyTorch Part 3: Writing the Base nn.Module скачать в хорошем качестве

Lets Build Our Own PyTorch Part 3: Writing the Base nn.Module

Every Pytorch module begins with inheriting nn.Module. To make our framework as similar as possible, we will be implementing our own base nn.Module that we will be using for everything else to come! This is not that interesting a video, mostly just housekeeping and taking a look at what is going on inside frameworks like PyTorch! (sorry for the slightly broken voice, had a bit of a sore throat!) Previous Videos: Building an Array Mechanism:    • Lets Build our own PyTorch Part 1: Running...   Autograd From Scratch:    • Lets Build our own PyTorch Part 2: A Fully...   Timestamps: 00:00:00 - Recap what we have done so far 00:06:50 - Modules vs Functionals 00:08:15 - Creating the package structure 00:12:05 - Starting the Module Class 00:15:00 - What is _dict_ in python classes 00:16:45 - Python class inheritance and super() 00:18:35 - Complete the _setattr_ 00:21:00 - Implement .parameters() 00:24:11 - Python Generators 00:26:09 - Accounting for parameter (weight) sharing 00:29:00 - Implement non-deduplicated .parameters() 00:30:30 - Implement normal and non-dedup .named_parameters() 00:36:50 - Implement .register_buffer() 00:38:00 - Persistent vs non-Persistent Buffers 00:39:27 - Add in all other access methods for buffers 00:39:58 - Implement .to() to move tensors btwn devices 00:42:42 - Implement .apply() for future weight init 00:14:15 - Implement _repr_ for pretty printing 00:45:37 - Assign .forward() to __call__() method 00:46:45 - Create the state_dict() method to get weights of model 00:49:20 - Create the load_state_dict() method to load weights into model 00:53:30 - Toggle train vs eval modes 00:54:45 - Implement nn.ModuleList() and nn.Sequential() 01:01:20 - What's next? Socials! X   / data_adventurer   Instagram   / nixielights   Linkedin   / priyammaz   Discord   / discord   🚀 Github: https://github.com/priyammaz 🌐 Website: https://www.priyammazumdar.com/