Exploring Deep Learning Framework PyTorch
summary
Claims that PyTorch gives dynamic computation graphs built at runtime and lets code-run define backprop while TensorFlow uses static graphs defined before run. Demonstrates tensor ops, torch.device and torch.nn.DataParallel for multi-GPU, notes CUDA copies are costly and require torch.cuda.synchronize. Reports serialization tradeoffs: recommended save_state_dict vs pickling for resumes, and ONNX plus planned torch.jit/Caffe2 export for production and mobile. RNNs run faster on PyTorch.
AI-written summary · about
chapters (23)
- 0:00Intro
- 0:36About Stephanie
- 1:24About Algorithm
- 2:12Sign up
- 2:28Agenda
- 3:29About PyTorch
- 5:49Advantages of PyTorch
- 7:19Tensors
- 8:39Computational graphs
- 9:42Dynamic computational graphs
- 11:48AutoGrad
- 15:40Static Graphs
- 16:57Code Example
- 17:26Flow Notes
- 19:49GPU Support
- 22:06GPU Example
- 23:13Moving Content
- 24:35GPU Utilization
- 26:00Serialization
- 29:06Open Neural Network Exchange
- 37:54Promo Code
- 38:54Elevator Speech
- 41:54Karis vs PyTorch
related talks
36mPyTorch: Fast Differentiable Dynamic Graphs in Python
47mGet to Know PyTorch - A Cutting Edge Open Source AI Framework for Deep Learning
32mPyTorch Demystified, Why Did I Switch
22mIncreasing the Impact of AI Through Better Software
54mKeynote: PyTorch 2.1 Technical Deep Dive
49mDynamic Neural Networks using PyTorch
32mDeep Learning Battle: TensorFlow 2 vs. PyTorch
19mGet to know PyTorch - a cutting edge open source AI framework for Machine Learning