Switched from TensorFlow to PyTorch last month after 3 failed model deployments

Used to fight with TensorFlow's static graphs for hours until a coworker at Google showed me PyTorch's eager execution and now my prototyping time dropped from 2 days to 4 hours - has anyone else noticed a big speed difference when switching frameworks?

2 comments

2 Comments

barbara_taylor8326d ago

Isn't the slower training speed with dynamic graphs a pretty big hidden cost though?

miles27726d ago

Is it REALLY that much slower though? I've seen benchmarks where the difference is like 10-15% at most and for most projects that's nothing. Honestly @barbara_taylor83, you save way more time on debugging and prototyping than you lose on raw speed.