The State of Machine Learning Frameworks in 2019 (thegradient.pub)

A 2019 analysis of major ML conferences shows a clear shift: PyTorch surged from minority to majority in research, with 2019 papers showing PyTorch dominant at CVPR (~69%), NAACL/ACL (75%+), and leading at ICLR/ICML. Growth rates were dramatic (e.g., PyTorch up 192–450% across conferences while TensorFlow stagnated or declined in several venues). Despite this research momentum, industry still favors TensorFlow—driven by legacy inertia and production features—so the landscape is split: PyTorch for fast experimentation, TensorFlow for deployment. The technical tug-of-war is narrowing: PyTorch added a JIT and TorchScript (an IR produced by tracing or scripting) to enable static-graph optimization and non‑Python deployment; TensorFlow adopted eager execution (plus tf.function and Autograph) to improve usability. Both approaches trade off: tracing can miss unexecuted control flow, and scripting or autograph requires reinterpretation or restrictions on Python, so converting research code to production remains nontrivial. The broader implication is that researcher preference, tooling maturity (debuggability, API design), and production capability (no‑Python runtimes, mobile via TF Lite, serving infrastructure) will decide future dominance. The contest now hinges on whether TensorFlow 2.0 can match PyTorch’s usability without losing production strengths, and how quickly PyTorch can close remaining gaps for enterprise deployment.