ML frameworks in 2019: analysis of AI research papers shows TensorFlow is the platform of choice in industry, but most researchers are now using PyTorch
Since deep learning regained prominence in 2012, many machine learning frameworks have clamored to become the new favorite among researchers and industry practitioners. Tweets: @biztechpm , @m_galtier , @gradientpub , @schrep , @reiinakano , @danilbaibak , @m_deff , and @xamat Tweets: Roger Bullivant / @biztechpm : Very good analyses of PyTorch, Tensorflow and the future of ML frameworks by @cHHillee on @gradientpub. The conclusion: Pytorch now owns research due to ease of use, while TF owns industry (i.e. mobile & server) but both have room for improvement. https://thegradient.pub/... pic.twitter.com/lAlI3nhrjd Mathieu Galtier / @m_galtier : An expert blog post about the dynamics of ML frameworks: tensorflow is stalling while pytorch is rising sharply in particular for research. I would never have guessed that a year ago! A proof that nice APIs can drive powerful change https://thegradient.pub/... @gradientpub : The war between ML frameworks has raged on since the rebirth of deep learning. Who is winning? @cHHillee's data analysis shows clear trends: PyTorch is winning dramatically among researchers, while Tensorflow still dominates industry. #PyTorch #Tensorflow https://thegradient.pub/... Mike Schroepfer / @schrep : PyTorch long ago became the tool of choice for AI research at fb. It's now the tool of choice for production as well. https://thegradient.pub/... Reiichiro Nakano / @reiinakano : https://thegradient.pub/... ""ML frameworks don't just enable machine learning research, they enable and restrict the ideas that researchers are able to easily explore."" Perhaps the next big step forward is something completely inexpressible in the TF/PyTorch paradigm. Danylo Baibak / @danilbaibak : #TensorFlow vs #PyTorch - In research PyTorch >> TF - For small teams PyTorch >> TF - For fast product delivery PyTorch >> TF - For corporations TF > PyTorch - For edge computing/mobile TF > PyTorch - For production in general, soon PyTorch ~ TF https://thegradient.pub/... Michal Defferrard / @m_deff : TF has lost the battle for researchers a while ago. Only latency keeps it around. Industry will inevitably follow (building missing PyTorch prod tools). https://twitter.com/... Xavier Amatriain / @xamat : “My analysis suggests that researchers are abandoning TensorFlow and flocking to PyTorch in droves. Meanwhile in industry, Tensorflow is currently the platform of choice, but that may not be true for long.” https://thegradient.pub/...
Context & Ripple Effects
The 2019 framework split documented by The Gradient — TensorFlow owning industry deployment, PyTorch owning the research bench — was the pivot point of a contest that had run since deep learning's 2012 resurgence. The prior coverage framed ML as a potential ubiquitous enabling layer on the scale of relational databases, which is why control of the default framework mattered beyond developer convenience.
First-order effects
- Google's TensorFlow held the production stack — mobile and server — but the paper analysis showed new ML graduates and labs standardizing on PyTorch, meaning Google was winning deployments while losing the talent pipeline that decides the next generation of deployments.
- Facebook's PyTorch gained de facto control of research mindshare, converting ease-of-use into an ecosystem advantage that compounds with every paper published on it.
Second-order effects
- Research-first adoption converted into production migration, most visibly when OpenAI announced it would migrate to PyTorch for future projects, pulling a flagship lab off TensorFlow and pressuring other industry users to justify staying.
- The split forced framework strategy to split along the research-to-production boundary, with Google leaning on its deployment footprint while Meta's framework rode research momentum toward industry.
Third-order effects
- PyTorch's reliance on Nvidia's CUDA tied framework dominance to hardware, and the decade retrospective records both PyTorch overtaking TensorFlow and the erosion of the CUDA monopoly — framework and chip control becoming intertwined layers of the ML stack.
- If the pattern holds, the durable position in ML software is whoever owns the deployment and inference layer, not the training framework — which is the battleground the later coverage of Google, AWS, Microsoft, Nvidia, Meta and Databricks integration strategies maps out in detail.
The trend: ML framework competition is resolving through the research-to-production pipeline: research mindshare converts into industry standardization over a multi-year lag, with framework choice increasingly entangled with hardware and inference control.