www.teamblind.com
Dear Google ML, why don't you throw away this TensorFlow shit. | Tech Industry - Blind
Excerpt
Would you please elaborate what do you mean by I am slow. I am using TensorFlow since November 2015 from the day it is launched. I have contacted 100 of users, even the people who developed this library don't like it. ... The abstract syntax tree is great choice for creating distributed model architecture but what about user experience. Why don't they just threw away the static graph execution and adopted the dynamic graph? Why don't they integrated auto differentiation since inception of eager execution. Why this big mess which was created by lots of open source community managed properly. It is a fiasco. It is terrible. It is waste of human resources and people time. What to do with this Tensoflow 1.x and why they are still publishing in TensorFlow 1.x if it is decommissioned. … What exactly is the complaint? Performance? Expressiveness? Ease of use? Backward compatibility? Change in syntax/associated nonperf details? Which of these are bad, and which of these have been affected by 1.x-> 2.x move? I would say all of them, their Data loader is terribly slow. Not at all performance compared to caffe, they keep changing and introducing new API regularly. Horrible debugging and what the hell is tf.keras. why don't just keep one layer API then this all non sense. Keep changing compiler, trying to Integrate their proprietary TPUs. Even they changed from lazy execution to eager execution, under the hood it is still a pile of mess. … TF 2 is backwards compatible at the graph level. TF 2 runs TF 1 graphs I think with frameworks that’s always a problem , tensorflow will keep increasing the functionalities and making it comparable for prod use ... their main goal is to push everyone to use GCP and hence added tensorflow lite and many others for mobile devices and hand held devices.... Pytorch is great for prototyping but still long way to go for prod robustness Google doesnt understand people or maintaining a customer relationship. That’s the issue with only hiring nerds
Source URL
https://www.teamblind.com/post/Dear-Google-ML-why-dont-you-throw-away-this-TensorFlow-shit-AemN78aORelated Pain Points
Poor backward compatibility management across TensorFlow 1.x to 2.x transition
7TensorFlow's transition from 1.x to 2.x involved breaking changes and continued support for deprecated 1.x versions, creating confusion about which version to use and wasting developer time.
Static Computational Graph Rigidity
6TensorFlow's static computational graph model requires developers to define the entire computational graph before execution, which is less flexible than dynamic graph alternatives like PyTorch and challenging for complex, evolving models.
Complex Debugging Mechanisms
5TensorFlow's debugging mechanisms are complex and not straightforward, making it quite tricky to debug code with problems, particularly around sessions and variables management.
Overhead in Data Preprocessing and Loading
5TensorFlow exhibits overhead in data preprocessing and loading operations, creating performance bottlenecks in the overall model training pipeline.
Lack of auto-differentiation integration in early TensorFlow
5Auto differentiation was not integrated from the inception of eager execution in TensorFlow, requiring users to work around this limitation and causing confusion about the framework's capabilities.
Lack of direction and fragmented product vision
5TensorFlow's public face has grown without clear strategic direction. Multiple competing initiatives (XLA, TFDBG, etc.) are announced constantly without cohesion, making it difficult for external developers to understand the intended evolution.