What needs improvement with TensorFlow?

Providing more control by allowing users to build custom functions would make TensorFlow a better option. It currently offers inbuilt functions, however, having the ability to implement custom libraries would enhance its usefulness for enterprise-level applications. The process of creating models could be more user-friendly. We encountered version mismatch errors while using the product. It sometimes does not integrate well with other libraries in Python, which can be problematic. Additionally, it can be less intuitive when creating neural networks than PyTorch. The versatility of the concept is undeniable, but it can pose a challenge for developers unfamiliar with machine learning. For newcomers to the field, the learning curve can be steep, often requiring about a year of dedicated effort. Real-time capability needs enhancement, particularly in dynamic environments. Additionally, community support is crucial, and it should be more robust and accessible for remote assistance. … There are a lot of problems, such as integrating our custom code. In my experience model tuning has been a bit difficult to edit and tune the graph model for best performance. We have to go into the model but we do not have a model viewer for quick access. There should be better integration and standardization with different operating systems. We need to always convert from one model to another and there is not a single standardized model output that we could use on different platforms, such as Intel x56, x64 based, AR-based, or Apple M1 chips. … I tend to find it to be a bit too much orientated to AI itself for other use cases, which is fine — that's what it's designed for. Personally, I find it to be a bit too much AI-oriented. Overall, the solution has been quite helpful. I can't recall missing any features when I was using it. I know this is out of the scope of TensorFlow, however, every time I've sent a request, I had to renew the model into RAM and they didn't make that prediction or inference. This makes the point for the request that much longer. If they could provide anything to help in this part, it will be very great. … In terms of improvement, we always look for ways they can optimize the model, accelerate the speed and the accuracy, and how can we optimize with our different techniques. There are various techniques available in TensorFlow. Maintaining accuracy is an area they should work on. When there are more and more objects involved with the model, the models get confused. So maintaining the accuracy and speed with the number of classes is the biggest area for improvement. It is a major challenge that we are seeing right now and we are trying to solve the problem. … TensorFlow is primarily geared towards Python community at present. JavaScript is a different thing and all the websites, web apps and all the mobile apps are built-in JavaScript. JavaScript is the core of that. What can be improved with TensorFlow is how it can mix in. How the JavaScript developers can use TensorFlow. There's a huge gap currently. If you are a web developer, then using Machine Learning with TF is not as straightforward as using a regular Javascript library by reading its documentation. TensorFlow should provide a way to do that easily. … But even while using TensorFlow, it uses two threads. For example, if I have a high powered CPU, I cannot use it. For example with my laptop, I have a high-powered CPU and I'm using Ubuntu, but my GPU is not recognized. So I can use the CPU, but it's not really optimized for this purpose. Huge calculations require GPU's. I think that could be the second thing that could be optimized. I think TensorFlow 2 has huge improvements over TensorFlow 1. However, it would be really nice if we can actually somehow use the code written in TensorFlow 1, to incorporate it into TensorFlow 2. It generates a lot of errors and you have to change a lot of code and settings. What we can optimize is to actually have consistency between the versions. So TensorFlow 2 is actually a different product, to TensorFlow 1. It doesn't allow for fast the proto-typing. So usually when we do proto-typing we will start with PyTorch and then once we have a good model that we trust, we convert it into TensorFlow. So definitely, TensorFlow is not very flexible.