Shulou(Shulou.com)06/01 Report--

How to write code that works on both PyTorch and Tensorflow? for this problem, this article introduces the corresponding analysis and solution in detail, hoping to help more partners who want to solve this problem to find a more simple and feasible way.

Foolbox is a Python library built on top of EagerPy. The library is rewritten in EagerPy instead of NumPy for development in PyTorch and TensorFlow, and there is only one code base with no code duplication. Foolbox is a library for adversarial attacks on machine learning models.

The importance of frame irrelevance

In order to solve the differences between frames, the author explores syntactic deviations. In the case of PyTorch, the gradient using In-place requires the use of * * _ grad_ (), while back propagation is called using backward** ().

However, TensorFlow provides an advanced manager and functions like tape.gradient to query gradients. Even at the syntactic level, the two frameworks are very different. For example, for parameters, dim vs axis; for functions, sum vs reduce_sum.

This is where "EagerPy" comes into play. It addresses the difference between PyTorch and TensorFlow by providing a unified API that transparently maps to various underlying frameworks without computing overhead.

"EagerPy allows you to write code that automatically uses PyTorch, TensorFlow, JAX and NumPy."

EagerPy focuses on Eager execution, the researchers write, and its approach is transparent, allowing users to combine framework-independent EagerPy code with framework-specific code.

The similar characteristics of eager execution module and PyTorch introduced by TensorFlow make eager execution mainstream and the framework more similar. However, despite these similarities between PyTorch and TensorFlow2, it is not easy to write framework-independent code. At the grammatical level, the api used for automatic differentiation in these frameworks is different.

Automatic differentiation refers to the algorithm for solving differential equations. Its working principle is chain rule, that is to say, solving the derivative of a function can be reduced to basic mathematical operations (addition, subtraction, multiplication, division). These arithmetic operations can be represented in a graphical format. In particular, EagerPy uses a functional method to automatically distinguish.

Here is a code snippet from the document:

Import eagerpy as epx = ep.astensor (x) def loss_fn (x): # this function accepts and returns an eager tensor return x.square (). Sum () print (loss_fn (x)) # PyTorchTensor (tensor (14.)) print (ep.value_and_grad (loss_fn, x))

First define the first function, and then differentiate according to its input. It is then passed to ep.value_and_grad to get the value of the function and its gradient.

In addition, the norm function can now be used with native tensors and arrays in PyTorch, TensorFlow, JAX, and NumPy with little overhead compared to native code. It also applies to the GPU tensor.

Import torchnorm (torch.tensor ([1,2,3.])) import tensorflow as tfnorm (tf.constant ([1.,2.,3.]))

In summary, EagerPy is designed to provide the following features:

Provides a unified API for rapid execution

Maintain the native performance of the framework

Fully linkable API

Comprehensive type checking support

The researchers claim that these attributes make it easier and more secure to use them than the underlying framework-specific api. Despite these changes and improvements, the team behind EagerPy ensured that eagerpy API followed the standards set by NumPy, PyTorch, and JAX.

Getting started with EagerPy:

Use pip to install the latest version from PyPI:

Python3-m pip install eagerpyimport eagerpy as epdef norm (x): X = ep.astensor (x) result = x.square (). Sum (). Sqrt () return result.raw 's answer to the question on how to write code that works on both PyTorch and Tensorflow is shared here. I hope the above content can be of some help to you. If you still have a lot of doubts to solve, you can follow the industry information channel for more related knowledge.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.