Which of the following is the best lasagne, keras, pylearn2, and nolearn deep learning libraries?

It is best to compare lasagne, keras, pylearn2, and nolearn. I have already selected theano for tensor and symbolic computing frameworks. Which of the above databases is better? First, the document should be as detailed as possible. Second, the architecture should be clear, and the Inheritance and call should be convenient. It is best to compare lasagne, keras, pylearn2, and nolearn. I have already selected theano for tensor and symbolic computing frameworks. Which of the above databases is better?
First, the document should be as detailed as possible. Second, the architecture should be clear, and the Inheritance and call should be convenient. Reply: Keras is a favorite of various Python-based libraries and has a good wrapper for various computing modes. Therefore, it is very helpful to implement some standard models. I talked to Francois about Keras's clear positioning: Keras is similar to the API layer. The computing and optimization of the backend engine are separated from the front-end model, therefore, you can change the backend as needed.

Why is it a good thing to switch the backend? Because all Theano-based libraries have a potential problem, that is, the Compilation Time is very slow and Theano is born... So if you like Python, consider cgt:
Computation Graph Toolkit
Real-time compilation is very fast. Simply put, there is Theano function, without Theano's weakness.

Finally, let me spoof it. The Python-based library actually has a decaf This is called a more thorough abandon of the building than Pylearn2 to stop Development (do not know which one is the strongest, only know which pylearn2 is the worst. It took me about a month to learn about it. It was a nightmare to stop development. First, Pylearn2 can also rank among them...
The subject is probably Baidu's "experience post" several hundred years ago ..

Second, the subject asks "which database to use ". From the perspective of "installation force", we must get started with Theao and make up DNN. However, the answer is really awkward and the learning curve is huge...

Then, what about lasagne, keras, passage? A bunch of databases are all Jesus standing on theano's shoulders... they greatly reduced the difficulty of getting started with deep learning algorithms .. in other words, it is a silly call. In fact, almost all of them are used.

I remember that this year's Deep Learning summit in London had a great opportunity to get a class comparison chart of the python library (searching for the source image...). I clearly remember that lasagne was the first place to be stable. From the perspective of my current contact, lasagne has indeed captured the hearts of a group of in-depth experts in Europe and America with its rigorous architecture logic and strong adaptability. However, its method naming and calling of various terms and habits and terminology are different from those I have come into contact... so I really don't want to accept it... (In short, its documentation does not match my own gas field ~)

Also, I am a face viewer... the free template on the lasagne homepage...
So I decided to choose Keras ~
At least his homepage has a style topic ~
Years of Experience tell me
The quality of companies with reliable face filter values will not be too bad ~

I was a little touched by the Community activity and maintenance efforts of Keras. Basically, issue can provide a perfect answer within 24 hours... the only question is that keras maintenance staff often fail to understand and often turn off issue after self-thinking answers .. however, they have not actually solved the problem... fortunately, all the children who use keras know to search for open and close issue... otherwise, the real Nima asked a question, and the second turned back to his face, "silly d, you don't understand this", and then let the sea sink.

Of course, the best weapon is your own "Private Database ". After all, you have a lot of customized things for your own projects, especially keras, which is still very basic. They pay too much attention to "use for Dummies", resulting in limited functionality. If you have some whimsy, such as creating a new pipeline for CNN, preprocessing the image, and then sending it to CNN, their existing methods won't solve it.

So, the best solution is to fork them on Github, open a hanging branch, start a hanging name, and modify some new ways of thinking amazing.

The ending is no wonder
F * k in minutes ..
Or change the world in minutes ..
No ~ Look at the statistics on github, such as the number of consumers, issues, and stars. First, let's go where there are too many people. If we don't use them, let's get a little fewer people, I chose keras myself in this way.
It is mentioned that pylearn2 is a loss task, the main developers acknowledge that all the engineering bugs were made by a bunch of nolearn + Theano + lasagne. If you ask this question, I guess mxnet will blow up soon. This problem is recursion, until stack explosion! Pylearn2 has stopped development and won't be noticed. If it is mainly used to use customized modules, keras is extremely convenient and easy to use, and the update frequency is good. Besides theano, tensorflow is now supported, if you have any questions, you can ask them at keras-users or github. lasagne has never used it. blocks can call theano directly from computationgraph. function, which is convenient to use in combination with the code written by theano, and there is also the attention module (I just watched blocks for this purpose, and I tried to write it using keras, which is almost exhausted ), however, the fuel module for reading data is quite complicated. The current version is only updated to 0.1.1. The configuration environment is more complicated than keras, which is for reference only. In addition, it is recommended to take a look at mxnet, which has been roughly tested. The memory usage is low and the compilation speed is much faster than theano, however, it is easier to implement custom functions than to implement theano-based features. keras should be the most easy-to-use deep learning framework. The features of pure python make it very convenient to read and modify the source code, after all, python is more simple than C ++. The documentation is also complete. Currently, Theano + keras and Tensorflow + Keras are supported.
Make a small advertisement and create a group to discuss keras and deep learning. Welcome to add 119427073. Welcome to God and Tom!