It is better to have a comparison of these lasagne,keras,pylearn2,nolearn, tensor and symbolic calculation framework I have chosen to use Theano, the top of the library with which good?
First of all, the document is as detailed as possible, its secondary structure is clear, the inheritance and the invocation is convenient.
Reply content:
Python-based libraries personal favorite is the Keras, for a variety of computing models have a good wrapper, so if you want to implement some standard model is the hand. I talked to Francois, his position on Keras is very clear: Keras is a similar API layer exists, the back-end engine calculation and optimization is separate from the front-end model, so you can change the different backend as needed.
Why is the backend convertible a good thing? Because all Theano-based libraries have a potential problem, that is, the compilation time is very slow, Theano born ... So if the Lord likes Python, let's look at CGT:
Computation Graph Toolkit
Real-time compilation is quick, simply said to have Theano function, no Theano soft rib.
Finally, let me spoof a bit, the Python-based library actually has a decaf
, this is called more thoroughly than PYLEARN2 abandoned building stop development (escape do not know which home the strongest, only know pylearn2 the worst. It took about one months to learn, it was a nightmare, fortunately stopped development. First of all, Pylearn2 can also be ranked among them ...
The main topic is Baidu has a hundreds of years ago, "experience post" it.
Second, the main question is "which library to use." If from the "Loading force" to say that it is necessary to simply get started Theao himself happy to make up DNN is a roar. But the answer is really bad and learning curve giant TM outrageous ...
Then, what lasagne,keras,passage ah, and so a heap of libraries are standing on the shoulders of Theano Jesus. They greatly weaken the difficulty of getting a deep learning algorithm to get started. In other words: a stupid call. In fact, whichever is the same.
Remember this year's Deep Learning Summit in London a good person po past a class comparison of a Python library (searching the original image ...) I remember clearly that lasagne was the first place to be steady. From the level of my current contact, lasagne really with its rigorous framework logic and strong adaptability, and even has captured the European and American a bunch of deep Daniel's hearts. But its methods of naming and invoking various idioms and terminology are different from the genres I come into contact with ... So I don't want to accept it. (simply say that it's documentation and my own gas field does not match ~)
Also, I am a face-looking person ... lasagne homepage that set of free templates. I can't bear to look straight.
So I decided to choose the keras~
At least his homepage has a style theme ~
Years of experience tell me
Yan Value Reliable company quality is not too bad oh ~
Keras's community activity and maintenance is really a bit of a touch to me. Basically issue can do the perfect solution within 24h ... The only problem is that Keras's maintenance is often self-righteous, and the answer is closed issue. But they didn't really solve the problem ... Fortunately with Keras children all know the open and close issue all search a circle ... Otherwise, the real thing is to ask a question to be a second back to a face "silly d you do not understand" and then went to the sink.
The best weapon, of course, is your own "private library". After all, your own project has a lot of customization, especially the Keras itself is very elementary. They pay too much attention to "use for fools", which makes many functions very limited. If you have some fantastic ideas, like a new pipeline for CNN, a pre-treatment for the pictures, and then a CNN. Their existing methods is not to be solved.
So, the best solution is to fork them on GitHub, to open a hanging hanging branch, a hanging hanging name, modify some of their own feel surprised for the new play of Heaven and Man.
The end is no wonder
minute by person F*k.
or change the world in minutes.
Yes, look at the number of the issues on GitHub, and the stars and so on, first go to People's places, with a less suitable and more people, I myself use this way to choose Keras
Mention PYLEARN2 is a failure to do, the main developers admit that the engineering bug all a bunch of nolearn+theano+lasagne you ask questions here, I guess a little bit of mxnet is coming. The problem is recursion until the stack explodes! PYLEARN2 has stopped development, did not pay attention to, if mainly in order to use custom good module, Keras extremely convenient, easy to get started, update frequency is also good, now in addition to Theano also support TensorFlow, there are problems can be asked in keras-users or GitHub ; Lasagne no use, blocks can directly from the written computationgraph () Call Theano.function, so with the Theano write code to use very convenient, but also have attention module (I just see blocks for this, I tried to write with Keras, simply exhausted), but it read the module fuel feel quite complex, now the version is updated to 0.1.1, the configuration environment is more troublesome than Keras, for reference only. Also recommended to pay attention to mxnet, rough tried, video memory occupancy rate is low, compile faster than Theano much faster, but want to implement a custom function compared to Theano-based will be troublesome keras should be the most easy to use the deep learning framework, The pure Python feature makes it easy to read and modify source code, after all, Python is not as easy to know as C + +. The documentation is also quite complete and currently supports Theano+keras and Tensorflow+keras.
Make a small ad, build a discussion of keras and deep learning groups, welcome to add 119427073, welcome to the big God and small white
