A source of inspiration
Very grateful to the landlord, under quarantine for classroom Feedback needs to make the epidemic a large classroom Feedback generator . For the school was forced to write this almost every day of our Feedback can be described as the last straw, click on the link immediately generate a random what they thought of you, both the length of three segments, the number of words from short to long 50-500 is not Wait. Shown below, it is a randomly generated word Feedback 878, also with the back of a large theme class.
It looks good, these no nutrition, no value, then why are so easy to use it? Schools and teachers can fool requirements, saving us a lot of time.
Yet a closer look at what they thought were all generated three sections, but there is no link between any two segments, the first paragraph of the way to prevent the virus, after a while went on a hero, basically very relevance low, and after generating more than a few found some passages are exactly alike. Can be inferred, but integrated manner behind Feedback is randomly selected three segments in the huge corpus and random combinations (combinations may exclude a portion of the read barrier). If there is a strict check re-review is, of course, but it's a lie.
However, if not in the Feedback section of the unit, but in a unit of it? If the result is not the corpus generated random combinations, but the neural network model through training and generate huge corpus of segments of it? Will there be more logical relationship between the sentence and the sentence, will write this after an automatic generator of it once and for all?
On this basis, if by topic of randomly generated, the most typical refueling draft Games, Another example is an application to join the party, and his impressions party school education requirements, leading hypnotic speeches on major rally, moved to China ten VIP thoughts ...... we found this automatic generator can have a role in a variety of occasions.
In addition, last year we project the music is automatically generated by the RNN is not very relevant with this one? Not only generated statements, but also randomly generated music, and even generate random combinations of elements in the picture!
While randomly generated out of the results are basically the people themselves and certainly not created out of text, music, photography comparable, but often we need is not the most beautiful thing that we need to escape, save some on the "write thoughts" first time to better put other work. And at the same time, we will also get inspiration on the product randomly generated for creation.
As Less is More, Copy to Createwhat I want to say, and have been, in most cases, create itself from copy to copy others elements created, replicated everywhere in life elements. And transformation, random combinations of the elements of the last generates its own unique product.
Second, the module
1. Corpus generator
demand:
- The general idea is classified by topic. Some common themes include: Feedback, refueling draft, application, love letters, writing, etc., in the large classification as well as specific small class, a specific scene.
- After clicking on a specific user can specify the classification need to generate a paragraph of words, ranging from 100 to 1000, default 500 words.
- In the next generation of text, you can attach a picture related to the topic at the same time.
Implementation:
- Original method: collect on the network a lot corpus, and label classification, the user selects a topic, random selection and integration is expected to be by segment.
- Neural Networks: RNN trained by an automatically generated model. It can be generated without duplication, while logically associated segment certain semantics. Reference aesthetic sentence generator .
- Additional pictures: advance can be ready for a lot of pictures categorized by subject.
2. Music Generator
demand:
- Users can click on a musical classification: quiet, delicate, cheerful, deep, simple, ornate; or is classified by style: Baroque, classical, romantic, impression, pop, rock and many more.
- The user can specify the length of the music produced, ranging from the 10s to the 60s, 30s default
Implementation:
- Basically small projects in accordance with Andrew Ng teacher depth learning courses on the line, on-line are described in detail: the use of neural network LSTM Music Synthesizer
- However, note that we have chosen a different theme music model is trained, can be classified emotionally, but also from the style, the music scene is divided into several categories, respectively, the model training.
The image generator
demand:
- Users can choose from a range of elements in the collection are not mutually exclusive elements (for example: lawn + cat + afternoon), and finally generate a picture that contains all the selected elements, combinations between elements also need to consider how the most beautiful.
- Support for small categories, such as cats have a black cat, white cat, calico cats.
- The support element is preferably generated in the image translation, scaling, rotation, stretching operation, to ensure real-time interactivity.
Implementation:
- Original method: find a lot of material on the web, label classification, the user selects a topic, random combinations of materials, similar to the corpus generator.
- Neural Networks: Worst GAN image generator , with more randomness. Single elements are generated by way of GAN, to ensure that any two pictures exactly the same elements do not appear.
Third, expand
The following extensions belong extension, the future possible.
1. Automatic reptiles
If the randomly generated not only support, but also keep the material fresh download to update the database on the network the better. Every day a large number of incoming data on the network, if there is a self-crawled the page and classify the elements of automation program works behind the scenes to ensure that the site will be continually fresh and maintainability.
2. Lightweight
Of course, the page itself regardless of the platform, in the mobile terminal, computer terminal, any one operating system can be stable landing. This is exactly the ideal situation, and that is all users need only click on a link, they flew into a web page, does not require much overhead. The purpose of this software is exactly what you want users to be able to use in any occasion.
3. Knowledge Mapping
Like Baidu's search, like Enter the keyword, the search will immediately pop up the corresponding web pages, images. Random generator may provide an input search box based on user needs, to automatically generate the required object. Rather than fixed several options.
4. Functional Integration
Pa for spectral functions matting function and an image of the music may be integrated as random generator which allows the user to upload the elements as the material produced.