Create readme.md

readme.md

@@ -0,0 +1,117 @@
+![ShowAndTell](https://github.com/litleCarl/ShowAndTell/blob/master/DemoImages/showAndTell.png)
+
+<p align="center">
+<a href="https://developer.apple.com/swift"><img src="https://img.shields.io/badge/language-swift4-f48041.svg?style=flat"></a>
+<a href="https://developer.apple.com/ios"><img src="https://img.shields.io/badge/platform-iOS%2011%2B-blue.svg?style=flat"></a>
+<a href="https://github.com/wxxsw/GSMessages/tree/1.0.0"><img src="https://img.shields.io/badge/release-1.0.0-blue.svg"></a>
+</p>
+
+# Show and Tell: A Neural Image Caption Generator 
+
+## Brief
+
+***Pull requests and issues:*** 
+@litleCarl
+
+A CoreML implementation of the image-to-text model described in the paper:
+
+"Show and Tell: Lessons learned from the 2015 MSCOCO Image Captioning
+Challenge."
+
+Oriol Vinyals, Alexander Toshev, Samy Bengio, Dumitru Erhan.
+
+*IEEE transactions on pattern analysis and machine intelligence (2016).*
+
+Full text available at: http://arxiv.org/abs/1609.06647
+
+## Demo
+<img src="https://github.com/LitleCarl/ShowAndTell/blob/master/DemoImages/demo_1.png" width="150" ><img src="https://github.com/LitleCarl/ShowAndTell/blob/master/DemoImages/demo_2.png" width="150" ><img src="https://github.com/LitleCarl/ShowAndTell/blob/master/DemoImages/demo_3.png" width="150" ><img src="https://github.com/LitleCarl/ShowAndTell/blob/master/DemoImages/demo_4.png" width="150" >
+
+
+## Usage
+
+### Simple use
+```Swift
+let showAndTell = ShowAndTell()
+let results = showAndTell.predict(image: uiimage2predict, beamSize: 3, maxWordNumber: 30)
+```
+
+
+
+```Swift
+// Parameter explaination
+//    image:         The image to be used to generate the caption.
+//    beamSize:      Max caption count in result to be reserved in beam search.(Affect the performance greatly)
+//    maxWordNumber: Max number of words in a sentence to be predicted.
+class ShowAndTell {
+  ...
+  func predict(image: UIImage, beamSize: Int = 3, maxWordNumber:Int = 20) -> PriorityQueue<Caption>
+  ...
+}
+```
+
+## Benchmark (Tested on iPhone 7+, welcome pr for more decices)
+<table>
+<tr><th>maxWordNumber = 20 </th><th>maxWordNumber = 30</th></tr>
+<tr><td>
+
+beamSize | Time (ms)
+---- | ---
+1  | 480.12
+2  | 845.78
+3  | 1443.82
+4  | 2001.30
+5  | 2648.48
+6  | 3158.53
+7  | 4179.14
+8  | 4861.66
+9  | 6003.65
+10 | 7087.97
+11 | 8134.95
+12 | 9627.79
+
+</td><td>
+
+beamSize | Time (ms)
+---- | ---
+1 | 451.12
+2 | 1194.65
+3 | 1965.27
+4 | 2971.92
+5 | 3798.28
+6 | 4391.35
+7 | 5714.87
+8 | 6937.60
+9 | 8482.03
+10 | 10421.52
+11 | 12460.80
+12 | 13777.67
+</td></tr> </table>
+
+### Line chart for Time vs Beam Size (When `maxWordNumber = 30`)
+<img src="https://github.com/LitleCarl/ShowAndTell/blob/master/DemoImages/chart_of_beam_size" >
+
+## Original Model
+This coreml model is exported from keras which is trained with MSCOCO dataset for about 40k epoches. And presently it is not in the state of art yet. You may not use this in production.
+I trained the dataset with only one GTX Force 1080Ti for about 48 hours and currently don't have more time to train on it.Hope for community to keep on it.
+
+## Requirements
+- iOS 11.0+
+- Xcode 9.0+ (Swift 4.x)
+
+## TODO 
+- Train on the dataset to 100k epoches. (currently 40k)
+- Open source origin model based on Keras which is trained with.
+
+## Thanks for third party lib in demo
+- [GSMessages](https://github.com/wxxsw/GSMessages)
+- [Swift-PriorityQueue](https://github.com/Bouke/Swift-PriorityQueue/)
+
+## Contact
+- 曹佳鑫 （tsao）An iOS developer with experience in deep learning living in Shanghai.
+- Pull requests and issues are welcome.
+- Mail: [email protected]
+
+## License
+
+ShowAndTell is available under the MIT license. See the LICENSE file for more info.