1. first version able to run mobilenet

example/mbnetv1/README.md

@@ -0,0 +1,122 @@
+# Maix-EMC Convert  MobileNetV1 Model
+This example explain how to convert TL's mobilenet model to K210's kmodel, and run it on MaixPy environment.
+
+## Model network modify
+mobilenetv1.py is adjusted from tensorlayer's file:
+```
+/usr/local/lib/python3.6/dist-packages/tensorlayer/models/mobilenetv1.py
+```
+we change two things:
+1. TF's default padding is right-bottom, K210 only support  zero around padding, 
+so we need manual add ZeroPad2d Layer before conv layer which stride=(2,2)
+**Note:** after this adjust, we should retrain the network, while I tested the result, orignal weights data still have same result of top5, and top1 probability changed less then 3%, so we skip the retrain step at the moment.
+2. orignal TL's model don't support alpha param(default 1.0), we add it on. 
+It is because K210 only have ~5MB ram for model in C environment, ~3.5MB ram in micropython environment.  
+1.0 version cost 4.2MB memory, not suit for micropython, so we choose 0.75 version, it takes about 2.7MB.
+
+## Get model  weight file
+We use the weights data from **keras**.
+Save_keras_weight.py will save keras mobilenet weights into npz file, store in 'params'.
+You can choose your alpha value to get different model size.
+|  alpha   |  size(MB)   |  Top-1 Accuracy | Top-5 Accuracy |
+| --- | --- | --- | ---|
+|  1.0   |  4.24   | 70.9 | 89.9 |
+|  0.75   |  2.59   | 68.4 | 88.2 |
+|  0.5   |  1.34   | 63.3 | 84.9 |
+|  0.25   |  0.47   | 49.8 | 74.2 |
+
+## Convert TL model to K210's Kmodel
+First you need put some imagenet pictures into **mbnetv1_dataset** folder, it is used for **quantization**.
+Then you just need run mbnetv1_to_kmodel.py to get mobilenet's kmodel.
+In this file, you will see the step:
+1. build the mobilenet model in TL: 
+```
+mobilenetv1 = MobileNetV1(pretrained=True, alpha=0.75)
+```
+2. save the model into kmodel:
+```
+emc.save_kmodel(mobilenetv1, kmodel_name, './mbnetv1_dataset', dataset_func='img_0_1', quant_func='minmax', quant_bit=8, version=3, sm_flag=True)
+```
+API:
+```
+save_kmodel(network, filepath, dataset_dir, dataset_func='img_0_1', quant_func='minmax', quant_bit=8, version=3, sm_flag=False)
+```
+
+|  Parameter   |  Intro |
+| --- | --- | 
+|network|  TL model network|
+|filepath|  output kmodel's filepath|
+|dataset_dir|  the path you store pictures for quantization|
+|dataset_func| the dataset loader function, convert image to 0~1, or -1~1, or 0~255|
+|quant_func|   quantization functions,  choose from minmax, kld|
+|quant_bit|    default 8bit|
+|version| kmodel version, default 3|
+|sm_flag|  add softmax layer at the end.|
+
+## Run kmodel in MaixPy
+We have several kinds of k210 boards, called Maix boards, and we port convenient micropython environment called **MaixPy**.
+[https://github.com/sipeed/MaixPy](https://github.com/sipeed/MaixPy)
+[https://maixpy.sipeed.com/](https://maixpy.sipeed.com/)    
+
+### Burn MaixPy Firmware
+MobileNet kmodel cost about 2.7MB RAM, the “full” MaixPy can’t fit it in, we need the minimal version of MaixPy (strip most openmv function and misc functions)
+**MaixPy Firmware Download **from(please use >0.3.2 version or master version):
+[http://dl.sipeed.com/MAIX/MaixPy/release/](http://dl.sipeed.com/MAIX/MaixPy/release/)
+The **firmware download tool**:
+[https://github.com/sipeed/kflash\_gui](https://github.com/sipeed/kflash_gui)
+
+### Burn kmodel
+And you can burn the **kmodel** you have generate before:  mbnetv1.kfpkg.
+kfpkg is the file zip kmodel and flash-list.json, used by kflash_gui.
+flash-list.json indicate the kmodel address in flash, choose postion in 0x200000~0xa00000  (2~10MB offset)
+We defaultly use 0x200000.
+
+### Add label text
+In addition, we need label list to identify number to 1000 class name.
+Put labels.txt into MicroSD or Flash's filesystem.
+
+### Run MobileNet on MaixPy!
+Open terminal of Serial port(use minicom or somthing),  Press Ctrl+E goto paste mode, and Press Ctrl+D to run the code.
+Or use our MaixPy IDE:
+[http://dl.sipeed.com/MAIX/MaixPy/ide/](http://dl.sipeed.com/MAIX/MaixPy/ide/)
+
+~~~
+import sensor, image, lcd, time
+import KPU as kpu
+lcd.init()
+sensor.reset()
+sensor.set_pixformat(sensor.RGB565)
+sensor.set_framesize(sensor.QVGA)
+sensor.set_windowing((224, 224))
+sensor.set_vflip(1)
+sensor.run(1)
+lcd.clear()
+lcd.draw_string(100,96,"MobileNet Demo")
+lcd.draw_string(100,112,"Loading labels...")
+f=open('labels.txt','r')
+labels=f.readlines()
+f.close()
+task = kpu.load(0x200000) 
+clock = time.clock()
+while(True):
+    img = sensor.snapshot()
+    clock.tick()
+    fmap = kpu.forward(task, img)
+    fps=clock.fps()
+    plist=fmap[:]
+    pmax=max(plist)	
+    max_index=plist.index(pmax)	
+    a = lcd.display(img, oft=(0,0))
+    lcd.draw_string(0, 224, "%.2f:%s                            "%(pmax, labels[max_index].strip()))
+    print(fps)
+a = kpu.deinit(task)
+
+~~~
+
+[![image](https://bbs.sipeed.com/uploads/default/optimized/1X/a9329fc053909faca7f34d2500f4f5ec2d576e50_2_668x500.jpeg)
+Video:
+[https://www.bilibili.com/video/av46664014](https://www.bilibili.com/video/av46664014)
+
+We can see it identify husky picture correctly~  
+And we can see fps in serial terminal is about 26fps.  
+You can make it faster by boost CPU and KPU freq.  

example/mbnetv1/flash-list.json

@@ -0,0 +1,10 @@
+{
+  "version": "0.1.0",
+  "files": [
+    {
+      "address": 0x200000,
+      "bin": "mbnetv1.kmodel",
+      "sha256Prefix": false
+    }
+  ]
+}