Release 1

.gitignore

@@ -1,3 +1,4 @@
 __pycache__
+.ipynb_checkpoints
 .vscode
 *.csv

Host/PerformanceCounter.py

@@ -2,6 +2,10 @@
 
 class PerformanceCounter:
     def __init__(self):
+        self.Reset()
+
+
+    def Reset(self):
         self.Count = 0
         self.LastTick = time.perf_counter()
 

Host/SensorChart.py

@@ -20,19 +20,41 @@ def Run(self, proc):
         GLUT.glutMainLoop()
 
 
-    def Draw(self, data=None):
+    def Draw(self, data=None, data_his=None):
         GL.glClear(GL.GL_COLOR_BUFFER_BIT)
         GL.glBegin(GL.GL_LINES)
 
+        YScale = lambda y: y / 5 * 0.00001 
+
+        # Draw Bars
         if data != None:
             for i in range(len(data)):
                 x = i * 0.1 - 0.9
-                y = 0.75 + data[i] / 4 * 0.00001
+                y = 0.75 + YScale(data[i])
                 c = AxisColor[i % 3]
 
                 GL.glColor3f(c[0], c[1], c[2])
                 GL.glVertex3f(x, 0.75, 0)
                 GL.glVertex3f(x, y, 0)
 
+        # Draw Curves
+        if data_his != None and len(data_his) > 0:
+            size = len(data_his)
+            for i in range(6):
+                c = AxisColor[i % 3]
+                y0 = 0.1 if i > 2 else -0.6
+                GL.glColor3f(c[0], c[1], c[2])
+
+                px = -1
+                py = y0 + YScale(data_his[0][i])   
+                for k in range(size - 1):
+                    x = -1 + (k + 1) *  2 / 120
+                    y =  y0 + YScale(data_his[k + 1][i])
+                    GL.glVertex3f(px, py, 0)
+                    GL.glVertex3f(x, y, 0)
+                    px = x
+                    py = y
+
+
         GL.glEnd()
         GL.glFlush()

Host/SensorProcess.py

@@ -1,23 +1,25 @@
-##
-
 from asyncio.windows_events import NULL
 import socket
+from unittest import skip
 from PerformanceCounter import PerformanceCounter
 from SensorChart import SensorChart
 from DataFile import DataFile
 
-raw_data = []
-
 class UdpServer:
     def __init__(self):
+        self.Sample = []
+        self.Data = []
+        self.Idle = True
+        self.SampleCount = 0
+
         self.Counter = PerformanceCounter()
         self.Chart = SensorChart()
         self.DataFile = DataFile()
 
         self.Server = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
         self.Server.setsockopt(socket.SOL_SOCKET, socket.SO_BROADCAST, 1)
         self.Server.setsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF, 65536 * 16)
-        self.Server.settimeout(1)
+        self.Server.settimeout(0.5)
 
         ip = self.GetLocalIP("192.168.10")
         print(ip)
@@ -30,30 +32,55 @@ def Run(self):
 
 
     def Recv(self):
-        raw_data = [0] * 9
+        self.Data = [0] * 9
 
         try:
-            data, address = self.Server.recvfrom(36)
+            data, address = self.Server.recvfrom(40)
 
             #print(len(data), data)
             reader = lambda p: int.from_bytes(data[p:p + 4], byteorder="little", signed=True)
 
+            c = reader(0)    
             for i in range(9):
-               raw_data[i]  = reader(i * 4)
+               self.Data[i]  = reader((i + 1) * 4)
+
+            if self.Idle: 
+                self.Sample = []
+                self.Idle = False
+
+            while self.AddData(c, self.Data[:]) < c:
+                continue 
+
 
         except socket.timeout:
-            raw_data = []
+            if not self.Idle:
+                if len(self.Sample) == 120:
+                    self.SampleCount += 1
+                    for d in self.Sample:
+                        self.DataFile.Write(d)
+                else:
+                    self.Sample = []
+
+                print("Sample: ", self.SampleCount)
+
             self.DataFile.Close()
+            self.Counter.Reset()
+            self.Idle = True
 
         except Exception as e:
-            raw_data = []
+            self.Data = []
             print("error", e.args)
 
-        [count, t] = self.Counter.Frame()
-        print(count, t, raw_data)
 
-        self.DataFile.Write(raw_data)
-        self.Chart.Draw(raw_data)
+    def AddData(self, c, d):
+            self.Sample.append(d)
+
+            [count, t] = self.Counter.Frame()
+            print(count, int(t * 1000), c, d)
+
+            self.Chart.Draw(d, self.Sample)
+
+            return count
 
 
     def GetLocalIP(self, mask):

Host/tinyml.py

@@ -21,7 +21,7 @@ def CreateModel():
     model.add(layers.Dense(2, activation='softmax'))
 
     opt_adam = keras.optimizers.Adam()
-    model.compile(optimizer=opt_adam, loss='categorical_crossentropy', metrics=['categorical_accuracy'])   
+    model.compile(optimizer=opt_adam, loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])   
     model.summary()
 
     return model
@@ -37,12 +37,14 @@ def PrepareModel(model):
     y_values = np.random.uniform(low=0, high=1, size=(SAMPLES, 2))
     #y_values = np.random.randn(*y_values.shape)
 
-    Train(model, x_values, y_values)
+    return x_values, y_values
 
 
 def TrainModel(model, x, y):
+    SampleCount = len(x)
+
     # split into train, validation, test
-    TRAIN_SPLIT =  int(0.8 * SAMPLES)
+    TRAIN_SPLIT =  int(0.8 * SampleCount)
     x_train, x_validate = np.split(x, [TRAIN_SPLIT, ])
     y_train, y_validate = np.split(y, [TRAIN_SPLIT, ])
 
@@ -52,18 +54,18 @@ def TrainModel(model, x, y):
 def ConvertModel(model):
     converter = tf.lite.TFLiteConverter.from_keras_model(model)
     tflite_model = converter.convert()
+
+    return tflite_model
 
+def SaveModel(model):
     with open("model.tflite", "wb") as f:
-        f.write(tflite_model)
+        f.write(model)
 
 
 # Function: Convert some hex value into an array for C programming
-def Hex2H(model, h_model_name):
-    converter = tf.lite.TFLiteConverter.from_keras_model(model)
-    tflite_model = converter.convert()
-
+def Hex2H(model, h_model_name):   
     c_str = ''
-    model_len = len(tflite_model)
+    model_len = len(model)
 
     # Create header guard
     c_str += '#ifndef ' + h_model_name.upper() + '_H\n'
@@ -75,7 +77,7 @@ def Hex2H(model, h_model_name):
     # Declare C variable
     c_str += 'unsigned char ' + h_model_name + '[] = {'
     hex_array = []
-    for i, val in enumerate(tflite_model) :
+    for i, val in enumerate(model) :
         # Construct string from hex
         hex_str = format(val, '#04x')
 
@@ -93,23 +95,26 @@ def Hex2H(model, h_model_name):
     c_str += '#endif //' + h_model_name.upper() + '_H\n'
 
     # Write TFLite model to a C source (or header) file
-    with open(h_model_name + '.h', 'w') as file:
+    base_path = os.path.dirname(__file__)
+    with open(base_path + "/" + h_model_name + '.h', 'w') as file:
         file.write(c_str)   
+        file.flush()
 
 
-def ReadData(file, v):
+def ReadDataFile(file, v):
     size = SAMPLES_PER_GESTURE * 6
 
     dataX = np.empty([0, size])
-    dataY = np.empty([0, 2])
+    dataY = np.empty([0,])
+    # dataY = np.empty([0, 2])
 
     base_path = os.path.dirname(__file__)
     file = open(base_path + "/data/" + file, "r")
 
     data = []
     for line in file.readlines():
         items = line.split()
-        values = [int(v) for v in items ]
+        values = [int(v) / 32768 for v in items ]
         data += values
 
     count = len(data)
@@ -120,13 +125,32 @@ def ReadData(file, v):
             tmp = np.expand_dims(tmp, axis=0)
 
             dataX = np.concatenate((dataX, tmp), axis=0)
-            dataY = np.concatenate((dataY, [[0, 1]] if v == 0 else [[1, 0]]))
+            dataY = np.append(dataY, v)
+            #dataY = np.concatenate((dataY, [[0, 1]] if v == 0 else [[1, 0]]))
+
+    return dataX, dataY
+
+
+def ReadTrainingData():
+    circle_x, circle_y = ReadDataFile("circle.csv", 1)
+    cross_x, cross_y = ReadDataFile("cross.csv", 0)
+
+    dataX = np.concatenate((circle_x, cross_x), axis=0)
+    dataY = np.concatenate((circle_y, cross_y), axis=0)
 
     return dataX, dataY
 
-ReadData("circle.csv", 1)
+if __name__ == '__main__':
+    Model = CreateModel()
+    #PrepareModel(Model)
+
+    x, y = ReadTrainingData()
+    TrainModel(Model, x, y)
+
+    print("==== Convert Model")
+    tfModel = ConvertModel(Model)
+
+    print("==== Make Header File")
+    Hex2H(tfModel, "model")
 
-Model = CreateModel()
-PrepareModel(Model)
-ConvertModel(Model)
-Hex2H(Model, "model")
+    print("==== Model Build Finished")

README.md

@@ -1,19 +1,39 @@
 # TinyML for Predict Gesture  on ESP32-C3
----
+
 ## 概述
+基于6轴传感器数据采样（加速度计和陀螺仪），识别两种手势动作
 ## 硬件
 ![image](Images/beetle.jpg)
 ![image](Images/mu9250.jpg)
-## 流程
+
 ## 框架
 Tensorflow
 Keras
 Tensorflow Lite Micro
 ## 环境
 ### Python
+Python 3.8
+Tensorflow
+PyOpenGL 
 
 ### Arduino
-ESP32 Arduino
-Tensorflow Lite-ESP32
+ESP32 Arduino 2.0.3
+TensorFlowLite_ESP32 0.9.0
+
+## 设计
+### 数据
+1.2s 数据长度，采样率 100/s
+
+### 模型
+线性全连接网络
+720-向量输入，2-向量输出
+
+## 流程
+数据采集
+预处理
+训练
+推理应用
 
-## 代码
+## 代码
+Host
+prediect_gesture