[Python:Refractor] Add Image train Demo and move dirs

pymnn/INSTALL.md

@@ -0,0 +1,10 @@
+# PC 上安装
+仅支持python3
+
+1、进入pymnn/pip_package 目录
+2、python3 build_deps.py
+3、sudo python3 setup.py install
+
+# 示例
+1、v2接口/推理：pymnn/examples/MNNEngineDemo
+2、训练/ 压缩：pymnn/examples/MNNTrain

pymnn/examples/MNNTrain/Image/mobilenetCls4.py

@@ -0,0 +1,85 @@
+import MNN
+import MNN.var as var
+c_train = MNN.c_train
+nn = c_train.cnn
+F = MNN.expr
+data = c_train.data
+import time
+
+import sys
+class Net(MNN.train.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+        modelFile = sys.argv[1]
+        print(modelFile)
+        varMap = F.load_dict(modelFile)
+        inputVar = varMap['input']
+        outputVar = varMap['MobilenetV2/Logits/AvgPool']
+        self.net = c_train.load_module([inputVar], [outputVar], True)
+        self.fc = nn.conv(1280, 4, [1, 1])
+    def forward(self, x):
+        x = self.net(x)
+        x = self.fc(x)
+        x = F.softmax(F.reshape(F.convert(x, F.NCHW), [0, -1]))
+        return x
+
+scale = [0.00784314, 0.00784314, 0.00784314, 0.00784314]
+mean = [127.5, 127.5, 127.5, 0]
+
+imageConfig = data.image.config(MNN.cv.BGR, 224, 224, scale, mean, [1.0, 1.0], False)
+picturePath = sys.argv[2]
+print(picturePath)
+txtPath = sys.argv[3]
+imageDataset = data.image.image_label(picturePath, txtPath, imageConfig, False)
+imageLoader = imageDataset.create_loader(10, True, True, 0)
+
+def trainFunc(loader, net, opt):
+    loader.reset()
+    net.train(True)
+    t0 = time.time()
+    iter_number = loader.iter_number()
+    for i in range(0, iter_number):
+        example = loader.next()[0]
+        data = example[0][0]
+        label = F.reshape(example[1][0], [-1])
+        data = F.convert(data, F.NC4HW4)
+        predict = net(data)
+        target = F.one_hot(F.cast(label, F.int), var.int(4), var.float(1.0), var.float(0.0))
+        loss = c_train.loss.CrossEntropy(predict, target)
+        if i % 10 == 0:
+            print(i, loss.read(), iter_number)
+        opt.step(loss)
+    t1 = time.time()
+    cost = t1 - t0
+    print("Epoch cost: %.3f" %cost)
+    F.save(net.parameters(), "cache/temp.snapshot")
+
+def testFunc(loader, net):
+    loader.reset()
+    net.train(False)
+    iter_number = loader.iter_number()
+    correct = 0
+    for i in range(0, iter_number):
+        example = loader.next()[0]
+        data = example[0][0]
+        label = F.reshape(example[1][0], [-1])
+        data = F.convert(data, F.NC4HW4)
+        predict = net(data)
+        predict = F.argmax(predict, 1)
+        accu = F.reduce_sum(F.equal(predict, F.cast(label, F.int)), [], False)
+        correct += accu.read()[0]
+    print("Accu: ", correct * 100.0 / loader.size(), "%")
+
+net = Net()
+net.loadParameters(F.load("cache/temp.snapshot"))
+opt = c_train.SGD(0.0001, 0.9);
+opt.append(net.parameters())
+F.setThreadNumber(4)
+testTxt = sys.argv[4]
+testDataset = data.image.image_label(picturePath, testTxt, imageConfig, False)
+testLoader = testDataset.create_loader(10, True, False, 0)
+
+for epoch in range(0, 10):
+    testFunc(testLoader, net)
+    trainFunc(imageLoader, net, opt)
+

pymnn/examples/MNNTrain/Image/quanByMSE.py

@@ -0,0 +1,65 @@
+import MNN
+import MNN.var as var
+c_train = MNN.c_train
+nn = c_train.cnn
+F = MNN.expr
+data = c_train.data
+import time
+
+import sys
+modelFile = sys.argv[1]
+print(modelFile)
+
+varMap = F.load_dict(modelFile)
+inputVar = varMap['sub_7']
+outputVar = varMap['ResizeBilinear_3']
+net = c_train.load_module([inputVar], [outputVar], True)
+c_train.compress.quantize(net, 8,  c_train.compress.PerChannel, c_train.compress.MovingAverage)
+checkNet = c_train.load_module([inputVar], [outputVar], False)
+
+scale = [0.00784314, 0.00784314, 0.00784314, 0.00784314]
+mean = [127.5, 127.5, 127.5, 0]
+
+imageConfig = data.image.config(MNN.cv.BGR, 257, 257, scale, mean, [1.0, 1.0], False)
+picturePath = sys.argv[2]
+print(picturePath)
+imageDataset = data.image.image_no_label(picturePath, imageConfig)
+imageLoader = imageDataset.create_loader(5, True, True, 0)
+
+def trainFunc(loader, net, checkNet, opt):
+    loader.reset()
+    net.train(True)
+    t0 = time.time()
+    iter_number = loader.iter_number()
+    for i in range(0, iter_number):
+        example = loader.next()[0]
+        data = example[0][0]
+        data = F.convert(data, F.NC4HW4)
+        p0 = net(data)
+        p1 = checkNet(data)
+        p0 = F.reshape(F.convert(p0, F.NCHW), [0, -1])
+        p1 = F.reshape(F.convert(p1, F.NCHW), [0, -1])
+        loss = c_train.loss.MSE(p0, p1)
+        opt.step(loss)
+        if i % 10 == 0:
+            print(loss.read())
+    t1 = time.time()
+    cost = t1 - t0
+    print("Epoch cost: %.3f" %cost)
+    F.save(net.parameters(), "cache/temp.snapshot")
+
+
+opt = c_train.SGD(0.000000000, 0.9);
+opt.append(net.parameters())
+
+for epoch in range(0, 1):
+    trainFunc(imageLoader, net, checkNet, opt)
+
+net.train(False)
+testInput = F.placeholder([1, 3, 257, 257], F.NC4HW4)
+testInput.set_name("data")
+testOutput = net(testInput)
+testOutput.set_name("prob");
+quanName = "temp.quan.mnn"
+print("Save to " + quanName)
+F.save([testOutput], quanName)

pymnn/examples/MNNTrain/Mnist/learn.py

@@ -1,74 +1,65 @@
-import MNN.train as train
-import MNNPy.train
-import MNN.train.cnn as nn
-import MNN.expr as F
+import MNN
+import MNN.var as var
+c_train = MNN.c_train
+nn = c_train.cnn
+F = MNN.expr
+data = c_train.data
 import time
-import MNN.train.data as data
 
-class Net(MNNPy.train.Module):
+class Net(MNN.train.Module):
     def __init__(self):
         super(Net, self).__init__()
-        self.conv1 = nn.Conv(1, 20, [5, 5])
-        self.conv2 = nn.Conv(20, 50, [5, 5])
-        self.fc1 = nn.Linear(800, 500)
-        self.fc2 = nn.Linear(500, 10)
+        self.conv1 = nn.conv(1, 20, [5, 5])
+        self.conv2 = nn.conv(20, 50, [5, 5])
+        self.fc1 = nn.linear(800, 500)
+        self.fc2 = nn.linear(500, 10)
 
     def forward(self, x):
-        x = F.Relu(self.conv1(x))
-        x = F.MaxPool(x, [2, 2], [2, 2])
-        x = F.Relu(self.conv2(x))
-        x = F.MaxPool(x, [2, 2], [2, 2])
-        x = F.Convert(x, F.NCHW)
-        x = F.Reshape(x, [0, -1])
-        x = F.Relu(self.fc1(x))
+        x = F.relu(self.conv1(x))
+        x = F.max_pool(x, [2, 2], [2, 2])
+        x = F.relu(self.conv2(x))
+        x = F.max_pool(x, [2, 2], [2, 2])
+        x = F.convert(x, F.NCHW)
+        x = F.reshape(x, [0, -1])
+        x = F.relu(self.fc1(x))
         x = self.fc2(x)
-        x = F.Softmax(x, 1)
+        x = F.softmax(x, 1)
         return x
 
 
-def initFloat(value):
-    res = F.Input([], F.NCHW, F.float)
-    res.write([value])
-    res.fix(F.Const)
-    return res
-def initInt(value):
-    res = F.Input([], F.NCHW, F.int)
-    res.write([value])
-    res.fix(F.Const)
-    return res
 
 def testFunc(loader, net):
     loader.reset()
     net.train(False)
-    iterNumber = loader.iterNumber()
+    iter_number = loader.iter_number()
     correct = 0
-    for i in range(0, iterNumber):
+    for i in range(0, iter_number):
         example = loader.next()[0]
         data = example[0][0]
         label = example[1][0]
 
-        data = F.Multiply(F.Cast(data, F.float), initFloat(1.0/255.0))
+        data = F.cast(data, F.float) * var.float(1.0/255.0)
         predict = net(data)
-        predict = F.ArgMax(predict, 1)
-        accu = F.ReduceSum(F.Equal(predict, F.Cast(label, F.int)), [], False)
+        predict = F.argmax(predict, 1)
+        accu = F.reduce_sum(F.equal(predict, F.cast(label, F.int)), [], False)
         correct += accu.read()[0]
-    print(correct * 1.0 / loader.size())
+    print("Accu: ", correct * 100.0 / loader.size(), "%")
 
 
 def trainFunc(loader, net, opt):
     loader.reset()
     net.train()
     t0 = time.time()
-    iterNumber = loader.iterNumber()
-    for i in range(0, iterNumber):
+    iter_number = loader.iter_number()
+    for i in range(0, iter_number):
         example = loader.next()[0]
         data = example[0][0]
         label = example[1][0]
 
-        data = F.Multiply(F.Cast(data, F.float), initFloat(1.0/255.0))
+        data = F.cast(data, F.float) * var.float(1.0/255.0)
         predict = net(data)
-        target = F.OneHot(F.Cast(label, F.int), initInt(10), initFloat(1.0), initFloat(0.0))
-        loss = train.loss.CrossEntropy(predict, target)
+        target = F.one_hot(F.cast(label, F.int), var.int(10), var.float(1.0), var.float(0.0))
+        loss = c_train.loss.CrossEntropy(predict, target)
         opt.step(loss)
         if i % 100 == 0:
             print(loss.read())
@@ -79,14 +70,17 @@ def trainFunc(loader, net, opt):
 
 
 net = Net()
-opt = train.SGD(0.01, 0.9)
+opt = c_train.SGD(0.01, 0.9)
 net.loadParameters(F.load("cache/temp.snapshot"))
 opt.append(net.parameters())
 
-mnistDataset = data.mnist.create("/Users/jiangxiaotang/data/mnist", data.mnist.Train)
-trainLoader = mnistDataset.createLoader(64, True, True, 0)
-testmnistDataset = data.mnist.create("/Users/jiangxiaotang/data/mnist", data.mnist.Test)
-testLoader = mnistDataset.createLoader(10, True, False, 0)
+import sys
+mnistDataPath = sys.argv[1]
+
+mnistDataset = data.mnist.create(mnistDataPath, data.mnist.Train)
+trainLoader = mnistDataset.create_loader(64, True, True, 0)
+testmnistDataset = data.mnist.create(mnistDataPath, data.mnist.Test)
+testLoader = mnistDataset.create_loader(10, True, False, 0)
 
 F.setThreadNumber(4)
 for epoch in range(0, 10):
@@ -94,7 +88,7 @@ for epoch in range(0, 10):
     # Save Model
     fileName = 'cache/%d.mnist.mnn' %epoch
     net.train(False)
-    predict = net.forward(F.Input([1, 1, 28, 28], F.NC4HW4))
+    predict = net.forward(F.placeholder([1, 1, 28, 28], F.NC4HW4))
     print("Save to " + fileName)
     F.save([predict], fileName)
     testFunc(testLoader, net)

pymnn/pip_package/MNN/__init__.py

@@ -0,0 +1,4 @@
+from _mnncengine import *
+from . import train
+from . import tools
+

pymnn/pip_package/MNN/tools/__init__.py

@@ -0,0 +1,3 @@
+from . import mnn, mnnops, mnnquant, mnnvisual, mnnconvert 
+from . import utils
+from . import mnn_fb