百度Paddle速查_CPU和GPU的mnist预测训练_模型导出_模型导入再预测_导出onnx并预测

需要做点什么

方便广大烟酒生研究生、人工智障炼丹师算法工程师快速使用百度PaddelPaddle，所以特写此文章，默认使用者已有基本的深度学习概念、数据集概念。

系统环境

python 3.7.4
paddlepaddle-gpu 2.2.2
paddle2onnx 0.9.1
onnx 1.9.0
onnxruntime-gpu 1.9.0

数据准备

MNIST数据集csv文件是一个42000x785的矩阵
42000表示有42000张图片
785中第一列是图片的类别(0,1,2,..,9),第二列到最后一列是图片数据向量 (28x28的图片张成784的向量)， 数据集长这个样子:

1 0 0 0 0 0 0 0 0 0 ..
0 0 0 0 0 0 0 0 0 0 ..
1 0 0 0 0 0 0 0 0 0 ..

1. 导入需要的包

import osimport onnximport paddleimport numpy as npimport pandas as pdimport onnxruntime as ortimport paddle.nn.functional as Ffrom paddle.metric import Accuracyfrom paddle.static import InputSpecfrom sklearn.metrics import accuracy_score

2. 参数准备

N_EPOCH = 2N_BATCH = 64N_BATCH_NUM = 250S_DATA_PATH = r"mnist_train.csv"S_PADDLE_MODEL_PATH = r"cnn_model"S_ONNX_MODEL_PATH = r"cnn_model_batch%d.onnx" % N_BATCHS_DEVICE, N_DEVICE_ID, S_DEVICE_FULL = "gpu", 0, "gpu:0"# S_DEVICE, N_DEVICE_ID, S_DEVICE_FULL = "cpu", 0, "cpu"paddle.set_device(S_DEVICE_FULL)

运行输出:

CUDAPlace(0)

3. 读取数据

df = pd.read_csv(S_DATA_PATH, header=None)print(df.shape)np_mat = np.array(df)print(np_mat.shape)X = np_mat[:, 1:]Y = np_mat[:, 0]X = X.astype(np.float32) / 255X_train = X[:N_BATCH * N_BATCH_NUM]X_test = X[N_BATCH * N_BATCH_NUM:]Y_train = Y[:N_BATCH * N_BATCH_NUM]Y_test = Y[N_BATCH * N_BATCH_NUM:]X_train = X_train.reshape(X_train.shape[0], 1, 28, 28)X_test = X_test.reshape(X_test.shape[0], 1, 28, 28)print(X_train.shape)print(Y_train.shape)print(X_test.shape)print(Y_test.shape)class MnistDataSet(paddle.io.Dataset):    def __init__(self, X, Y):        self.l_data, self.l_label = [], []        for i in range(X.shape[0]):            self.l_data.append(X[i, :, :, :])            self.l_label.append(Y[i])    def __getitem__(self, index):        return self.l_data[index], self.l_label[index]    def __len__(self):        return len(self.l_data)train_loader = paddle.io.DataLoader(MnistDataSet(X_train, Y_train),  batch_size=N_BATCH, shuffle=True)test_loader = paddle.io.DataLoader(MnistDataSet(X_test, Y_test), batch_size=N_BATCH, shuffle=False)

运行输出

(42000, 785)(42000, 785)(16000, 1, 28, 28)(16000,)(26000, 1, 28, 28)(26000,)

4. 模型构建

class Net(paddle.nn.Layer):    def __init__(self):        super(Net, self).__init__()        self.encoder = paddle.nn.Sequential(paddle.nn.Conv2D(1, 16, 3, 1),                                            paddle.nn.MaxPool2D(2),                                            paddle.nn.Flatten(1),                                            paddle.nn.Linear(2704, 128),                                            paddle.nn.ReLU(),                                            paddle.nn.Linear(128, 10))    def forward(self, x):        out = self.encoder(x)        return out

5. 模型训练和保存

print("model train")model = paddle.Model(Net(), InputSpec([None, 1, 28, 28], 'float32', 'x'), InputSpec([None, 10], 'float32', 'x'))model.prepare(paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters()), paddle.nn.CrossEntropyLoss(), Accuracy())model.fit(train_loader,          test_loader,          epochs=N_EPOCH,          batch_size=N_BATCH,          save_dir=S_PADDLE_MODEL_PATH + "_iter",          verbose=1)model.save(S_PADDLE_MODEL_PATH + "_final_model")print()# model.save(S_PADDLE_MODEL_PATH) # Model save

运行输出

model trainThe loss value printed in the log is the current step, and the metric is the average value of previous steps.Epoch 1/2step  30/250 [==>...........................] - loss: 0.3036 - acc: 0.7531 - ETA: 1s - 6ms/stepstep 250/250 [==============================] - loss: 0.3151 - acc: 0.9073 - 4ms/step          save checkpoint at D:\Document\_Code_Py\ai_fast_handbook\cnn_model_iter\0Eval begin...step 407/407 [==============================] - loss: 0.0230 - acc: 0.9330 - 2ms/step          - loss: 0.1698 - acc: 0.9315 - ETA: 0s - 2ms/ - loss: 0.3643 - acc: 0.9326 - ETA: 0s - 2mEval samples: 26000Epoch 2/2step 250/250 [==============================] - loss: 0.0744 - acc: 0.9642 - 3ms/step          save checkpoint at D:\Document\_Code_Py\ai_fast_handbook\cnn_model_iter\1Eval begin...step 407/407 [==============================] - loss: 0.0614 - acc: 0.9575 - 2ms/step          Eval samples: 26000save checkpoint at D:\Document\_Code_Py\ai_fast_handbook\cnn_model_iter\final

6.模型预测

print("model pred")model.evaluate(test_loader, batch_size=N_BATCH, verbose=1)print()

运行输出

model predEval begin...step 407/407 [==============================] - loss: 0.0614 - acc: 0.9575 - 2ms/step          - loss: 0.2162 - acc: 0.9559 - ETAEval samples: 26000

7.模型加载和加载模型使用

print("load model and pred test data")model_load = paddle.Model(Net(), InputSpec([None, 1, 28, 28], 'float32', 'x'), InputSpec([None, 10], 'float32', 'x'))# model_load.load(S_PADDLE_MODEL_PATH + "_iter/final")model_load.load(S_PADDLE_MODEL_PATH + "_final_model")model_load.prepare(paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters()), paddle.nn.loss.CrossEntropyLoss(), Accuracy())model_load.evaluate(test_loader, batch_size=N_BATCH, verbose=1)print()

运行输出

load model and pred test dataEval begin...step 407/407 [==============================] - loss: 0.0614 - acc: 0.9575 - 2ms/step          - loss: 0.1656 - acc: 0.9561 - ETA:Eval samples: 26000

8.导出ONNX

x_spec = InputSpec([None, 1, 28, 28], 'float32', 'x')paddle.onnx.export(Net(), S_ONNX_MODEL_PATH, input_spec=[x_spec])

运行输出

2022-03-24 08:08:21 [INFO]	ONNX model saved in cnn_model_batch64.onnx.onnx

8. 加载ONNX并运行

S_DEVICE = "cuda" if S_DEVICE == "gpu" else S_DEVICEmodel = onnx.load(S_ONNX_MODEL_PATH + ".onnx")print(onnx.checker.check_model(model))  # Check that the model is well formedprint(onnx.helper.printable_graph(model.graph))  # Print a human readable representation of the graphls_input_name, ls_output_name = [input.name for input in model.graph.input], [output.name for output in model.graph.output]print("input name ", ls_input_name)print("output name ", ls_output_name)s_input_name = ls_input_name[0]x_input = X_train[:N_BATCH * 2, :, :, :].astype(np.float32)ort_val = ort.OrtValue.ortvalue_from_numpy(x_input, S_DEVICE, N_DEVICE_ID)print("val device ", ort_val.device_name())print("val shape ", ort_val.shape())print("val data type ", ort_val.data_type())print("is_tensor ", ort_val.is_tensor())print("array_equal ", np.array_equal(ort_val.numpy(), x_input))providers = 'CUDAExecutionProvider' if S_DEVICE == "cuda" else 'CPUExecutionProvider'print("providers ", providers)ort_session = ort.InferenceSession(S_ONNX_MODEL_PATH + ".onnx", providers=[providers])  # gpu运行ort_session.set_providers([providers])outputs = ort_session.run(None, {s_input_name: ort_val})print("sess env ", ort_session.get_providers())print(type(outputs))print(outputs[0])'''For example ['CUDAExecutionProvider', 'CPUExecutionProvider']    means execute a node using CUDAExecutionProvider if capable, otherwise execute using CPUExecutionProvider.'''

运行输出

Nonegraph paddle-onnx (  %x[FLOAT, -1x1x28x28]) {  %conv2d_2.w_0 = Constant[value = <Tensor>]()  %conv2d_2.b_0 = Constant[value = <Tensor>]()  %linear_4.w_0 = Constant[value = <Tensor>]()  %linear_4.b_0 = Constant[value = <Tensor>]()  %linear_5.w_0 = Constant[value = <Tensor>]()  %linear_5.b_0 = Constant[value = <Tensor>]()  %conv2d_3.tmp_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%x, %conv2d_2.w_0)  %Constant_0 = Constant[value = <Tensor>]()  %Reshape_0 = Reshape(%conv2d_2.b_0, %Constant_0)  %conv2d_3.tmp_1 = Add(%conv2d_3.tmp_0, %Reshape_0)  %pool2d_0.tmp_0 = MaxPool[kernel_shape = [2, 2], pads = [0, 0, 0, 0], strides = [2, 2]](%conv2d_3.tmp_1)  %Shape_0 = Shape(%pool2d_0.tmp_0)  %Slice_0 = Slice[axes = [0], ends = [1], starts = [0]](%Shape_0)  %Constant_1 = Constant[value = <Tensor>]()  %Concat_0 = Concat[axis = 0](%Slice_0, %Constant_1)  %flatten_3.tmp_0 = Reshape(%pool2d_0.tmp_0, %Concat_0)  %linear_6.tmp_0 = MatMul(%flatten_3.tmp_0, %linear_4.w_0)  %linear_6.tmp_1 = Add(%linear_6.tmp_0, %linear_4.b_0)  %relu_0.tmp_0 = Relu(%linear_6.tmp_1)  %linear_7.tmp_0 = MatMul(%relu_0.tmp_0, %linear_5.w_0)  %linear_7.tmp_1 = Add(%linear_7.tmp_0, %linear_5.b_0)  return %linear_7.tmp_1}input name  ['x']output name  ['linear_7.tmp_1']val device  cudaval shape  [128, 1, 28, 28]val data type  tensor(float)is_tensor  Truearray_equal  Trueproviders  CUDAExecutionProvidersess env  ['CUDAExecutionProvider', 'CPUExecutionProvider']<class 'list'>[[ 0.763783   -0.16668957 -0.16518936 ...  0.07235195 -0.01643395   0.06049304] [ 1.8068395  -0.74552214  0.3836273  ...  0.75880224 -0.88902843   0.32921085] [ 0.2381373  -0.14879732 -0.21634206 ... -0.06579521 -0.461351   0.15305203] ... [ 0.97004616  0.07693841  0.05774391 ...  0.21991295  0.07179791  -0.22383693] [ 0.5787286  -0.34370935 -0.12914304 ... -0.03083546 -0.01817408  -0.5147962 ] [ 0.60808766 -0.12549599 -0.32095248 ... -0.32175955 -0.03176413  -0.06790417]]

你甚至不愿意Start的Github

ai_fast_handbook