需要做点什么
方便广大研究生、算法工程师快速使用keras，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
tensorflow 2.6.0
keras 2.6.0
onnx 1.9.0
onnxruntime-gpu 1.9.0
tf2onnx 1.9.3
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:
1000000000 ..
0000000000
1000000000
4000000000
0000000000
0000000000
7000000000
3000000000
5000000000
3000000000
8000000000
9000000000
1000000000
3000000000
3000000000
1000000000
2000000000
0000000000
1. 导入需要的包
import os
import onnx
import keras
import logging
import subprocess
import numpy as np
import pandas as pd
import tensorflow as tf
import onnxruntime as ort
from sklearn.metrics import accuracy_score
from keras.models import Sequential, Model, load_model, save_model
from keras.layers import Dense, Activation, Dropout, Conv2D, Flatten, MaxPool2D, Input, Conv1D
from keras.utils.np_utils import to_categorical
tf.autograph.set_verbosity(0)
logging.getLogger("tensorflow").setLevel(logging.ERROR)
2. 参数准备
N_EPOCH = 1
N_BATCH = 64
N_BATCH_NUM = 500
S_DATA_PATH = r"mnist_train.csv"
S_KERAS_MODEL_DIR_PATH = r"cnn_keras"
S_KERAS_MODEL_PATH = r"cnn_keras.h5"
S_ONNX_MODEL_PATH = r"cnn_keras.onnx"
S_DEVICE, N_DEVICE_ID, S_DEVICE_FULL = "cuda", 0, "cuda:0" # 使用gpu
# S_DEVICE, N_DEVICE_ID, S_DEVICE_FULL = "cpu", 0, "cpu" # 没有gpu请反注释这行以使用CPU
if S_DEVICE == "cpu":
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
3. 读取数据
df = pd.read_csv(S_DATA_PATH, header=None)
np_mat = np.array(df)
print(df.shape)
print(np_mat.shape)
X = np_mat[:, 1:]
Y = np_mat[:, 0]
X = X.astype(np.float32) / 255
X_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], 28, 28, 1)
X_test = X_test.reshape(X_test.shape[0], 28, 28, 1)
Y_train = to_categorical(Y_train, num_classes=10)
Y_test = to_categorical(Y_test, num_classes=10)
print(X_train.shape)
print(Y_train.shape)
print(X_test.shape)
print(Y_test.shape)
运行输出
(42000, 785)
(42000, 785)
(32000, 28, 28, 1)
(32000, 10)
(10000, 28, 28, 1)
(10000, 10)
4. 模型构建
x_in = Input(shape=(28, 28, 1)) # 图像维度必须是 w h c
x = Conv2D(filters=32, kernel_size=(3, 3))(x_in)
x = MaxPool2D(pool_size=(2, 2))(x)
x = Dropout(0.2)(x)
x = Flatten()(x)
x = Dense(128)(x)
x = Activation('relu')(x)
x = Dense(10)(x)
y = Activation('softmax')(x)
model = Model(x_in, y)
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
print(model.summary())
运行输出
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 28, 28, 1)] 0
_________________________________________________________________
conv2d (Conv2D) (None, 26, 26, 32) 320
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 13, 13, 32) 0
_________________________________________________________________
dropout (Dropout) (None, 13, 13, 32) 0
_________________________________________________________________
flatten (Flatten) (None, 5408) 0
_________________________________________________________________
dense (Dense) (None, 128) 692352
_________________________________________________________________
activation (Activation) (None, 128) 0
_________________________________________________________________
dense_1 (Dense) (None, 10) 1290
_________________________________________________________________
activation_1 (Activation) (None, 10) 0
=================================================================
Total params: 693,962
Trainable params: 693,962
Non-trainable params: 0
_________________________________________________________________
None
5. 模型训练和保存
model.fit(X_train,
Y_train,
epochs=N_EPOCH,
batch_size=N_BATCH,
verbose=1,
validation_data=(X_test, Y_test))
score = model.evaluate(X_test, Y_test, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
save_model(model, S_KERAS_MODEL_PATH)
运行输出
486/500 [============================>.] - ETA: 0s - loss: 0.2873 - accuracy: 0.9144
500/500 [==============================] - 4s 3ms/step - loss: 0.2837 - accuracy: 0.9155 - val_loss: 0.1352 - val_accuracy: 0.9616
Test score: 0.13516278564929962
Test accuracy: 0.9616000056266785
6.模型加载和加载模型使用
load_model = load_model(S_KERAS_MODEL_PATH)
print("load model ok")
score = load_model.evaluate(X_test, Y_test, verbose=0)
print('load model Test score:', score[0])
print('load model Test accuracy:', score[1])
运行输出
load model ok
load model Test score: 0.13516278564929962
load model Test accuracy: 0.9616000056266785
7.导出ONNX
s_cmd = 'python -m tf2onnx.convert --keras %s --output %s' % (S_KERAS_MODEL_PATH, S_ONNX_MODEL_PATH)
print(s_cmd)
print(os.system(s_cmd))
# proc = subprocess.run(s_cmd.split(), check=True)
# print(proc.returncode)
运行输出
python -m tf2onnx.convert --keras G:\Data\task_model_out\_tmp_out\cnn_keras.h5 --output G:\Data\task_model_out\_tmp_out\cnn_keras.onnx
0
8. 加载ONNX并运行
model = onnx.load(S_ONNX_MODEL_PATH)
print(onnx.checker.check_model(model)) # Check that the model is well formed
print(onnx.helper.printable_graph(model.graph)) # Print a human readable representation of the graph
ls_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, 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])
运行输出
None
graph tf2onnx (
%input_1:0[FLOAT, unk__17x28x28x1]
) initializers (
%new_shape__15[INT64, 4]
%model/dense_1/MatMul/ReadVariableOp:0[FLOAT, 128x10]
%model/dense_1/BiasAdd/ReadVariableOp:0[FLOAT, 10]
%model/dense/MatMul/ReadVariableOp:0[FLOAT, 5408x128]
%model/dense/BiasAdd/ReadVariableOp:0[FLOAT, 128]
%model/conv2d/Conv2D/ReadVariableOp:0[FLOAT, 32x1x3x3]
%model/conv2d/BiasAdd/ReadVariableOp:0[FLOAT, 32]
%const_fold_opt__16[INT64, 2]
) {
%model/conv2d/BiasAdd__6:0 = Reshape(%input_1:0, %new_shape__15)
%model/conv2d/BiasAdd:0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], strides = [1, 1]](%model/conv2d/BiasAdd__6:0, %model/conv2d/Conv2D/ReadVariableOp:0, %model/conv2d/BiasAdd/ReadVariableOp:0)
%model/max_pooling2d/MaxPool:0 = MaxPool[kernel_shape = [2, 2], strides = [2, 2]](%model/conv2d/BiasAdd:0)
%model/max_pooling2d/MaxPool__12:0 = Transpose[perm = [0, 2, 3, 1]](%model/max_pooling2d/MaxPool:0)
%model/flatten/Reshape:0 = Reshape(%model/max_pooling2d/MaxPool__12:0, %const_fold_opt__16)
%model/dense/MatMul:0 = MatMul(%model/flatten/Reshape:0, %model/dense/MatMul/ReadVariableOp:0)
%model/dense/BiasAdd:0 = Add(%model/dense/MatMul:0, %model/dense/BiasAdd/ReadVariableOp:0)
%model/activation/Relu:0 = Relu(%model/dense/BiasAdd:0)
%model/dense_1/MatMul:0 = MatMul(%model/activation/Relu:0, %model/dense_1/MatMul/ReadVariableOp:0)
%model/dense_1/BiasAdd:0 = Add(%model/dense_1/MatMul:0, %model/dense_1/BiasAdd/ReadVariableOp:0)
%Identity:0 = Softmax[axis = 1](%model/dense_1/BiasAdd:0)
return %Identity:0
}
input name ['input_1:0']
output name ['Identity:0']
val device cuda
val shape [128, 28, 28, 1]
val data type tensor(float)
is_tensor True
array_equal True
providers CUDAExecutionProvider
sess env ['CUDAExecutionProvider', 'CPUExecutionProvider']
<class 'list'>
[[1.0287621e-04 9.9524093e-01 5.0408958e-04 ... 6.5664819e-05
3.8182980e-03 1.2303158e-05]
[9.9932754e-01 2.7173186e-08 3.5315077e-04 ... 3.0959238e-06
8.5986117e-05 3.6047477e-06]
[1.1101285e-05 9.9719965e-01 3.8205151e-04 ... 1.2267688e-03
7.8595197e-04 4.0839368e-05]
...
[2.8337089e-02 1.5399084e-05 2.1733245e-01 ... 1.5945830e-05
2.1134425e-02 1.7111158e-03]
[1.7888090e-06 3.3868539e-06 5.2631256e-04 ... 9.9888057e-01
5.4794059e-06 5.5255485e-04]
[4.1398227e-05 1.0462944e-06 5.5901739e-03 ... 3.1221823e-09
6.6847453e-04 7.8918066e-07]]