机器视觉整合图像处理与机械控制技术，能提升工业生产灵活性与自动化水平，国产自主可控开源硬件飞腾派凭借强劲算力、良好兼容性与可扩展性，可高效支撑这类视觉应用的开发与部署。

飞腾派识别数字

项目准备

硬件规格：飞腾派

操作系统：飞腾OS（飞腾派OS系统下载：‍https://www.iceasy.com/cloud/Phytium?pid=1877647545995448322）

具体操作：

一、安装Tensorflow lite

安装pip

LeNet = keras.models.Sequential(    [        keras.layers.Conv2D(            input_shape=(32, 32, 1), filters=6, kernel_size=(3, 3), activation="relu"        ),        keras.layers.AveragePooling2D(),        keras.layers.Conv2D(filters=16, kernel_size=(3, 3), activation="relu"),        keras.layers.AveragePooling2D(),        keras.layers.Flatten(),        keras.layers.Dense(units=120, activation="relu"),        keras.layers.Dense(units=84, activation="relu"),        keras.layers.Dense(units=10, activation="softmax"),    ])

安装tflite-runtime

def read_mnist(images_path: str, labels_path: str):    with gzip.open(labels_path, "rb") as labelsFile:        labels = np.frombuffer(labelsFile.read(), dtype=np.uint8, offset=8)    with gzip.open(images_path, "rb") as imagesFile:        length = len(labels)        # Load flat 28x28 px images (784 px), and convert them to 28x28 px        features = (            np.frombuffer(imagesFile.read(), dtype=np.uint8, offset=16)            .reshape(length, 784)            .reshape(length, 28, 28, 1)        )        # LeNet architecture accepts a 32x32 pixel images as input,        # mnist data is 28x28 pixels. We simply pad the images with zeros to overcome        features = np.pad(features, ((0, 0), (2, 2), (2, 2), (0, 0)), "constant")    return features, labelsx_train, y_train = read_mnist(    "train-images-idx3-ubyte.gz", "train-labels-idx1-ubyte.gz")x_test, y_test = read_mnist("t10k-images-idx3-ubyte.gz", "t10k-labels-idx1-ubyte.gz")x_train, x_test = x_train / 255.0, x_test / 255.0

二、训练LeNet模型

LeNet模型由Yann LeCun在1998年设计并提出的，是一种非常经典的卷积神经网络模型，先使用Tensorflow API实现这个模型

loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)LeNet.compile(optimizer='adam',              loss=loss_fn,              metrics=['accuracy'])LeNet.fit(x_train, y_train, epochs=5)

载入数据

def read_mnist(images_path: str, labels_path: str):    with gzip.open(labels_path, "rb") as labelsFile:        labels = np.frombuffer(labelsFile.read(), dtype=np.uint8, offset=8)    with gzip.open(images_path, "rb") as imagesFile:        length = len(labels)        # Load flat 28x28 px images (784 px), and convert them to 28x28 px        features = (            np.frombuffer(imagesFile.read(), dtype=np.uint8, offset=16)            .reshape(length, 784)            .reshape(length, 28, 28, 1)        )        # LeNet architecture accepts a 32x32 pixel images as input,        # mnist data is 28x28 pixels. We simply pad the images with zeros to overcome        features = np.pad(features, ((0, 0), (2, 2), (2, 2), (0, 0)), "constant")    return features, labelsx_train, y_train = read_mnist(    "train-images-idx3-ubyte.gz", "train-labels-idx1-ubyte.gz")x_test, y_test = read_mnist("t10k-images-idx3-ubyte.gz", "t10k-labels-idx1-ubyte.gz")x_train, x_test = x_train / 255.0, x_test / 255.0

设置损失函数以及优化器开始训练

loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)LeNet.compile(optimizer='adam',              loss=loss_fn,              metrics=['accuracy'])LeNet.fit(x_train, y_train, epochs=5)

评估模型

LeNet.evaluate(x_test,  y_test, verbose=2)

1875/1875 [==============================] - 9s 3ms/step - loss: 0.2255 - accuracy: 0.9328Epoch 2/51875/1875 [==============================] - 6s 3ms/step - loss: 0.0749 - accuracy: 0.9767Epoch 3/51875/1875 [==============================] - 6s 3ms/step - loss: 0.0529 - accuracy: 0.9836Epoch 4/51875/1875 [==============================] - 5s 3ms/step - loss: 0.0408 - accuracy: 0.9869Epoch 5/51875/1875 [==============================] - 5s 3ms/step - loss: 0.0329 - accuracy: 0.9895313/313 - 1s - loss: 0.0341 - accuracy: 0.9897 - 628ms/epoch - 2ms/step

将模型导出为TFLite格式

# convert modelconverter = tf.lite.TFLiteConverter.from_keras_model(LeNet)tflite_model = converter.convert()# Save the model.with open('lenet5.tflite', 'wb') as f:  f.write(tflite_model)
  

在飞腾派部署LeNet模型

首先设置track装饰器，让我们了解飞腾派的推理速度

import timeimport logging as logimport syslog.basicConfig(format='[ %(levelname)s ] %(message)s', level=log.INFO, stream=sys.stdout)def timer(label:str, ms_:bool=True):    def out_wrapper(func):        def wrapper(*args, **kwargs):            st = time.perf_counter()            ret = func(*args, **kwargs)            measure = time.perf_counter() - st            if ms_:                measure *= 1e3            tag = "ms" if ms_ else "s"            log.info(f"{label} elapsed {(measure):.3f} {tag}.")            return ret        return wrapper    return out_wrapper
    

推理函数

@timer("infer")def infer(model:str, input_blob:np.array) -> int:    interpreter = tflite.Interpreter(model_path=model, num_threads=4)    interpreter.allocate_tensors()    input_details = interpreter.get_input_details()    # Get input and output tensors.    output_details = interpreter.get_output_details()    interpreter.set_tensor(input_details[0]['index'], input_blob)    interpreter.invoke()    return np.argmax(interpreter.get_tensor(output_details[0]['index']))
    

运行结果

数字识别全部正确！而且识别速度特别快，轻松满足各类应用需求！

飞腾派免费试用！轻松带走定制文创礼品！

➡️点击查看活动详情

（内附测评分享小tips）