mnist

MNIST

Now take MNIST model as an example to illustrate the use of gxDNN.

Generate NPU File

Generate Frozen PB

In order to easily obtain the input node and output node, we need to name the input node and output node:

change x = tf.placeholder(tf.float32, [None, 784])

to

x = tf.placeholder(tf.float32, [None, 784], name="input_x")

change

y = tf.add(tf.matmul(x, W), b)

to

y = tf.add(tf.matmul(x, W), b, name="result")

To generate a PB file, add the following at the end of the file:

from tensorflow.python.framework import graph_util
constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ["result"])
with tf.gfile.FastGFile('./generated_pb/mnist_with_ckpt.pb', mode='wb') as f:
    f.write(constant_graph.SerializeToString())

For the modified document, see here。 After running python mnist.py, mnist_with_ckpt.pb will be generated under 'generated_pb' file.

Edit the NPU Config File

Edit the config file config.yaml:

CORENAME: LEO                               # 芯片型号 LEO/LEO_MPW
PB_FILE: ./generated_pb/mnist_with_ckpt.pb  # 输入的pb文件
OUTPUT_FILE: mnist.npu                      # 输出的NPU文件名
NPU_UNIT: NPU64                             # NPU设备类型
COMPRESS: false                             # 压缩模型
COMPRESS_QUANT_BITS: 8                      # 量化成8bits
COMPRESS_TYPE: LINEAR                       # 线性压缩
OUTPUT_TYPE: raw                            # NPU文件的类型
INPUT_OPS:
        input_x: [1, 784]                   # 输入结点名称和数据维度，每运行一次输入数据为1x784，即一幅图
OUTPUT_OPS: [result]                        # 输出结点名称 

Compile

Compile using gxnpuc

gxnpuc config.yaml

Generate mnist.npu.

Run Model

After the NPU file is generated, we need to call the API provided by NationalChip to deploy the model to the gx8010 development board.

Call API Process

1. Open NPU device.
2. Load the model and input data, run the model and get the output data.
3. Close NPU device.

MNIST Example

Please refer to the code here。The program requires the user to input a binary file with 28x28 pixel values and output the recognized number.

Several binary test files are stored in the directory images. The content is the normalized pixel value of 28x28.

Execute make generate executable test_mnist.elf.

Run test_mnist.elf, the results are as follows:

./test_mnist.elf images/image0
Digit: 7
./test_mnist.elf images/image1
Digit: 2
./test_mnist.elf images/image2
Digit: 1
./test_mnist.elf images/image3
Digit: 0
./test_mnist.elf images/image4
Digit: 4
./test_mnist.elf images/image5
Digit: 1