TFLM(Tensorflow lite micro)验证嵌入式端模型运行，直截了当做法是：对比PC端和嵌入式端运行的tflite模型的输入输出。笔者就TinyML的HelloWorld例程，实践了PC端tflite模型运行情况和MCU端RT1062部署后运行情况。
- mcu： nxp rt1062
- ide： MCUXpresso IDE
- TinyML framework： TensorflowLiteMicro
1. 嵌入式端：输入输出打印(float print)
1） IDE支持浮点数打印
NXP IDE MCUXpresso支持float打印比较简单，如下设置：
设置浮点打印支持2） 打印输出代码
代码语言：javascript复制void model_run()
{
float position = static_cast<float>(inference_count) /
static_cast<float>(kInferencesPerCycle);
float x = position * kXrange;
、
int8_t x_quantized = x / input->params.scale input->params.zero_point;
input->data.int8[0] = x_quantized;
MODEL_RunInference();
int8_t y_quantized = output->data.int8[0];
float y = (y_quantized - output->params.zero_point) * output->params.scale;
printf("inference: %d > x: %f > x_q: %d, y_q: %d > y: %f rn",
inference_count,x,x_quantized,y_quantized, y);
inference_count = 1;
if (inference_count >= kInferencesPerCycle) inference_count = 0;
}运行后串口输出日志信息：
代码语言：javascript复制input zero_porint: -128 , scale: 0.024574
output zero_porint: 4 , scale: 0.008472
inference: 0 > x: 0.000000 > x_q: -128, y_q: 4 > y: 0.000000
inference: 1 > x: 0.314159 > x_q: -115, y_q: 48 > y: 0.372770
inference: 2 > x: 0.628319 > x_q: -102, y_q: 70 > y: 0.559154
inference: 3 > x: 0.942478 > x_q: -89, y_q: 103 > y: 0.838731
inference: 4 > x: 1.256637 > x_q: -76, y_q: 118 > y: 0.965812
inference: 5 > x: 1.570796 > x_q: -64, y_q: 127 > y: 1.042060
inference: 6 > x: 1.884956 > x_q: -51, y_q: 117 > y: 0.957340
inference: 7 > x: 2.199115 > x_q: -38, y_q: 101 > y: 0.821787
inference: 8 > x: 2.513274 > x_q: -25, y_q: 67 > y: 0.533738
inference: 9 > x: 2.827433 > x_q: -12, y_q: 32 > y: 0.237217
inference: 10 > x: 3.141593 > x_q: 0, y_q: 5 > y: 0.008472
inference: 11 > x: 3.455752 > x_q: 12, y_q: -32 > y: -0.304993
inference: 12 > x: 3.769912 > x_q: 25, y_q: -59 > y: -0.533738
inference: 13 > x: 4.084070 > x_q: 38, y_q: -88 > y: -0.779427
inference: 14 > x: 4.398230 > x_q: 50, y_q: -110 > y: -0.965812
inference: 15 > x: 4.712389 > x_q: 63, y_q: -127 > y: -1.109837
inference: 16 > x: 5.026548 > x_q: 76, y_q: -112 > y: -0.982756
inference: 17 > x: 5.340708 > x_q: 89, y_q: -84 > y: -0.745539
inference: 18 > x: 5.654867 > x_q: 102, y_q: -59 > y: -0.533738
inference: 19 > x: 5.969026 > x_q: 114, y_q: -38 > y: -0.355825
inference: 0 > x: 0.000000 > x_q: -128, y_q: 4 > y: 0.000000
inference: 1 > x: 0.314159 > x_q: -115, y_q: 48 > y: 0.3727702. PC端：串口日志解析
代码语言：javascript复制import numpy as np
import pandas as pd
import re
def log_to_df(log_file, pat, x_match_idx, y_match_idx):
x = []
y = []
with open(log_file,'r') as fs:
lines = fs.readlines()
for line in lines:
rlt = re.search(pat,line)
if rlt:
x.append(rlt.group(x_match_idx))
y.append(rlt.group(y_match_idx))
df = pd.DataFrame({
'x':x,
'y':y
})
df['x'] = df['x'].astype(float)
df['y'] = df['y'].astype(float)
return df
log_file = '../data/sine_model_out.txt'
pat ='inference: (d ) > x: ([-]?d .d ) > x_q: ([-]?d ), y_q: ([-]?d ) > y: ([-]?d.d )'
x_match_idx = 2
y_match_idx = 5
df = log_to_df(log_file, pat, x_match_idx, y_match_idx)
dflog_to_df使用re解析得到输入输出
df3. PC端：输入输出对比验证
代码语言：javascript复制import tensorflow as tf
import numpy as np
def predict_tflite(tflite_model, x_test):
# Prepare the test data
x_test_ = x_test.copy()
x_test_ = x_test_.reshape((x_test.size, 1))
x_test_ = x_test_.astype(np.float32)
# Initialize the TFLite interpreter
if isinstance(tflite_model, str):
interpreter = tf.lite.Interpreter(model_path=tflite_model)
else:
interpreter = tf.lite.Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()[0]
output_details = interpreter.get_output_details()[0]
# If required, quantize the input layer (from float to integer)
input_scale, input_zero_point = input_details["quantization"]
if (input_scale, input_zero_point) != (0.0, 0):
x_test_ = x_test_ / input_scale input_zero_point
x_test_ = x_test_.astype(input_details["dtype"])
# Invoke the interpreter
y_pred = np.empty(x_test_.size, dtype=output_details["dtype"])
for i in range(len(x_test_)):
interpreter.set_tensor(input_details["index"], [x_test_[i]])
interpreter.invoke()
y_pred[i] = interpreter.get_tensor(output_details["index"])[0]
# If required, dequantized the output layer (from integer to float)
output_scale, output_zero_point = output_details["quantization"]
if (output_scale, output_zero_point) != (0.0, 0):
y_pred = y_pred.astype(np.float32)
y_pred = (y_pred - output_zero_point) * output_scale
return y_pred
def cross_validate_tflite(df, tflite_model_path):
df['pred_pc'] = predict_tflite(tflite_model_path,df['x'].to_numpy().astype(np.float32))
df[:20].set_index('x',drop=True).plot()
return df
tflite_model_path = 'hello_world/train/models/model.tflite'
cross_validate_tflite(df, tflite_model_path)
输出完美匹配
其他
helloworld例程是简单回归模型，输入输出都是一维，相对简单，直接。对于图像分类问题，输入图像矩阵，输出各个类别概率，只能比较输出。另外对比基于相同的输入才有意义，这就要把样例保存到flash或者sd卡，通过fatfs相同在运行时进行识别，这样也只是覆盖小部分样例测试，效率较低。完善的解决方案，应该是通过网络进行pc端和嵌入式端的实时通讯和调试，实现硬件在环的测试。NXP eIQ AI平台工具链实现了这一过程，其本质应该是嵌入式端实现了tcp server之类的。
