Toybrick

标题: RK3399Pro入门教程（9）MNIST RKNN量化教程 [打印本页]
作者: peng    时间: 2019-9-20 08:34
标题: RK3399Pro入门教程（9）MNIST RKNN量化教程
[attach]570[/attach]

一.搭建网络
    model.py,同tensorflow官网mnist的例子差不多，不懂的可以去官网看下官网代码解析
  1. import tensorflow as tf

  3. #这里输入采用28x28方便之后进行rknn量化
  4. x = tf.placeholder("float", [None, 28,28],name='x')
  5. y_ = tf.placeholder("float", [None,10],name='y_')
  6. keep_prob = tf.placeholder("float", name='keep_prob')
  7. def weight_variable(shape,name):
  8.     initial = tf.truncated_normal(shape, stddev=0.1)
  9.     return tf.Variable(initial,name=name)

  11. def bias_variable(shape,name):
  12.     initial = tf.constant(0.1, shape=shape)
  13.     return tf.Variable(initial,name=name)

  15. def conv2d(x, W):
  16.     return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

  18. def max_pool_2x2(x):
  19.     return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
  20.                         strides=[1, 2, 2, 1], padding='SAME')

  22. # convolution layer
  23. def lenet5_layer(input, weight, bias,weight_name,bias_name):
  24.     W_conv = weight_variable(weight,weight_name)
  25.     b_conv = bias_variable(bias,bias_name)
  26.     h_conv = tf.nn.relu(conv2d(input, W_conv) + b_conv)
  27.     return max_pool_2x2(h_conv)
  28. # connected layer
  29. def dense_layer(layer, weight, bias,weight_name,bias_name):
  30.     W_fc = weight_variable(weight,weight_name)
  31.     b_fc = bias_variable(bias,bias_name)
  32.     return tf.nn.relu(tf.matmul(layer, W_fc) + b_fc)

  34. def build_model(is_training):
  35.     #first conv
  36.     x_image = tf.reshape(x, [-1,28,28,1])
  37.     W_conv1 = [5, 5, 1, 32]
  38.     b_conv1 = [32]
  39.     layer1 = lenet5_layer(x_image,W_conv1,b_conv1,'W_conv1','b_conv1')
  40.     #second conv
  41.     W_conv2 = [5, 5, 32, 64]
  42.     b_conv2 = [64]
  43.     layer2 = lenet5_layer(layer1,W_conv2,b_conv2,'W_conv2','b_conv2')
  44.     #third conv
  45.     W_fc1 = [7 * 7 * 64, 1024]
  46.     b_fc1 = [1024]
  47.     layer2_flat = tf.reshape(layer2, [-1, 7*7*64])
  48.     layer3 = dense_layer(layer2_flat,W_fc1,b_fc1,'W_fc1','b_fc1')
  49.     #softmax
  50.     W_fc2 = weight_variable([1024, 10],'W_fc2')
  51.     b_fc2 = bias_variable([10],'b_fc2')
  52.     if is_training:
  53.         #dropout
  54.         h_fc1_drop = tf.nn.dropout(layer3, keep_prob)
  55.         finaloutput=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2,name="y_conv")
  56.     else:
  57.         finaloutput=tf.nn.softmax(tf.matmul(layer3, W_fc2) + b_fc2,name="y_conv")
  58.     print('finaloutput:', finaloutput)
  59.     return finaloutput
复制代码
二.训练网络
    train.py,这里代码兼容了tf伪量化的代码，这里我们把create_training_graph()传入的参数is_quantify设为False就可以了，由于mnist拿到的train和test数据shape都是（784，),这里定义了一个reshape_batch函数把train时的batch以及test时的输入都reshape成（28,28），具体代码如下：
  1. # -*- coding=utf-8 -*-
  2. import tensorflow as tf
  3. from tensorflow.examples.tutorials.mnist import input_data
  4. from model import build_model, x, y_, keep_prob

  6. mnist = input_data.read_data_sets("../MNIST_data/", one_hot=True)
  7. def create_training_graph(is_quantify):
  8.     #创建训练图，加入create_training_graph：
  9.     g = tf.get_default_graph()   # 给create_training_graph的参数，默认图
  10.     #调用网络定义，也就是拿到输出
  11.     y_conv = build_model(is_training=True)    #这里的is_training设置为True，因为前面模型定义写了训练时要用到dropout
  12.     #损失函数
  13.     cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))
  14.     print('cost:', cross_entropy)
  15.     if is_quantify:
  16.         # 加入 create_training_graph函数，注意位置要在loss之后， optimize之前
  17.         tf.contrib.quantize.create_training_graph(input_graph=g, quant_delay=0)
  18.     #  optimize
  19.     optimize = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
  20.     #计算准确率
  21.     correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
  22.     # 给出识别准确率[这会返回我们一个布尔值的列表.为了确定哪些部分是正确的，我们要把它转换成浮点值，然后再示均值。 比如, [True, False, True, True] 会转换成 [1,0,1,1] ，从而它的准确率就是0.75.]   
  23.     accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

  25.     # 返回所需数据，供训练使用
  26.     return dict(
  27.         x=x,
  28.         y=y_,
  29.         keep_prob=keep_prob,
  30.         optimize=optimize,
  31.         cost=cross_entropy,
  32.         correct_prediction=correct_prediction,
  33.         accuracy=accuracy,
  34.     )

  36. def reshape_batch(batch):
  37.     rebatch = []
  38.     for item in batch:
  39.         b = item.reshape(28,28)
  40.         rebatch.append(b)
  41.     return rebatch
  42. #开始训练
  43. def train_network(graph,ckpt,point_dir,pbtxt):
  44.     # 初始化
  45.     init = tf.global_variables_initializer()
  46.     # 调用Saver函数保存所需文件
  47.     saver = tf.train.Saver()
  48.     # 创建上下文，开始训练sess.run(init)
  49.     with tf.Session() as sess:
  50.         sess.run(init)
  51.         # 一共训练两万次，准确率达到百分99以上
  52.         for i in range(20000):
  53.         # 每次处理50张图片
  54.             batch = mnist.train.next_batch(50)
  55.             # 每100次保存并打印一次准确率等
  56.             if i % 100 == 0:
  57.             # feed_dict喂数据，数据全reshape成28x28
  58.                 train_accuracy = sess.run([graph['accuracy']], feed_dict={
  59.                                                                            graph['x']:reshape_batch(batch[0]),    # batch[0]存的图片数据
  60.                                                                            graph['y']:batch[1],    # batch[1]存的标签
  61.                                                                            graph['keep_prob']: 1.0})
  62.                 print("step %d, training accuracy %g"%(i, train_accuracy[0]))
  63.             sess.run([graph['optimize']], feed_dict={
  64.                                                        graph['x']:reshape_batch(batch[0]),
  65.                                                        graph['y']:batch[1],
  66.                                                        graph['keep_prob']:0.5})
  67.         test_accuracy = sess.run([graph['accuracy']], feed_dict={
  68.                                                                   graph['x']: reshape_batch(mnist.test.images),
  69.                                                                   graph['y']: mnist.test.labels,
  70.                                                                   graph['keep_prob']: 1.0})
  71.         print("Test accuracy %g" % test_accuracy[0])
  72.         # 保存ckpt(checkpoint)和pbtxt。记得把路径改成自己的路径
  73.         saver.save(sess, ckpt)
  74.         tf.train.write_graph(sess.graph_def,point_dir,pbtxt, True)
  75.         print(tf.trainable_variables())
  76.         print(tf.get_variable('W_fc2',[1024, 10]).value)


  79. if __name__ == "__main__":
  80.     ckpt = './checkpoint/mnist.ckpt'
  81.     point_dir = './checkpoint'
  82.     pbtxt = 'mnist.pbtxt'
  83.     g1 = create_training_graph(False)
  84.     train_network(g1,ckpt,point_dir,pbtxt)

复制代码三.保存网络参数
    freese.py,将网络中参数和变量从ckpt中读出来，保存为pb文件，同上一步一样，将frozen函数的is_quantify设为False就可以了：
  1. import tensorflow as tf
  2. import os.path
  3. from model import build_model
  4. from tensorflow.python.framework import graph_util

  6. # 创建推理图
  7. def create_inference_graph():
  8.     """Build the mnist model for evaluation."""
  9. # 调用网络，Create an output to use for inference.
  10.     logits = build_model(is_training=False)
  11.     return logits
  12.     # # 得到分类输出  
  13.     # tf.nn.softmax(logits, name='output')
  14. def load_variables_from_checkpoint(sess, start_checkpoint):
  15.     """Utility function to centralize checkpoint restoration.
  16.     Args:
  17.       sess: TensorFlow session.
  18.       start_checkpoint: Path to saved checkpoint on disk.
  19.     """
  20.     saver = tf.train.Saver(tf.global_variables())
  21.     saver.restore(sess, start_checkpoint)

  23. def frozen(is_quantify,ckpt,pbtxt):
  24.     # Create the model and load its weights.
  25.     init = tf.global_variables_initializer()
  26.     with tf.Session() as sess:
  27.         sess.run(init)
  28. # 推理图
  29.         logits = create_inference_graph()  
  30. # 加入create_eval_graph()，转化为tflite可接受的格式。以下语句中有路径的，记得改路径。
  31.         if is_quantify:
  32.             tf.contrib.quantize.create_eval_graph()
  33.         load_variables_from_checkpoint(sess, ckpt)
  34.         # Turn all the variables into inline constants inside the graph and save it.
  35. # 固化 frozen：ckpt + pbtxt
  36.         frozen_graph_def = graph_util.convert_variables_to_constants(
  37.             sess, sess.graph_def, ['y_conv'])
  38. # 保存最终的pb模型
  39.         tf.train.write_graph(
  40.             frozen_graph_def,
  41.             os.path.dirname(pbtxt),
  42.             os.path.basename(pbtxt),
  43.             as_text=False)
  44.         tf.logging.info('Saved frozen graph to %s', pbtxt)

  46. if __name__ == "__main__":
  47.     ckpt = './checkpoint/mnist.ckpt'
  48.     pbtxt = 'mnist_frozen_graph.pb'
  49.     frozen(False,ckpt,pbtxt)
  50.     #is_quantify False   mnist_frozen_graph_not_28x28.pb
  51.     # ckpt = './checkpoint_not/mnist.ckpt'
  52.     # pbtxt = 'mnist_frozen_graph_not.pb'
  53.     # frozen(False,ckpt,pbtxt)
  54.     # ckpt = './test/mnist.ckpt'
  55.     # pbtxt = 'test.pb'
  56.     # frozen(False,ckpt,pbtxt)
复制代码四.将pb模型转为rknn
     由于量化rknn模型需要相应图片集，因此我们先要获取相应的数据集进入mnist数据目录下，解压t10k-images-idx3-ubyte.gz，然后运行get_image.py,将原先压缩的数据转为图片，同时得到量化需要的dataset.txt文件。
    get_image.py：
  1. import struct
  2. import numpy as np
  3. #import matplotlib.pyplot as plt
  4. import PIL.Image
  5. from PIL import Image
  6. import os

  8. os.system("mkdir ../MNIST_data/mnist_test")
  9. filename='../MNIST_data/t10k-images.idx3-ubyte'
  10. dataset = './dataset.txt'
  11. binfile=open(filename,'rb')
  12. buf=binfile.read()
  13. index=0
  14. data_list = []
  15. magic,numImages,numRows,numColumns=struct.unpack_from('>IIII',buf,index)
  16. index+=struct.calcsize('>IIII')
  17. for image in range(0,numImages):
  18.     im=struct.unpack_from('>784B',buf,index)
  19.     index+=struct.calcsize('>784B')
  20.     im=np.array(im,dtype='uint8')
  21.     im=im.reshape(28,28)
  22.     im=Image.fromarray(im)
  23.     im.save('../MNIST_data/mnist_test/test_%s.jpg'%image,'jpeg')
  24.     data_list.append('../MNIST_data/mnist_test/test_%s.jpg\n'%image)
  25. with open(dataset,'w+') as ff:
  26.     ff.writelines(data_list)
复制代码rknn_transfer.py:
  1. from rknn.api import RKNN

  3. def common_transfer(pb_name,export_name):
  4.         ret = 0
  5.         #看具体log 传入verbose=True
  6.         rknn = RKNN()
  7.         #灰度图无需此步操作
  8.         # rknn.config(channel_mean_value='', reorder_channel='')
  9.         print('--> Loading model')

  11.         ret = rknn.load_tensorflow(
  12.                 tf_pb='./mnist_frozen_graph.pb',
  13.                 inputs=['x'],
  14.                 outputs=['y_conv'],
  15.                 input_size_list=[[28,28,1]])
  16.         if ret != 0:
  17.                 print('load_tensorflow error')
  18.                 rknn.release()
  19.                 return ret
  20.         print('done')
  21.         print('--> Building model')
  22.         rknn.build(do_quantization=False)
  23.         print('done')
  24.         # 导出保存rknn模型文件
  25.         rknn.export_rknn('./mnist.rknn')
  26.         # Release RKNN Context
  27.         rknn.release()
  28.         return ret

  30. def quantify_transfer(pb_name,dataset_name,export_name):
  31.         ret = 0
  32.         print(pb_name,dataset_name,export_name)
  33.         rknn = RKNN()
  34.         rknn.config(channel_mean_value='', reorder_channel='',quantized_dtype='dynamic_fixed_point-8')
  35.         print('--> Loading model')
  36.         ret = rknn.load_tensorflow(
  37.                 tf_pb=pb_name,
  38.                 inputs=['x'],
  39.                 outputs=['y_conv'],
  40.                 input_size_list=[[28,28,1]])
  41.         if ret != 0:
  42.                 print('load_tensorflow error')
  43.                 rknn.release()
  44.                 return ret
  45.         print('done')
  46.         print('--> Building model')
  47.         rknn.build(do_quantization=True,dataset=dataset_name)
  48.         print('done')
  49.         # 导出保存rknn模型文件
  50.         rknn.export_rknn(export_name)
  51.         # Release RKNN Context
  52.         rknn.release()
  53.         return ret
  54. if __name__ == '__main__':
  55.         #pb转化为rknn模型
  56.         pb_name = './mnist_frozen_graph.pb'
  57.         export_name = './mnist.rknn'
  58.         ret = common_transfer(pb_name,export_name)
  59.         if ret != 0:
  60.                 print('======common transfer error !!===========')
  61.         else:
  62.                 print('======common transfer ok !!===========')
  63.         dataset_name = './dataset.txt'
  64.         export_name = './mnist_quantization.rknn'
  65.         #pb转化为量化的rknn模型
  66.         quantify_transfer(pb_name,dataset_name,export_name)
  67.         if ret != 0:
  68.                 print('======quantization transfer 10000 error !!===========')
  69.         else:
  70.                 print('======quantization transfer 10000 ok !!===========')

复制代码五.对比pb和rknn的推理结果，比较他们的准确度
     分别运行tf_predict.py,rknn_predict.py得到tf模型,rknn模型,量化的rknn模型的运行结果:
     tf_predict.py
  1. #! -*- coding: utf-8 -*-
  2. from __future__ import absolute_import, unicode_literals
  3. from tensorflow.examples.tutorials.mnist import input_data
  4. import tensorflow as tf

  6. mnist = input_data.read_data_sets("../MNIST_data/", one_hot=True)
  7. origin_test = mnist.test.images
  8. reshape_test = []
  9. for t in origin_test:
  10.     b = t.reshape(28,28)
  11.     reshape_test.append(b)
  12. for length in [100,500,1000,10000]:
  13.     with tf.Graph().as_default():
  14.         output_graph_def = tf.GraphDef()
  15.         output_graph_path = './mnist_frozen_graph.pb'

  17.         with open(output_graph_path, 'rb') as f:
  18.             output_graph_def.ParseFromString(f.read())
  19.             _ = tf.import_graph_def(output_graph_def, name="")
  20.      
  21.         with tf.Session() as sess:
  22.             sess.run(tf.global_variables_initializer())
  23.             input = sess.graph.get_tensor_by_name("x:0")
  24.             output = sess.graph.get_tensor_by_name("y_conv:0")
  25.             y_conv_2 = sess.run(output, feed_dict={input:reshape_test[0:length]})
  26.             y_2 = mnist.test.labels[0:length]
  27.             print("first image:",y_conv_2[0])
  28.             correct_prediction_2 = tf.equal(tf.argmax(y_conv_2, 1), tf.argmax(y_2, 1))
  29.             accuracy_2 = tf.reduce_mean(tf.cast(correct_prediction_2, "float"))
  30.             print('%d:'%length,"check accuracy %g" % sess.run(accuracy_2))
复制代码rknn_predict.py
  1. import numpy as np
  2. from PIL import Image
  3. from rknn.api import RKNN
  4. import cv2
  5. from tensorflow.examples.tutorials.mnist import input_data
  6. import tensorflow as tf

  8. mnist = input_data.read_data_sets("../MNIST_data/", one_hot=True)
  9. print(mnist.test.images[0].shape)
  10. # 解析模型的输出，获得概率最大的手势和对应的概率
  11. def get_predict(probability):
  12.     data = probability[0][0]
  13.     data = data.tolist()
  14.     max_prob = max(data)
  15.     return data.index(max_prob), max_prob
  16. # return data.index(max_prob), max_prob;
  17. def load_model(model_name):
  18.     # 创建RKNN对象
  19.     rknn = RKNN()
  20.     # 载入RKNN模型
  21.     print('-->loading model')
  22.     rknn.load_rknn(model_name)
  23.     print('loading model done')
  24.     # 初始化RKNN运行环境
  25.     print('--> Init runtime environment')
  26.     ret = rknn.init_runtime()
  27.     if ret != 0:
  28.        print('Init runtime environment failed')
  29.        exit(ret)
  30.     print('done')
  31.     return rknn
  32. def predict(rknn,length):
  33.     acc_count = 0
  34.     for i in range(length):
  35.         # im = mnist.test.images[i]
  36.         im = Image.open("../MNIST_data/mnist_test/test_%d.jpg"%i)   # 加载图片
  37.         im = im.resize((28,28),Image.ANTIALIAS)
  38.         im = np.asarray(im)
  39.         outputs = rknn.inference(inputs=[im])
  40.         pred, prob = get_predict(outputs)
  41.         if i ==0:
  42.             print(outputs)
  43.             print(prob)
  44.             print(pred)
  45.         if i ==100 or i ==500 or i ==1000 or i ==10000:
  46.             result = float(acc_count)/i
  47.             print('result%d:'%i,result)
  48.         if list(mnist.test.labels[i]).index(1) == pred:
  49.             acc_count += 1
  50.     result = float(acc_count)/length
  51.     print('result:',result)
  52.     # acc_count = 0
  53.     # length = len(mnist.test.images)
  54.     # for i in range(length):
  55.         # im = mnist.test.images[i]# 加载图片
  56.         # outputs = rknn.inference(inputs=[im])   # 运行推理，得到推理结果
  57.         # pred, prob = get_predict(outputs)     # 将推理结果转化为可视信息
  58.         # if i%100 == 0:
  59.             # print(prob)
  60.             # print(pred)
  61.             # print(acc_count)
  62.             # print(list(mnist.test.labels[i]).index(1))
  63.         # if list(mnist.test.labels[i]).index(1) == pred:
  64.             # acc_count += 1
  65.     # result = float(acc_count)/length
  66.     # print('result:',result)
  67. if __name__=="__main__":
  68.     #此处要改成相应的量化或者非量化rknn模型
  69.     model_name = './mnist.rknn'
  70.     length = 10000
  71.     rknn = load_model(model_name)
  72.     predict(rknn,length)

  74.     rknn.release()
复制代码得到最终的结果对比图表如下:


作者: peng    时间: 2019-9-20 08:43
本帖最后由 peng 于 2019-9-27 14:50 编辑

《量化训练》关键的两个函数为：tf.contrib.quantize.create_training_graph，tf.contrib.quantize.create_eval_graph(input_graph=g)，在上一节中代码中已经提供了，具体可看以及几个步骤：第一步，修改train.py,将create_training_graph()的is_quantify设为True.然后重新训练模型。
  1. if __name__ == "__main__":
  2. ckpt = './checkpoint_fake/mnist.ckpt'
  3. point_dir = './checkpoint_fake'
  4. pbtxt = 'mnist_fake.pbtxt'
  5. g1 = create_training_graph(True)
  6. train_network(g1,ckpt,point_dir,pbtxt)
复制代码第二步，修改freese.py,将frozen()的is_quantify设为True.重新固化成pb文件，
  1. if __name__ == "__main__":
  2. ckpt = './checkpoint_fake/mnist.ckpt'
  3. pbtxt = 'mnist_frozen_graph_fake.pb'
  4. frozen(True,ckpt,pbtxt)
复制代码
第三步，使用toco工具将pb模型转化为全量化的tflite模型(输入，输出和权值都变成int8类型)

  1. #!/bin/sh

  3. toco \
  4. --graph_def_file=mnist_frozen_graph_fake.pb \
  5. --output_file=mnist_fakequantize.tflite \
  6. --output_format=TFLITE \
  7. --inference_type=QUANTIZED_UINT8 \
  8. --input_shapes=1,28,28 \
  9. --input_arrays=x \
  10. --output_arrays=y_conv \
  11. --mean_values=0 \
  12. --std_dev_values=256 \
  13. --change_concat_input_ranges=false --allow_custom_ops
复制代码

第四步，将tflite模型转为rknn，修改rknn_transfer.py:

  1. def tflite_transfer():
  2. rknn = RKNN()
  3. print('--> Loading model')
  4. ret = rknn.load_tflite(model = './mnist_fakequantize.tflite')
  5. print('done')
  6. print('--> Building model')
  7. rknn.build(do_quantization=False)
  8. print('done')
  9. # 导出保存rknn模型文件
  10. rknn.export_rknn('./mnist_quantization_fake.rknn')
  11. # Release RKNN Context
  12. rknn.release()
复制代码
  1. if __name__ == '__main__':
  2. tflite_transfer()
复制代码第五步，分别运行tflite_predict.py,和rknn_predict.py，对比tflite和rknn推理的结果和准确度：
tflite_predict.py
  1. mport tensorflow as tf
  2. from tensorflow.examples.tutorials.mnist import input_data
  3. from PIL import Image

  5. mnist = input_data.read_data_sets("../../MNIST_data/", one_hot=True)
  6. length = 100
  7. # 加载模型并分配张量
  8. interpreter = tf.lite.Interpreter(model_path="mnist_fakequantize.tflite")
  9. interpreter.allocate_tensors()
  10. # 获取输入输出张量
  11. input_details = interpreter.get_input_details()
  12. print(input_details)
  13. output_details = interpreter.get_output_details()
  14. print(output_details)



  18. acc_count = 0
  19. for i in range(length):
  20. #im = mnist.test.images[i]
  21. im = Image.open("../../MNIST_data/mnist_test/test_%d.jpg"%i) # 加载图片
  22. im = im.resize((28,28),Image.ANTIALIAS)
  23. im = np.asarray(im)
  24. # print(im.dtype)
  25. input_shape = input_details[0]['shape']
  26. input_data = im.reshape(1,28,28)
  27. # print(input_data.dtype)
  28. interpreter.set_tensor(input_details[0]['index'], input_data)

  30. interpreter.invoke()
  31. output_data = interpreter.get_tensor(output_details[0]['index'])
  32. # print(output_data)
  33. # print(output_data.argmax(axis=1)[0])
  34. if i <3:
  35. print(output_data)

  37. if list(mnist.test.labels[i]).index(1) == output_data.argmax(axis=1)[0]:
  38. acc_count += 1
  39. result = float(acc_count)/length
  40. print('result:',result)
复制代码rknn_predict.py
  1. import numpy as np
  2. from PIL import Image
  3. from rknn.api import RKNN
  4. import cv2
  5. from tensorflow.examples.tutorials.mnist import input_data
  6. import tensorflow as tf

  8. mnist = input_data.read_data_sets("../MNIST_data/", one_hot=True)
  9. print(mnist.test.images[0].shape)
  10. # 解析模型的输出，获得概率最大的手势和对应的概率
  11. def get_predict(probability):
  12. data = probability[0][0]
  13. data = data.tolist()
  14. max_prob = max(data)
  15. return data.index(max_prob), max_prob
  16. # return data.index(max_prob), max_prob;
  17. def load_model(model_name):
  18. # 创建RKNN对象
  19. rknn = RKNN()
  20. # 载入RKNN模型
  21. print('-->loading model')
  22. rknn.load_rknn(model_name)
  23. print('loading model done')
  24. # 初始化RKNN运行环境
  25. print('--> Init runtime environment')
  26. ret = rknn.init_runtime()
  27. if ret != 0:
  28. print('Init runtime environment failed')
  29. exit(ret)
  30. print('done')
  31. return rknn
  32. def predict(rknn,length):
  33. acc_count = 0
  34. for i in range(length):
  35. im = mnist.test.images[i]
  36. # im = Image.open("../MNIST_data/mnist_test/test_%d.jpg"%i) # 加载图片
  37. # im = im.resize((28,28),Image.ANTIALIAS)
  38. # im = np.asarray(im)
  39. im = im.reshape(1,28,28)
  40. outputs = rknn.inference(inputs=[im])
  41. pred, prob = get_predict(outputs)
  42. if i ==0:
  43. print(outputs)
  44. print(prob)
  45. print(pred)
  46. if i ==100 or i ==500 or i ==1000 or i ==10000:
  47. result = float(acc_count)/i
  48. print('result%d:'%i,result)
  49. if list(mnist.test.labels[i]).index(1) == pred:
  50. acc_count += 1
  51. result = float(acc_count)/length
  52. print('result:',result)
  53. if __name__=="__main__":
  54. #此处要改成相应的量化或者非量化rknn模型
  55. model_name = './mnist_quantization_fake.rknn'
  56. length = 10000
  57. rknn = load_model(model_name)
  58. predict(rknn,length)

  60. rknn.release()
复制代码最终运行结果如下表所示：
[attach]603[/attach]
作者: puyanan    时间: 2019-9-21 14:50
本帖最后由 puyanan 于 2019-9-21 14:52 编辑

楼主，你好，这个例子里 测试模型没有dropout层，输入只有一个，正好rknn只支持单输入。如果一个网络里的每一个conv2D都有BN层，最后还有Dropout层。tensorflow的pb文件中输入有两个，“input”和“phase_train”，其中phase_train节点输出连接每一个conv2D，怎么把这种情况改成“单输入”呢？      修改模型，将phase_train固定为Flase再重新固化成pb模型，请问用什么工具实现呢？
作者: jefferyzhang    时间: 2019-9-23 09:36
puyanan 发表于 2019-9-21 14:50
楼主，你好，这个例子里 测试模型没有dropout层，输入只有一个，正好rknn只支持单输入。如果一个网络里的每 ...

rknntoolkit 1.2 已经支持多输入了，不用太纠结单输入问题。
bn和dropout层都是训练层，推理模式时候freeze过程就会被tf固定参数，也不存在多输入问题。
作者: RK用户    时间: 2019-10-16 17:17
本帖最后由 RK用户 于 2019-10-18 16:09 编辑

我运行了训练代码
报错
python3 train.py

/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/dtypes.py:516: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint8 = np.dtype([("qint8", np.int8, 1)])
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/dtypes.py:517: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_quint8 = np.dtype([("quint8", np.uint8, 1)])
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/dtypes.py:518: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint16 = np.dtype([("qint16", np.int16, 1)])
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/dtypes.py:519: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_quint16 = np.dtype([("quint16", np.uint16, 1)])
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/dtypes.py:520: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint32 = np.dtype([("qint32", np.int32, 1)])
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  np_resource = np.dtype([("resource", np.ubyte, 1)])
/usr/local/lib/python3.6/dist-packages/tensorboard/compat/tensorflow_stub/dtypes.py:541: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint8 = np.dtype([("qint8", np.int8, 1)])
/usr/local/lib/python3.6/dist-packages/tensorboard/compat/tensorflow_stub/dtypes.py:542: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_quint8 = np.dtype([("quint8", np.uint8, 1)])
/usr/local/lib/python3.6/dist-packages/tensorboard/compat/tensorflow_stub/dtypes.py:543: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint16 = np.dtype([("qint16", np.int16, 1)])
/usr/local/lib/python3.6/dist-packages/tensorboard/compat/tensorflow_stub/dtypes.py:544: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_quint16 = np.dtype([("quint16", np.uint16, 1)])
/usr/local/lib/python3.6/dist-packages/tensorboard/compat/tensorflow_stub/dtypes.py:545: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint32 = np.dtype([("qint32", np.int32, 1)])
/usr/local/lib/python3.6/dist-packages/tensorboard/compat/tensorflow_stub/dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  np_resource = np.dtype([("resource", np.ubyte, 1)])
WARNING:tensorflow:From /work/PythonTensorFlow/pymnist_rknn/model.py:7: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.

WARNING:tensorflow:From train.py:11: read_data_sets (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use alternatives such as official/mnist/dataset.py from tensorflow/models.
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:260: maybe_download (from tensorflow.contrib.learn.python.learn.datasets.base) is deprecated and will be removed in a future version.
Instructions for updating:
Please write your own downloading logic.
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:262: extract_images (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use tf.data to implement this functionality.
Extracting ./MNIST_data/train-images-idx3-ubyte.gz
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:267: extract_labels (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use tf.data to implement this functionality.
Extracting ./MNIST_data/train-labels-idx1-ubyte.gz
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:110: dense_to_one_hot (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use tf.one_hot on tensors.
Extracting ./MNIST_data/t10k-images-idx3-ubyte.gz
Extracting ./MNIST_data/t10k-labels-idx1-ubyte.gz
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:290: DataSet.__init__ (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use alternatives such as official/mnist/dataset.py from tensorflow/models.
WARNING:tensorflow:From train.py:14: The name tf.get_default_graph is deprecated. Please use tf.compat.v1.get_default_graph instead.

WARNING:tensorflow:From /work/PythonTensorFlow/pymnist_rknn/model.py:11: The name tf.truncated_normal is deprecated. Please use tf.random.truncated_normal instead.

WARNING:tensorflow:From /work/PythonTensorFlow/pymnist_rknn/model.py:22: The name tf.nn.max_pool is deprecated. Please use tf.nn.max_pool2d instead.

WARNING:tensorflow:From /work/PythonTensorFlow/pymnist_rknn/model.py:57: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.
Instructions for updating:
Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.
finaloutput: Tensor("y_conv:0", shape=(?, 10), dtype=float32)
WARNING:tensorflow:From train.py:18: The name tf.log is deprecated. Please use tf.math.log instead.

cost: Tensor("Neg:0", shape=(), dtype=float32)
WARNING:tensorflow:From train.py:24: The name tf.train.AdamOptimizer is deprecated. Please use tf.compat.v1.train.AdamOptimizer instead.

WARNING:tensorflow:From train.py:50: The name tf.global_variables_initializer is deprecated. Please use tf.compat.v1.global_variables_initializer instead.

WARNING:tensorflow:From train.py:52: The name tf.train.Saver is deprecated. Please use tf.compat.v1.train.Saver instead.

2019-10-16 17:01:13.549217: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-10-16 17:01:13.574709: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2893325000 Hz
2019-10-16 17:01:13.574937: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x19eb300 executing computations on platform Host. Devices:
2019-10-16 17:01:13.574957: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): <undefined>, <undefined>
2019-10-16 17:01:13.672102: W tensorflow/compiler/jit/mark_for_compilation_pass.cc:1412] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set that envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA is active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.
step 0, training accuracy 0.16
step 100, training accuracy 0.84
step 200, training accuracy 0.9
step 300, training accuracy 0.92
step 400, training accuracy 1
step 500, training accuracy 0.98
step 600, training accuracy 0.96
step 700, training accuracy 0.94
step 800, training accuracy 0.96
step 900, training accuracy 1
step 1000, training accuracy 0.98
step 1100, training accuracy 0.96
step 1200, training accuracy 0.96
step 1300, training accuracy 0.96
step 1400, training accuracy 0.96
step 1500, training accuracy 1
step 1600, training accuracy 1
step 1700, training accuracy 0.92
step 1800, training accuracy 0.94
step 1900, training accuracy 0.96
step 2000, training accuracy 0.98
step 2100, training accuracy 0.96
step 2200, training accuracy 1
step 2300, training accuracy 1
step 2400, training accuracy 1
step 2500, training accuracy 0.98
step 2600, training accuracy 1
step 2700, training accuracy 0.96
step 2800, training accuracy 0.96
step 2900, training accuracy 0.98
step 3000, training accuracy 0.98
step 3100, training accuracy 0.96
step 3200, training accuracy 0.96
step 3300, training accuracy 0.98
step 3400, training accuracy 1
step 3500, training accuracy 1
step 3600, training accuracy 0.96
step 3700, training accuracy 1
step 3800, training accuracy 1
step 3900, training accuracy 1
step 4000, training accuracy 0.98
step 4100, training accuracy 0.96
step 4200, training accuracy 1
step 4300, training accuracy 1
step 4400, training accuracy 0.98
step 4500, training accuracy 0.96
step 4600, training accuracy 1
step 4700, training accuracy 1
step 4800, training accuracy 1
step 4900, training accuracy 0.96
step 5000, training accuracy 1
step 5100, training accuracy 1
step 5200, training accuracy 1
step 5300, training accuracy 1
step 5400, training accuracy 0.98
step 5500, training accuracy 1
step 5600, training accuracy 0.98
step 5700, training accuracy 1
step 5800, training accuracy 1
step 5900, training accuracy 1
step 6000, training accuracy 1
step 6100, training accuracy 0.98
step 6200, training accuracy 0.96
step 6300, training accuracy 0.98
step 6400, training accuracy 0.94
step 6500, training accuracy 1
step 6600, training accuracy 0.98
step 6700, training accuracy 0.98
step 6800, training accuracy 1
step 6900, training accuracy 1
step 7000, training accuracy 1
step 7100, training accuracy 0.98
step 7200, training accuracy 1
step 7300, training accuracy 1
step 7400, training accuracy 1
step 7500, training accuracy 1
step 7600, training accuracy 1
step 7700, training accuracy 1
step 7800, training accuracy 1
step 7900, training accuracy 1
step 8000, training accuracy 0.98
step 8100, training accuracy 1
step 8200, training accuracy 1
step 8300, training accuracy 0.98
step 8400, training accuracy 1
step 8500, training accuracy 0.98
step 8600, training accuracy 1
step 8700, training accuracy 0.98
step 8800, training accuracy 1
step 8900, training accuracy 1
step 9000, training accuracy 0.98
step 9100, training accuracy 1
step 9200, training accuracy 1
step 9300, training accuracy 1
step 9400, training accuracy 1
step 9500, training accuracy 1
step 9600, training accuracy 1
step 9700, training accuracy 1
step 9800, training accuracy 1
step 9900, training accuracy 1
2019-10-16 17:11:55.047290: W tensorflow/core/framework/allocator.cc:107] Allocation of 1003520000 exceeds 10% of system memory.
2019-10-16 17:11:55.743553: W tensorflow/core/framework/allocator.cc:107] Allocation of 250880000 exceeds 10% of system memory.
2019-10-16 17:11:55.986352: W tensorflow/core/framework/allocator.cc:107] Allocation of 501760000 exceeds 10% of system memory.
terminate called after throwing an instance of 'std::bad_alloc'
  what():  std::bad_alloc
Aborted (core dumped)
野指针
这个有人遇到吗是不是我设置错了

内存不足的问题

作者: yuys    时间: 2019-11-8 15:16
rk3399pro上运行rknn_transfer.py:失败，这个不能再板子上运行吗？
作者: yuys    时间: 2019-11-8 18:12
在pc上运行成功，我想问mnist.rknn 6.25MB； mnist_frozen_graph.pb为12.4MB，为啥common_transfer转换后权重为pb权重大小的一半呢？
作者: yuys    时间: 2019-11-8 18:44
量化quantized_dtype='dynamic_fixed_point-8'的模型是mnist_quantization.rknn大小为3.13MB是mnist.rknn（6.25MB）权重大小的一半； mnist.rknn 6.25MB是 mnist_frozen_graph.pb（12.4MB）权重大小的一半；
是不是mnist_frozen_graph.pb对应数据类型：float 32;
mnist.rknn对应数据类型是int16；
mnist_quantization.rknn对应数据类型是int8?
作者: w910820618    时间: 2019-12-16 10:51
  1. from rknn.api import RKNN

  3. INPUT_SIZE = 128

  5. if __name__ == '__main__':
  6.     rknn = RKNN()

  8.     rknn.config(channel_mean_value='0 0 0 255', reorder_channel='0 1 2')
  9.     print('--> Loading model')
  10.     rknn.load_tensorflow(tf_pb='mars-small128.pb',
  11.             inputs=['images'], outputs=['features'], input_size_list=[[128, 64, 3]])

  13.     print('done')

  15.     print('--> Building model')
  16.     rknn.build(do_quantization=False)
  17.     print('done')

  19.     rknn.export_rknn('./mars-small128.rknn')
  20.     rknn.release()
复制代码
作者: mahongwei    时间: 2020-4-17 23:26

2020-04-17 23:23:52.630269: W tensorflow/compiler/jit/mark_for_compilation_pass.cc:1412] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set that envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA is active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.
pyfunc exception: Cannot feed value of shape (1, 28, 28, 1) for Tensor 'x:0', which has shape '(?, 28, 28)'
Traceback (most recent call last):
  File "/home/ol/.local/lib/python3.5/site-packages/rknn/base/acuitylib/converter/pyfunc_builder.py", line 30, in pyfunc_call
    return function_builder(func_str)(self, node, tensor)
  File "<string>", line 9, in r_fc_tensor_dot_rule_get_param_axis_pre_condition
  File "rknn/base/acuitylib/converter/convert_tf.py", line 134, in rknn.base.acuitylib.converter.convert_tf.convert_tf.shape_pick
  File "rknn/base/acuitylib/converter/tensorflowloader.py", line 205, in rknn.base.acuitylib.converter.tensorflowloader.TF_Graph_Preprocess.query_tensors
  File "rknn/base/acuitylib/converter/tf_util.py", line 371, in rknn.base.acuitylib.converter.tf_util.TFProto_Util.query_tensor
  File "rknn/base/acuitylib/converter/tf_util.py", line 374, in rknn.base.acuitylib.converter.tf_util.TFProto_Util.query_tensor
  File "rknn/base/acuitylib/converter/tf_util.py", line 407, in rknn.base.acuitylib.converter.tf_util.TFProto_Util.query_tensor
  File "/home/ol/.local/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 950, in run
    run_metadata_ptr)
  File "/home/ol/.local/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 1149, in _run
    str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (1, 28, 28, 1) for Tensor 'x:0', which has shape '(?, 28, 28)'


说我的输入不对？这个怎么解决啊
作者: Bone1    时间: 2021-4-21 21:00
RK用户 发表于 2019-10-16 17:17
我运行了训练代码
报错
python3 train.py

为了避免内存问题，参考对训练集的处理方法，将一次喂入神经网络的测试数据量减小，从而减小内存分配。
作者: gaoxukkk    时间: 2024-3-10 10:21
请问在进行mnist模型的量化时 出现：could not convert string to float 这是为什么



欢迎光临 Toybrick (https://t.rock-chips.com/) Powered by Discuz! X3.3