rk3399 pro不支持 一维卷积

用tensorflow conv2d 将卷积核由5*5改成1*5，pb到rknn模型转换没有问题 ，但是在调用做推理是，加载1*64的数据报错

问问题要把错误信息贴出来

报错信息是：Aborted (core dumped)，日志是空的模型结构：

1. import tensorflow as tf
   
2. 
   
3. def weight_variable(shape):
   
4.         tf.set_random_seed(1)
   
5.         return tf.Variable(tf.truncated_normal(shape, stddev=0.1))
   
6. def bias_variable(shape):
   
7.         return tf.Variable(tf.constant(0.0, shape=shape))
   
8. 
   
9. def conv2d(x, W):
   
10.         return tf.nn.conv2d(x, W, strides=[1,1,1,1], padding='SAME')
    
11. 
    
12. def max_pool_2x2(z):
    
13.         return tf.nn.max_pool(z, ksize=[1,1,2,1], strides=[1,1,2,1], padding='SAME')
    
14. X = tf.placeholder(tf.float32, [None, 1, 64, 1], name="input_x")
    
15. y = tf.placeholder(tf.float32, [None, 10], name="input_y")
    
16. kp = tf.placeholder_with_default(1.0, shape=(), name="keep_prob")
    
17. lam = tf.placeholder(tf.float32, name="lamda")
    
18.         #conv1
    
19. W_conv1 = weight_variable([1,5,1,32])
    
20. b_conv1 = bias_variable([32])
    
21. z1 = tf.nn.relu(conv2d(X, W_conv1) + b_conv1)
    
22. maxpool1 = max_pool_2x2(z1) #max_pool1完后maxpool1维度为[?,32,32,32]
    
23. 
    
24.         #conv2
    
25. W_conv2 = weight_variable([1,5,32,64])
    
26. b_conv2 = bias_variable([64])
    
27. z2 = tf.nn.relu(conv2d(maxpool1, W_conv2) + b_conv2)
    
28. maxpool2 = max_pool_2x2(z2) #max_pool2,shape [?,16,16,64]
    
29. print("maxpool2.shape",maxpool2.get_shape())
    
30. 
    
31.         #conv3  效果比较好的一次模型是没有这一层，只有两次卷积层，隐藏单元100，训练20次
    
32.         # W_conv3 = weight_variable([5, 5, 64, 128])
    
33.         # b_conv3 = bias_variable([128])
    
34.         # z3 = tf.nn.relu(conv2d(maxpool2, W_conv3) + b_conv3)
    
35.         # maxpool3 = max_pool_2x2(z3)  # max_pool3,shape [?,8,8,128]
    
36. 
    
37.         #full connection1
    
38. W_fc1 = weight_variable([1*16*64, 200])
    
39. b_fc1 = bias_variable([200])
    
40. maxpool2_flat = tf.reshape(maxpool2, [-1, 2*16*64])
    
41. z_fc1 = tf.nn.relu(tf.matmul(maxpool2_flat, W_fc1) + b_fc1)
    
42. #z_fc1_drop = tf.nn.dropout(z_fc1, keep_prob=kp)
    
43. 
    
44. #softmax layer
    
45. W_fc2 = weight_variable([200, 10])
    
46. b_fc2 = bias_variable([10])
    
47. z_fc2 = tf.add(tf.matmul(z_fc1, W_fc2),b_fc2, name="outlayer")
    
48. prob = tf.nn.softmax(z_fc2, name="probability")
    
49. pred = tf.argmax(prob, 1, output_type="int32", name="predict")
    
50. 
    
51. init = tf.global_variables_initializer()
    
52. with tf.Session() as sess:
    
53.         sess.run(init)
    
54.         saver = tf.train.Saver({'W_conv1':W_conv1, 'b_conv1':b_conv1, 'W_conv2':W_conv2, 'b_conv2':b_conv2,
    
55.                                         'W_fc1':W_fc1, 'b_fc1':b_fc1, 'W_fc2':W_fc2, 'b_fc2':b_fc2})
    
56.         saver.save(sess, "make_model/cnn_model.ckpt")
    
57.         output_graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, output_node_names=['predict'])
    
58.         with tf.gfile.FastGFile('make_model_tl/digital_gesture.pb', mode='wb') as f:
    
59.                 f.write(output_graph_def.SerializeToString())

复制代码转换模型的代码是：

1. from rknn.api import RKNN
   
2. 
   
3. INPUT_SIZE = 64
   
4. 
   
5. if __name__ == '__main__':
   
6.         # 创建RKNN执行对象
   
7.         rknn = RKNN()
   
8. 
   
9.         #初始化RKNN运行环境，指定host=’rk3399pro’，后续的操作将会运行在NPU上
   
10.         print('--> Init runtime environment')
    
11.         ret = rknn.init_runtime()
    
12.         if ret != 0:
    
13.                 print('Init runtime environment failed')
    
14.                 exit(ret)
    
15.                 print('done')
    
16. 
    
17.         # 配置模型输入，用于NPU对数据输入的预处理
    
18.         # channel_mean_value='0 0 0 255'，那么模型推理时，将会对RGB数据做如下转换
    
19.         # (R - 0)/255, (G - 0)/255, (B - 0)/255。推理时，RKNN模型会自动做均值和归一化处理
    
20.         # reorder_channel=’0 1 2’用于指定是否调整RBG顺序，设置成0 1 2即按输入的RGB顺序不做调整
    
21.         # reorder_channel=’2 1 0’表示交换0和2通道，如果输入是RGB，将会被调整为BGR。如果是BGR将会被
    
22.         # 调整为RGB
    
23.         #rknn.config(channel_mean_value='0 0 0 255', reorder_channel='0 1 2')
    
24.         rknn.config(channel_mean_value='0 0 0 1',reorder_channel='0 1 2')
    
25.         # 加载TensorFlow模型
    
26.         # tf_pb='../digital_gesture_recognition/model_2500/digital_gesture.pb'指定待转换的TensorFlow模型
    
27.         # inputs指定模型中的输入节点
    
28.         # outputs指定模型中输出节点
    
29.         # input_size_list指定模型输入的大小
    
30.         print('--> Loading model')
    
31.         rknn.load_tensorflow(tf_pb='make_model_tl/digital_gesture.pb',
    
32.                                 inputs=['input_x'],
    
33.                                 outputs=['outlayer'],
    
34.                                 input_size_list=[[1, 64, 1]])
    
35.         print('done')
    
36. 
    
37.         # 创建解析pb模型
    
38.         # do_quantization=False指定不进行量化
    
39.         # 量化会减小模型的体积和提升运算速度，但是会有精度的丢失
    
40.         print('--> Building model')
    
41.         rknn.build(do_quantization=False)
    
42.         print('done')

复制代码用pb推理成功的代码：

1. import numpy as np
   
2. 
   
3. #read image
   
4. def img_to_mat(path):
   
5.     mat = np.arange(64).reshape(1,1,64,1)
   
6.     return mat
   
7. #load pb
   
8. def load_model(pb_path):
   
9.     with tf.gfile.GFile(pb_path, "rb") as f:
   
10.         graph_def = tf.GraphDef()
    
11.         graph_def.ParseFromString(f.read())
    
12.         with tf.Graph().as_default() as graph:
    
13.             tf.import_graph_def(graph_def, name="")
    
14.             return graph
    
15. 
    
16. #do inference
    
17. def predict(graph):
    
18.     mat = img_to_mat("out.jpg")
    
19.     x = graph.get_tensor_by_name("input_x:0")
    
20.     outlayer = graph.get_tensor_by_name("outlayer:0")
    
21.     with tf.Session(graph=graph) as sess:
    
22.         np.set_printoptions(suppress=True)
    
23.         out= sess.run([outlayer],feed_dict={x:mat})
    
24.         return out
    
25. if __name__=="__main__":
    
26.     graph = load_model('make_model_tl/digital_gesture.pb')
    
27.     out = predict(graph)
    
28.     print(out)

复制代码rknn推理失败的代码：

1. from rknn.api import RKNN
   
2. from PIL import Image
   
3. import numpy as np
   
4. 
   
5. #read image
   
6. def img_to_mat(path):
   
7.     mat = np.arange(64).reshape(1,64)
   
8.     return mat
   
9. #load rknn
   
10. def load_model(rknn_path):
    
11.     rknn = RKNN(verbose=True, verbose_file='./log.txt')
    
12.      # 载入RKNN模型
    
13.     print('-->loading model')
    
14.     rknn.load_rknn(rknn_path)
    
15.     print('loading model done')
    
16. 
    
17.     # 初始化RKNN运行环境
    
18.     print('--> Init runtime  environment')
    
19.     ret = rknn.init_runtime()
    
20.     if ret != 0:
    
21.         print('Init runtime environment failed')
    
22.         exit(ret)
    
23.     print('done')
    
24. 
    
25.     return rknn
    
26. #do inference
    
27. def predict(rknn):
    
28.     mat = img_to_mat("out.jpg")
    
29.     print("mat.shape",mat.shape)
    
30.     #mat = mat.reshape(64,1)
    
31.    # print("mat.shape",mat.shape)
    
32.     out=rknn.inference(inputs=[mat])
    
33.     print('ooo')
    
34.     return out
    
35. if __name__=="__main__":
    
36.     rknn = load_model('make_model_tl/digital_gesture.rknn')
    
37.     out = predict(rknn)
    
38.     print(out)

复制代码结论：是不支持这种1*n的卷积核吗？

troy 发表于 2019-3-28 15:24
问问题要把错误信息贴出来

您好，我把code和相关信息贴出来了