Failed to get convolution algorithm解决：keras中model.predict都报错：...

东方耀

解决：keras中model.predict都报错：显存不足


Failed to get convolution algorithm.
    This is probably because cuDNN failed to initialize,

解释：允许显存动态增长，默认情况下tf和keras会分配所有的显存


如果有多块显卡：还需指定：
import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
# 选择哪一块gpu,如果是-1，就是调用cpu
os.environ['CUDA_VISIBLE_DEVICES'] = "0,1"

有个细节要注意：函数allow_gpu_memory_growth()应该放在模型构建build_model所在的文件最前面哦！

用这个就不标红了：


from tensorflow.python.keras.backend import set_session

1. 
   
2. # python 3.6 and tensorflow (both 1.x and 2.0)
   
3. def allow_gpu_memory_growth(log_device_placement=True):
   
4.     """
   
5.     Allow dynamic memory growth (by default, tensorflow allocates all gpu memory).
   
6.     This sometimes fixes the
   
7.     <<Error : Failed to get convolution algorithm.
   
8.     This is probably because cuDNN failed to initialize,
   
9.     so try looking to see if a warning log message was printed above>>.
   
10.     May hurt performance slightly (see https://www.tensorflow.org/guide/gpu).
    
11. 
    
12.     Usage: Run before any other code.
    
13. 
    
14.     :param log_device_placement: set True to log device placement (on which device the operation ran)
    
15.     :return:None
    
16.     """
    
17.     # set_session 标红不用管 Sets the global TensorFlow session
    
18.     from tensorflow.compat.v1.keras.backend import set_session
    
19.     import tensorflow as tf
    
20.     config = tf.compat.v1.ConfigProto()
    
21.     config.gpu_options.allow_growth = True  # dynamically grow the memory used on the GPU
    
22.     config.log_device_placement = log_device_placement
    
23.     sess = tf.compat.v1.Session(config=config)
    
24.     set_session(sess)

复制代码

遇到这样一个问题：
直接用pycharm启动服务 并使用 没问题
但是用终端启动服务，则报错：显存不够 显存明显是够的
报错：Failed to get convolution algorithm. This is probably because cuDNN failed to initialize,
问题一定在：两者启动的环境配置不一样竟然解决了：我看到一个提示说 cudnn一个是7.4.0  一个是7.6.0 说什么不一致  发现我本机安装的是7.4.0
然后我升级为7.6.0 就成功啦！ 当然还得加allow_gpu_memory_growth()这个函数  cuda的版本是10.0的

这次要改成这样的：

1. import os
   
2. import tensorflow as tf
   
3. import keras
   
4. os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
   
5. # 选择哪一块gpu,如果是-1，就是调用cpu
   
6. os.environ['CUDA_VISIBLE_DEVICES'] = "0"
   
7. 
   
8. 
   
9. # python 3.6 and tensorflow (both 1.x and 2.0)
   
10. def allow_gpu_memory_growth(log_device_placement=True):
    
11.     """
    
12.     Allow dynamic memory growth (by default, tensorflow allocates all gpu memory).
    
13.     This sometimes fixes the
    
14.     <<Error : Failed to get convolution algorithm.
    
15.     This is probably because cuDNN failed to initialize,
    
16.     so try looking to see if a warning log message was printed above>>.
    
17.     May hurt performance slightly (see https://www.tensorflow.org/guide/gpu).
    
18. 
    
19.     Usage: Run before any other code.
    
20. 
    
21.     :param log_device_placement: set True to log device placement (on which device the operation ran)
    
22.     :return:None
    
23.     """
    
24.     # set_session 标红不用管 Sets the global TensorFlow session
    
25.     # from tensorflow.compat.v1.keras.backend import set_session
    
26.     # from tensorflow.python.keras.backend import set_session
    
27.     config = tf.compat.v1.ConfigProto()
    
28.     config.gpu_options.allow_growth = True  # dynamically grow the memory used on the GPU
    
29.     config.log_device_placement = log_device_placement
    
30.     sess = tf.compat.v1.Session(config=config)
    
31.     # 一定要拿到全局的session来配置 因为模型预测会用到keras.backend.get_session()
    
32.     keras.backend.set_session(session=sess)
    
33. 
    
34. 
    
35. allow_gpu_memory_growth()
    
36. 
    
37. # 函数allow_gpu_memory_growth()应该放在这句的前面啊
    
38. from hyperlpr_py3 import pipline
    
39. 
    
40. # 导入车牌识别库
    
41. 
    
42. 
    
43. app = Flask(__name__)

复制代码

东方耀

os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
# 选择哪一块gpu,如果是-1，就是调用cpu
os.environ['CUDA_VISIBLE_DEVICES'] = "0"


# python 3.6 and tensorflow (both 1.x and 2.0)
def allow_gpu_memory_growth(log_device_placement=True):
    """
    Allow dynamic memory growth (by default, tensorflow allocates all gpu memory).
    This sometimes fixes the
    <<Error : Failed to get convolution algorithm.
    This is probably because cuDNN failed to initialize,
    so try looking to see if a warning log message was printed above>>.
    May hurt performance slightly (see https://www.tensorflow.org/guide/gpu).

    Usage: Run before any other code.

    :param log_device_placement: set True to log device placement (on which device the operation ran)
    :return:None
    """
    # set_session 标红不用管 Sets the global TensorFlow session
    # from tensorflow.compat.v1.keras.backend import set_session
    # from tensorflow.python.keras.backend import set_session
    config = tf.compat.v1.ConfigProto()
    config.gpu_options.allow_growth = True  # dynamically grow the memory used on the GPU
    config.log_device_placement = log_device_placement
    sess = tf.compat.v1.Session(config=config)
    # 一定要拿到全局的session来配置 因为模型预测会用到keras.backend.get_session()
    keras.backend.set_session(session=sess)


allow_gpu_memory_growth()
# 函数allow_gpu_memory_growth()应该放在这句的前面啊