如何解决Keras中的“分段错误（核心转储”错误）

Question

I'm having issues with Keras. Basically, it gives me the following error "Segmentation fault (core dumped)" when I try to fit a model with a conv2d layer.

My code works on the CPU. It also works without any conv2d layers (even though it's ineffective for my use case). I've got cuda, cudnn, and tensorflow installed. I've tried reinstalling keras and tensorflow.

Code:

def model_build():
    model = Sequential()
    model.add(Conv2D(input_shape = (env_size()[0], env_size()[1], 1), filters=4, kernel_size=(3,3), strides=1, activation=swisher))
    model.add(Conv2D(filters=4, kernel_size=(5,5), strides=1, activation=swisher))
    model.add(Conv2D(filters=4, kernel_size=(5,5), strides=1, activation=swisher))
    model.add(Conv2D(filters=4, kernel_size=(5,5), strides=1, activation=swisher))
    model.add(Flatten())
    model.add(Dense(128, activation='softmax'))
    model.add(Dense(4, activation='softmax'))
    return model

if __name__ == '__main__':
    y = model_build()
    y.compile(loss = "mean_squared_error", optimizer = 'adam')
    y.fit(x=env(), y = np.array([[0,0,0,0]])

Error:

Using TensorFlow backend.
Epoch 1/1
2019-03-27 05:52:27.687323: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-03-27 05:52:27.789975: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-03-27 05:52:27.790819: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1411] Found device 0 with properties:
name: GeForce RTX 2060 major: 7 minor: 5 memoryClockRate(GHz): 1.83
pciBusID: 0000:01:00.0
totalMemory: 5.73GiB freeMemory: 5.40GiB
2019-03-27 05:52:27.790834: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1490] Adding visible gpu devices: 0
2019-03-27 05:52:28.068080: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-03-27 05:52:28.068115: I tensorflow/core/common_runtime/gpu/gpu_device.cc:977]      0
2019-03-27 05:52:28.068121: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 0:   N
2019-03-27 05:52:28.068487: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1103] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 5147 MB memory) -> physical GPU (device: 0, name: GeForce RTX 2060, pci bus id: 0000:01:00.0, compute capability: 7.5)
2019-03-27 05:52:28.177752: W tensorflow/core/framework/allocator.cc:113] Allocation of 518619136 exceeds 10% of system memory.
2019-03-27 05:52:28.337277: W tensorflow/core/framework/allocator.cc:113] Allocation of 518619136 exceeds 10% of system memory.
2019-03-27 05:52:28.500486: W tensorflow/core/framework/allocator.cc:113] Allocation of 518619136 exceeds 10% of system memory.
2019-03-27 05:52:28.586280: W tensorflow/core/framework/allocator.cc:113] Allocation of 518619136 exceeds 10% of system memory.
2019-03-27 05:52:28.675738: W tensorflow/core/framework/allocator.cc:113] Allocation of 518619136 exceeds 10% of system memory.
Segmentation fault (core dumped)

EDIT:

Self-contained example.

import numpy as np
import keras

model = keras.models.Sequential() #Sequential model type.
model.add(keras.layers.Conv2D(filters=1, kernel_size=(3,3), strides = 1, activation="sigmoid")) #Convolutional layer.
model.add(keras.layers.Flatten()) #Flatten layer.
model.add(keras.layers.Dense(4)) #Dense layer of 4 units.
model.compile(loss='mean_squared_error', optimizer='adam') #compile model.
y = np.random.rand(1,4) #Random expected output
x = np.random.rand(1, 38, 21, 1) # Random input.
model.fit(x, y) #And fit...

EDIT2:

Keras version: 'v2.1.6-tf'

Tensorflow-GPU version: 'v1.12'

Python version: 'v3.5.2'

CUDA version: 'v9.0.176'

CUDNN version: 'v7.2.1.38-1+cuda9.0

Ubuntu version: 'v16.04'

Answer 1

It seems that your GPU does not have enough memory. Your model does not seem to be too big, so I would guess that the problem comes from the line:

y.fit(x=env(), y = np.array([[0,0,0,0]])

The output of env() might be too big to be handle by your GPU memory.

Answer 2

Your MWE works fine for me (if I add , input_shape=(38, 21, 1) to the first convolution layer):

import numpy as np
import keras

model = keras.models.Sequential() #Sequential model type.
model.add(keras.layers.Conv2D(filters=1, kernel_size=(3,3), strides = 1, activation="sigmoid", input_shape=(38, 21, 1))) #Convolutional layer.
model.add(keras.layers.Flatten()) #Flatten layer.
model.add(keras.layers.Dense(4)) #Dense layer of 4 units.
model.compile(loss='mean_squared_error', optimizer='adam') #compile model.
y = np.random.rand(2, 4) #Random expected output
x = np.random.rand(2, 38, 21, 1) # Random input.
model.fit(x, y)

That means that your issue must come from your system or installation.

Looking at the compatibility chart of tensorflow shows that your python, tensorflow and CUDA versions should be compatible.

For your configuration the cuDNN version 7.0.x is recommended. The cuDNN version 7.2 that you are using is probably incompatible. Try installing / using cuDNN 7.0.x .