Keras sees my GPU but doesn't use it when training a neural network

Question

My GPU is not used by Keras/TensorFlow.

To try to make my GPU working with tensorflow, I installed tensorflow-gpu via pip (I am using Anaconda on Windows)

I have nvidia 1080ti

print(tf.test.is_gpu_available())

True

print(tf.config.experimental.list_physical_devices())

[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'), 
 PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

I tied

physical_devices = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], True)

but it didnt help

sess = tf.compat.v1.Session(config=tf.compat.v1.ConfigProto(log_device_placement=True))
print(sess)

Device mapping:
/job:localhost/replica:0/task:0/device:GPU:0 -> device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1

<tensorflow.python.client.session.Session object at 0x000001A2A3BBACF8>

only warning from tf:

W tensorflow/stream_executor/cuda/redzone_allocator.cc:312] Internal: Invoking ptxas not supported on Windows 

entire log:

2019-10-18 20:06:26.094049: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_100.dll
2019-10-18 20:06:35.078225: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
2019-10-18 20:06:35.090832: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library nvcuda.dll
2019-10-18 20:06:35.180744: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.683
pciBusID: 0000:01:00.0
2019-10-18 20:06:35.185505: I tensorflow/stream_executor/platform/default/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.
2019-10-18 20:06:35.189328: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1746] Adding visible gpu devices: 0
2019-10-18 20:06:35.898592: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1159] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-10-18 20:06:35.901683: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1165]      0
2019-10-18 20:06:35.904235: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1178] 0:   N
2019-10-18 20:06:35.906687: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1304] Created TensorFlow device (/device:GPU:0 with 8784 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1)
2019-10-18 20:06:38.694481: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.683
pciBusID: 0000:01:00.0
2019-10-18 20:06:38.700482: I tensorflow/stream_executor/platform/default/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.
2019-10-18 20:06:38.704020: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1746] Adding visible gpu devices: 0
[I 20:06:47.324 NotebookApp] Saving file at /Untitled.ipynb
2019-10-18 20:07:22.227110: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.683
pciBusID: 0000:01:00.0
2019-10-18 20:07:22.246012: I tensorflow/stream_executor/platform/default/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.
2019-10-18 20:07:22.261643: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1746] Adding visible gpu devices: 0
2019-10-18 20:07:22.272150: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1159] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-10-18 20:07:22.275457: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1165]      0
2019-10-18 20:07:22.277980: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1178] 0:   N
2019-10-18 20:07:22.316260: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1304] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 8784 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1)
Device mapping:
/job:localhost/replica:0/task:0/device:GPU:0 -> device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1
2019-10-18 20:07:32.986802: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.683
pciBusID: 0000:01:00.0
2019-10-18 20:07:32.990509: I tensorflow/stream_executor/platform/default/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.
2019-10-18 20:07:32.993763: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1746] Adding visible gpu devices: 0
2019-10-18 20:07:32.995570: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1159] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-10-18 20:07:32.997920: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1165]      0
2019-10-18 20:07:32.999435: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1178] 0:   N
2019-10-18 20:07:33.001380: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1304] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 8784 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1)
2019-10-18 20:07:36.048204: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudnn64_7.dll
2019-10-18 20:07:37.971703: W tensorflow/stream_executor/cuda/redzone_allocator.cc:312] Internal: Invoking ptxas not supported on Windows
Relying on driver to perform ptx compilation. This message will be only logged once.
2019-10-18 20:07:38.576861: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cublas64_100.dll

also tried reinstalling tensorflow-gpu with pip

Why I think GPU doesnt work? - Because my python kernel uses CPU 99%, RAM 99% and sometimes GPU ~7% but most of time its 0
I use custom data generator but now its only selects batches and resizes them(skimage.io.resize) 1 epoch ~ 44s Also has strange behavior of freezing in random points every ~10 samples and freezes hardly on last sample(37/38)(~10-15 sec)

Edit:

I post my custom datagen here

train_gen = DataGenerator(x = x_train,
                              y = y_train,
                              batch_size = 128,
                              target_shape = (100, 100, 3), 
                              sample_std = False,
                              feature_std = False,
                              proj_parameters = None,
                              blur_parameters = None,
                              nois_parameters = None,
                              flip_parameters = None,
                              gamm_parameters = None)

validation is same

Update:

So its a generator that couses the problem, but how i can fix it?
I used only skimage and numpy operations

Answer 1

The logs are showing that the GPU does get used. You are almost certainly running into an IO bottleneck: your GPU is processing whatever the CPU is throwing at it way faster than the CPU can load and preprocess it. This is very common in deep learning, and there are ways to address it.

We cannot provide a lot of help without knowing more about your data pipeline (byte size of a batch, preprocessing steps, ...), and how the data is stored. One typical way to speed things up is to store the data is a binary format, like TFRecords, so that the CPU can load it faster. See the official documentation for this.

---

Edit: I quickly went through your input pipeline. The issue is very likely to indeed by IO:

• You should run the preprocessing steps on the GPU as well, plenty of the augmentation techniques you use are implemented in tf.image. If you can, you should think about using Tensorflow 2.0, because it includes Keras and there are plenty of helpers in there as well.
• Checkout the tf.data.Dataset API, it has plenty of helpers to load all the data in different threads, which can roughly speed up the process by the number of cores you have.
• You should store your images as TFRecords. This is likely to speed up the loading by an order of magnitude if your input images are smallish.
• You could probably try larger batch sizes as well, I'm thinking your images are probably really small.