Check failed: work_element_count > 0 (0 vs. 0)
1I'm trying to train a mask rcnn model using Keras on my own dataset on a p2.xlarge EC2 aws instance.
When I launch the training, after a few steps of training:
Epoch 1/1 2/1000 [..............................] - ETA: 4:27:49 - loss: 5.1578 - rpn_class_loss: 0.0937 - rpn_bbox_loss: 0.6471 - mrcnn_class_loss: 2.6594 - mrcnn_bbox_loss: 1.1266 - mrcnn_mask_loss:0.6311I get this error message:
2018-05-02 13:44:56.193439: F ./tensorflow/core/util/cuda_launch_config.h:127] Check failed: work_element_count > 0 (0 vs. 0)My images are relatively small (~100Kb), few images in my dataset (~150).
The config I am using is as follow:
class CustomConfig(Config): """Configuration for training on the custom dataset. Derives from the base Config class """ # Give the configuration a recognizable name NAME = "blabla" # We use a GPU with 12GB memory, which can fit two images. # Adjust down if you use a smaller GPU. IMAGES_PER_GPU = 2 # Number of classes (including background) NUM_CLASSES = 11Any suggestion ? Thanks !
I downgraded my tensorflow-gpu package to 1.7.0 and it worked
This error has nothing to do with CUDA, it is something internal to Tensorflow, and I have edited the question accordingly
这个问题真奇怪，很多地方说版本问题，没有请自实验，因为我的机器不是所有tensorflow的程序都不能跑，所以不敢随便变化，只是在程序中加了
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
这个设置就ok了，奇怪不奇怪-_--_-
页:
[1]