我一直在尝试训练刚好超过100张图像的R-CNN（https://github.com/kbardool/keras-frcnn）。我尝试使用以下行（https://www.floydhub.com/laurynasg/projects/faster_r-cnn/3/）在Floydhub上进行培训：
floyd run --gpu --env tensorflow-1.9:py2 --data laurynasg/datasets/faster_r-cnn_images/1:images 'python train_frcnn.py -o simple -p train.txt'
我的问题是训练序列看起来很奇怪而且很慢。首先，它会迭代一千次，然后给出一些消息：
Average number of overlapping bounding boxes from RPN = 9.603 for 1000 previous iterations
2018-08-29 04:19:49,839 INFO -
2018-08-29 04:19:49,839 INFO - 1000/1000 [==============================] - 6421s 6s/step - rpn_cls: 2.1735 - rpn_regr: 0.1228 - detector_cls: -0.1509 - detector_regr: -0.1189
2018-08-29 04:19:49,839 INFO - Mean number of bounding boxes from RPN overlapping ground truth boxes: 9.607
2018-08-29 04:19:49,840 INFO - Classifier accuracy for bounding boxes from RPN: 0.952875
2018-08-29 04:19:49,840 INFO - Loss RPN classifier: 0.308176488076
2018-08-29 04:19:49,840 INFO - Loss RPN regression: 0.0141665605258
2018-08-29 04:19:49,840 INFO - Loss Detector classifier: 0.115702805009
2018-08-29 04:19:49,840 INFO - Loss Detector regression: 0.0648736124858
2018-08-29 04:19:49,840 INFO - Elapsed time: 1045.54498291
2018-08-29 04:19:49,841 INFO - Total loss decreased from 0.745386148392 to 0.502919466096, saving weights
2018-08-29 04:19:49,841 INFO - Exception: `save_weights` requires h5py.
2018-08-29 04:19:50,832 INFO -
在下一个纪元开始之前，这似乎要重复随机次数（从1到5）。同样，单次迭代大约需要1秒，即使在Nvidia K80卡上，这也将使2000个纪元的整个训练变得非常缓慢。完整的日志在这里：https://www.floydhub.com/api/v1/resources/vkumzAf8WEGdJkYyF6oQMY?content=true 我在做错什么吗，因为我敢肯定它不会花这么长时间