在 GPU 上运行的 tensorflow 迁移学习问题

Question

我正在尝试运行 retrain.py 脚本（可在此处获得：https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/image_retraining/retrain.py）。我注意到以第 747 行开头的部分是在 CPU 上执行的，而默认应该是 GPU。因此，我添加了以下行来强制它在 GPU 上工作：

`with tf.device("/gpu:0"): 
  (train_step, cross_entropy, bottleneck_input, ground_truth_input, final_tensor) = add_final_training_ops(len(image_lists.keys()),
                                      FLAGS.final_tensor_name,
                                      bottleneck_tensor)`

它会导致以下错误：

'tensorflow.python.framework.errors.InvalidArgumentError: Cannot assign a device to node 'gradients/Mean_grad/Prod': Could not satisfy explicit device specification '/device:GPU:0' because no supported kernel for GPU devices is available
     [[Node: gradients/Mean_grad/Prod = Prod[T=DT_INT32, keep_dims=false, _device="/device:GPU:0"](gradients/Mean_grad/Shape_2, gradients/Mean_grad/range_1)]]
Caused by op u'gradients/Mean_grad/Prod', defined at:
File "retrain_tensorboard_pickle_mean.py", line 921, in <module>
tf.app.run()
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/platform/app.py", line 30, in run
sys.exit(main(sys.argv))
File "retrain_tensorboard_pickle_mean.py", line 839, in main
(train_step, cross_entropy, bottleneck_input, ground_truth_input, label_ground_truth_input, final_tensor) = add_final_training_ops(len(image_lists.keys()), FLAGS.final_tensor_name, bottleneck_tensor)
File "retrain_tensorboard_pickle_mean.py", line 686, in add_final_training_ops
cross_entropy_mean)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/optimizer.py", line 190, in minimize
colocate_gradients_with_ops=colocate_gradients_with_ops)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/optimizer.py", line 241, in compute_gradients
colocate_gradients_with_ops=colocate_gradients_with_ops)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gradients.py", line 481, in gradients
in_grads = _AsList(grad_fn(op, *out_grads))
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/math_grad.py", line 91, in _MeanGrad
factor = (math_ops.reduce_prod(input_shape) //
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/math_ops.py", line 810, in reduce_prod
keep_dims, name=name)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_math_ops.py", line 1115, in _prod
keep_dims=keep_dims, name=name)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/op_def_library.py", line 655, in apply_op
op_def=op_def)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2146, in create_op
original_op=self._default_original_op, op_def=op_def)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1154, in __init__
self._traceback = _extract_stack()

...which was originally created as op u'Mean', defined at:
File "retrain_tensorboard_pickle_mean.py", line 921, in <module>
tf.app.run()
[elided 1 identical lines from previous traceback]
File "retrain_tensorboard_pickle_mean.py", line 839, in main
(train_step, cross_entropy, bottleneck_input, ground_truth_input, label_ground_truth_input, final_tensor) = add_final_training_ops(len(image_lists.keys()), FLAGS.final_tensor_name, bottleneck_tensor)
File "retrain_tensorboard_pickle_mean.py", line 681, in add_final_training_ops
cross_entropy_mean = tf.reduce_mean(cross_entropy)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/math_ops.py", line 783, in reduce_mean
keep_dims, name=name)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_math_ops.py", line 973, in _mean
keep_dims=keep_dims, name=name)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/op_def_library.py", line 655, in apply_op
op_def=op_def)  
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2146, in create_op
original_op=self._default_original_op, op_def=op_def)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1154, in __init__
self._traceback = _extract_stack()

我发现here 可能是一个问题，mean 没有在 GPU 上实现，但另一方面，github 上有一个 commit 修复了 GPU 上的计数平均值。

上一部分，例如生成瓶颈（第 744 行）在 GPU 上完美运行，甚至不需要强制它。

如果有任何帮助，我将不胜感激！

贾斯蒂娜

Answer 1

这已在b874e2c 中得到修复，不错