when running the 'train.py' in src, there are some errors.

tadax / srgan

SRGAN implemetation with TensorFlow

83 stars 31 forks source link

when running the 'train.py' in src, there are some errors. #5

Closed DunguLock closed 7 years ago

DunguLock commented 7 years ago

These errors as follows: File "/home/project/srgan-maset/src/train.py", line 87, in train() File "/home/project/srgan-maset/src/train.py", line 17, in train model = SRGAN() File "/home/project/srgan-maset/src/srgan.py", line 19, in init x = pixel_shuffle_layer(x, 2, 64)
File "../utils/layer.py", line 71, in pixel_shuffle_layer xc = tf.split(3, n_split, x) File “ home/anaconda2/python2.7/site-packages/tensorflow/python/ops/gen_array_ops.py”, line 3426, in _split num_split=num_split, name=name) File "home/anaconda2/python2.7/site-pacakges/tensorflow/python/framework/op_def_library.py", line 509, in apply_op (prefix, dtype.as_type(input_arg.type.name)) TypeError: Input 'split_dim' of 'Split' Op has type float32 that does not match except type int32.

I guess that this error is caused by using python2. Can you help me to do with this error? Thank you! @tadax

tadax commented 7 years ago

This is because the order of parameters are different in previous versions of TensorFlow.

for versions < 0.12.0 tf.split(axis, num_or_size_splits, value)
for versions > 0.12.0 tf.split(value, num_or_size_splits, axis)

By the way, tf.concat need to be changed too.

for versions < 0.12.0 tf.concat(axis, values)
for versions > 0.12.0 tf.concat(values, axis)

Updated source code is coming soon.

DunguLock commented 7 years ago

many thanks!

DunguLock commented 7 years ago

The results that you put in this web seems not good, there always some bad patches in the sr-images, have you do some work to this problem?

tadax commented 7 years ago

I modified as follows.

Fixed adversarial loss.
Upgraded TF Version 0.9 to 1.1.
Pretrained VGG19 with ImageNet dataset instead of CIFAR-100 dataset.

The result is a significant improvement over the previous one. See the results.

DunguLock commented 7 years ago

The results have improved a alot comparing to the old sr-images, but there are some blue point in these result as before. I think you should do some works on this problem! last, Thanks for your works and sharing!