How to set global random seed?

[lfwin]

I try to solve this bug #180, but it would be more convenient if we can make random generator stable, and i used method in http://stackoverflow.com/questions/36288235/how-to-get-stable-results-with-tensorflow-setting-random-seed and changed chainer.py, but this does not work! Does tflearn has some function to set a globle seed for random generator?

edit:I locate the bugs in https://github.com/tflearn/tflearn/issues/180， because image read by PIL maybe size (299, 299) but right one should be (299, 299, 3)

[aymericdamien]

Have you try to use tflearn.init_graph(seed=...)?
About the other link, is it an error with PIL? Are your images grayscale? maybe reshaping them should works (you will need to edit imagePreloader source).

[lfwin]

Hi,
1:
I change code based on residual_network_cifar_10.py, but i believe init_graph(seed=...) function not used when running residual_network_cifar_10.py, because i wanted to change gpu memory fraction and modified Trainer class init function as follows
` self.incr_global_step = tf.assign(self.global_step,
tf.add(self.global_step, 1))

        config = None
        tflearn_conf = tf.get_collection(tf.GraphKeys.GRAPH_CONFIG)

        if tflearn_conf == []:
            config = init_graph(gpu_memory_fraction=0.65)
        if tflearn_conf:
            config = tflearn_conf[0]`

and this actually change gpu memory fraction.
In the same way, I set seed config = init_graph(seed=1, gpu_memory_fraction=0.65) It not work, because loss changes every time I run residual_network_cifar_10.py.

2:
After read img, i add these
if img.mode == 'L': img = convert_color(img, 'RGB')
to check is grayscale and transform into RGB and right now can work except loss is nan.
`class ImagePreloader(Preloader):
def init(self, array, image_shape, normalize=True, grayscale=False):
fn = lambda x: self.preload(x, image_shape, normalize, grayscale)
super(ImagePreloader, self).init(array, fn)

def preload(self, path, image_shape, normalize=True, grayscale=False):
    img = load_image(path)
    if img.mode == 'L':
        img = convert_color(img, 'RGB')
    width, height = img.size

    if width != image_shape[0] or height != image_shape[1]:
        img = resize_image(img, image_shape[0], image_shape[1])
    if grayscale:
        img = convert_color(img, 'L')

    img = pil_to_nparray(img)

    if normalize:
        img /= 255.
    return img`

3
After above modification, the loss is nan, how should i debug? code is based on residual_network_cifar_10.py except input use preloader and some image size relevant params
Edit:Solved by reduce image size from (299, 299) to (220, 200) and batch size is 16, my GPU memory is 12G, maybe memory is too small or some other reasons?

[silgon]

I'm also having the same issue. Any solutions yet?