Expected a_in <= b_in to be true, but got false

1

Hi,
I have got this error

Traceback (most recent call last):
File “train.py”, line 15, in
model = CreateModel(opt)
File “model/model_Loader.py”, line 15, in CreateModel
model.init_weights()
File “model/DRGAN.py”, line 204, in init_weights
self.G.apply(weights_init_normal)
File “/home/pirl/anaconda3/envs/dr-gan-torch1-2/lib/python3.6/site-packages/torch/nn/modules/module.py”, line 293, in apply
module.apply(fn)
File “/home/pirl/anaconda3/envs/dr-gan-torch1-2/lib/python3.6/site-packages/torch/nn/modules/module.py”, line 293, in apply
module.apply(fn)
File “/home/pirl/anaconda3/envs/dr-gan-torch1-2/lib/python3.6/site-packages/torch/nn/modules/module.py”, line 293, in apply
module.apply(fn)
[Previous line repeated 2 more times]
File “/home/pirl/anaconda3/envs/dr-gan-torch1-2/lib/python3.6/site-packages/torch/nn/modules/module.py”, line 294, in apply
fn(self)
File “model/Component.py”, line 49, in weights_init_normal
init.uniform_(m.weight.data, 1.0, 0.02)
File “/home/pirl/anaconda3/envs/dr-gan-torch1-2/lib/python3.6/site-packages/torch/nn/init.py”, line 88, in uniform_
return no_grad_uniform(tensor, a, b)
File “/home/pirl/anaconda3/envs/dr-gan-torch1-2/lib/python3.6/site-packages/torch/nn/init.py”, line 14, in no_grad_uniform
return tensor.uniform_(a, b)
RuntimeError: Expected a_in <= b_in to be true, but got false. (Could this error message be improved? If so, please report an enhancement request to PyTorch.)

File “train.py”, line 15, in is

model = CreateModel(opt)

File “model/model_Loader.py”, line 15, in CreateModel is

model.init_weights()

File “model/DRGAN.py”, line 204, in init_weights is

203 def init_weights(self):
204 self.G.apply(weights_init_normal)
205 self.D.apply(weights_init_normal)

I got this pytorch code from

and try to run it.

According to github-site, the required pytorch version is 0.2, however my torch version is 1.2, the latest version.
I want to run in the current version without changing the version.

Any help will be greatly appreciated!!!

2

I’m having trouble understanding the logic in the original code. In both version 0.2 and 1.2 of PyTorch, calling .uniform_(a, b) on a tensor requires a <= b, as the error states, and the call uses a = 1.0 and b = 0.02.

This might be a mistake in the original code, but I’m not familiar with GAN initialization standards / approaches for batch norm layers. Have a look here for the torchvision provided ResNet model.

1 Like
3

Sorry for late comment :frowning:
Thx for your advice, I finally solve the problem.

I change the value of a & b like

init.uniform_(m.weight.data, 1.0, 0.02)
to
init.uniform_(m.weight.data, 0.02, 1.0)

After that, I install the torchvision using in different way

conda install torchvision -c pytorch
instead of
pip install torchvision

and I work :slight_smile:
I really appreciate your help!!!

2 Likes