luee-c / figr Goto Github PK

Hi, I tried to reproduce your results on MNIST dataset, but the model seems to diverge.
I simply clone your repo and run python train.py --dataset Mnist.

Here are my loss curves and generated samples.

Can you share your loss curves and generated samples during training?
(I'm using PyTorch v1.1)

In validation_run method in train.py, at line 99 and 115 np.expand_dims method create an unnecessary dimension. It causes the following error.

  File "train.py", line 223, in <module>
    env.training()
  File "train.py", line 165, in training
    self.validation_run()
  File "train.py", line 120, in validation_run
    self.writer.add_image('Validation_generated', img, self.eps)
  File ".../tensorboardX/writer.py", line 427, in add_image
    image(tag, img_tensor, dataformats=dataformats), global_step, walltime)
  File ".../tensorboardX/summary.py", line 211, in image
    tensor = convert_to_HWC(tensor, dataformats)
  File ".../tensorboardX/utils.py", line 98, in convert_to_HWC
    assert(len(tensor.shape) == len(input_format)), "size of input tensor and input format are different"
AssertionError: size of input tensor and input format are different

add_image method has the dataformats parameter and its default value is CHW (channel-height-width). The code sends a four-dimensional tensor to the add_image method. example: (1, 1, 32, 32) [not necessary dimension and causes an error, CHANNEL, HEIGHT, WIDTH]

I tried with all supported datasets.

python train.py --dataset Omniglot
python train.py --dataset Mnist
python train.py --dataset FIGR8

I will send a pull request that resolves this issue.

how can I train with a custom dataset or vocformat dataset?

Hi, I tried to reproduce your results on MNIST dataset. But while training, it shows segmentation fault at line 83 of train.py (discriminator_loss.backward() ).
All the requirements were successfully installed.

Showing the output error shown while running the code: