Hi there,

Do you have any pre-trained model for the Cityscapes example that could be made available online?

Thanks

Once I set '-- input_scale_size' not to 64, an error will be reported. It seems that the discriminator cannot adapt to the size of the input image. How did you solve the problem?
"ValueError: Target and input must have the same number of elements. target nelement (2) != input nelement (338)"

Thanks for the PyTorch implementation of DiscoGAN.
I am having troubles running it with multiple GPUs though.
I think the first problem is in the config.py file, where it casts the num_gpu argument to a bool instead of an int.

Traceback (most recent call last):
  File "main.py", line 41, in <module>
    main(config)
  File "main.py", line 33, in main
    trainer.train()
  File "/home/***/Documents/DiscoGAN-pytorch-master/trainer.py", line 187, in train
    x_AB = self.G_AB(x_A).detach()
  File "/home/***/.pyenv/versions/2.7.13/lib/python2.7/site-packages/torch/nn/modules/module.py", line 206, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/***/Documents/DiscoGAN-pytorch-master/models.py", line 45, in forward
    return nn.parallel.data_parallel(self.main, x, gpu_ids)
  File "/home/***/.pyenv/versions/2.7.13/lib/python2.7/site-packages/torch/nn/parallel/data_parallel.py", line 101, in data_parallel
    replicas = replicate(module, device_ids[:len(inputs)])
  File "/home/***/.pyenv/versions/2.7.13/lib/python2.7/site-packages/torch/nn/parallel/replicate.py", line 10, in replicate
    params = list(network.parameters())
AttributeError: 'function' object has no attribute 'parameters'

Any idea what could be the cause?

getting this message:

0%| | 1/50000 [00:03<55:08:42, 3.97s/it]C:\Users\Shadow\Anaconda3\lib\site-packages\torch\nn\modules\loss.py:512: UserWarning: Using a target size (torch.Size([200])) that is different to the input size (torch.Size([200, 1])) is deprecated. Please ensure they have the same size.

As of now I don't have any GPU to train.

I get the following warning message;

[!] Sampled dataset from A and B have different # of data. Try resampling...

From what I understand it is not necessary that both data sets are equal in size?
Both datasets have pictures with 256x256 pixels.

DiscoGAN paper they assert that
"to avoid costly pairing, we address
the task of discovering cross-domain relations
given unpaired data"
But, I don't know why this implementation of DiscoGAN requires paired data?
(as far as I know/see, in the dataset downloaded )
Or, is it just to simplify the input process?

class Generator(nn.Module):

def __init__(self, input_size, output_size, hidden_dims):

    super(Generator, self).__init__()

    self.layers = []

    

    prev_dim = input_size

    for hidden_dim in hidden_dims:

        self.layers.append(nn.Linear(prev_dim, hidden_dim))

        self.layers.append(nn.ReLU(True))

        prev_dim = hidden_dim

    self.layers.append(nn.Linear(prev_dim, output_size))

    

    self.layer_module = ListModule(*self.layers)

    

def forward(self, x):

    out = x

    for layer in self.layers:

        out = layer(out)

    return out

​

class Discriminator(nn.Module):

def __init__(self, input_size, output_size, hidden_dims):

    super(Discriminator, self).__init__()

    self.layers = []

    

    prev_dim = input_size

    for idx, hidden_dim in enumerate(hidden_dims):

        self.layers.append(nn.Linear(prev_dim, hidden_dim))

        self.layers.append(nn.ReLU(True))

        prev_dim = hidden_dim

        

    self.layers.append(nn.Linear(prev_dim, output_size))

    self.layers.append(nn.Sigmoid())

    

    self.layer_module = ListModule(*self.layers)

​

def forward(self, x):

    out = x

    for layer in self.layers:

        out = layer(out)

    return out.view(-1, 1)

​

network

hidden_dim = 128

g_num_layer = 3

d_num_layer = 5

​

G_AB = Generator(2, 2, [hidden_dim] * g_num_layer)

G_BA = Generator(2, 2, [hidden_dim] * g_num_layer)

​

D_A = Discriminator(2, 1, [hidden_dim] * d_num_layer)

D_B = Discriminator(2, 1, [hidden_dim] * d_num_layer)

​

G_AB.cuda()

G_BA.cuda()

D_A.cuda()

D_B.cuda()

​

optimizer

lr = 0.0002

beta1 = 0.5

beta2 = 0.999

​

d = nn.MSELoss()

bce = nn.BCELoss()

​

optimizer_d = torch.optim.Adam(

chain(D_A.parameters(), D_B.parameters()), lr=lr, betas=(beta1, beta2))

optimizer_g = torch.optim.Adam(

chain(G_AB.parameters(), G_BA.parameters()), lr=lr, betas=(beta1, beta2))

​

training

num_epoch = 50000

​

real_label = 1

fake_label = 0

​

real_tensor = Variable(torch.FloatTensor(batch_size).cuda())

_ = real_tensor.data.fill_(real_label)

print(real_tensor.sum())

​

fake_tensor = Variable(torch.FloatTensor(batch_size).cuda())

_ = fake_tensor.data.fill_(fake_label)

print(fake_tensor.sum())

RuntimeError Traceback (most recent call last)
in ()
77
78 real_tensor = Variable(torch.FloatTensor(batch_size).cuda())
---> 79 _ = real_tensor.data.fill_(real_label)
80 print(real_tensor.sum())
81

RuntimeError: cuda runtime error (8) : invalid device function at /py/conda-bld/pytorch_1493677666423/work/torch/lib/THC/generic/THCTensorMath.cu:15

It downsizes my dataset image to the height of 68px while training

when I test my model, the problem 'unexpected key "module.layer_module.0.weight" in state_dict' happened. Did you meet the problem?

Looks like you crash if width and height isn't equal on custom datasets. I worked around by forcing my 640x360 images to 640x640.

Hello Sir,

I downloaded your code and maps-dataset..
And when I started I met some error.

...
Traceback (most recent call last):
  File "main.py", line 41, in <module>
    main(config)
  File "main.py", line 33, in main
    trainer.train()
  File "/itsme/TESTBOARD/additional_networks/GAN/pytorch_DiscoGAN_carpedm20/trainer.py", line 247, in train
    format(step, self.max_step, l_d.data[0], l_g.data[0]))
IndexError: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python number

Thanks..

Hi,

I've been trying to run the example code (on the maps dataset):

python main.py --dataset=maps --num_gpu=4

I get the error below related to the NCCL library. I'm trying to run this on 4 K80 GPUs.

Any suggestions on what could be causing this and what a solution could be?

pix2pix processing: 100%|#######################| 1096/1096 [00:00<00:00, 178591.97it/s]
pix2pix processing: 100%|#######################| 1096/1096 [00:00<00:00, 213732.43it/s]
[] MODEL dir: logs/maps_2017-10-26_20-36-34
[] PARAM path: logs/maps_2017-10-26_20-36-34/params.json
0%| | 0/500000 [00:00<?, ?it/s]

Traceback (most recent call last):
File "main.py", line 41, in
main(config)
File "main.py", line 33, in main
trainer.train()
File "/home/nbserver/DiscoGAN-pytorch/trainer.py", line 193, in train
x_AB = self.G_AB(x_A).detach()
File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 224, in
call
result = self.forward(*input, **kwargs)
File "/usr/local/lib/python2.7/dist-packages/torch/nn/parallel/data_parallel.py", line
59, in forward
replicas = self.replicate(self.module, self.device_ids[:len(inputs)])
File "/usr/local/lib/python2.7/dist-packages/torch/nn/parallel/data_parallel.py", line
64, in replicate
return replicate(module, device_ids)
File "/usr/local/lib/python2.7/dist-packages/torch/nn/parallel/replicate.py", line 12,
in replicate
param_copies = Broadcast(devices)(*params)
File "/usr/local/lib/python2.7/dist-packages/torch/nn/parallel/_functions.py", line 19
, in forward
outputs = comm.broadcast_coalesced(inputs, self.target_gpus)
File "/usr/local/lib/python2.7/dist-packages/torch/cuda/comm.py", line 54, in broadcas
t_coalesced
results = broadcast(_flatten_tensors(chunk), devices)
File "/usr/local/lib/python2.7/dist-packages/torch/cuda/comm.py", line 24, in broadcas
t
nccl.broadcast(tensors)
File "/usr/local/lib/python2.7/dist-packages/torch/cuda/nccl.py", line 182, in broadca
st
comm = communicator(inputs)
File "/usr/local/lib/python2.7/dist-packages/torch/cuda/nccl.py", line 133, in communi
cator
_communicators[key] = NcclCommList(devices)
File "/usr/local/lib/python2.7/dist-packages/torch/cuda/nccl.py", line 106, in _init
_
check_error(lib.ncclCommInitAll(self, len(devices), int_array(devices)))
File "/usr/local/lib/python2.7/dist-packages/torch/cuda/nccl.py", line 118, in check_e
rror
raise NcclError(status)
torch.cuda.nccl.NcclError: System Error (2)

Hello,
please can you show how to test the final model on a single image
Thank you
Lafi

I tried out my own dataset in data/mydata (with the A and B folders) but I get the following error:

~/DiscoGAN-pytorch$ python main.py --dataset=mydata --num_gpu=1
[*] MODEL dir: logs/mydata_2017-03-21_14-49-25
[*] PARAM path: logs/mydata_2017-03-21_14-49-25/params.json
Traceback (most recent call last):
  File "main.py", line 41, in <module>
    main(config)
  File "main.py", line 33, in main
    trainer.train()
  File "/home/bart/DiscoGAN-pytorch/trainer.py", line 161, in train
    valid_x_A, valid_x_B = self._get_variable(A_loader.next()), self._get_variable(B_loader.next())
  File "/home/bart/anaconda3/envs/Python36/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 174, in __next__
    return self._process_next_batch(batch)
  File "/home/bart/anaconda3/envs/Python36/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 198, in _process_next_batch
    raise batch.exc_type(batch.exc_msg)
RuntimeError: Traceback (most recent call last):
  File "/home/bart/anaconda3/envs/Python36/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 32, in _worker_loop
    samples = collate_fn([dataset[i] for i in batch_indices])
  File "/home/bart/anaconda3/envs/Python36/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 68, in default_collate
    return torch.stack(batch, 0)
  File "/home/bart/anaconda3/envs/Python36/lib/python3.6/site-packages/torch/functional.py", line 56, in stack
    return torch.cat(list(t.unsqueeze(dim) for t in sequence), dim)
RuntimeError: inconsistent tensor sizes at /data/users/soumith/miniconda2/conda-bld/pytorch-0.1.10_1488755368782/work/torch/lib/TH/generic/THTensorMath.c:2548