Running on custom dataset
jyopari opened this issue · comments
Hello,
How would I train this using my own dataset?
Thanks
HI @jyopari,
Please check this repository which contains some examples for other datasets -> https://github.com/akanimax/pro_gan_pytorch-examples.
Also there is an ongoing issue which could be helpful for you. -> akanimax/pro_gan_pytorch-examples#2.
Please feel free to ask if you have any questions.
I know I am due writing a detailed document for the API usage 😄. Have just been busy with other work.
Hope this helps for now.
Best regards,
@akanimax
Good to know that you got it to work 😄 👍. Yes, sure: GAN_GEN_SHADOW is the model which contains the stable weights for the Exponential Moving Averaging of the Generator parameters. OPTIM contains the adam optimizer state and 0 is a depth of (4_x_4) and 1 is a depth of (8_x_8). Please feel free to ask if you have any further questions.
Best regards,
@akanimax
@akanimax or @jyopari, sorry to dig this up, i'm trying out this on a colab notebook and want to use it with my own dataset. I'm trying this to load my data:
`def setup_data():
"""
setup the CIFAR-10 dataset for training the CNN
:param batch_size: batch_size for sgd
:param num_workers: num_readers for data reading
:param download: Boolean for whether to download the data
:return: classes, trainloader, testloader => training and testing data loaders
"""
# data setup:
transforms = tv.transforms.ToTensor()
data = tv.datasets.ImageFolder(root=data_path,
transform=transforms)
trainset = th.utils.data.DataLoader(data, shuffle=True, num_workers=1)
return trainset`
but running the pro-gan gives an error:
`Starting the training process ...
Currently working on Depth: 0
Current resolution: 4 x 4
Epoch: 1
TypeError Traceback (most recent call last)
in ()
22 epochs=num_epochs,
23 fade_in_percentage=fade_ins,
---> 24 batch_sizes=batch_sizes
25 )
26 # ======================================================================
2 frames
/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py in _process_next_batch(self, batch)
606 raise Exception("KeyError:" + batch.exc_msg)
607 else:
--> 608 raise batch.exc_type(batch.exc_msg)
609 return batch
610
TypeError: Traceback (most recent call last):
File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/_utils/worker.py", line 99, in _worker_loop
samples = collate_fn([dataset[i] for i in batch_indices])
File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/_utils/worker.py", line 99, in
samples = collate_fn([dataset[i] for i in batch_indices])
TypeError: 'DataLoader' object does not support indexing`
Is there any example to load a folder with images, that's what i have: dir1/dir2/10k images...
Thanks,
Tiago
@talvasconcelos, Seems to me that you are not using the pyTorch's ImageFolder dataset properly. Could you please provide the complete error stack for this error?
I've solved it, i was messing the depth for my image size, 4 instead of 6. Also i pulled the examples DataLoader module to load the images, it's working now. I get an error when setting loss='hinge', though, going with the default!
I just feel that i got better results from DCGAN... need to experiment more!
@talvasconcelos, ProGAN definitely gives a lot better results than DCGAN at higher resolutions. Only caveat is that you need to set the progressive schedule (no. of epochs per resolution) very carefully. Otherwise, the whole training gets messed up. You could also try our new paper which addresses this caveat MSG-GAN. There is no hyperparameter tuning required and it is very easy to use.
Could you please share the error that you are receiving for loss = hinge? I'll address this in the next version of the package as well.
Giving it another try, with longer training.
Trying now with parameters (using example form git):
depth = 6
num_epochs = [200, 200, 200, 300, 500, 300]
fade_ins = [50, 50, 50, 50, 50, 50]
batch_sizes = [256, 256, 256, 128, 64, 32]
latent_size = 256
Dataset = dl.FlatDirectoryImageDataset
dataset = Dataset(
data_dir=data_path,
transform=dl.get_transform()
)
print("total examples in training: ", len(dataset))
pro_gan.train(
dataset=dataset,
epochs=num_epochs,
fade_in_percentage=fade_ins,
batch_sizes=batch_sizes,
feedback_factor=1,
log_dir=f"{base_dir}/logs/",
sample_dir=f"{base_dir}/images/",
save_dir=f"{base_dir}/models/")
this is the error i get when i set loss='hinge':
`TypeError Traceback (most recent call last)
in ()
33 log_dir=f"{base_dir}/logs/",
34 sample_dir=f"{base_dir}/images/",
---> 35 save_dir=f"{base_dir}/models/", loss="hinge")
36
37 # :param loss: the loss function to be used
TypeError: train() got an unexpected keyword argument 'loss'`
Will definitely take a look at MSG-GAN, and try to run it on colab.
also I tried the lfw.conf file, but i get this error. 'EasyDict' object has no attribute 'folder_distributed'
@jyopari how did you solve this?
also I tried the lfw.conf file, but i get this error. 'EasyDict' object has no attribute 'folder_distributed'
@jyopari how did you solve this?
This is solved by adding a line to configs/mnist.conf
folder_distributed: False # whether images are distributed among foldersClosing due to inactivity. Cheers!