'VQModel' object has no attribute 'be_unconditional'

[ChristianFJung]

Following @rom1504 's instructions for a custom dataset. I have no classes and have trained a model that is producing reconstructions. When attempting to sample by running sample_fast.py, I am receiving this error.

Working with z of shape (1, 256, 16, 16) = 65536 dimensions.
loaded pretrained LPIPS loss from taming/modules/autoencoder/lpips/vgg.pth
VQLPIPSWithDiscriminator running with hinge loss.
Logging to logs/2021-06-25T12-27-06_custom_vqgan/samples/top_k_250_temp_1.00_top_p_1.0/234637
Traceback (most recent call last):
  File "scripts/sample_fast.py", line 262, in <module>
    run(logdir, model, opt.batch_size, opt.temperature, opt.top_k, unconditional=model.be_unconditional,
  File "/home/virginia/dalle/dall2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 778, in __getattr__
    raise ModuleAttributeError("'{}' object has no attribute '{}'".format(
torch.nn.modules.module.ModuleAttributeError: 'VQModel' object has no attribute 'be_unconditional'

[michael-gc]

It seems like I got the same issue!

I have trained a custom VQGAN model on a custom dataset using the config custom_vqgan.yaml only with training_images_list_file and test_images_list_file modified. And the reconstructions are good.

But when sampling using any of these scripts make_samples.py, sample_conditional, or sample_fast.py with the logs path provided, as introduced in the readme, there are errors! When I dig in, I found that these errors occur because there are no corresponding attributes claimed in the VQModel in vqgan.py, while they are claimed in the Net2NetTransformer in cond_transformer.py.

My vqgan model is trained by python main.py --base configs/custom_vqgan.yaml -t True --gpus 0,1 following @rom1504

1. when running python sample_conditional.py -r ../logs/2021-06-28T22-32-47_custom_vqgan/ --outdir ../results I got errors:

Working with z of shape (1, 256, 16, 16) = 65536 dimensions.
loaded pretrained LPIPS loss from taming/modules/autoencoder/lpips/vgg.pth
VQLPIPSWithDiscriminator running with hinge loss.
Traceback (most recent call last):
  File "/home/michael/vqgan-clip/taming-transformers/scripts/sample_conditional.py", line 355, in <module>
    run_conditional(model, dsets)
  File "/home/michael/.conda/envs/vqgan/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 26, in decorate_context
    return func(*args, **kwargs)
  File "/home/michael/vqgan-clip/taming-transformers/scripts/sample_conditional.py", line 70, in run_conditional
    x = model.get_input("image", example).to(model.device)
  File "/home/michael/vqgan-clip/taming-transformers/taming/models/vqgan.py", line 77, in get_input
    x = batch[k]
TypeError: string indices must be integers

It turns out that the .get_input of class class VQModel(pl.LightningModule) is defined as

    def get_input(self, batch, k):
        x = batch[k]
        if len(x.shape) == 3:
            x = x[..., None]
        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format)
        return x.float()

the variables"image", examplepassed in are wrong. It should be def get_input(self, k, batch)
However this is correct in the class Net2NetTransformer(pl.LightningModule) module.

2. when running python make_samples.py.py -r ../logs/2021-06-28T22-32-47_custom_vqgan/ --outdir ../results got similar error

Working with z of shape (1, 256, 16, 16) = 65536 dimensions.
loaded pretrained LPIPS loss from taming/modules/autoencoder/lpips/vgg.pth
VQLPIPSWithDiscriminator running with hinge loss.
Missing Keys in State Dict: []
Unexpected Keys in State Dict: []
Global step: 107168
Writing samples to  ../results/107168_100_1.0
Dataset:  CustomTrain
  0%|                                                 | 0/12960 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "/home/michael/vqgan-clip/taming-transformers/scripts/make_samples.py", line 292, in <module>
    run_conditional(model, dsets, outdir, opt.top_k, opt.temperature)
  File "/home/michael/.conda/envs/vqgan/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 26, in decorate_context
    return func(*args, **kwargs)
  File "/home/michael/vqgan-clip/taming-transformers/scripts/make_samples.py", line 31, in run_conditional
    x = model.get_input("image", example).to(model.device)
  File "/home/michael/vqgan-clip/taming-transformers/taming/models/vqgan.py", line 77, in get_input
    x = batch[k]
TypeError: string indices must be integers

3. when running python sample_fast.py.py -r ../logs/2021-06-28T22-32-47_custom_vqgan/ --outdir ../results got same error as @ChristianFJung

So how to fix it? If we cannot set the model's target to be taming.models.vqgan.VQModel in the config? Did I do something wrong?

Much appreciated for your help!

[rom1504]

hi,
if you want to sample with this model you need to train a transformer model too, not only the vqgan.
some more details in this comment #54 (comment)
I think the errors you're getting are due to providing the wrong model kind for the sample script

[EAdnds]

Following @rom1504 's instructions for a custom dataset. I have no classes and have trained a model that is producing reconstructions. When attempting to sample by running sample_fast.py, I am receiving this error.

Working with z of shape (1, 256, 16, 16) = 65536 dimensions.
loaded pretrained LPIPS loss from taming/modules/autoencoder/lpips/vgg.pth
VQLPIPSWithDiscriminator running with hinge loss.
Logging to logs/2021-06-25T12-27-06_custom_vqgan/samples/top_k_250_temp_1.00_top_p_1.0/234637
Traceback (most recent call last):
  File "scripts/sample_fast.py", line 262, in <module>
    run(logdir, model, opt.batch_size, opt.temperature, opt.top_k, unconditional=model.be_unconditional,
  File "/home/virginia/dalle/dall2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 778, in __getattr__
    raise ModuleAttributeError("'{}' object has no attribute '{}'".format(
torch.nn.modules.module.ModuleAttributeError: 'VQModel' object has no attribute 'be_unconditional'

Hi,
Did you solve this problem?
I have the same one and I didn't understand how to fix it.

[alex-pv01]

I am dealing with the same problem.

[Echo-jyt]

I am dealing with the same problem.
Hi，have you solved it?