How do set y neq x

[bentoayr]

Hello,
I am trying to build a denoising auto encoder.
Is there a way to make the target data be different from the input data?
In particular, I want the input data to be noisy and the target data to be clean data.
In math, instead of having min_f L(x,f(x)) for some loss function I want to have min_f L(x,f(x + noise))?
Cheers !

[clementchadebec]

Hello, sorry for the late reply. To do this you will need to amend a bit the implementation of the models themselves. The easiest way to do so is to fork this repo and installing it in editable model running

pip install -e .

Then you can amend the forward function so that the noisy data and clean data can be obtained from your datasets.

benchmark_VAE/src/pythae/models/vae/vae_model.py

Lines 65 to 87 in bb785de

	def forward(self, inputs: BaseDataset, **kwargs):
	"""
	The VAE model
	Args:
	inputs (BaseDataset): The training dataset with labels
	Returns:
	ModelOutput: An instance of ModelOutput containing all the relevant parameters
	"""
	x = inputs["data"]
	encoder_output = self.encoder(x)
	mu, log_var = encoder_output.embedding, encoder_output.log_covariance
	std = torch.exp(0.5 * log_var)
	z, eps = self._sample_gauss(mu, std)
	recon_x = self.decoder(z)["reconstruction"]
	loss, recon_loss, kld = self.loss_function(recon_x, x, mu, log_var, z)

I hope this helps,

Best,

Clément