Decode more than one molecule from one latent

[cankobanz]

Hello, thank you for sharing your work.

I have a question regarding the decoding process. I'm curious about how to decode more than one molecule from a specific latent space, especially those that are neighbors with high log likelihood relative to the returned molecule.

Currently, it seems that for the same latent input, the decoded output remains unchanged, and the sampling process doesn't seem to support starting from a specified latent position:

molecule-generation/molecule_generation/wrapper.py

Lines 156 to 165 in 48d532f

	def sample(self, num_samples: int) -> List[str]:
	"""Sample SMILES strings from the model.
	Args:
	num_samples: Number of samples to return.
	Returns:
	List of SMILES strings.
	"""
	return self.decode(self.sample_latents(num_samples))

I was considering customizing the sample method of the GeneratorWrapper to initiate from a specific latent point instead of starting from zeros. However, the provided checkpoint (GNN_Edge_MLP_MoLeR__2022-02-24_07-16-23_best.pkl) is configured for the VaeWrapper, not the GeneratorWrapper.

Note: I have taken into consideration your suggestion to add small noise to address this issue, as discussed in issue 40. However, my primary interest lies in exploring a more refined solution, specifically through adjusting the num_samples parameter here:

molecule-generation/molecule_generation/utils/moler_decoding_utils.py

Lines 64 to 72 in 48d532f

	num_samples = min(num_samples, num_choices) # Handle cases where we only have few candidates
	if sampling_mode == DecoderSamplingMode.GREEDY:
	# Note that this will return the top num_samples indices, but not in order:
	picked_indices = np.argpartition(logprobs, -num_samples)[-num_samples:]
	elif sampling_mode == DecoderSamplingMode.SAMPLING:
	p = np.exp(logprobs) # Convert to probabilities
	# We can only sample values with non-zero probabilities
	num_choices = np.sum(p > 0)
	num_samples = min(num_samples, num_choices)

Thank you in advance for your assistance.

[kmaziarz]

Sorry, I somehow missed your question!

I was considering customizing the sample method of the GeneratorWrapper to initiate from a specific latent point instead of starting from zeros.

Given that a generator-style model was trained always receiving an all-zeros input, this may not work. I would rather use the vae-style model which has been more thoroughly validated.

As you said, you could either go with perturbing the latent code, or make decoding randomized so that it can return different results from a single latent code. The latter option is not exposed in the model wrapper, but you could modify the sampling_mode argument in MoLeRInferenceServer directly (see here). However, I'm not sure if this would give better results than the perturbation-based approach; it would be best to try both and see which one works better empirically for your usecase.