Implement reparameterization

Finish reparameterize: given mu and logvar $= \log\sigma^2$ , draw $\varepsilon \sim \mathcal N(0, I)$ and return $z = \mu + \sigma \odot \varepsilon$ . Everything should work element-wise on tensors of shape (batch, latent_dim).

The reparameterization trick

main.pyPython

Running Python arrives with the isolated execution service

Check results

Static check: the code is not executed. We read the assignments and compare the result with a reference at random values.

Unit tests (once the runner is live)

Output shape matches munot run
Sample mean and variance are close to μ and σ²not run
Gradients reach mu and logvarnot run
Works with different batch sizesnot run
Does not create tensors on another devicenot run