Hello, I am very interested in your work, and this code is very helpful to me. However, I still have some issues about oatomobile/oatomobile/torch/networks/sequence.py
- The default output_shape of AutoregressiveFlow is (4,2), and the trajectory consisting of 40 points is obtained by interpolating the 4 output points. Why not just output more points？
- If the output_shape of AutoregressiveFlow is not (4,2), e.g., (16,2), the output_sizes of self._locscale may be not right, and the following code will error out:
dloc_scale = self._locscale(z)
dloc = dloc_scale[..., :2]
scale = F.softplus(dloc_scale[..., 2:]) + 1e-3
y_t = (y_tm1 + dloc) + scale * x_t
So I change the output_sizes of self._locscale to a fixed [32, 4], but I don't know if this is appropriate.
3. I'm not very familiar with Autoregressive Flow, so I'm confused about the process of calculating y and logabsdet. I would appreciate it if you could tell me which paper has a description of this part of the calculation process.
Hello, I am very interested in your work, and this code is very helpful to me. However, I still have some issues about oatomobile/oatomobile/torch/networks/sequence.py
So I change the output_sizes of self._locscale to a fixed [32, 4], but I don't know if this is appropriate.
3. I'm not very familiar with Autoregressive Flow, so I'm confused about the process of calculating y and logabsdet. I would appreciate it if you could tell me which paper has a description of this part of the calculation process.