standardizing n_* #298
standardizing n_* #298
Conversation
vincentqb
force-pushed
the
numberof
branch
2 times, most recently
from
aa4c806 to
cbe41ce
Compare
|
We should be careful about the distinction between samples (or "time") and frames. The latter is often a small group of samples. |
|
|
| @@ -537,7 +537,7 @@ def lfilter(waveform, a_coeffs, b_coeffs): | |||
| Performs an IIR filter by evaluating difference equation. | |||
|
|
|||
| Args: | |||
| waveform (torch.Tensor): audio waveform of dimension of `(n_channel, n_frames)`. Must be normalized to -1 to 1. | |||
| waveform (torch.Tensor): audio waveform of dimension of `(channel, time)`. Must be normalized to -1 to 1. | |||
| @@ -558,37 +557,37 @@ def lfilter(waveform, a_coeffs, b_coeffs): | |||
|
|
|||
| device = waveform.device | |||
| dtype = waveform.dtype | |||
| n_channels, n_frames = waveform.size() | |||
| n_channel, n_sample = waveform.size() | |||
There was a problem hiding this comment.
For consistency, channel and frame, rather than n_channel and n_sample?
There was a problem hiding this comment.
We use n_* to refer to the actual value, and * to the name of the dimension. Based on this convention, this would be ok. We could decide to change the convention itself, but I'd leave that to a separate PR if at all.
| n_order = a_coeffs.size(0) | ||
| assert(n_order > 0) | ||
|
|
||
| # Pad the input and create output | ||
| padded_waveform = torch.zeros(n_channels, n_frames + n_order - 1, dtype=dtype, device=device) | ||
| padded_waveform = torch.zeros(n_channel, n_sample + n_order - 1, dtype=dtype, device=device) |
| padded_waveform[:, (n_order - 1):] = waveform | ||
| padded_output_waveform = torch.zeros(n_channels, n_frames + n_order - 1, dtype=dtype, device=device) | ||
| padded_output_waveform = torch.zeros(n_channel, n_sample + n_order - 1, dtype=dtype, device=device) |
| a0_repeated = torch.ones(n_channels, dtype=dtype, device=device) * a_coeffs[0] | ||
| ones = torch.ones(n_channels, n_frames, dtype=dtype, device=device) | ||
| a0_repeated = torch.ones(n_channel, dtype=dtype, device=device) * a_coeffs[0] | ||
| ones = torch.ones(n_channel, n_sample, dtype=dtype, device=device) |
There was a problem hiding this comment.
I saw frame, time, sample are used. Do they refer to the same thing?
There was a problem hiding this comment.
Before this PR, time was used to indicate time dimension, and n_frames and n_samples were used inconsistently to refer to the number of time sample available. Here, I'm suggesting to use frame (and n_frame) to refer to "groups of samples" (e.g. in STFT, we have rolling windows of samples to operate on), and "n_sample" to the number of samples in a "time" dimension.
We could also decide to change "time" dimension to "sample" dimension, but that does not seem to have been a source of confusion, so I would avoid changing this in this PR.
We want to clarify the naming of
n_*variables, initiated by #293. Documentation of standardization in README.