Add Learnable Query #19
Conversation
Keep the latent space as the 4D tensor and then linearize at the end, the same as what the Perceiver models do
Change prints to Debug statements
| x = torch.randn((4, 6, 12, 16, 16)) | ||
| out = query_creator(x) | ||
| # Output is flattened, so should be [B, T*H*W, C] | ||
| assert out.shape == (4, 16 * 16 * 6, 803) |
There was a problem hiding this comment.
Sorry, I probably just dont understand, where does the 803 come from?
There was a problem hiding this comment.
That is the number of channels of the channel_dim so 32 + the number of Fourier Features, which for this is 771
There was a problem hiding this comment.
And the number of Fourier Features comes from (num input axis)((num_freq_bands2)+1) so for this 3*(128*2+1) = 771
This isn't copied from anywhere, just the general idea from using a random latent space to generate predictions from GANs, somewhat the Skillful Nowcasting GAN paper I guess |
jacobbieker
moved this from In progress
to Review in progress
in National Grid Nowcasting: WP1
| _LOG.setLevel(logging.WARN) | ||
|
|
||
|
|
||
| class LearnableQuery(torch.nn.Module): |
There was a problem hiding this comment.
On the topic of position encoding for the queries...
I think the position encoding needs to be 'internally consistent' across each entire example. That is, the position encoding should be consistent across data inputs and the queries so the model can see that, say, the last timestep of the recent history in the input is immediately before the first timestep of the query. And that, in general, the model can see that the timeseries of recent history and the timeseries of queries are two parts of a contiguous timeseries.
To give a concrete example: If the recent history in the input spans 11:00 to 11:55, and the query is for 12:00 to 12:55, then we want the model to see that the first timestep of the query is 5 minutes after the last timestep of the input... if that makes sense?!
One way to do this might be to encode the positions once for all timesteps in the example (i.e. the concatenation of the recent history and the forecast timesteps), and then concatenate the last forecast_timesteps of the position encoding to the queries? Or something like that?! Not sure what's best!
Perhaps this could wait for a future PR though!
This is just a hunch, of course!
And, ultimately, we might want the position encoding to include both the relative position ("the ith element in the array") and the absolute position in time and space ("2pm in South London"). But we can worry about that later :)
There was a problem hiding this comment.
Yeah, I might add that in a follow up PR, but it should be fairly easy to add those as options. I think it's a good idea to try at least!
|
|
||
| Args: | ||
| channel_dim: Channel dimension for the output of the network | ||
| query_shape: The final shape of the query, generally, the (T, H, W) of the output |
There was a problem hiding this comment.
Sorry for being slow but what's the T in (T, H, W)? Time?
There was a problem hiding this comment.
Yes, its the number of timesteps
| elif conv_layer == "2d": | ||
| conv = torch.nn.Conv2d | ||
| else: | ||
| raise ValueError(f"Value for 'layer' is {conv_layer} which is not one of '3d', '2d'") |
Pull Request
Description
This adds a Learnable Query constructor that creates a query with some randomness to possibly be able to "ensemble" predictions through querying the Perceiver output multiple times.
Fixes issue #16
How Has This Been Tested?
Unit tests
Checklist: