in wall-to-wall example

[neerajk]

Hey,
I'm trying to recreate the wall to wall example, but after we create the datacube and try to run it through the encoder.
I'm getting a error of wrong shapes

EinopsError: Error while processing repeat-reduction pattern "B D -> B L D".
Input tensor shape: torch.Size([48]). Additional info: {'L': 1024}.
Wrong shape: expected 2 dims. Received 1-dim tensor.

---

EinopsError Traceback (most recent call last)
File /opt/anaconda3/envs/claymodel/lib/python3.11/site-packages/einops/einops.py:522, in reduce(tensor, pattern, reduction, **axes_lengths)
521 shape = backend.shape(tensor)
--> 522 recipe = _prepare_transformation_recipe(pattern, reduction, axes_names=tuple(axes_lengths), ndim=len(shape))
523 return _apply_recipe(
524 backend, recipe, cast(Tensor, tensor), reduction_type=reduction, axes_lengths=hashable_axes_lengths
525 )

File /opt/anaconda3/envs/claymodel/lib/python3.11/site-packages/einops/einops.py:365, in _prepare_transformation_recipe(pattern, operation, axes_names, ndim)
364 if ndim != len(left.composition):
--> 365 raise EinopsError(f"Wrong shape: expected {len(left.composition)} dims. Received {ndim}-dim tensor.")
366 left_composition = left.composition

EinopsError: Wrong shape: expected 2 dims. Received 1-dim tensor.

During handling of the above exception, another exception occurred:

EinopsError Traceback (most recent call last)
Cell In[87], line 7
1 #Run the model
2
3 #pass the datacube prepared above to the model to create embeddings.
4 #this will create 1 embedding vector for each image we downloaded
6 with torch.no_grad():
...
--> 533 raise EinopsError(message + "\n {}".format(e))

EinopsError: Error while processing repeat-reduction pattern "B D -> B L D".
Input tensor shape: torch.Size([48]). Additional info: {'L': 1024}.
Wrong shape: expected 2 dims. Received 1-dim tensor.

what exactly is the issue?

[neerajk]

@brunosan
this fixed the dims but still throwing a size error, when running through the model encoder

RuntimeError Traceback (most recent call last)
Cell In[120], line 7
1 #Run the model
2
3 #pass the datacube prepared above to the model to create embeddings.
4 #this will create 1 embedding vector for each image we downloaded
6 with torch.no_grad():
----> 7 unmsk_patch, unmsk_idx, msk_idx, msk_matrix = model.model.encoder(datacube)

File /opt/anaconda3/envs/claymodel/lib/python3.11/site-packages/torch/nn/modules/module.py:1773, in Module._wrapped_call_impl(self, *args, **kwargs)
1771 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1772 else:
-> 1773 return self._call_impl(*args, **kwargs)

File /opt/anaconda3/envs/claymodel/lib/python3.11/site-packages/torch/nn/modules/module.py:1784, in Module._call_impl(self, *args, **kwargs)
1779 # If we don't have any hooks, we want to skip the rest of the logic in
1780 # this function, and just call forward.
1781 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1782 or _global_backward_pre_hooks or _global_backward_hooks
1783 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1784 return forward_call(*args, **kwargs)
1786 result = None
1787 called_always_called_hooks = set()
...
101 if func in _device_constructors() and kwargs.get("device") is None:
102 kwargs["device"] = self.device
--> 103 return func(*args, **kwargs)

RuntimeError: The size of tensor a (1024) must match the size of tensor b (1020) at non-singleton dimension 2

[tcourat]

What is the datacube you are using ?
I had the same issue, but it was because the time and latlon features were not normalized properly.

Indeed, there should be 4 values for time and 4 values for latlon in the datacube, given by the output of normalization function :

normalize_timestamp(date):
   week = date.isocalendar().week * 2 * np.pi / 52
   hour = date.hour * 2 * np.pi / 24

   return (math.sin(week), math.cos(week)), (math.sin(hour), math.cos(hour))

# Prep lat/lon embedding using the
def normalize_latlon(lat, lon):
   lat = lat * np.pi / 180
   lon = lon * np.pi / 180

   return (math.sin(lat), math.cos(lat)), (math.sin(lon), math.cos(lon))

[neerajk]

What is the datacube you are using ? I had the same issue, but it was because the time and latlon features were not normalized properly.

normalization looks good

datacube = {
'platform': platform,
'time' : time_tensor,
'latlon' : latlon_tensor,
'pixels' : pixels.to(device),
'gsd' : torch.tensor(stack.gsd.values, device=device),
'waves' : torch.tensor(waves, device=device),
}

this was the datacube that i used, and yes we need 4 values from time and lat lon, since we are returning 4 values from each of these functions you have showed above

This issue was fixed by issue #376 check it out, instead of using np.hstack, they suggested to use np.columnstack and it worked, there was some dims issue