Hi! I might be wrong about this since I'm fairly new to the domain, so perhaps I've misunderstood something.
I'm trying to pass a dummy Sentinel-2 instance through the model. To do this, I'm using the provided code to import the model and then construct a sample compatible with Galileo input:
from single_file_galileo import Encoder as SingleFileEncoder
import torch
from pathlib import Path
DATA_FOLDER = Path("data")
sf_model = SingleFileEncoder.load_from_folder(
DATA_FOLDER / "models/nano", device=torch.device("cpu")
)
from src.data.utils import S2_BANDS, construct_galileo_input
from src.masking import MaskedOutput
t, h, w = 2, 4, 4
s2 = torch.randn((t, h, w, len(S2_BANDS)))
masked_output = construct_galileo_input(s2=s2, normalize=True)The problem I'm facing is related to tensor dimensions. For example, the shape of masked_output.space_time_x is torch.Size([2, 4, 4, 13]) -> [T, H, W, len(SPACE_TIME_BANDS)]. Similarly, masked_output.months.shape -> torch.Size([4]) . Essentially, the batch dimension B is missing?
This issue becomes apparent when passing masked_output through the model:
output = sf_model(
masked_output.space_time_x,
masked_output.space_x,
masked_output.time_x,
masked_output.static_x,
masked_output.space_time_mask,
masked_output.space_mask,
masked_output.time_mask,
masked_output.static_mask,
masked_output.months,
patch_size=2
)This results in the following error:
File /galileo/single_file_galileo.py:922, in Encoder.apply_linear_projection(self, s_t_x, sp_x, t_x, st_x, s_t_m, sp_m, t_m, st_m, patch_size)
922 b, h, w, t, _ = s_t_x.shape
ValueError: not enough values to unpack (expected 5, got 4)
This makes sense given the missing batch dimension B.
I considered solving this issue by applying an unsqueeze operation to manually introduce a batch dimension, but I'm unsure if this approach is appropriate.
Again, it's entirely possible I've misunderstood something. Any guidance or clarification would be greatly appreciated. Thanks!
Hi! I might be wrong about this since I'm fairly new to the domain, so perhaps I've misunderstood something.
I'm trying to pass a dummy Sentinel-2 instance through the model. To do this, I'm using the provided code to import the model and then construct a sample compatible with Galileo input:
The problem I'm facing is related to tensor dimensions. For example, the shape of
masked_output.space_time_xistorch.Size([2, 4, 4, 13])->[T, H, W, len(SPACE_TIME_BANDS)]. Similarly,masked_output.months.shape->torch.Size([4]). Essentially, the batch dimensionBis missing?This issue becomes apparent when passing
masked_outputthrough the model:This results in the following error:
This makes sense given the missing batch dimension
B.I considered solving this issue by applying an
unsqueezeoperation to manually introduce a batch dimension, but I'm unsure if this approach is appropriate.Again, it's entirely possible I've misunderstood something. Any guidance or clarification would be greatly appreciated. Thanks!