Input Type and weight type not match

[Chenchao-is-CC]

Hello,

First, thank you for developing such a useful tool!

Recently, I have been using this package to detect smFISH spots from a 3D dataset (dtype: uint16). My data is stored as an OME-Zarr file because it is a large 3D tissue dataset. I see that your package already supports Zarr files, which is great.

However, I encountered the following error:
RuntimeError: Input type (torch.cuda.DoubleTensor) and weight type (torch.cuda.FloatTensor) should be the same

I am not sure what causes this issue. I asked ChatGPT for advice, and here is the recommendation it provided:
In 'spotiflow.py'
In the tiled loop case)

# the origin code  (in line 1028)
tile = torch.from_numpy(tile)
# revised code
tile = torch.from_numpy(tile).float()

As for non-tiled case

# the origin code (in line 867)
img_t = torch.from_numpy(x).to(device).unsqueeze(0)
# revised code
img_t = torch.from_numpy(x).float().to(device).unsqueeze(0)

Finally, after modifying the code, it works correctly. I hope this information can provide a useful hint to help fix this bug. Please double-check the code, as sometimes ChatGPT’s suggestions may not be entirely accurate.

[AlbertDominguez]

Hi @Chenchao-is-CC, thanks for reporting this and for the fix hint!

So I can try to replicate it, are you using the CLI with spotiflow-predict directly with the OME-Zarr, or are you using the Python API? Also which version of Spotiflow are you running? You can check it by running pip show spotiflow in the command line.

[Chenchao-is-CC]

I am using the Python API for spot detection. The standard usage is based on the predict_3d.py script, but it does not support OME-Zarr files. To work around this, I use the imread_wrapped function within your package to read the dataset and then pass the resulting dask.array directly to the Spotiflow class. For reference, I am using Spotiflow version 0.6.1.

[AlbertDominguez]

Unfortunately I can't seem to replicate with the tweaks you are mentioning, seems to work for me. Are you using the default normalizer, or are you giving a explicit one to the predict call? It'd be great if you can share the relevant part from your script so I can take a look!

[Chenchao-is-CC]

Thank you so much for your timely help in solving this problem. I’m sharing an example code snippet below, which I hope will be useful for you.

from spotiflow.model import Spotiflow
from spotiflow.utils import imread_wrapped, write_coords_csv
from pathlib import Path

dataset_path = Path(r"D:\Documents\Python\jupyterFile\spotiflow_test\test_dataset\C1_sample-03_ch1tp0_Rnd2_Gfp_Gfp_reg.zarr")

da_dataset = imread_wrapped(dataset_path, channels=None, zarr_component="/0")

pretrain_model = "smfish_3d"

model = Spotiflow.from_pretrained(pretrain_model)
max_yx_tile_size = 256
max_z_tile_size = 32
min_distance = 1
n_tiles = tuple(max(s//g, 1) for s, g in zip(da_dataset.shape, (max_z_tile_size, max_yx_tile_size, max_yx_tile_size)))

spot, detials = model.predict(img=da_dataset,
                              n_tiles=n_tiles,
                              subpix=True,
                              min_distance=min_distance,
                              device="auto")

spot_results_fn = dataset_path.parent.joinpath(f"{dataset_path.stem}.csv")
write_coords_csv(spot, spot_resulfs_fn)

[AlbertDominguez]

Hi @Chenchao-is-CC,

Thanks for the snippet! I've given another look and tried but still cannot replicate. I'm unsure if it's because of a different zarr format potentially. My main "concern" is that in the code, the array should be cast to np.float32 already by the predict function:

spotiflow/spotiflow/model/spotiflow.py

Line 816 in fc03b01

img = img.astype(np.float32)

Also the normalize_dask function that's called should also cast to np.float32:

spotiflow/spotiflow/utils/utils.py

Lines 379 to 387 in fc03b01

	def normalize_dask(
	x: da.Array,
	pmin: float = 1.0,
	pmax: float = 99.8,
	eps: float = 1e-20,
	max_samples: int = 1e5,
	clip: bool = False,
	dtype: Optional[np.dtype] = np.float32,
	) -> da.Array:

The fix you're proposing is correct as it casts the double into a float, but it should already be cast to that data type by then, so I'm a bit hesitant to simply add another cast because it might he masking another problem out. I'm unsure where the mismatch could be coming from, I'm probably missing something. It'd be great if you could add some debug statements in parts of the code to check the data type at that point, like right before the line of the fix you are proposing, and report them to me if you have the time.

Would you also let know the versions you're using for the dask, numpy, and zarr packages, and also of Python itself?

A couple of other things that are independent but might be helpful just in case:

• I had to transpose the axis to channel-last format (ZYXC) in the snippet you gave me as i'm working with OME-Zarrs. Make sure that the data !
• For the number of tiles I also had to change the computation a bit due to the axis order. You can also try to remove the number of tiles argument for now as there's some sensible defaults that should hopefully work.

[Chenchao-is-CC]

1. Error message (before cast the dtype):

spot, detials = model.predict(img=da_dataset,
                              n_tiles=n_tiles,
                              subpix=True,
                              min_distance=min_distance,
                              device="auto")
INFO:spotiflow.model.spotiflow:Will use device: cuda
INFO:spotiflow.model.spotiflow:Predicting with prob_thresh = [0.4], min_distance = 1
INFO:spotiflow.model.spotiflow:Peak detection mode: fast
INFO:spotiflow.model.spotiflow:Image shape (1594, 4385, 4649)
INFO:spotiflow.model.spotiflow:Predicting with (49, 17, 18) tiles
INFO:spotiflow.model.spotiflow:Normalizing...
INFO:spotiflow.model.spotiflow:Padding to shape (1600, 4416, 4672, 1)
Predicting tiles:   0%|          | 0/14076 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\IPython\core\interactiveshell.py", line 3699, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-5-ba3cf89742d7>", line 1, in <module>
    spot, detials = model.predict(img=da_dataset,
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\spotiflow\model\spotiflow.py", line 1038, in predict
    out = self(img_t)
          ^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\spotiflow\model\spotiflow.py", line 256, in forward
    x = self._downsampler(x)
        ^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\conv.py", line 717, in forward
    return self._conv_forward(input, self.weight, self.bias)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\conv.py", line 712, in _conv_forward
    return F.conv3d(
           ^^^^^^^^^
RuntimeError: Input type (torch.cuda.DoubleTensor) and weight type (torch.cuda.FloatTensor) should be the same

2. OME-Zarr data structure：
   And as for ome-zarr file data, I forgot one thing to tell you. For convenience when loading in Fiji, my OME-Zarr files use Zarr format version 2. Each channel is stored in an independent OME-Zarr file. Therefore, each file contains only a 3D array with shape (Z, Y, X) — not (C, Z, Y, X).

3. Environment details

• python=3.12.11
• dask=2025.9.1
• numpy=2.1.2.
• Spotiflow=0.6.1

4. Ignore argument 'n_tiles':

spot, detials = model.predict(img=da_dataset,
                              #n_tiles=n_tiles,
                              subpix=True,
                              min_distance=min_distance,
                              device="auto")
INFO:spotiflow.model.spotiflow:Will use device: cuda
INFO:spotiflow.model.spotiflow:Predicting with prob_thresh = [0.4], min_distance = 1
INFO:spotiflow.model.spotiflow:Peak detection mode: fast
INFO:spotiflow.model.spotiflow:Image shape (1594, 4385, 4649)
INFO:spotiflow.model.spotiflow:Predicting with (25, 18, 19) tiles
INFO:spotiflow.model.spotiflow:Normalizing...
INFO:spotiflow.model.spotiflow:Padding to shape (1600, 4416, 4672, 1)
Predicting tiles:   0%|          | 0/7038 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\IPython\core\interactiveshell.py", line 3699, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-6-9331d0d47ab2>", line 1, in <module>
    spot, detials = model.predict(img=da_dataset,
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\spotiflow\model\spotiflow.py", line 1038, in predict
    out = self(img_t)
          ^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\spotiflow\model\spotiflow.py", line 256, in forward
    x = self._downsampler(x)
        ^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\conv.py", line 717, in forward
    return self._conv_forward(input, self.weight, self.bias)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\chenchao\.conda\envs\spotiflow\Lib\site-packages\torch\nn\modules\conv.py", line 712, in _conv_forward
    return F.conv3d(
           ^^^^^^^^^
RuntimeError: Input type (torch.cuda.DoubleTensor) and weight type (torch.cuda.FloatTensor) should be the same

[Chenchao-is-CC]

@AlbertDominguez
By the way, if I convert the OME-Zarr file into TIFF format, Spotiflow works perfectly without modifying the original code.
This suggests that the issue might be related to the OME-Zarr format (version 2).