Pretrained `sccnn` models seem to be broken

[Jiaqi-Lv]

Python version = 3.12
TIAToolBox version = 1.6.0
OS: Ubuntu 22.04.5 LTS

The pretrained model sccnn-conic and sccnn-crchisto seem to be broken. They only output 0 for all input.

Example:

from tiatoolbox.models.architecture import get_pretrained_model
import torch

sccnn_model = get_pretrained_model(pretrained_model="sccnn-conic")[0]
sccnn_model.eval()
sccnn_model.to('cuda')

with torch.no_grad():
    test_input = torch.randn(1, 3, 31, 31).to('cuda')
    test_output = sccnn_model(test_input)
    print(test_output.shape)
    print(torch.max(test_output), torch.min(test_output))

Output:

torch.Size([1, 1, 13, 13])
tensor(0., device='cuda:0') tensor(0., device='cuda:0')

I encounter this "bug" when developing the new NucleusDetector Engine #967 I have tested this multiple times and every time it just outputs 0 for whatever input. Not sure what's going on with this model.

[shaneahmed]

Have you tested

tiatoolbox/tests/models/test_arch_sccnn.py

Lines 26 to 58 in 1614a1f

	def test_functionality(remote_sample: Callable) -> None:
	"""Functionality test for SCCNN.
	Test the functionality of SCCNN model for inference at the patch level.
	"""
	sample_wsi = str(remote_sample("wsi1_2k_2k_svs"))
	reader = WSIReader.open(sample_wsi)
	# * test fast mode (architecture used in PanNuke paper)
	patch = reader.read_bounds(
	(30, 30, 61, 61),
	resolution=0.25,
	units="mpp",
	coord_space="resolution",
	)
	batch = torch.from_numpy(patch)[None]
	model = _load_sccnn(name="sccnn-crchisto")
	output = model.infer_batch(
	model,
	batch,
	device=select_device(on_gpu=env_detection.has_gpu()),
	)
	output = model.postproc(output[0])
	assert np.all(output == [[8, 7]])
	model = _load_sccnn(name="sccnn-conic")
	output = model.infer_batch(
	model,
	batch,
	device=select_device(on_gpu=env_detection.has_gpu()),
	)
	output = model.postproc(output[0])
	assert np.all(output == [[7, 8]])

on develop or master?

[Jiaqi-Lv]

Have you tested

tiatoolbox/tests/models/test_arch_sccnn.py

Lines 26 to 58 in 1614a1f

def test_functionality(remote_sample: Callable) -> None:
"""Functionality test for SCCNN.

 Test the functionality of SCCNN model for inference at the patch level. 

 """ 
 sample_wsi = str(remote_sample("wsi1_2k_2k_svs")) 
 reader = WSIReader.open(sample_wsi) 

 # * test fast mode (architecture used in PanNuke paper) 
 patch = reader.read_bounds( 
     (30, 30, 61, 61), 
     resolution=0.25, 
     units="mpp", 
     coord_space="resolution", 
 ) 
 batch = torch.from_numpy(patch)[None] 
 model = _load_sccnn(name="sccnn-crchisto") 
 output = model.infer_batch( 
     model, 
     batch, 
     device=select_device(on_gpu=env_detection.has_gpu()), 
 ) 
 output = model.postproc(output[0]) 
 assert np.all(output == [[8, 7]]) 
 model = _load_sccnn(name="sccnn-conic") 
 output = model.infer_batch( 
     model, 
     batch, 
     device=select_device(on_gpu=env_detection.has_gpu()), 
 ) 
 output = model.postproc(output[0]) 
 assert np.all(output == [[7, 8]]) 

on develop or master?

Ok I'll try this. I couldn't test it because the server for remote samples was down