Description details¶
In [4]:
root = Path().home() /"Desktop/demo"
!ls {root}
README.md __pycache__ cover.png model.py model_teaser.py rdf.yaml sample_input.tif sample_output.tif test_input.npy test_output.npy weights.pt weights_torchscript.pt
In [5]:
show(root / "test_input.npy", root / "test_output.npy")
In [6]:
show(root / "cover.png")
Source Code of the model architecture¶
In [7]:
from IPython.display import Code
Code(filename=root / "model_teaser.py", language='python')
Out[7]:
import torch.nn as nn
class HyLFM_Net(nn.Module):
def __init__(
self,
*,
z_out: int,
nnum: int,
kernel2d: int = 3,
conv_per_block2d: int = 2,
...
):
...
def forward(self, x):
x = self.channel_from_lf(x)
x = self.res2d(x)
x = self.conv2d(x)
x = self.c2z(x)
x = self.res3d(x)
x = self.conv3d(x)
...
return x
Model Weights¶
In [8]:
pytorch_weights = torch.load(root / "weights.pt", weights_only=False)
pprint([(k, tuple(v.shape)) for k, v in pytorch_weights.items()][:4] + ["..."])
[('res2d.0.block.0.conv.weight', (768, 361, 3, 3)),
('res2d.0.block.0.conv.bias', (768,)),
('res2d.0.block.1.conv.weight', (768, 768, 3, 3)),
('res2d.0.block.1.conv.bias', (768,)),
'...']
Model Input¶
In [ ]:
input_descr = InputTensorDescr(
id=TensorId("lf"),
axes=[BatchAxis(),
ChannelAxis(channel_names=[Identifier("lf")]),
SpaceInputAxis(
id=AxisId('y'),
size=ParameterizedSize(min=190, step=19),
scale=4,
concatenable=False),
SpaceInputAxis(
id=AxisId('x'),
size=ParameterizedSize(min=190, step=19),
scale=4,
concatenable=False),
],
test_tensor=FileDescr(source=root / "test_input.npy"),
sample_tensor=FileDescr(source=root / "sample_input.tif"),
data=IntervalOrRatioDataDescr(type="float32"),
preprocessing = [ScaleRangeDescr(
kwargs=ScaleRangeKwargs(
axes= (AxisId('y'), AxisId('x')),
max_percentile= 99.8,
min_percentile= 5.0,
)
)]
)
Model Output¶
In [ ]:
from bioimageio.spec.model.v0_5 import OutputTensorDescr, SizeReference, SpaceOutputAxis
output_descr = OutputTensorDescr(
id=TensorId("prediction"),
description="predicted volume of fluorescence signal",
axes=[BatchAxis(),
ChannelAxis(channel_names=[Identifier("prediction")]),
SpaceOutputAxis(id=AxisId("z"), size=49),
SpaceOutputAxis(id=AxisId('y'),
scale=19,
size=SizeReference(tensor_id=TensorId("lf"),
axis_id=AxisId("y"),
offset=-16)),
SpaceOutputAxis(id=AxisId('x'),
scale=19,
size=SizeReference(tensor_id=TensorId("lf"),
axis_id=AxisId("x"),
offset=-16)),
],
test_tensor=FileDescr(source=root / "test_output.npy"),
sample_tensor=FileDescr(source=root / "sample_output.tif"),
data=IntervalOrRatioDataDescr(type="float32"),
)
Model Architecture¶
In [ ]:
from bioimageio.spec.model.v0_5 import (
ArchitectureFromFileDescr,
Version,
)
pytorch_version = Version(torch.__version__)
pytorch_architecture = ArchitectureFromFileDescr(
source=root / "model.py",
callable=Identifier("HyLFM_Net"),
kwargs=dict(
c_in_3d=64,
last_kernel2d=5,
nnum=19,
z_out=49,
)
)
Create a model description¶
In [ ]:
my_model_descr = ModelDescr(
name="HyLFM-Net-stat",
description=(
"HyLFM-Net trained on static images of arrested medaka hatchling hearts. "
+ "The network reconstructs a volumentric image from a given light-field."
),
covers=[root / "cover.png"],
authors=[
Author(
name="Fynn Beuttenmueller",
affiliation="EMBL Heidelberg",
github_user="fynnbe",
orcid=OrcidId("0000-0002-8567-6389"))
],
cite=[
CiteEntry(
text=(
"Beuttenmueller, Wagner, N., F., Norlin, N. et al."
+ " Deep learning-enhanced light-field imaging with continuous validation."
+ " Nat Methods 18, 557–563 (2021)."
),
doi=Doi("10.1038/s41592-021-01136-0"),
)
],
license=LicenseId("MIT"),
...
In [ ]:
...
documentation=root / "README.md",
git_repo=HttpUrl("https://github.com/kreshuklab/hylfm-net"),
tags= [
"light-field-microscopy",
"pytorch",
"fluorescence-light-microscopy",
"image-reconstruction",
"nuclei",
"hylfm",
],
training_data=LinkedDataset(id=DatasetId("uplifting-ice-cream")),
inputs=[input_descr],
outputs=[output_descr],
weights=WeightsDescr(
pytorch_state_dict=PytorchStateDictWeightsDescr(
source=root / "weights.pt",
architecture=pytorch_architecture,
pytorch_version=pytorch_version,
),
torchscript=TorchscriptWeightsDescr(
source=root / "weights_torchscript.pt",
pytorch_version=pytorch_version,
parent="pytorch_state_dict", # these weights were converted from the pytorch_state_dict weights ones.
),
),
attachments=[FileDescr(source=root / "model_teaser.py")],
)
Test a described model¶
In [ ]:
from bioimageio.core import test_model
validation_summary = test_model(my_model_descr)
In [14]:
validation_summary.display()
| ✔️ | bioimageio validation passed |
|---|---|
| source | in-memory |
| format version | model 0.5.3 |
| bioimageio.spec | 0.5.3.3 |
| bioimageio.core | 0.6.10 |
| ❓ | location | detail |
|---|---|---|
| ✔️ | initialized ModelDescr to describe model 0.5.3 | |
| ✔️ | Has expected resource type | |
| ✔️ | Reproduce test outputs from test inputs (pytorch_state_dict) | |
| ✔️ | Run pytorch_state_dict inference for inputs with batch_size: 1 and size parameter n: 0 | |
| ✔️ | Run pytorch_state_dict inference for inputs with batch_size: 2 and size parameter n: 0 | |
| ✔️ | Run pytorch_state_dict inference for inputs with batch_size: 1 and size parameter n: 1 | |
| ✔️ | Run pytorch_state_dict inference for inputs with batch_size: 2 and size parameter n: 1 | |
| ✔️ | Run pytorch_state_dict inference for inputs with batch_size: 1 and size parameter n: 2 | |
| ✔️ | Run pytorch_state_dict inference for inputs with batch_size: 2 and size parameter n: 2 | |
| ✔️ | Reproduce test outputs from test inputs (torchscript) | |
| ✔️ | Run torchscript inference for inputs with batch_size: 1 and size parameter n: 0 | |
| ✔️ | Run torchscript inference for inputs with batch_size: 2 and size parameter n: 0 | |
| ✔️ | Run torchscript inference for inputs with batch_size: 1 and size parameter n: 1 | |
| ✔️ | Run torchscript inference for inputs with batch_size: 2 and size parameter n: 1 | |
| ✔️ | Run torchscript inference for inputs with batch_size: 1 and size parameter n: 2 | |
| ✔️ | Run torchscript inference for inputs with batch_size: 2 and size parameter n: 2 |