BSD68
Find me on Github The BSD68 dataset was adapted from K. Zhang et al (TIP, 2017) and is composed of natural images. The noise was artificially added, allowing for quantitative comparisons with the ground truth, one of the benchmark used in many denoising publications. Here, we check the performances of Noise2Void.
In [1]:
Copied!
# Imports necessary to execute the code
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import tifffile
from careamics import CAREamist
from careamics.config import create_n2v_configuration
from careamics.utils.metrics import scale_invariant_psnr
from careamics_portfolio import PortfolioManager
from PIL import Image
# Imports necessary to execute the code from pathlib import Path import matplotlib.pyplot as plt import numpy as np import tifffile from careamics import CAREamist from careamics.config import create_n2v_configuration from careamics.utils.metrics import scale_invariant_psnr from careamics_portfolio import PortfolioManager from PIL import Image
Import the dataset¶
The dataset can be directly downloaded using the careamics-portfolio package, which uses pooch to download the data.
In [ ]:
Copied!
# instantiate data portfolio manage
portfolio = PortfolioManager()
# and download the data
root_path = Path("./data")
files = portfolio.denoising.N2V_BSD68.download(root_path)
# create paths for the data
data_path = Path(root_path / "denoising-N2V_BSD68.unzip/BSD68_reproducibility_data")
train_path = data_path / "train"
val_path = data_path / "val"
test_path = data_path / "test" / "images"
gt_path = data_path / "test" / "gt"
# instantiate data portfolio manage portfolio = PortfolioManager() # and download the data root_path = Path("./data") files = portfolio.denoising.N2V_BSD68.download(root_path) # create paths for the data data_path = Path(root_path / "denoising-N2V_BSD68.unzip/BSD68_reproducibility_data") train_path = data_path / "train" val_path = data_path / "val" test_path = data_path / "test" / "images" gt_path = data_path / "test" / "gt"
Visualize data¶
In [3]:
Copied!
# load training and validation image and show them side by side
single_train_image = tifffile.imread(next(iter(train_path.rglob("*.tiff"))))[0]
single_val_image = tifffile.imread(next(iter(val_path.rglob("*.tiff"))))[0]
fig, ax = plt.subplots(1, 2, figsize=(10, 5))
ax[0].imshow(single_train_image, cmap="gray")
ax[0].set_title("Training Image")
ax[1].imshow(single_val_image, cmap="gray")
ax[1].set_title("Validation Image")
# load training and validation image and show them side by side single_train_image = tifffile.imread(next(iter(train_path.rglob("*.tiff"))))[0] single_val_image = tifffile.imread(next(iter(val_path.rglob("*.tiff"))))[0] fig, ax = plt.subplots(1, 2, figsize=(10, 5)) ax[0].imshow(single_train_image, cmap="gray") ax[0].set_title("Training Image") ax[1].imshow(single_val_image, cmap="gray") ax[1].set_title("Validation Image")
Out[3]:
Text(0.5, 1.0, 'Validation Image')
Train with CAREamics¶
The easiest way to use CAREamics is to create a configuration and a CAREamist.
Create configuration¶
The configuration can be built from scratch, giving the user full control over the various parameters available in CAREamics. However, a straightforward way to create a configuration for a particular algorithm is to use one of the convenience functions.
In [4]:
Copied!
config = create_n2v_configuration(
experiment_name="bsd68_n2v",
data_type="tiff",
axes="SYX",
patch_size=(64, 64),
batch_size=64,
num_epochs=100,
)
print(config)
config = create_n2v_configuration( experiment_name="bsd68_n2v", data_type="tiff", axes="SYX", patch_size=(64, 64), batch_size=64, num_epochs=100, ) print(config)
{'algorithm_config': {'algorithm': 'n2v',
'loss': 'n2v',
'lr_scheduler': {'name': 'ReduceLROnPlateau',
'parameters': {}},
'model': {'architecture': 'UNet',
'conv_dims': 2,
'depth': 2,
'final_activation': 'None',
'in_channels': 1,
'independent_channels': True,
'n2v2': False,
'num_channels_init': 32,
'num_classes': 1},
'optimizer': {'name': 'Adam',
'parameters': {'lr': 0.0001}}},
'data_config': {'axes': 'SYX',
'batch_size': 64,
'data_type': 'tiff',
'patch_size': [64, 64],
'transforms': [{'flip_x': True,
'flip_y': True,
'name': 'XYFlip',
'p': 0.5},
{'name': 'XYRandomRotate90', 'p': 0.5},
{'masked_pixel_percentage': 0.2,
'name': 'N2VManipulate',
'roi_size': 11,
'strategy': 'uniform',
'struct_mask_axis': 'none',
'struct_mask_span': 5}]},
'experiment_name': 'bsd68_n2v',
'training_config': {'accumulate_grad_batches': 1,
'check_val_every_n_epoch': 1,
'checkpoint_callback': {'auto_insert_metric_name': False,
'mode': 'min',
'monitor': 'val_loss',
'save_last': True,
'save_top_k': 3,
'save_weights_only': False,
'verbose': False},
'enable_progress_bar': True,
'gradient_clip_algorithm': 'norm',
'max_steps': -1,
'num_epochs': 100,
'precision': '32'},
'version': '0.1.0'}
Train¶
A CAREamist can be created using a configuration alone, and then be trained by using the data already loaded in memory.
In [ ]:
Copied!
# instantiate a CAREamist
careamist = CAREamist(source=config)
# train
careamist.train(
train_source=train_path,
val_source=val_path,
)
# instantiate a CAREamist careamist = CAREamist(source=config) # train careamist.train( train_source=train_path, val_source=val_path, )
Predict with CAREamics¶
Prediction is done with the same CAREamist used for training.
In [ ]:
Copied!
prediction = careamist.predict(
source=test_path,
axes="YX",
tile_size=(128, 128),
tile_overlap=(48, 48),
batch_size=1,
)
prediction = careamist.predict( source=test_path, axes="YX", tile_size=(128, 128), tile_overlap=(48, 48), batch_size=1, )
Visualize the prediction¶
In [7]:
Copied!
# Show two images
noises = [tifffile.imread(f) for f in sorted(test_path.glob("*.tiff"))]
gts = [tifffile.imread(f) for f in sorted(gt_path.glob("*.tiff"))]
# images to show
images = np.random.choice(range(len(noises)), 3)
fig, ax = plt.subplots(3, 3, figsize=(15, 15))
fig.tight_layout()
for i in range(3):
pred_image = prediction[images[i]].squeeze()
psnr_noisy = scale_invariant_psnr(gts[images[i]], noises[images[i]])
psnr_result = scale_invariant_psnr(gts[images[i]], pred_image)
ax[i, 0].imshow(noises[images[i]], cmap="gray")
ax[i, 0].title.set_text(f"Noisy\nPSNR: {psnr_noisy:.2f}")
ax[i, 1].imshow(pred_image, cmap="gray")
ax[i, 1].title.set_text(f"Prediction\nPSNR: {psnr_result:.2f}")
ax[i, 2].imshow(gts[images[i]], cmap="gray")
ax[i, 2].title.set_text("Ground-truth")
# Show two images noises = [tifffile.imread(f) for f in sorted(test_path.glob("*.tiff"))] gts = [tifffile.imread(f) for f in sorted(gt_path.glob("*.tiff"))] # images to show images = np.random.choice(range(len(noises)), 3) fig, ax = plt.subplots(3, 3, figsize=(15, 15)) fig.tight_layout() for i in range(3): pred_image = prediction[images[i]].squeeze() psnr_noisy = scale_invariant_psnr(gts[images[i]], noises[images[i]]) psnr_result = scale_invariant_psnr(gts[images[i]], pred_image) ax[i, 0].imshow(noises[images[i]], cmap="gray") ax[i, 0].title.set_text(f"Noisy\nPSNR: {psnr_noisy:.2f}") ax[i, 1].imshow(pred_image, cmap="gray") ax[i, 1].title.set_text(f"Prediction\nPSNR: {psnr_result:.2f}") ax[i, 2].imshow(gts[images[i]], cmap="gray") ax[i, 2].title.set_text("Ground-truth")
Compute metrics¶
In [8]:
Copied!
psnrs = np.zeros((len(prediction), 1))
for i, (pred, gt) in enumerate(zip(prediction, gts)):
psnrs[i] = scale_invariant_psnr(gt, pred.squeeze())
print(f"PSNR: {psnrs.mean():.2f} +/- {psnrs.std():.2f}")
print("Reported PSNR: 27.71")
psnrs = np.zeros((len(prediction), 1)) for i, (pred, gt) in enumerate(zip(prediction, gts)): psnrs[i] = scale_invariant_psnr(gt, pred.squeeze()) print(f"PSNR: {psnrs.mean():.2f} +/- {psnrs.std():.2f}") print("Reported PSNR: 27.71")
PSNR: 26.39 +/- 2.59 Reported PSNR: 27.71
Create cover¶
In [9]:
Copied!
# create a cover image
im_idx = 3
cv_image_noisy = noises[im_idx]
cv_image_pred = prediction[im_idx].squeeze()
# create image
cover = np.zeros((256, 256))
(height, width) = cv_image_noisy.shape
assert height > 256
assert width > 256
# normalize train and prediction
norm_noise = (cv_image_noisy - cv_image_noisy.min()) / (
cv_image_noisy.max() - cv_image_noisy.min()
)
norm_pred = (cv_image_pred - cv_image_pred.min()) / (
cv_image_pred.max() - cv_image_pred.min()
)
# fill in halves
cover[:, : 256 // 2] = norm_noise[
height // 2 - 256 // 2 : height // 2 + 256 // 2, width // 2 - 256 // 2 : width // 2
]
cover[:, 256 // 2 :] = norm_pred[
height // 2 - 256 // 2 : height // 2 + 256 // 2, width // 2 : width // 2 + 256 // 2
]
# plot the single image
plt.imshow(cover, cmap="gray")
# save the image
im = Image.fromarray(cover * 255)
im = im.convert("L")
im.save("BSD68_Noise2Void.jpeg")
# create a cover image im_idx = 3 cv_image_noisy = noises[im_idx] cv_image_pred = prediction[im_idx].squeeze() # create image cover = np.zeros((256, 256)) (height, width) = cv_image_noisy.shape assert height > 256 assert width > 256 # normalize train and prediction norm_noise = (cv_image_noisy - cv_image_noisy.min()) / ( cv_image_noisy.max() - cv_image_noisy.min() ) norm_pred = (cv_image_pred - cv_image_pred.min()) / ( cv_image_pred.max() - cv_image_pred.min() ) # fill in halves cover[:, : 256 // 2] = norm_noise[ height // 2 - 256 // 2 : height // 2 + 256 // 2, width // 2 - 256 // 2 : width // 2 ] cover[:, 256 // 2 :] = norm_pred[ height // 2 - 256 // 2 : height // 2 + 256 // 2, width // 2 : width // 2 + 256 // 2 ] # plot the single image plt.imshow(cover, cmap="gray") # save the image im = Image.fromarray(cover * 255) im = im.convert("L") im.save("BSD68_Noise2Void.jpeg")
In [11]:
Copied!
# Export the model
careamist.export_to_bmz(
path_to_archive="bsd68_n2v_model.zip",
friendly_model_name="BSD68_N2V",
input_array=noises[im_idx][np.newaxis, ..., :256, :256],
authors=[{"name": "CAREamics authors", "affiliation": "Human Technopole"}],
)
# Export the model careamist.export_to_bmz( path_to_archive="bsd68_n2v_model.zip", friendly_model_name="BSD68_N2V", input_array=noises[im_idx][np.newaxis, ..., :256, :256], authors=[{"name": "CAREamics authors", "affiliation": "Human Technopole"}], )
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0] /localscratch/miniforge3/envs/careamics/lib/python3.10/site-packages/pytorch_lightning/trainer/connectors/data_connector.py:424: The 'predict_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` to `num_workers=3` in the `DataLoader` to improve performance.
Predicting: | | 0/? [00:00<?, ?it/s]
computing SHA256 of inputs.npy (result: 4e99b89c03663aafbcdd7b1a92d0db4e3421c05ea63f68163e7a5d61775095c8): 100%|██████████| 3/3 [00:00<00:00, 1180.05it/s] computing SHA256 of outputs.npy (result: a79cf7edbaee8320b4c19d755cc2a4d7945681e8e6522f45a9c056c54b6a47ef): 100%|██████████| 3/3 [00:00<00:00, 1695.35it/s] computing SHA256 of environment.yml (result: d743faa28b30f054d6caa699271006e27d3cda48e07c02cf092b4e0732830574): 100%|██████████| 1/1 [00:00<00:00, 867.31it/s] computing SHA256 of weights.pth (result: ca2b5bafccb5d5d4f35b557d17c973a2a0cce8c51515d4867092970e278ae849): 100%|██████████| 16/16 [00:00<00:00, 5814.32it/s] computing SHA256 of config.yml (result: f57e31e3392743a721fd489d29628c89a23687563470a5bf46fb81825ee80b9e): 100%|██████████| 1/1 [00:00<00:00, 912.60it/s] 2024-09-19 16:15:00.647 | Level 30 | bioimageio.spec._internal.field_warning:issue_warning:149 - documentation: No '# Validation' (sub)section found in /home/melisande.croft/.careamics/README.md. 2024-09-19 16:15:00.707 | INFO | bioimageio.core._resource_tests:_test_model_inference:147 - starting 'Reproduce test outputs from test inputs (pytorch_state_dict)' /localscratch/miniforge3/envs/careamics/lib/python3.10/site-packages/bioimageio/core/model_adapters/_pytorch_model_adapter.py:44: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature. state: Any = torch.load( 2024-09-19 16:15:00.946 | INFO | bioimageio.core._resource_tests:_test_model_inference_parametrized:242 - Testing inference with 2 different input tensor sizes /localscratch/miniforge3/envs/careamics/lib/python3.10/site-packages/bioimageio/core/model_adapters/_pytorch_model_adapter.py:44: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature. state: Any = torch.load( computing SHA256 of config.yml (result: f57e31e3392743a721fd489d29628c89a23687563470a5bf46fb81825ee80b9e): 100%|██████████| 1/1 [00:00<00:00, 299.21it/s] computing SHA256 of inputs.npy (result: 4e99b89c03663aafbcdd7b1a92d0db4e3421c05ea63f68163e7a5d61775095c8): 100%|██████████| 3/3 [00:00<00:00, 719.27it/s] computing SHA256 of outputs.npy (result: a79cf7edbaee8320b4c19d755cc2a4d7945681e8e6522f45a9c056c54b6a47ef): 100%|██████████| 3/3 [00:00<00:00, 748.40it/s] computing SHA256 of environment.yml (result: d743faa28b30f054d6caa699271006e27d3cda48e07c02cf092b4e0732830574): 100%|██████████| 1/1 [00:00<00:00, 243.50it/s] computing SHA256 of weights.pth (result: ca2b5bafccb5d5d4f35b557d17c973a2a0cce8c51515d4867092970e278ae849): 100%|██████████| 16/16 [00:00<00:00, 3131.98it/s]
In [ ]:
Copied!