Experiments - Microsplit/Musplit on JUMP-QC dataset

[ShataDg]

JUMPQC-43

• Training set: 20 images/channel,
• Channels: DNA, AGP
• Epochs: 5
• Noise models: ch0 and ch1 (from HTLIF24 example data)

Step size was 2048 for this run

The following appeared between epochs 0 and 1

Metric val_loss improved. New best score: 0.002
[neptune] [error  ] Error occurred during asynchronous operation processing: X-coordinates (step) must be strictly increasing for series attribute: training/epoch. Invalid point: 2047.0

Same error also appeared between epochs 1 and 2

Metric val_loss improved by 0.001 >= min_delta = 1e-06. New best score: 0.002
[neptune] [error  ] Error occurred during asynchronous operation processing: X-coordinates (step) must be strictly increasing for series attribute: training/epoch. Invalid point: 4095.0

Proper prediction images were not generated

[ShataDg]

JUMPQC-44

• Training set: 20 images/channel,
• Channels: DNA, AGP
• Epochs: 5
• Noise models: ch1 and ch2 (from HTLIF24 example data)

model = VAEModule(algorithm_config=experiment_config)

expects a Ch0.npz

---------------------------------------------------------------------------
FileNotFoundError                         Traceback (most recent call last)
Cell In[11], [line 1](vscode-notebook-cell:?execution_count=11&line=1)
----> [1](vscode-notebook-cell:?execution_count=11&line=1) model = VAEModule(algorithm_config=experiment_config)

File ~/anaconda3/envs/microsplit/lib/python3.11/site-packages/careamics/lightning/lightning_module.py:335, in VAEModule.__init__(self, algorithm_config)
    332 self.model: nn.Module = model_factory(self.algorithm_config.model)
    334 # create loss function
--> [335](https://file+.vscode-resource.vscode-cdn.net/Users/sdasgupt/Documents/microsplit/microsplit/MicroSplit-reproducibility/examples/2D/custom_dataset/~/anaconda3/envs/microsplit/lib/python3.11/site-packages/careamics/lightning/lightning_module.py:335) self.noise_model: Optional[NoiseModel] = noise_model_factory(
    336     self.algorithm_config.noise_model
    337 )
    339 self.noise_model_likelihood: Optional[NoiseModelLikelihood] = (
    340     likelihood_factory(
    341         config=self.algorithm_config.noise_model_likelihood,
    342         noise_model=self.noise_model,
    343     )
    344 )
    346 self.gaussian_likelihood: Optional[GaussianLikelihood] = likelihood_factory(
    347     self.algorithm_config.gaussian_likelihood
    348 )

File ~/anaconda3/envs/microsplit/lib/python3.11/site-packages/careamics/models/lvae/noise_models.py:101, in noise_model_factory(model_config)
     99 if nm.path:
    100     if nm.model_type == "GaussianMixtureNoiseModel":
--> [101](https://file+.vscode-resource.vscode-cdn.net/Users/sdasgupt/Documents/microsplit/microsplit/MicroSplit-reproducibility/examples/2D/custom_dataset/~/anaconda3/envs/microsplit/lib/python3.11/site-packages/careamics/models/lvae/noise_models.py:101)         noise_models.append(GaussianMixtureNoiseModel(nm))
    102     else:
    103         raise NotImplementedError(
    104             f"Model {nm.model_type} is not implemented"
    105         )

File ~/anaconda3/envs/microsplit/lib/python3.11/site-packages/careamics/models/lvae/noise_models.py:271, in GaussianMixtureNoiseModel.__init__(self, config)
    268 self.device = torch.device("cpu")
    270 if config.path is not None:
--> [271](https://file+.vscode-resource.vscode-cdn.net/Users/sdasgupt/Documents/microsplit/microsplit/MicroSplit-reproducibility/examples/2D/custom_dataset/~/anaconda3/envs/microsplit/lib/python3.11/site-packages/careamics/models/lvae/noise_models.py:271)     params = np.load(config.path)
    272 else:
    273     params = config.model_dump(exclude_none=True)

File ~/anaconda3/envs/microsplit/lib/python3.11/site-packages/numpy/lib/npyio.py:427, in load(file, mmap_mode, allow_pickle, fix_imports, encoding, max_header_size)
    425     own_fid = False
    426 else:
--> [427](https://file+.vscode-resource.vscode-cdn.net/Users/sdasgupt/Documents/microsplit/microsplit/MicroSplit-reproducibility/examples/2D/custom_dataset/~/anaconda3/envs/microsplit/lib/python3.11/site-packages/numpy/lib/npyio.py:427)     fid = stack.enter_context(open(os_fspath(file), "rb"))
    428     own_fid = True
    430 # Code to distinguish from NumPy binary files and pickles.

FileNotFoundError: [Errno 2] No such file or directory: 'noise_models/noise_model_Ch0.npz'

I renamed Ch2.npz as Ch0.npz

The same warning appeared during this run

Metric val_loss improved. New best score: 0.002
[neptune] [error  ] Error occurred during asynchronous operation processing: X-coordinates (step) must be strictly increasing for series attribute: training/epoch. Invalid point: 2047.0

Proper prediction was not generated

[ShataDg]

JUMPQC-45

Trying musplit to avoid using noise models

• Training set: 20 images/channel,
• Channels: DNA, AGP
• Epochs: 5

Musplit algorithm does not use noise models unlike denoisesplit and denoisesplit_musplit

Even though the algorithm does not use a noise model, it requires a noise_model_path in experiment_params

# setting up train, validation, and test data configs
train_data_config, val_data_config, test_data_config = get_data_configs(
    image_size=PATCH_SIZE,
    num_channels=NUM_CHANNELS
)

# setting up MicroSplit parametrization
experiment_params = get_microsplit_parameters(
    algorithm="musplit", # changed to "musplit"
    img_size=PATCH_SIZE,
    batch_size=BATCH_SIZE,
    num_epochs=EPOCHS,
    multiscale_count=3,
    noise_model_path=NM_PATH,
    target_channels=NUM_CHANNELS,
)

# create the dataset
train_dset, val_dset, _, data_stats = create_train_val_datasets(
    datapath=DATA_PATH,
    train_config=train_data_config,
    val_config=val_data_config,
    test_config=val_data_config,
    load_data_func=get_train_val_data,
)

Commenting out noise_model_path=NM_PATH produces the following error

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[7], [line 8](vscode-notebook-cell:?execution_count=7&line=8)
      2 train_data_config, val_data_config, test_data_config = get_data_configs(
      3     image_size=PATCH_SIZE,
      4     num_channels=NUM_CHANNELS
      5 )
      7 # setting up MicroSplit parametrization
----> [8](vscode-notebook-cell:?execution_count=7&line=8) experiment_params = get_microsplit_parameters(
      9     algorithm="musplit", # changed to "musplit"
     10     img_size=PATCH_SIZE,
     11     batch_size=BATCH_SIZE,
     12     num_epochs=EPOCHS,
     13     multiscale_count=3,
     14     target_channels=NUM_CHANNELS,
     15 )
     17 # create the dataset
     18 train_dset, val_dset, _, data_stats = create_train_val_datasets(
     19     datapath=DATA_PATH,
     20     train_config=train_data_config,
   (...)     23     load_data_func=get_train_val_data,
     24 )

TypeError: get_microsplit_parameters() missing 1 required positional argument: 'noise_model_path'

I provided the noise_model_path, and now it shows that the model will use GaussianLikelihood only

[GaussianLikelihood] PredLVar:pixelwise LowBLVar:-5.0

The output looks like the following when it uses noise models

[GaussianMixtureNoiseModel] min_sigma: 200.0
[GaussianMixtureNoiseModel] min_sigma: 200.0
[MultiChannelNoiseModel] Nmodels count:2
[GaussianLikelihood] PredLVar:pixelwise LowBLVar:-5.0

Output when training run is initiated

  | Name                | Type               | Params | Mode 
-------------------------------------------------------------------
0 | model               | LadderVAE          | 3.6 M  | train
1 | gaussian_likelihood | GaussianLikelihood | 0      | train
-------------------------------------------------------------------
3.6 M     Trainable params
0         Non-trainable params
3.6 M     Total params
14.270    Total estimated model params size (MB)
317       Modules in train mode
0         Modules in eval mode

Predictions

[ShataDg]

NEXT STEPS

• 1. Increase dataset
• 2. Increase epochs
• 3. Hyperparameter tuning
• 4. Prioritize testing
• 5. Check metrics that predict notebook provides
• 6. Try saving each FOV as a four-channel image (real and predicted images)
• 7. Check error images and compute stats
• 8. Check metrics already logged in Neptune
• 9. Think about displays and artifacts to check for
• 10. Check performance for good and bad images
• 11. Start with Broad data from one microscope (then expand to others)
• 12. Come up with a metric for QC
• 13. (Last step) Compare with previous results from Arpit (metric based baselines)

[ShataDg]

JUMPQC-47

Musplit with a higher number of epochs

• Training set: 20 images/channel,
• Channels: DNA, AGP
• Epochs: 20

Predictions

[ShataDg]

Running inference on model trained at JUMPQC-47

Previous attempt to run inference failed probably because the wrong checkpoint was used.
Modified the code to use the bets epoch of the most recent checkpoint.

TESTING ON GOOD_IMAGES_TRAIN USING VALIDATION DATA

RANDOMLY SELECTED FOREGROUND LOCATIONS OF THE ABOVE IMAGES

MMSE PREDICTIONS

[ShataDg]

@bethac07
Short update: I managed to save the predictions for the test split as two channel images, but will work on saving them as four channels (original channels included). I also tried saving predictions for the whole batch, which is obviously a larger number of tiles compared to only the test and val splits, but the kernel crashed. I will give this another try. I will also change the code such that the predictions can be saved with the original image file names.