Add files using upload-large-folder tool

NASH_steato/dynamic/scheduler_config.json

@@ -0,0 +1,19 @@
+{
+  "_class_name": "DDIMScheduler",
+  "_diffusers_version": "0.33.1",
+  "beta_end": 0.02,
+  "beta_schedule": "linear",
+  "beta_start": 0.0001,
+  "clip_sample": false,
+  "clip_sample_range": 1.0,
+  "dynamic_thresholding_ratio": 0.995,
+  "num_train_timesteps": 3000,
+  "prediction_type": "v_prediction",
+  "rescale_betas_zero_snr": false,
+  "sample_max_value": 1.0,
+  "set_alpha_to_one": true,
+  "steps_offset": 0,
+  "thresholding": true,
+  "timestep_spacing": "leading",
+  "trained_betas": null
+}

NASH_steato/my_conf/my_training_conf.py

@@ -0,0 +1,194 @@
+from datetime import datetime
+
+from omegaconf import MISSING
+
+###################################################################################################
+############################################ Base conf ############################################
+###################################################################################################
+# These are generic classes that need full instantiation
+# pylint: disable=unused-import
+from GaussianProxy.conf.training_conf import (
+    Accelerate,
+    AccelerateLaunchArgs,
+    Checkpointing,
+    Config,
+    DataLoader,
+    DDIMSchedulerConfig,
+    Evaluation,
+    ForwardNoising,  # noqa: F401
+    InvertedRegeneration,  # noqa: F401
+    IterativeInvertedRegeneration,  # noqa: F401
+    MetricsComputation,  # noqa: F401
+    SimilarityWithTrainData,
+    SimpleGeneration,  # noqa: F401
+    Slurm,
+    Training,
+)
+
+# pylint: enable=unused-import
+
+###################################################################################################
+########################################## Defaults conf ##########################################
+###################################################################################################
+defaults = [
+    {"dataset": "NASH_steatosis/NASH_steatosis_fully_ordered"},
+    "hydra/job_logging/custom",
+    "_self_",
+]
+
+# fmt: off
+
+# ------------------------------------------- Job launch ------------------------------------------
+now = datetime.now().strftime("%Y-%m-%d--%H-%M-%S")
+slurm = Slurm(
+    enabled          = True,
+    monitor          = False,
+    account          = "icr@h100",
+    partition        = None,
+    constraint       = "h100",
+    qos              = "t3",
+    nodes            = 1,
+    num_gpus         = 4,
+    max_num_requeue  = 4,
+    total_job_time   = 20 * 60,
+    output_folder    = "${hydra:run.dir}" + f"/{now}_%j",
+    email            = "tboyer@bio.ens.psl.eu",
+    job_launch_delay = None,
+)
+
+accelerate_launch_args = AccelerateLaunchArgs(
+    machine_rank      = 0,
+    num_machines      = 1,
+    gpu_ids           = "all",
+    rdzv_backend      = "static",
+    same_network      = "true",
+    mixed_precision   = "bf16",
+    num_processes     = slurm.num_gpus,
+    main_process_port = 29503,
+    dynamo_backend    = "inductor",
+)
+
+accelerate = Accelerate(
+    launch_args = accelerate_launch_args,
+    offline     = True,
+)
+
+# ---------------------------------------------- Data ---------------------------------------------
+data_loader = DataLoader(
+    num_workers           = 4,
+    train_prefetch_factor = 4,
+    pin_memory            = True,
+    persistent_workers    = True,
+)
+
+# -------------------------------------------- Training -------------------------------------------
+training = Training(
+    gradient_accumulation_steps = 1,
+    train_batch_size            = 128 - 16,
+    max_grad_norm               = 1,
+    nb_time_samplings           = 1_000_000,
+    unpaired_data               = False,
+)
+
+checkpointing = Checkpointing(
+    checkpoints_total_limit  = 3,
+    resume_from_checkpoint   = True,
+    checkpoint_every_n_steps = 5000,
+    chckpt_base_path         = MISSING,
+)
+
+# ------------------------------------------- Evaluation ------------------------------------------
+# naming convention is lowercase + underscore; has to be respected for debug args modification
+metrics_compute = MetricsComputation(
+    nb_samples_to_gen_per_time     = "adapt half aug",
+    batch_size                     = 512,
+    nb_diffusion_timesteps         = 50,
+    selected_times                 = [0, 1, 2, 3],
+    augmentations_for_metrics_comp = ["RandomHorizontalFlip", "RandomVerticalFlip", "RandomRotationSquareSymmetry"],
+)
+
+simple_generation = SimpleGeneration(
+    nb_diffusion_timesteps = 50,
+    n_rows_displayed       = 4, # TODO: merge training & evaluation configs
+    nb_generated_samples   = 16, # TODO: merge training & evaluation configs
+)
+
+inverted_regeneration = InvertedRegeneration(
+    nb_diffusion_timesteps     = 50,
+    n_rows_displayed           = 8,  # TODO: merge training & evaluation configs
+    nb_generated_samples       = 16, # TODO: merge training & evaluation configs
+    nb_video_times_in_parallel = 8,  # TODO: merge training & evaluation configs TODO: not used in training!
+    nb_video_timesteps         = 50, # TODO: merge training & evaluation configs
+)
+
+sim_with_train = SimilarityWithTrainData( # must be put after metrics_compute!
+    nb_generated_samples   = -1,  # TODO: not used
+    batch_size             = 2048,
+    nb_batches_shown       = -1,  # TODO: not used
+    n_rows_displayed       = -1,  # TODO: not used
+    nb_diffusion_timesteps = -1,   # TODO: not used
+)
+
+evaluation = Evaluation(
+    every_n_opt_steps      = 25_000,
+    batch_size             = 16, # TODO: remove this and use config from above
+    nb_video_timesteps     = 50, # TODO: remove this and use config from above
+    strategies             = [simple_generation, inverted_regeneration, metrics_compute, sim_with_train],
+)
+
+# ------------------------------------------- Diffusion -------------------------------------------
+dynamic = DDIMSchedulerConfig(
+    num_train_timesteps    = 3000,
+    clip_sample            = False,
+    clip_sample_range      = 1,
+    thresholding           = True,
+    sample_max_value       = 1,
+    prediction_type        = "v_prediction",
+    rescale_betas_zero_snr = False,
+    timestep_spacing       = "leading",
+)
+
+# ---------------------------------------------- Model --------------------------------------------
+from my_conf.net.net_128_3 import net, time_encoder  # noqa: E402
+
+# ------------------------------------------ Final Config -----------------------------------------
+config = Config(
+    # defaults
+    defaults                      = defaults,
+    # model
+    dynamic                       = dynamic,
+    net                           = net,
+    time_encoder                  = time_encoder,
+    # script
+    launcher_script_parent_folder = "/linkhome/rech/genlxz01/ufc43hj/sources/GaussianProxy",
+    script                        = "train",
+    # experiment variables
+    exp_parent_folder             = "/lustre/fsn1/projects/rech/icr/ufc43hj/experiments",
+    project                       = MISSING,
+    run_name                      = MISSING,
+    # hydra
+    hydra                         = {"run": {"dir": "${exp_parent_folder}/${project}/${run_name}"}},
+    # slurm
+    slurm                         = slurm,
+    # accelerate
+    accelerate                    = accelerate,
+    # misc.
+    debug                         = False,
+    profile                       = False,
+    tmpdir_location               = None,
+    # experiment tracker
+    logger                        = "wandb",
+    entity                        = "thomasboyer",
+    # checkpointing
+    checkpointing                 = checkpointing,
+    # dataset
+    dataset                       = MISSING,
+    # dataloaders
+    dataloaders                   = data_loader,
+    # training
+    training                      = training,
+    # evaluation
+    evaluation                    = evaluation,
+    # optimizer
+    learning_rate                 = 1e-4,
+)

NASH_steato/net/config.json

@@ -0,0 +1,64 @@
+{
+  "_class_name": "UNet2DConditionModel",
+  "_diffusers_version": "0.33.1",
+  "act_fn": "silu",
+  "addition_embed_type": null,
+  "addition_embed_type_num_heads": 64,
+  "addition_time_embed_dim": null,
+  "attention_head_dim": 8,
+  "attention_type": "default",
+  "block_out_channels": [
+    64,
+    128,
+    256
+  ],
+  "center_input_sample": false,
+  "class_embed_type": null,
+  "class_embeddings_concat": false,
+  "conv_in_kernel": 3,
+  "conv_out_kernel": 3,
+  "cross_attention_dim": 64,
+  "cross_attention_norm": null,
+  "down_block_types": [
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D"
+  ],
+  "downsample_padding": 1,
+  "dropout": 0.0,
+  "dual_cross_attention": false,
+  "encoder_hid_dim": null,
+  "encoder_hid_dim_type": null,
+  "flip_sin_to_cos": true,
+  "freq_shift": 0,
+  "in_channels": 3,
+  "layers_per_block": 2,
+  "mid_block_only_cross_attention": null,
+  "mid_block_scale_factor": 1,
+  "mid_block_type": "UNetMidBlock2DCrossAttn",
+  "norm_eps": 1e-05,
+  "norm_num_groups": 32,
+  "num_attention_heads": null,
+  "num_class_embeds": null,
+  "only_cross_attention": false,
+  "out_channels": 3,
+  "projection_class_embeddings_input_dim": null,
+  "resnet_out_scale_factor": 1.0,
+  "resnet_skip_time_act": false,
+  "resnet_time_scale_shift": "default",
+  "reverse_transformer_layers_per_block": null,
+  "sample_size": 128,
+  "time_cond_proj_dim": null,
+  "time_embedding_act_fn": null,
+  "time_embedding_dim": null,
+  "time_embedding_type": "positional",
+  "timestep_post_act": null,
+  "transformer_layers_per_block": 1,
+  "up_block_types": [
+    "CrossAttnUpBlock2D",
+    "CrossAttnUpBlock2D",
+    "CrossAttnUpBlock2D"
+  ],
+  "upcast_attention": false,
+  "use_linear_projection": false
+}

NASH_steato/video_time_encoder/config.json

@@ -0,0 +1,8 @@
+{
+  "_class_name": "VideoTimeEncoding",
+  "_diffusers_version": "0.33.1",
+  "downscale_freq_shift": 1.0,
+  "encoding_dim": 128,
+  "flip_sin_to_cos": true,
+  "time_embed_dim": 64
+}

biotine/dynamic/scheduler_config.json

@@ -0,0 +1,19 @@
+{
+  "_class_name": "DDIMScheduler",
+  "_diffusers_version": "0.33.1",
+  "beta_end": 0.02,
+  "beta_schedule": "linear",
+  "beta_start": 0.0001,
+  "clip_sample": false,
+  "clip_sample_range": 1.0,
+  "dynamic_thresholding_ratio": 0.995,
+  "num_train_timesteps": 3000,
+  "prediction_type": "v_prediction",
+  "rescale_betas_zero_snr": false,
+  "sample_max_value": 1.0,
+  "set_alpha_to_one": true,
+  "steps_offset": 0,
+  "thresholding": true,
+  "timestep_spacing": "leading",
+  "trained_betas": null
+}

biotine/my_conf/my_inference_conf.py

@@ -0,0 +1,95 @@
+# ruff: noqa: F401
+
+from datetime import datetime
+from pathlib import Path
+
+from torch import bfloat16, float16, float32
+
+from GaussianProxy.conf.inference_conf import InferenceConfig, ProfileConfig
+from GaussianProxy.conf.training_conf import (
+    Accelerate,
+    AccelerateLaunchArgs,
+    ForwardNoising,
+    ForwardNoisingLinearScaling,
+    InversionRegenerationOnly,
+    InvertedRegeneration,
+    IterativeInvertedRegeneration,
+    MetricsComputation,
+    SimilarityWithTrainData,
+    SimpleGeneration,
+    Slurm,
+)
+
+# -------------------------------------------- Dataset --------------------------------------------
+from my_conf.dataset.biotine_png_128_inference import dataset
+
+# --------------------------------------------- Model ---------------------------------------------
+root_experiments_path = Path("/lustre/fsn1/projects/rech/icr/ufc43hj/experiments")
+project_name = "GaussianProxy"
+run_name = "biotine_all_paired_new_jz"
+
+# ------------------------------------------ Evaluations ------------------------------------------
+eval_strats = [
+    # InvertedRegeneration(
+    #     nb_diffusion_timesteps=100,
+    #     name="InvertedRegeneration_100_diffsteps_no_SNR_leading_f32_J_14_fld_2",
+    #     nb_generated_samples=64,
+    #     plate_name_to_simulate="J_14_fld_2",
+    #     nb_video_times_in_parallel=3,
+    #     nb_video_timesteps=19,
+    #     n_rows_displayed=8,
+    # ),
+    MetricsComputation(
+        nb_samples_to_gen_per_time="adapt aug",
+        batch_size=512 + 32,
+        nb_diffusion_timesteps=100,
+        selected_times=[1, 5, 10, 15, 19],
+        name="MetricsComputation_100_diffsteps_no_SNR_leading_f32_adapt_aug",
+        regen_images=False,
+    ),
+    # InversionRegenerationOnly(
+    #     nb_diffusion_timesteps=100,
+    #     name="InversionRegenerationOnly_100_diffsteps_no_SNR_leading_f32",
+    #     nb_generated_samples=64,
+    #     plate_name_to_simulate="M_13_fld_3",
+    #     n_rows_displayed=8,
+    # )
+]
+
+
+# ------------------------------------------ Profiling --------------------------------------------
+# fmt: off
+profile_conf = ProfileConfig(
+    enabled             = False,
+    record_shapes       = False,
+    profile_memory      = True,
+    with_stack          = True,
+    with_flops          = False,
+    export_chrome_trace = False,
+)
+
+
+# ------------------------------------------ Final Config -----------------------------------------
+inference_conf = InferenceConfig(
+    # Choose the experiment (trained model weights)
+    root_experiments_path       = root_experiments_path,
+    project_name                = project_name,
+    run_name                    = run_name,
+    # Choose a custom scheduler
+    scheduler_config            = Path("my_conf", "scheduler", "DDIM_3k_vpred_tresh_leading.json"),
+    # Output directory (where to put the generated images / tensors)
+    output_dir                  = Path(root_experiments_path, project_name, run_name, "inferences"),
+    # Device
+    device                      = "cuda:2",
+    # Optimizations
+    compile                     = True,
+    dtype                       = float32,
+    # Data
+    dataset                     = dataset,
+    # Evaluations 
+    evaluation_strategies       = eval_strats, # pyright: ignore[reportArgumentType]
+    # Profiling
+    profiling                   = profile_conf,
+    # Temp Dir
+    tmpdir_location             = "/tmp",
+)

biotine/my_conf/net/net_128_3_big.py

@@ -0,0 +1,22 @@
+from GaussianProxy.conf.training_conf import UNet2DConditionModelConfig, TimeEncoderConfig
+
+cross_attn_dim = 64
+
+net = UNet2DConditionModelConfig(
+    sample_size=128,
+    in_channels=3,
+    out_channels=3,
+    down_block_types=("DownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D"),
+    up_block_types=("CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"),
+    block_out_channels=(128, 256, 512),
+    layers_per_block=2,
+    act_fn="silu",
+    cross_attention_dim=cross_attn_dim,
+)
+
+time_encoder = TimeEncoderConfig(
+    encoding_dim=128,
+    time_embed_dim=cross_attn_dim,
+    flip_sin_to_cos=True,
+    downscale_freq_shift=1,
+)

biotine/my_conf/net/net_256_3_20M.py

@@ -0,0 +1,22 @@
+from GaussianProxy.conf.training_conf import UNet2DConditionModelConfig, TimeEncoderConfig
+
+cross_attn_dim = 64
+
+net = UNet2DConditionModelConfig(
+    sample_size=256,
+    in_channels=3,
+    out_channels=3,
+    down_block_types=("DownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D"),
+    up_block_types=("CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"),
+    block_out_channels=(64, 128, 224),
+    layers_per_block=2,
+    act_fn="silu",
+    cross_attention_dim=cross_attn_dim,
+)
+
+time_encoder = TimeEncoderConfig(
+    encoding_dim=128,
+    time_embed_dim=cross_attn_dim,
+    flip_sin_to_cos=True,
+    downscale_freq_shift=1,
+)

biotine_unpaired/my_conf/dataset/BBBC021/BBBC021_196_docetaxel.yaml

@@ -0,0 +1,16 @@
+name: BBBC021_196_docetaxel
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/docetaxel
+data_shape: [3, 196, 196]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8

biotine_unpaired/my_conf/dataset/BBBC021/BBBC021_196_docetaxel_fully_ordered.yaml

@@ -0,0 +1,23 @@
+name: BBBC021_196_docetaxel_fully_ordered
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/docetaxel
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/BBBC021_196_docetaxel__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [3, 196, 196]
+
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/BBBC021/BBBC021_196_docetaxel_hard_aug_inference.py

@@ -0,0 +1,4 @@
+from GaussianProxy.conf.dataset.BBBC021_196_hard_aug_inference import dataset
+
+dataset.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/docetaxel_hard_augmented"
+dataset.name += "_docetaxel"

biotine_unpaired/my_conf/dataset/BBBC021/BBBC021_196_docetaxel_inference.py

@@ -0,0 +1,4 @@
+from GaussianProxy.conf.dataset.BBBC021_196_inference import BBBC021_196_inference as BBBC021_196_docetaxel_inference
+
+BBBC021_196_docetaxel_inference.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/docetaxel"
+BBBC021_196_docetaxel_inference.name += "_docetaxel"

biotine_unpaired/my_conf/dataset/BBBC021/BBBC021_196_docetaxel_skip_half_doses_fully_ordered.yaml

@@ -0,0 +1,24 @@
+name: BBBC021_196_docetaxel_fully_ordered_skip_half_doses
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/docetaxel
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/BBBC021_196_docetaxel_skip_half_doses__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [3, 196, 196]
+
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+    - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+      dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+    - _target_: torchvision.transforms.Normalize
+      mean: [0.5, 0.5, 0.5]
+      std: [0.5, 0.5, 0.5]
+    - _target_: torchvision.transforms.RandomHorizontalFlip
+    - _target_: torchvision.transforms.RandomVerticalFlip
+    - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+selected_dists: ["0.0003", "0.003", "0.03", "0.3"]
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/BBBC021/BBBC021_196_docetaxel_skip_many_doses_fully_ordered.yaml

@@ -0,0 +1,24 @@
+name: BBBC021_196_docetaxel_fully_ordered_skip_many_doses
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/docetaxel
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/BBBC021_196_docetaxel_skip_many_doses__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [3, 196, 196]
+
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+    - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+      dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+    - _target_: torchvision.transforms.Normalize
+      mean: [0.5, 0.5, 0.5]
+      std: [0.5, 0.5, 0.5]
+    - _target_: torchvision.transforms.RandomHorizontalFlip
+    - _target_: torchvision.transforms.RandomVerticalFlip
+    - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+selected_dists: ["0.0003", "0.001", "1.0"]
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/BBBC021/BBBC021_196_nocodazole_fully_ordered.yaml

@@ -0,0 +1,23 @@
+name: BBBC021_196_nocodazole_fully_ordered
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/nocodazole
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/BBBC021/196x196/BBBC021_196_nocodazole__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [3, 196, 196]
+
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/ChromaLive6h/ChromaLive6h_3ch_png.yaml

@@ -0,0 +1,19 @@
+name: chromaLive6h_3ch_png_patches_380px
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/20231017ChromaLive_6hr_4ch/MIP_normalized/paired_dataset/patches
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ] # move to [-1:1]
+    std: [ 0.5, 0.5, 0.5 ]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [ 0, 255 ]
+expected_dtype: torch.uint8
+selected_dists:

biotine_unpaired/my_conf/dataset/ChromaLive6h/ChromaLive6h_3ch_png_fully_ordered.yaml

@@ -0,0 +1,21 @@
+name: chromaLive6h_3ch_png_patches_380px_fully_ordered
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/ChromaLive_6hr_4ch/MIP_normalized/patches_380px
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/ChromaLive_6hr_4ch/MIP_normalized/chromaLive6h_3ch_png_patches_380px__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ] # move to [-1:1]
+    std: [ 0.5, 0.5, 0.5 ]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [ 0, 255 ]
+expected_dtype: torch.uint8
+selected_dists:
+fully_ordered: true

biotine_unpaired/my_conf/dataset/ChromaLive6h/ChromaLive6h_3ch_png_hard_aug.yaml

@@ -0,0 +1,16 @@
+name: chromaLive6h_3ch_png_patches_380px_hard_aug
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/20231017ChromaLive_6hr_4ch/MIP_normalized/paired_dataset/patches_380px_hard_augmented
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ] # move to [-1:1]
+    std: [ 0.5, 0.5, 0.5 ]
+expected_initial_data_range: [ 0, 255 ]
+expected_dtype: torch.uint8
+selected_dists:

biotine_unpaired/my_conf/dataset/ChromaLive6h/ChromaLive6h_4ch_tif.yaml

@@ -0,0 +1,19 @@
+name: chromaLive6h_4ch_tif_patches_380px
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/20231017ChromaLive_6hr_4ch/MIP_normalized/paired_dataset/patches
+data_shape: [ 4, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5, 0.5 ] # move to [-1:1]
+    std: [ 0.5, 0.5, 0.5, 0.5 ]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [ 0, 65536 ]
+expected_dtype: torch.uint16
+selected_dists: [ 'time_1', 'time_3', 'time_5', 'time_7', 'time_9', 'time_11', 'time_13' ]

biotine_unpaired/my_conf/dataset/ChromaLive6h/chromalive6h_3ch_png_hard_aug_inference.py

@@ -0,0 +1,33 @@
+from torch import float32
+from torchvision.transforms import Compose, ConvertImageDtype, Normalize, Resize
+
+from GaussianProxy.conf.training_conf import DataSet, DatasetParams
+from GaussianProxy.utils.data import ImageDataset
+
+DEFINITION = 128
+NUMBER_OF_CHANNELS = 3
+
+transforms = Compose(
+    transforms=[
+        Resize(DEFINITION),
+        ConvertImageDtype(float32),
+        Normalize(mean=[0.5] * NUMBER_OF_CHANNELS, std=[0.5] * NUMBER_OF_CHANNELS),
+    ]
+)
+
+ds_params = DatasetParams(
+    file_extension="png",
+    key_transform=str,
+    sorting_func=lambda subdir: int(subdir.name.split("_")[1]),
+    dataset_class=ImageDataset,
+)
+
+dataset = DataSet(
+    name="chromaLive6h_3ch_png_patches_380px_hard_aug",
+    data_shape=(NUMBER_OF_CHANNELS, DEFINITION, DEFINITION),
+    transforms=transforms,
+    selected_dists=None,  # not used
+    expected_initial_data_range=(0, 255),
+    dataset_params=ds_params,
+    path="/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/20231017ChromaLive_6hr_4ch/MIP_normalized/paired_dataset/patches_hard_augmented",
+)

biotine_unpaired/my_conf/dataset/ChromaLive6h/chromalive6h_3ch_png_inference.py

@@ -0,0 +1,33 @@
+from torch import float32
+from torchvision.transforms import Compose, ConvertImageDtype, Normalize, Resize
+
+from GaussianProxy.conf.training_conf import DataSet, DatasetParams
+from GaussianProxy.utils.data import ImageDataset
+
+DEFINITION = 128
+NUMBER_OF_CHANNELS = 3
+
+transforms = Compose(
+    transforms=[
+        Resize(DEFINITION),
+        ConvertImageDtype(float32),
+        Normalize(mean=[0.5] * NUMBER_OF_CHANNELS, std=[0.5] * NUMBER_OF_CHANNELS),
+    ]
+)
+
+ds_params = DatasetParams(
+    file_extension="png",
+    key_transform=str,
+    sorting_func=lambda subdir: int(subdir.name.split("_")[1]),
+    dataset_class=ImageDataset,
+)
+
+dataset = DataSet(
+    name="chromaLive6h_3ch_png_patches_380px",
+    data_shape=(NUMBER_OF_CHANNELS, DEFINITION, DEFINITION),
+    transforms=transforms,
+    selected_dists=None,  # not used
+    expected_initial_data_range=(0, 255),
+    dataset_params=ds_params,
+    path="/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/20231017ChromaLive_6hr_4ch/MIP_normalized/paired_dataset/patches",
+)

biotine_unpaired/my_conf/dataset/ChromaLiveTL24h/ChromaLiveTL24h.yaml

@@ -0,0 +1,19 @@
+name: chromalive_tl_24h_380px
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/20230920ChromaLiveTL_24hr4ch/ch_4_3_1___norm_whole_ds_per_channel_per_zslice_0_99perc___patches_380
+data_shape: [3, 256, 256]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 256
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [0.5, 0.5, 0.5] # move to [-1:1]
+    std: [0.5, 0.5, 0.5]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+selected_dists: ['time_1', 'time_7', 'time_13', 'time_19', 'time_25', 'time_31', 'time_37', 'time_43', 'time_49', 'time_55', 'time_61', 'time_67', 'time_73', 'time_79', 'time_85', 'time_91', 'time_97', 'time_103', 'time_109', 'time_115', 'time_121', 'time_127', 'time_133', 'time_139', 'time_145']

biotine_unpaired/my_conf/dataset/DeepCycle/deepcycle_fully_ordered.yaml

@@ -0,0 +1,19 @@
+name: deepcycle_fully_ordered
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DeepCycle/128x128
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DeepCycle/deepcycle_brightfield_to_3D__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+data_shape: [4, 128, 128]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+    - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+      dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+    - _target_: torchvision.transforms.Normalize
+      mean: [0.5, 0.5, 0.5, 0.5]
+      std: [0.5, 0.5, 0.5, 0.5]
+    - _target_: torchvision.transforms.RandomHorizontalFlip
+    - _target_: torchvision.transforms.RandomVerticalFlip
+    - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+selected_dists:
+fully_ordered: true

biotine_unpaired/my_conf/dataset/DeepCycle/deepcycle_markers_crop_fully_ordered.yaml

@@ -0,0 +1,21 @@
+name: deepcycle_markers_fully_ordered
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DeepCycle/128x128
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DeepCycle/deepcycle_markers__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+data_shape: [4, 64, 64]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+    - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+      dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+    - _target_: torchvision.transforms.Normalize
+      mean: [0.5, 0.5, 0.5, 0.5]
+      std: [0.5, 0.5, 0.5, 0.5]
+    - _target_: torchvision.transforms.RandomHorizontalFlip
+    - _target_: torchvision.transforms.RandomVerticalFlip
+    - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+    - _target_: torchvision.transforms.CenterCrop
+      size: 64
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+selected_dists:
+fully_ordered: true

biotine_unpaired/my_conf/dataset/DeepCycle/deepcycle_markers_fully_ordered.yaml

@@ -0,0 +1,19 @@
+name: deepcycle_markers_fully_ordered
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DeepCycle/128x128
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DeepCycle/deepcycle_markers__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+data_shape: [4, 128, 128]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+    - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+      dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+    - _target_: torchvision.transforms.Normalize
+      mean: [0.5, 0.5, 0.5, 0.5]
+      std: [0.5, 0.5, 0.5, 0.5]
+    - _target_: torchvision.transforms.RandomHorizontalFlip
+    - _target_: torchvision.transforms.RandomVerticalFlip
+    - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+selected_dists:
+fully_ordered: true

biotine_unpaired/my_conf/dataset/Jurkat/Jurkat.yaml

@@ -0,0 +1,16 @@
+name: Jurkat
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/Jurkat/rgb_images_all_cell_cycles
+data_shape: [3, 66, 66]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8

biotine_unpaired/my_conf/dataset/Jurkat/Jurkat_brightfield_fully_ordered.yaml

@@ -0,0 +1,18 @@
+name: Jurkat_brightfield_fully_ordered
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/Jurkat/brightfield_reprocessed
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/Jurkat/Jurkat_brightfield__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+data_shape: [ 1, 66, 66 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [  0.5 ]
+    std: [ 0.5 ]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [ 0, 255 ]
+expected_dtype: torch.uint8
+fully_ordered: true

biotine_unpaired/my_conf/dataset/Jurkat/Jurkat_fully_ordered.yaml

@@ -0,0 +1,18 @@
+name: Jurkat_fully_ordered_dinov2_regs_giant_ds_preproc
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/Jurkat/rgb_images_all_cell_cycles
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/Jurkat/Jurkat__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+data_shape: [3, 66, 66]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+fully_ordered: true

biotine_unpaired/my_conf/dataset/Jurkat/Jurkat_inference.py

@@ -0,0 +1,3 @@
+from GaussianProxy.conf.dataset.Jurkat_inference import Jurkat_inference
+
+Jurkat_inference.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/Jurkat/rgb_images_all_cell_cycles"

biotine_unpaired/my_conf/dataset/Jurkat/Jurkat_inference_hard_aug.py

@@ -0,0 +1,3 @@
+from GaussianProxy.conf.dataset.Jurkat_inference import Jurkat_inference as dataset
+
+dataset.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/Jurkat/rgb_images_all_cell_cycles_hard_augmented"

biotine_unpaired/my_conf/dataset/NASH_fibrosis/NASH_fibrosis.yaml

@@ -0,0 +1,24 @@
+name: NASH_fibrosis
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/NASH/prepared_data/fibrosis
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  # Convert to float32 (and normalize to [0, 1]) before resizing
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  # Random crop from 299x299 to 192x192, then resize to 128x128
+  - _target_: torchvision.transforms.RandomCrop
+    size: 192
+  - _target_: torchvision.transforms.Resize
+    size: 128
+  # Normalize to [-1, 1]
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  # Random 8x square augmentations
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+selected_dists:
+expected_initial_data_range: [ 0, 255 ]

biotine_unpaired/my_conf/dataset/NASH_fibrosis/NASH_fibrosis_fully_ordered.yaml

@@ -0,0 +1,26 @@
+name: NASH_fibrosis_fully_ordered_dinov2_regs_giant_ds_preproc
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/NASH/fibrosis
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/NASH/NASH_fibrosis__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  # Convert to float32 (and normalize to [0, 1]) before resizing
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  # Random crop from 299x299 to 192x192, then resize to 128x128
+  - _target_: torchvision.transforms.RandomCrop
+    size: 192
+  - _target_: torchvision.transforms.Resize
+    size: 128
+  # Normalize to [-1, 1]
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  # Random 8x square augmentations
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+selected_dists:
+expected_initial_data_range: [ 0, 255 ]
+fully_ordered: true

biotine_unpaired/my_conf/dataset/NASH_fibrosis/NASH_fibrosis_inference.py

@@ -0,0 +1,3 @@
+from GaussianProxy.conf.dataset.NASH_fibrosis_inference import dataset
+
+dataset.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/NASH/prepared_data/fibrosis"

biotine_unpaired/my_conf/dataset/NASH_steatosis/NASH_steatosis.yaml

@@ -0,0 +1,24 @@
+name: NASH_steatosis
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/NASH/steatosis
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  # Convert to float32 (and normalize to [0, 1]) before resizing
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  # Random crop from 299x299 to 192x192, then resize to 128x128
+  - _target_: torchvision.transforms.RandomCrop
+    size: 192
+  - _target_: torchvision.transforms.Resize
+    size: 128
+  # Normalize to [-1, 1]
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  # Random 8x square augmentations
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+selected_dists:
+expected_initial_data_range: [ 0, 255 ]

biotine_unpaired/my_conf/dataset/NASH_steatosis/NASH_steatosis_fully_ordered.yaml

@@ -0,0 +1,26 @@
+name: NASH_steatosis_fully_ordered_dinov2_regs_giant_ds_preproc
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/NASH/steatosis
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/NASH/NASH_steatosis__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  # Convert to float32 (and normalize to [0, 1]) before resizing
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  # Random crop from 299x299 to 192x192, then resize to 128x128
+  - _target_: torchvision.transforms.RandomCrop
+    size: 192
+  - _target_: torchvision.transforms.Resize
+    size: 128
+  # Normalize to [-1, 1]
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  # Random 8x square augmentations
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+selected_dists:
+expected_initial_data_range: [ 0, 255 ]
+fully_ordered: true

biotine_unpaired/my_conf/dataset/NASH_steatosis/NASH_steatosis_inference.py

@@ -0,0 +1,3 @@
+from GaussianProxy.conf.dataset.NASH_steatosis_inference import dataset
+
+dataset.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/NASH/prepared_data/steatosis"

biotine_unpaired/my_conf/dataset/biotine/biotine_paired_same_nb_as_unpaired_fully_ordered.yaml

@@ -0,0 +1,28 @@
+name: biotine_paired_same_nb_as_unpaired_fully_ordered
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/patches_255
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/heldout_test_trajs_24__biotine_png__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+path_to_train_test_labels_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/paired_same_nb_as_unpaired_heldout_test_trajs_24__biotine_png__train_test_labels__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [ 3, 128, 128 ]
+
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  # ConvertImageDtype also scales to [0; 1] (from the *implicit* expected range that depends on the incoming dtype...)
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+
+selected_dists:
+
+expected_initial_data_range: [ 0, 255 ]
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/biotine/biotine_png_128.yaml

@@ -0,0 +1,19 @@
+name: biotine_png
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/patches_255
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  # ConvertImageDtype also scales to [0; 1] (from the *implicit* expected range that depends on the incoming dtype...)
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+selected_dists:
+expected_initial_data_range: [ 0, 255 ]

biotine_unpaired/my_conf/dataset/biotine/biotine_png_128_fully_ordered.yaml

@@ -0,0 +1,27 @@
+name: biotine_png_fully_ordered
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/patches_255
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/biotine_png__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [ 3, 128, 128 ]
+
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  # ConvertImageDtype also scales to [0; 1] (from the *implicit* expected range that depends on the incoming dtype...)
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+
+selected_dists:
+
+expected_initial_data_range: [ 0, 255 ]
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/biotine/biotine_png_128_hard_aug.yaml

@@ -0,0 +1,16 @@
+name: biotine_png_hard_aug
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/patches_255_hard_augmented
+data_shape: [ 3, 128, 128 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  # ConvertImageDtype also scales to [0; 1] (from the *implicit* expected range that depends on the incoming dtype...)
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+selected_dists:
+expected_initial_data_range: [ 0, 255 ]

biotine_unpaired/my_conf/dataset/biotine/biotine_png_128_hard_aug_inference.py

@@ -0,0 +1,3 @@
+from GaussianProxy.conf.dataset.biotine_png_128_hard_aug_inference import dataset
+
+dataset.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/patches_255_hard_augmented"

biotine_unpaired/my_conf/dataset/biotine/biotine_png_128_inference.py

@@ -0,0 +1,3 @@
+from GaussianProxy.conf.dataset.biotine_png_128_inference import dataset
+
+dataset.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/patches_255"

biotine_unpaired/my_conf/dataset/biotine/biotine_png_256.yaml

@@ -0,0 +1,18 @@
+name: biotine_png
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/patches_255
+data_shape: [ 3, 256, 256 ]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32}
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 256
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+selected_dists: [ 1, 5, 10, 15, 19 ]
+expected_initial_data_range: [ 0, 255 ]

biotine_unpaired/my_conf/dataset/biotine/biotine_unpaired_fully_ordered.yaml

@@ -0,0 +1,28 @@
+name: biotine_unpaired_fully_ordered
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/patches_255
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/unpaired_heldout_test_trajs_24__biotine_png__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+path_to_train_test_labels_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/biotine/3_channels_min_99_perc_normalized_rgb_stacks_png/unpaired_heldout_test_trajs_24__biotine_png__train_test_labels__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [ 3, 128, 128 ]
+
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 128
+  # ConvertImageDtype also scales to [0; 1] (from the *implicit* expected range that depends on the incoming dtype...)
+  - _target_: torchvision.transforms.ConvertImageDtype
+    dtype: ${torch_dtype:float32}
+  - _target_: torchvision.transforms.Normalize
+    mean: [ 0.5, 0.5, 0.5 ]
+    std: [ 0.5, 0.5, 0.5 ]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+
+selected_dists:
+
+expected_initial_data_range: [ 0, 255 ]
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/diabetic_retinopathy/diabetic_retinopathy.yaml

@@ -0,0 +1,18 @@
+name: diabetic_retinopathy
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DiabeticRetinopathy/prepared_dataset/train
+data_shape: [3, 256, 256]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.transforms.Resize
+    size: 256 # single int => image resized to (size * aspect_ratio, size) or (size, size * aspect_ratio) with aspect_ratio >= 1 preserved
+  - _target_: torchvision.transforms.v2.CenterCrop
+    size: 256 # square centered crop
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8

biotine_unpaired/my_conf/dataset/diabetic_retinopathy/diabetic_retinopathy_full_circle_augs_2048_crop_fully_ordered.yaml

@@ -0,0 +1,25 @@
+name: diabetic_retinopathy_2048_crop_fully_ordered
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DiabeticRetinopathy/prepared_dataset_full_circle_augmented
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DiabeticRetinopathy/balanced_classes__diabetic_retinopathy_2048_crop__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [3, 256, 256]
+
+transforms:
+  _target_: torchvision.transforms.v2.Compose
+  transforms:
+    - _target_: torchvision.transforms.v2.CenterCrop
+      size: 2048
+    - _target_: torchvision.transforms.v2.Resize
+      size: 256
+    - _target_: torchvision.transforms.v2.ToDtype
+      dtype: ${torch_dtype:float32}
+      scale: true
+    - _target_: torchvision.transforms.v2.Normalize
+      mean: [0.5, 0.5, 0.5]
+      std: [0.5, 0.5, 0.5]
+
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/diabetic_retinopathy/diabetic_retinopathy_full_circle_augs_2560_crop_fully_ordered.yaml

@@ -0,0 +1,25 @@
+name: diabetic_retinopathy_2560_crop_fully_ordered
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DiabeticRetinopathy/prepared_dataset_full_circle_augmented
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DiabeticRetinopathy/balanced_classes__diabetic_retinopathy_2560_crop__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [3, 256, 256]
+
+transforms:
+  _target_: torchvision.transforms.v2.Compose
+  transforms:
+    - _target_: torchvision.transforms.v2.CenterCrop
+      size: 2560
+    - _target_: torchvision.transforms.v2.Resize
+      size: 256
+    - _target_: torchvision.transforms.v2.ToDtype
+      dtype: ${torch_dtype:float32}
+      scale: true
+    - _target_: torchvision.transforms.v2.Normalize
+      mean: [0.5, 0.5, 0.5]
+      std: [0.5, 0.5, 0.5]
+
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/diabetic_retinopathy/diabetic_retinopathy_fully_ordered.yaml

@@ -0,0 +1,25 @@
+name: diabetic_retinopathy_fully_ordered
+
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DiabeticRetinopathy/train
+path_to_single_parquet: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DiabeticRetinopathy/diabetic_retinopathy__continuous_time_predictions__facebook_dinov2-with-registers-giant_dataset_preproc.parquet
+
+data_shape: [3, 256, 256]
+
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+    - _target_: torchvision.transforms.transforms.Resize
+      size: 256 # single int => image resized to (size * aspect_ratio, size) or (size, size * aspect_ratio) with aspect_ratio >= 1 preserved
+    - _target_: torchvision.transforms.v2.CenterCrop
+      size: 256 # square centered crop
+    - _target_: torchvision.transforms.RandomHorizontalFlip
+    - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+      dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+    - _target_: torchvision.transforms.Normalize
+      mean: [0.5, 0.5, 0.5]
+      std: [0.5, 0.5, 0.5]
+
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8
+
+fully_ordered: true

biotine_unpaired/my_conf/dataset/diabetic_retinopathy/diabetic_retinopathy_hard_aug_inference.py

@@ -0,0 +1,4 @@
+from GaussianProxy.conf.dataset.diabetic_retinopathy_inference import diabetic_retinopathy_inference as dataset
+
+dataset.path = "/lustre/fsn1/projects/rech/icr/ufc43hj/datasets/DiabeticRetinopathy/prepared_dataset/train_hard_augmented"
+dataset.name = "diabetic_retinopathy_inference_hard_augmented"

biotine_unpaired/my_conf/dataset/ependymal_context/ependymal_context.yaml

@@ -0,0 +1,16 @@
+name: ependymal_context
+path: /lustre/fsn1/projects/rech/icr/ufc43hj/datasets/ependymal/prepared_dataset_context
+data_shape: [3, 256, 256]
+transforms:
+  _target_: torchvision.transforms.transforms.Compose
+  transforms:
+  - _target_: torchvision.transforms.ConvertImageDtype # this also scales to [0; 1]
+    dtype: ${torch_dtype:float32} # passed dtype must be accessible as a "torch" attribute
+  - _target_: torchvision.transforms.Normalize
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+  - _target_: torchvision.transforms.RandomHorizontalFlip
+  - _target_: torchvision.transforms.RandomVerticalFlip
+  - _target_: GaussianProxy.utils.data.RandomRotationSquareSymmetry
+expected_initial_data_range: [0, 255]
+expected_dtype: torch.uint8