params_eval.yml

output_path: "./logs/output_${SLURM_JOB_ID}_${NOW}"
evaluation_path: "./evaluation/results"
evaluations: 8  # for multiple evaluations with majority vote and for LIDC number of samples
evaluation_vote_strategy: "confidence"  # One of ["majority", "confidence"]

dataset_file: datasets.lidc
dataset_split: 'val'  # One of ['val', 'test']
dataset_val_max_size: null # One of "null" (=full val size), or any number
class_weights: "uniform"  # One of ["uniform", "weighted"]
save_qualitative_results:
    enable: yes
    num_images: 3  # will be doubled for random imgs
    num_predictions: 3

dataset_pipeline_train: ["flip",  "resize", "colorjitter", "torchvision_normalise"]
dataset_pipeline_train_settings:
    target_size:  [128, 256]

dataset_pipeline_val: ["resize", "torchvision_normalise"]
dataset_pipeline_val_settings:
    target_size: [128, 256]
    return_original_labels: yes

evaluation:
    resolution: "original"
    evaluations: 1  # for multiple evaluations with majority vote
    evaluation_vote_strategy: "confidence"  # One of ["majority", "confidence"]

multigpu: no
distributed: no
mp_loaders: 0
batch_size: 2
max_epochs: 7500
wandb: no

polyak_alpha: 0.999
beta_schedule: "cosine" # One of ["cosine", "linear"]
beta_schedule_params:
    s: 0.008
time_steps: 250

backbone: "unet_openai"  # One of ["unetplus", "unet_openai", "resnet50", "resnet101"]

feature_cond_encoder:
    type: 'none' # ['dino', 'none']
    model: 'dino_vits8' # 'dino_vitb8
    channels: 384 # 768 for vitb, 384 for vits
    conditioning: "concat_pixels_concat_features"
    output_stride: 8  # at what output_stride in the unet to plug features
    scale: 'single'
    train: no
    source_layer: 11 # layer 11 is always the last layer of any dino vit model
    target_layer: 10 # at what output_stride in the unet to plug features


unet_openai:
    base_channels: 32
    channel_mult: null  # 9M: [1, 2, 2, 4, 5]
    attention_resolutions: [32, 16, 8]
    num_heads: 1  # Ignored if num_head_channels is not -1
    num_head_channels: 32  # If not -1, num_heads is automatically set to channels//num_head_channels
    softmax_output: yes # this is the default for build_model
    ce_head: no # adds an extra head that predicts logits (distinct from denoising head)

load_from: null