VisionLanguageGroup/MicroscopyMatching

1from dataclasses import dataclass, field
2from pathlib import Path
3from typing import Dict, List
4
5
6@dataclass
7class RunConfig:
8    # Guiding text prompt
9    prompt: str = "<task-prompt>"
10    # Which token indices to alter with attend-and-excite
11    token_indices: List[int] = field(default_factory=lambda: [2,5])
12    # Which random seeds to use when generating
13    seeds: List[int] = field(default_factory=lambda: [42])
14    # Path to save all outputs to
15    output_path: Path = Path('./outputs')
16    # Number of denoising steps
17    n_inference_steps: int = 50
18    # Text guidance scale
19    guidance_scale: float = 7.5
20    # Number of denoising steps to apply attend-and-excite
21    max_iter_to_alter: int = 25
22    # Resolution of UNet to compute attention maps over
23    attention_res: int = 16
24    # Whether to run standard SD or attend-and-excite
25    run_standard_sd: bool = False
26    # Dictionary defining the iterations and desired thresholds to apply iterative latent refinement in
27    thresholds: Dict[int, float] = field(default_factory=lambda: {0: 0.05, 10: 0.5, 20: 0.8})
28    # Scale factor for updating the denoised latent z_t
29    scale_factor: int = 20
30    # Start and end values used for scaling the scale factor - decays linearly with the denoising timestep
31    scale_range: tuple = field(default_factory=lambda: (1.0, 0.5))
32    # Whether to apply the Gaussian smoothing before computing the maximum attention value for each subject token
33    smooth_attentions: bool = True
34    # Standard deviation for the Gaussian smoothing
35    sigma: float = 0.5
36    # Kernel size for the Gaussian smoothing
37    kernel_size: int = 3
38    # Whether to save cross attention maps for the final results
39    save_cross_attention_maps: bool = False
40
41    def __post_init__(self):
42        self.output_path.mkdir(exist_ok=True, parents=True)
43