MCPcopy Index your code
hub / github.com/CompVis/diff2flow / TrainerModuleLatentFM

Class TrainerModuleLatentFM

diff2flow/trainer_module.py:28–723  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

26
27
28class TrainerModuleLatentFM(LightningModule):
29 def __init__(
30 self,
31 # models
32 fm_cfg: dict,
33 noising_step: int = -1,
34 start_from_noise: bool = False,
35 # first stage
36 first_stage: dict = None,
37 # lora
38 lora_cfg: dict = None,
39 # conditioning
40 cond_stage_cfg: dict = None,
41 context_key: str = None,
42 cond_dropout: float = 0.0,
43 conditioning_key: str = None,
44 # training
45 lr: float = 1e-4,
46 weight_decay: float = 0.,
47 sampling_steps: int = 50,
48 ema_rate: float = 0.999,
49 ema_update_every: int = 100,
50 ema_update_after_step: int = 1000,
51 use_ema_for_sampling: bool = True,
52 lr_scheduler_cfg: dict = None,
53 # logging
54 n_images_to_vis: int = 16,
55 log_grad_norm: bool = False,
56 metric_tracker_cfg: dict = None,
57 visualizer: dict = None,
58 ):
59 """
60 Args:
61 fm_cfg: Flow matching model config.
62 noising_step: Forward diffusion noising step with linear schedule
63 of Ho et al. Set to -1 to disable.
64 start_from_noise: Whether to start from noise with low-res image as
65 conditioning (FM) or directly from low-res image (IC-FM).
66 first_stage: First stage config, if None, identity is used.
67 lora_cfg: LoRA config, if None, no LoRA is used.
68 cond_stage_cfg: Conditioning stage config, if None, no conditioning is used.
69 context_key: Context conditioning signal, concatenated to the input.
70 conditioning_key: Key in the batch to use for conditioning.
71 cond_dropout: Dropout rate for conditioning.
72 lr: Learning rate.
73 weight_decay: Weight decay.
74 sampling_steps: Number of sampling steps for inference.
75 ema_rate: EMA rate.
76 ema_update_every: EMA update rate (every n steps).
77 ema_update_after_step: EMA update start after n steps.
78 use_ema_for_sampling: Whether to use the EMA model for sampling.
79 lr_scheduler_cfg: Learning rate scheduler config.
80 n_images_to_vis: Number of images to visualize.
81 log_grad_norm: Whether to log gradient norm.
82 metric_tracker_cfg: Metric tracker config, if None, no metrics are tracked.
83 visualizer: Visualizer config, if None, no visualization is done.
84 """
85 super().__init__()

Callers 1

mainFunction · 0.90

Calls

no outgoing calls

Tested by

no test coverage detected