File size: 5,516 Bytes
32da3e8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
"""Guidance utilities for eval-time model forward selection.



Supports CFG (classifier-free guidance) and IG (internal guidance).

GuidanceConfig lives in configs.stage2; this module provides

get_model_forward_fn() which selects the right forward method.

"""
from functools import partial

import torch

from configs.stage2 import GuidanceConfig


def forward_with_cfg(model, x, t, cfg_scale, cfg_interval=(0, 1), **condition_kwargs):
    """Forward pass with classifier-free guidance."""
    half = x[: len(x) // 2]
    combined = torch.cat([half, half], dim=0)
    model_out = model(combined, t, **condition_kwargs)
    if isinstance(model_out, tuple):
        # IG models return (full, base) tuple
        model_out = model_out[0]
    eps, rest = model_out[:, :model.in_channels], model_out[:, model.in_channels:]
    cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
    guid_t_min, guid_t_max = cfg_interval
    assert guid_t_min < guid_t_max, "cfg_interval should be (min, max) with min < max"
    t = t[: len(t) // 2]
    half_eps = torch.where(
        ((t >= guid_t_min) & (t <= guid_t_max)).view(-1, *[1] * (len(cond_eps.shape) - 1)),
        uncond_eps + cfg_scale * (cond_eps - uncond_eps), cond_eps
    )
    eps = torch.cat([half_eps, half_eps], dim=0)
    return torch.cat([eps, rest], dim=1)


def slice_context_kwargs(condition_kwargs, batch_size):
    half_kwargs = {}
    for key in condition_kwargs.keys():
        if condition_kwargs[key] is not None:
            half_kwargs[key] = condition_kwargs[key][:batch_size]
    return half_kwargs


def forward_with_internalguidance(model, x, t, ig_scale, ig_interval=(0, 1), **condition_kwargs):
    """Pure IG forward. Math: ig_output = base + ig_scale * (full - base)"""
    half = x[: len(x) // 2]
    t_half = t[: len(t) // 2]
    half_context_kwargs = slice_context_kwargs(condition_kwargs, half.shape[0])
    full_out, base_out = model(half, t_half, **half_context_kwargs)
    eps_full = full_out[:, :model.in_channels]
    eps_base = base_out[:, :model.in_channels]
    ig_t_min, ig_t_max = ig_interval
    assert ig_t_min < ig_t_max, "ig_interval should be (min, max) with min < max"
    ig_out = torch.where(
        ((t_half >= ig_t_min) & (t_half <= ig_t_max)).view(-1, *[1] * (eps_full.ndim - 1)),
        eps_base + ig_scale * (eps_full - eps_base),
        eps_full
    )
    return torch.cat([ig_out, ig_out], dim=0)


def forward_with_ig_and_cfg(

    model, x, t, ig_scale, cfg_scale, ig_interval=(0, 1), cfg_interval=(0, 1),

    uncond_ig_scale=None, **condition_kwargs

):
    """Combined IG + CFG. Expects doubled batch [cond, uncond].



    Args:

        uncond_ig_scale: IG scale for unconditional branch. Defaults to ig_scale.

    """
    uncond_ig_scale = ig_scale if uncond_ig_scale is None else uncond_ig_scale

    full_out, base_out = model(x, t, **condition_kwargs)

    eps_full = full_out[:, :model.in_channels]
    eps_base = base_out[:, :model.in_channels]

    full_c, full_u = eps_full.chunk(2, dim=0)
    base_c, base_u = eps_base.chunk(2, dim=0)
    t_half = t[: len(t) // 2]

    # Apply IG to cond/uncond branches
    ig_t_min, ig_t_max = ig_interval
    assert ig_t_min < ig_t_max, "ig_interval should be (min, max) with min < max"
    ig_cond = torch.where(
        ((t_half >= ig_t_min) & (t_half <= ig_t_max)).view(-1, *[1] * (full_c.ndim - 1)),
        base_c + ig_scale * (full_c - base_c),
        full_c
    )
    ig_uncond = torch.where(
        ((t_half >= ig_t_min) & (t_half <= ig_t_max)).view(-1, *[1] * (full_u.ndim - 1)),
        base_u + uncond_ig_scale * (full_u - base_u),
        full_u
    )

    # Apply CFG
    cfg_t_min, cfg_t_max = cfg_interval
    assert cfg_t_min < cfg_t_max, "cfg_interval should be (min, max) with min < max"
    out = torch.where(
        ((t_half >= cfg_t_min) & (t_half <= cfg_t_max)).view(-1, *[1] * (ig_cond.ndim - 1)),
        ig_uncond + cfg_scale * (ig_cond - ig_uncond),
        ig_cond
    )
    return torch.cat([out, out], dim=0)


def get_model_forward_fn(model, guid_cfg: GuidanceConfig):
    """Get the appropriate model forward function based on guidance config.



    Args:

        model: The stage2 model

        guid_cfg: Parsed guidance configuration



    Returns:

        Tuple of (model_fn, sample_kwargs)

    """
    if guid_cfg.use_ig and guid_cfg.use_cfg:
        # Combined IG + CFG
        model_fn = partial(forward_with_ig_and_cfg, model)
        sample_kwargs = dict(
            ig_scale=guid_cfg.ig.scale,
            cfg_scale=guid_cfg.cfg.scale,
            ig_interval=(guid_cfg.ig.t_min, guid_cfg.ig.t_max),
            cfg_interval=(guid_cfg.cfg.t_min, guid_cfg.cfg.t_max),
            uncond_ig_scale=guid_cfg.ig.unconditional_scale,
        )
    elif guid_cfg.use_ig:
        # IG only
        model_fn = partial(forward_with_internalguidance, model)
        sample_kwargs = dict(
            ig_scale=guid_cfg.ig.scale,
            ig_interval=(guid_cfg.ig.t_min, guid_cfg.ig.t_max),
        )
    elif guid_cfg.use_cfg:
        # CFG only
        model_fn = partial(forward_with_cfg, model)
        sample_kwargs = dict(
            cfg_scale=guid_cfg.cfg.scale,
            cfg_interval=(guid_cfg.cfg.t_min, guid_cfg.cfg.t_max),
        )
    else:
        # No guidance
        model_fn = model.forward
        sample_kwargs = dict()

    return model_fn, sample_kwargs