File size: 5,379 Bytes
3569ac3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
from gaussian_diffusion import get_named_beta_schedule
from respace import SpacedDiffusion, space_timesteps
from model import (
    InpaintText2ImUNet,
    SuperResInpaintText2ImUnet,
    SuperResText2ImUNet,
    Text2ImUNet,
)
from tokenizer.bpe import get_encoder


def model_and_diffusion_defaults():
    return dict(
        image_size=64,
        num_channels=192,
        num_res_blocks=3,
        channel_mult="",
        num_heads=1, # Editable
        num_head_channels=64, # Editable 64, 32, 18,
        num_heads_upsample=-1, # Editable
        attention_resolutions="32,16,8",
        dropout=.1, # Editable
        text_ctx=128,
        xf_width=512, # Non Editable
        xf_layers=16, # Non Editable
        xf_heads=8,  # Editable
        xf_final_ln=True,
        xf_padding=True,
        diffusion_steps=1000,  # Editable by steps of 100
        noise_schedule="squaredcos_cap_v2", # Editable can be linear or squaredcos_cap_v2
        timestep_respacing="",
        use_scale_shift_norm=True, # Non Editable
        resblock_updown=True, # Non Editable
        use_fp16=True, # Editable
        cache_text_emb=False,
        inpaint=False,
        super_res=False,
    )


def model_and_diffusion_defaults_upsampler():
    result = model_and_diffusion_defaults()
    result.update(
        dict(
            image_size=256,
            num_res_blocks=2,
            noise_schedule="linear",
            super_res=True,
        )
    )
    return result


def create_model_and_diffusion(
        image_size,
        num_channels,
        num_res_blocks,
        channel_mult,
        num_heads,
        num_head_channels,
        num_heads_upsample,
        attention_resolutions,
        dropout,
        text_ctx,
        xf_width,
        xf_layers,
        xf_heads,
        xf_final_ln,
        xf_padding,
        diffusion_steps,
        noise_schedule,
        timestep_respacing,
        use_scale_shift_norm,
        resblock_updown,
        use_fp16,
        cache_text_emb,
        inpaint,
        super_res,
):
    model = create_model(
        image_size,
        num_channels,
        num_res_blocks,
        channel_mult=channel_mult,
        attention_resolutions=attention_resolutions,
        num_heads=num_heads,
        num_head_channels=num_head_channels,
        num_heads_upsample=num_heads_upsample,
        use_scale_shift_norm=use_scale_shift_norm,
        dropout=dropout,
        text_ctx=text_ctx,
        xf_width=xf_width,
        xf_layers=xf_layers,
        xf_heads=xf_heads,
        xf_final_ln=xf_final_ln,
        xf_padding=xf_padding,
        resblock_updown=resblock_updown,
        use_fp16=use_fp16,
        cache_text_emb=cache_text_emb,
        inpaint=inpaint,
        super_res=super_res,
    )
    diffusion = create_gaussian_diffusion(
        steps=diffusion_steps,
        noise_schedule=noise_schedule,
        timestep_respacing=timestep_respacing,
    )
    return model, diffusion


def create_model(
        image_size,
        num_channels,
        num_res_blocks,
        channel_mult,
        attention_resolutions,
        num_heads,
        num_head_channels,
        num_heads_upsample,
        use_scale_shift_norm,
        dropout,
        text_ctx,
        xf_width,
        xf_layers,
        xf_heads,
        xf_final_ln,
        xf_padding,
        resblock_updown,
        use_fp16,
        cache_text_emb,
        inpaint,
        super_res,
):
    if channel_mult == "":
        if image_size == 256:
            channel_mult = (1, 1, 2, 2, 4, 4)
        elif image_size == 128:
            channel_mult = (1, 1, 2, 3, 4)
        elif image_size == 64:
            channel_mult = (1, 2, 3, 4)
        else:
            raise ValueError(f"unsupported image size: {image_size}")
    else:
        channel_mult = tuple(int(ch_mult) for ch_mult in channel_mult.split(","))
        assert 2 ** (len(channel_mult) + 2) == image_size

    attention_ds = []
    for res in attention_resolutions.split(","):
        attention_ds.append(image_size // int(res))

    if inpaint and super_res:
        model_cls = SuperResInpaintText2ImUnet
    elif inpaint:
        model_cls = InpaintText2ImUNet
    elif super_res:
        model_cls = SuperResText2ImUNet
    else:
        model_cls = Text2ImUNet
    return model_cls(
        text_ctx=text_ctx,
        xf_width=xf_width,
        xf_layers=xf_layers,
        xf_heads=xf_heads,
        xf_final_ln=xf_final_ln,
        tokenizer=get_encoder(),
        xf_padding=xf_padding,
        in_channels=3,
        model_channels=num_channels,
        out_channels=6,
        num_res_blocks=num_res_blocks,
        attention_resolutions=tuple(attention_ds),
        dropout=dropout,
        channel_mult=channel_mult,
        use_fp16=use_fp16,
        num_heads=num_heads,
        num_head_channels=num_head_channels,
        num_heads_upsample=num_heads_upsample,
        use_scale_shift_norm=use_scale_shift_norm,
        resblock_updown=resblock_updown,
        cache_text_emb=cache_text_emb,
    )


def create_gaussian_diffusion(
        steps,
        noise_schedule,
        timestep_respacing,
):
    betas = get_named_beta_schedule(noise_schedule, steps)
    if not timestep_respacing:
        timestep_respacing = [steps]
    return SpacedDiffusion(
        use_timesteps=space_timesteps(steps, timestep_respacing),
        betas=betas,
    )