Spaces:
Sleeping
Sleeping
add number of rows parameters
Browse files- app.py +2 -1
- cldm/glyph_control.py +0 -230
app.py
CHANGED
@@ -112,7 +112,8 @@ with block:
|
|
112 |
exec(f"""top_left_x_{i} = gr.Slider(label="Bbox Top Left x", minimum=0., maximum=1, value={0.35 - 0.25 * math.cos(math.pi * i)}, step=0.01) """)
|
113 |
exec(f"""top_left_y_{i} = gr.Slider(label="Bbox Top Left y", minimum=0., maximum=1, value={0.1 if i < 2 else 0.6}, step=0.01) """)
|
114 |
exec(f"""yaw_{i} = gr.Slider(label="Bbox Yaw", minimum=-180, maximum=180, value=0, step=5) """)
|
115 |
-
exec(f"""num_rows_{i} = gr.Slider(label="num_rows", minimum=1, maximum=4, value=1, step=1, visible=False) """)
|
|
|
116 |
|
117 |
with gr.Row():
|
118 |
with gr.Column():
|
|
|
112 |
exec(f"""top_left_x_{i} = gr.Slider(label="Bbox Top Left x", minimum=0., maximum=1, value={0.35 - 0.25 * math.cos(math.pi * i)}, step=0.01) """)
|
113 |
exec(f"""top_left_y_{i} = gr.Slider(label="Bbox Top Left y", minimum=0., maximum=1, value={0.1 if i < 2 else 0.6}, step=0.01) """)
|
114 |
exec(f"""yaw_{i} = gr.Slider(label="Bbox Yaw", minimum=-180, maximum=180, value=0, step=5) """)
|
115 |
+
# exec(f"""num_rows_{i} = gr.Slider(label="num_rows", minimum=1, maximum=4, value=1, step=1, visible=False) """)
|
116 |
+
exec(f"""num_rows_{i} = gr.Slider(label="num_rows", minimum=1, maximum=4, value=1, step=1) """)
|
117 |
|
118 |
with gr.Row():
|
119 |
with gr.Column():
|
cldm/glyph_control.py
DELETED
@@ -1,230 +0,0 @@
|
|
1 |
-
import torch.nn as nn
|
2 |
-
from ldm.modules.encoders.modules import OpenCLIPImageEmbedder, FrozenOpenCLIPEmbedder
|
3 |
-
from ldm.util import instantiate_from_config
|
4 |
-
import torch
|
5 |
-
from taming.models.vqgan import VQModelInterfaceEncoder, VQModel
|
6 |
-
from ldm.modules.attention import SpatialTransformer
|
7 |
-
from ldm.modules.attention import Normalize, BasicTransformerBlock#, exists
|
8 |
-
from ldm.modules.diffusionmodules.util import zero_module, identity_init_fc, conv_nd
|
9 |
-
from einops import rearrange
|
10 |
-
# from ldm.modules.diffusionmodules.openaimodel import TimestepEmbedSequential
|
11 |
-
def disabled_train(self, mode=True):
|
12 |
-
"""Overwrite model.train with this function to make sure train/eval mode
|
13 |
-
does not change anymore."""
|
14 |
-
return self
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
def make_zero_conv(in_channels, out_channels, kernel_size, stride=1, padding=0):
|
19 |
-
return zero_module(conv_nd(2, in_channels, out_channels, kernel_size, stride=stride, padding=padding))
|
20 |
-
|
21 |
-
|
22 |
-
class SpatialTransformer_v2(nn.Module):
|
23 |
-
"""
|
24 |
-
Transformer block for image-like data.
|
25 |
-
First, project the input (aka embedding)
|
26 |
-
and reshape to b, t, d.
|
27 |
-
Then apply standard transformer action.
|
28 |
-
Finally, reshape to image
|
29 |
-
NEW: use_linear for more efficiency instead of the 1x1 convs
|
30 |
-
"""
|
31 |
-
def __init__(self, in_channels, n_heads, d_head,
|
32 |
-
depth=1, dropout=0., context_dim=None,
|
33 |
-
disable_self_attn=False, use_linear=False,
|
34 |
-
use_checkpoint=True):
|
35 |
-
super().__init__()
|
36 |
-
# change:
|
37 |
-
# if exists(context_dim) and not isinstance(context_dim, list):
|
38 |
-
if not isinstance(context_dim, list):
|
39 |
-
context_dim = [context_dim]
|
40 |
-
self.in_channels = in_channels
|
41 |
-
inner_dim = n_heads * d_head
|
42 |
-
self.norm = Normalize(in_channels)
|
43 |
-
if not use_linear:
|
44 |
-
self.proj_in = nn.Conv2d(in_channels,
|
45 |
-
inner_dim,
|
46 |
-
kernel_size=1,
|
47 |
-
stride=1,
|
48 |
-
padding=0)
|
49 |
-
else:
|
50 |
-
self.proj_in = nn.Linear(in_channels, inner_dim)
|
51 |
-
|
52 |
-
self.transformer_blocks = nn.ModuleList(
|
53 |
-
[BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
|
54 |
-
disable_self_attn=disable_self_attn, checkpoint=use_checkpoint)
|
55 |
-
for d in range(depth)]
|
56 |
-
)
|
57 |
-
if not use_linear:
|
58 |
-
self.proj_out = zero_module(nn.Conv2d(inner_dim,
|
59 |
-
in_channels,
|
60 |
-
kernel_size=1,
|
61 |
-
stride=1,
|
62 |
-
padding=0))
|
63 |
-
else:
|
64 |
-
self.proj_out = zero_module(nn.Linear(inner_dim, in_channels)) # change: switch
|
65 |
-
self.use_linear = use_linear
|
66 |
-
|
67 |
-
def forward(self, x, context=None):
|
68 |
-
# note: if no context is given, cross-attention defaults to self-attention
|
69 |
-
if not isinstance(context, list):
|
70 |
-
context = [context]
|
71 |
-
b, c, h, w = x.shape
|
72 |
-
x_in = x
|
73 |
-
x = self.norm(x)
|
74 |
-
if not self.use_linear:
|
75 |
-
x = self.proj_in(x)
|
76 |
-
x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
|
77 |
-
if self.use_linear:
|
78 |
-
x = self.proj_in(x)
|
79 |
-
for i, block in enumerate(self.transformer_blocks):
|
80 |
-
x = block(x, context=context[i])
|
81 |
-
if self.use_linear:
|
82 |
-
x = self.proj_out(x)
|
83 |
-
x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
|
84 |
-
if not self.use_linear:
|
85 |
-
x = self.proj_out(x)
|
86 |
-
return x + x_in
|
87 |
-
|
88 |
-
class trans_glyph_emb(nn.Module):
|
89 |
-
def __init__(self,
|
90 |
-
type = "fc", # "conv", "attn"
|
91 |
-
input_dim = 256,
|
92 |
-
out_dim = 1024,
|
93 |
-
# fc
|
94 |
-
fc_init = "zero",
|
95 |
-
# conv/attn
|
96 |
-
conv_ks = 3,
|
97 |
-
conv_pad = 1,
|
98 |
-
conv_stride = 1,
|
99 |
-
# attn
|
100 |
-
ch = 512, # 1024
|
101 |
-
num_heads = 8, # 16
|
102 |
-
dim_head = 64,
|
103 |
-
use_linear_in_transformer = True,
|
104 |
-
use_checkpoint = False, #True,
|
105 |
-
):
|
106 |
-
super().__init__()
|
107 |
-
|
108 |
-
if type == "fc":
|
109 |
-
self.model = torch.nn.Linear(input_dim, out_dim)
|
110 |
-
if fc_init == "zero":
|
111 |
-
self.model = zero_module(self.model)
|
112 |
-
elif fc_init == "identity":
|
113 |
-
self.model = identity_init_fc(self.model)
|
114 |
-
elif type == "conv":
|
115 |
-
self.model = make_zero_conv(input_dim, out_dim, conv_ks, stride = conv_stride, padding = conv_pad)
|
116 |
-
elif type == "attn":
|
117 |
-
model = [
|
118 |
-
# nn.Conv2d(input_dim, ch, 3, stride = 1, padding = 1),
|
119 |
-
nn.Conv2d(input_dim, ch, conv_ks, stride = conv_stride, padding = conv_pad),
|
120 |
-
SpatialTransformer_v2( #SpatialTransformer(
|
121 |
-
ch, num_heads, dim_head, depth=1, context_dim=None, #ch,
|
122 |
-
disable_self_attn=False, use_linear=use_linear_in_transformer,
|
123 |
-
use_checkpoint=use_checkpoint, # False if the context is None
|
124 |
-
),
|
125 |
-
make_zero_conv(ch, out_dim, 1, stride = 1, padding = 0)
|
126 |
-
# make_zero_conv(ch, out_dim, conv_ks, stride = conv_stride, padding = conv_pad)
|
127 |
-
]
|
128 |
-
self.model = nn.Sequential(*model)
|
129 |
-
self.model_type = type
|
130 |
-
|
131 |
-
def forward(self, x):
|
132 |
-
if self.model_type == "fc":
|
133 |
-
# b, c, h, w = x.shape
|
134 |
-
x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
|
135 |
-
x = self.model(x)
|
136 |
-
# x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
|
137 |
-
# return x
|
138 |
-
else:
|
139 |
-
x = self.model(x)
|
140 |
-
x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
|
141 |
-
return x
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
class glyph_control(nn.Module):
|
146 |
-
def __init__(self,
|
147 |
-
image_encoder = "CLIP", # "VQGAN"
|
148 |
-
image_encoder_config = None,
|
149 |
-
fuse_way = "concat",
|
150 |
-
load_text_encoder = False,
|
151 |
-
text_encoder_config = None,
|
152 |
-
freeze_image_encoder = True,
|
153 |
-
trans_emb = False,
|
154 |
-
trans_emb_config = None,
|
155 |
-
# use_fp16 = False,
|
156 |
-
):
|
157 |
-
super().__init__()
|
158 |
-
if image_encoder_config is not None:
|
159 |
-
image_encoder_config.params.freeze = freeze_image_encoder
|
160 |
-
self.image_encoder = instantiate_from_config(image_encoder_config)
|
161 |
-
else:
|
162 |
-
if image_encoder == "CLIP":
|
163 |
-
self.image_encoder = OpenCLIPImageEmbedder(freeze=freeze_image_encoder)
|
164 |
-
elif image_encoder == "VQGAN":
|
165 |
-
print("VQGAN glyph image encoder is missing config")
|
166 |
-
raise ValueError
|
167 |
-
else:
|
168 |
-
print("Other types of glyph image encoder are not supported")
|
169 |
-
raise ValueError
|
170 |
-
|
171 |
-
if freeze_image_encoder:
|
172 |
-
self.freeze_imenc()
|
173 |
-
self.freeze_image_encoder = freeze_image_encoder
|
174 |
-
self.image_encoder_type = image_encoder
|
175 |
-
|
176 |
-
|
177 |
-
if load_text_encoder:
|
178 |
-
if text_encoder_config is None:
|
179 |
-
self.text_encoder = FrozenOpenCLIPEmbedder()
|
180 |
-
else:
|
181 |
-
self.text_encoder = instantiate_from_config(text_encoder_config)
|
182 |
-
self.fuse_way = fuse_way
|
183 |
-
# self.dtype = torch.float16 if use_fp16 else torch.float32
|
184 |
-
if trans_emb:
|
185 |
-
if trans_emb_config is not None:
|
186 |
-
self.trans_glyph_emb_model = instantiate_from_config(trans_emb_config)
|
187 |
-
else:
|
188 |
-
self.trans_glyph_emb_model = trans_glyph_emb()
|
189 |
-
else:
|
190 |
-
self.trans_glyph_emb_model = None
|
191 |
-
|
192 |
-
def freeze_imenc(self):
|
193 |
-
self.image_encoder = self.image_encoder.eval()
|
194 |
-
self.image_encoder.train = disabled_train
|
195 |
-
for param in self.image_encoder.parameters():
|
196 |
-
param.requires_grad = False
|
197 |
-
|
198 |
-
def forward(self, glyph_image, text = None, text_embed = None):
|
199 |
-
clgim_num_list = [img.shape[0] for img in glyph_image]
|
200 |
-
# image_embeds = self.image_encoder(torch.concat(glyph_image, dim=0))
|
201 |
-
gim_concat = torch.concat(glyph_image, dim=0)
|
202 |
-
image_embeds = self.image_encoder(gim_concat)
|
203 |
-
if self.trans_glyph_emb_model is not None:
|
204 |
-
image_embeds = self.trans_glyph_emb_model(image_embeds)
|
205 |
-
image_embeds = torch.split(image_embeds, clgim_num_list)
|
206 |
-
max_image_tokens = max(clgim_num_list)
|
207 |
-
pad_image_embeds = []
|
208 |
-
for image_embed in image_embeds:
|
209 |
-
if image_embed.shape[0] < max_image_tokens:
|
210 |
-
image_embed = torch.concat([
|
211 |
-
image_embed,
|
212 |
-
torch.zeros(
|
213 |
-
(max_image_tokens - image_embed.shape[0], *image_embed.shape[1:]), device=image_embed.device, dtype=image_embed.dtype, # add dtype
|
214 |
-
)], dim=0
|
215 |
-
)
|
216 |
-
pad_image_embeds.append(image_embed)
|
217 |
-
pad_image_embeds = torch.stack(pad_image_embeds, dim = 0)
|
218 |
-
if text_embed is None:
|
219 |
-
assert self.text_encoder, text is not None
|
220 |
-
text_embed = self.text_encoder(text)
|
221 |
-
if self.fuse_way == "concat":
|
222 |
-
assert pad_image_embeds.shape[-1] == text_embed.shape[-1]
|
223 |
-
if len(pad_image_embeds.shape) == 4:
|
224 |
-
b, _, _ , embdim = pad_image_embeds.shape
|
225 |
-
pad_image_embeds = pad_image_embeds.view(b, -1, embdim)
|
226 |
-
out_embed = torch.concat([text_embed, pad_image_embeds], dim= 1)
|
227 |
-
print("concat glyph_embed with text_embed:", out_embed.shape)
|
228 |
-
return out_embed
|
229 |
-
else:
|
230 |
-
raise ValueError("Not support other fuse ways for now!")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|