Upload 3 files
Browse files- config.json +38 -0
- got_vision_b.py +468 -0
- modeling_GOT.py +659 -0
config.json
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "ucaslcl/GOT-OCR2_0",
|
3 |
+
"architectures": [
|
4 |
+
"GOTQwenForCausalLM"
|
5 |
+
],
|
6 |
+
"auto_map": {
|
7 |
+
"AutoConfig": "modeling_GOT.GOTConfig",
|
8 |
+
"AutoModel": "modeling_GOT.GOTQwenForCausalLM"
|
9 |
+
},
|
10 |
+
"attention_dropout": 0.0,
|
11 |
+
"bos_token_id": 151643,
|
12 |
+
"eos_token_id": 151643,
|
13 |
+
"freeze_vision_tower": false,
|
14 |
+
"hidden_act": "silu",
|
15 |
+
"hidden_size": 1024,
|
16 |
+
"im_end_token": 151858,
|
17 |
+
"im_patch_token": 151859,
|
18 |
+
"im_start_token": 151857,
|
19 |
+
"image_token_len": 256,
|
20 |
+
"initializer_range": 0.02,
|
21 |
+
"intermediate_size": 2816,
|
22 |
+
"max_position_embeddings": 32768,
|
23 |
+
"max_window_layers": 21,
|
24 |
+
"model_type": "GOT",
|
25 |
+
"num_attention_heads": 16,
|
26 |
+
"num_hidden_layers": 24,
|
27 |
+
"num_key_value_heads": 16,
|
28 |
+
"rms_norm_eps": 1e-06,
|
29 |
+
"rope_theta": 1000000.0,
|
30 |
+
"sliding_window": 32768,
|
31 |
+
"tie_word_embeddings": true,
|
32 |
+
"torch_dtype": "bfloat16",
|
33 |
+
"transformers_version": "4.37.2",
|
34 |
+
"use_cache": true,
|
35 |
+
"use_im_start_end": true,
|
36 |
+
"use_sliding_window": false,
|
37 |
+
"vocab_size": 151860
|
38 |
+
}
|
got_vision_b.py
ADDED
@@ -0,0 +1,468 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as F
|
3 |
+
from typing import Optional, Tuple, Type
|
4 |
+
from functools import partial
|
5 |
+
import torch.nn as nn
|
6 |
+
from typing import Type
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
class MLPBlock(nn.Module):
|
11 |
+
def __init__(
|
12 |
+
self,
|
13 |
+
embedding_dim: int,
|
14 |
+
mlp_dim: int,
|
15 |
+
act: Type[nn.Module] = nn.GELU,
|
16 |
+
) -> None:
|
17 |
+
super().__init__()
|
18 |
+
self.lin1 = nn.Linear(embedding_dim, mlp_dim)
|
19 |
+
self.lin2 = nn.Linear(mlp_dim, embedding_dim)
|
20 |
+
self.act = act()
|
21 |
+
|
22 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
23 |
+
return self.lin2(self.act(self.lin1(x)))
|
24 |
+
|
25 |
+
|
26 |
+
|
27 |
+
class LayerNorm2d(nn.Module):
|
28 |
+
def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
|
29 |
+
super().__init__()
|
30 |
+
self.weight = nn.Parameter(torch.ones(num_channels))
|
31 |
+
self.bias = nn.Parameter(torch.zeros(num_channels))
|
32 |
+
self.eps = eps
|
33 |
+
|
34 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
35 |
+
u = x.mean(1, keepdim=True)
|
36 |
+
s = (x - u).pow(2).mean(1, keepdim=True)
|
37 |
+
x = (x - u) / torch.sqrt(s + self.eps)
|
38 |
+
x = self.weight[:, None, None] * x + self.bias[:, None, None]
|
39 |
+
return x
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
class ImageEncoderViT(nn.Module):
|
44 |
+
def __init__(
|
45 |
+
self,
|
46 |
+
img_size: int = 1024,
|
47 |
+
patch_size: int = 16,
|
48 |
+
in_chans: int = 3,
|
49 |
+
embed_dim: int = 768,
|
50 |
+
depth: int = 12,
|
51 |
+
num_heads: int = 12,
|
52 |
+
mlp_ratio: float = 4.0,
|
53 |
+
out_chans: int = 256,
|
54 |
+
qkv_bias: bool = True,
|
55 |
+
norm_layer: Type[nn.Module] = nn.LayerNorm,
|
56 |
+
act_layer: Type[nn.Module] = nn.GELU,
|
57 |
+
use_abs_pos: bool = True,
|
58 |
+
use_rel_pos: bool = False,
|
59 |
+
rel_pos_zero_init: bool = True,
|
60 |
+
window_size: int = 0,
|
61 |
+
global_attn_indexes: Tuple[int, ...] = (),
|
62 |
+
) -> None:
|
63 |
+
"""
|
64 |
+
Args:
|
65 |
+
img_size (int): Input image size.
|
66 |
+
patch_size (int): Patch size.
|
67 |
+
in_chans (int): Number of input image channels.
|
68 |
+
embed_dim (int): Patch embedding dimension.
|
69 |
+
depth (int): Depth of ViT.
|
70 |
+
num_heads (int): Number of attention heads in each ViT block.
|
71 |
+
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
|
72 |
+
qkv_bias (bool): If True, add a learnable bias to query, key, value.
|
73 |
+
norm_layer (nn.Module): Normalization layer.
|
74 |
+
act_layer (nn.Module): Activation layer.
|
75 |
+
use_abs_pos (bool): If True, use absolute positional embeddings.
|
76 |
+
use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
|
77 |
+
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
|
78 |
+
window_size (int): Window size for window attention blocks.
|
79 |
+
global_attn_indexes (list): Indexes for blocks using global attention.
|
80 |
+
"""
|
81 |
+
super().__init__()
|
82 |
+
self.img_size = img_size
|
83 |
+
|
84 |
+
self.patch_embed = PatchEmbed(
|
85 |
+
kernel_size=(patch_size, patch_size),
|
86 |
+
stride=(patch_size, patch_size),
|
87 |
+
in_chans=in_chans,
|
88 |
+
embed_dim=embed_dim,
|
89 |
+
)
|
90 |
+
|
91 |
+
self.pos_embed: Optional[nn.Parameter] = None
|
92 |
+
if use_abs_pos:
|
93 |
+
# Initialize absolute positional embedding with pretrain image size.
|
94 |
+
self.pos_embed = nn.Parameter(
|
95 |
+
torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim)
|
96 |
+
)
|
97 |
+
|
98 |
+
self.blocks = nn.ModuleList()
|
99 |
+
for i in range(depth):
|
100 |
+
block = Block(
|
101 |
+
dim=embed_dim,
|
102 |
+
num_heads=num_heads,
|
103 |
+
mlp_ratio=mlp_ratio,
|
104 |
+
qkv_bias=qkv_bias,
|
105 |
+
norm_layer=norm_layer,
|
106 |
+
act_layer=act_layer,
|
107 |
+
use_rel_pos=use_rel_pos,
|
108 |
+
rel_pos_zero_init=rel_pos_zero_init,
|
109 |
+
window_size=window_size if i not in global_attn_indexes else 0,
|
110 |
+
input_size=(img_size // patch_size, img_size // patch_size),
|
111 |
+
)
|
112 |
+
self.blocks.append(block)
|
113 |
+
|
114 |
+
self.neck = nn.Sequential(
|
115 |
+
nn.Conv2d(
|
116 |
+
embed_dim,
|
117 |
+
out_chans,
|
118 |
+
kernel_size=1,
|
119 |
+
bias=False,
|
120 |
+
),
|
121 |
+
LayerNorm2d(out_chans),
|
122 |
+
nn.Conv2d(
|
123 |
+
out_chans,
|
124 |
+
out_chans,
|
125 |
+
kernel_size=3,
|
126 |
+
padding=1,
|
127 |
+
bias=False,
|
128 |
+
),
|
129 |
+
LayerNorm2d(out_chans),
|
130 |
+
)
|
131 |
+
|
132 |
+
|
133 |
+
self.net_2 = nn.Conv2d(256, 512, kernel_size=3, stride=2, padding=1, bias=False)
|
134 |
+
self.net_3 = nn.Conv2d(512, 1024, kernel_size=3, stride=2, padding=1, bias=False)
|
135 |
+
|
136 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
137 |
+
x = self.patch_embed(x)
|
138 |
+
if self.pos_embed is not None:
|
139 |
+
x = x + self.pos_embed
|
140 |
+
|
141 |
+
for blk in self.blocks:
|
142 |
+
x = blk(x)
|
143 |
+
|
144 |
+
x = self.neck(x.permute(0, 3, 1, 2))
|
145 |
+
x = self.net_2(x)
|
146 |
+
x = self.net_3(x)
|
147 |
+
|
148 |
+
|
149 |
+
return x
|
150 |
+
|
151 |
+
|
152 |
+
class Block(nn.Module):
|
153 |
+
"""Transformer blocks with support of window attention and residual propagation blocks"""
|
154 |
+
|
155 |
+
def __init__(
|
156 |
+
self,
|
157 |
+
dim: int,
|
158 |
+
num_heads: int,
|
159 |
+
mlp_ratio: float = 4.0,
|
160 |
+
qkv_bias: bool = True,
|
161 |
+
norm_layer: Type[nn.Module] = nn.LayerNorm,
|
162 |
+
act_layer: Type[nn.Module] = nn.GELU,
|
163 |
+
use_rel_pos: bool = False,
|
164 |
+
rel_pos_zero_init: bool = True,
|
165 |
+
window_size: int = 0,
|
166 |
+
input_size: Optional[Tuple[int, int]] = None,
|
167 |
+
) -> None:
|
168 |
+
"""
|
169 |
+
Args:
|
170 |
+
dim (int): Number of input channels.
|
171 |
+
num_heads (int): Number of attention heads in each ViT block.
|
172 |
+
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
|
173 |
+
qkv_bias (bool): If True, add a learnable bias to query, key, value.
|
174 |
+
norm_layer (nn.Module): Normalization layer.
|
175 |
+
act_layer (nn.Module): Activation layer.
|
176 |
+
use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
|
177 |
+
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
|
178 |
+
window_size (int): Window size for window attention blocks. If it equals 0, then
|
179 |
+
use global attention.
|
180 |
+
input_size (tuple(int, int) or None): Input resolution for calculating the relative
|
181 |
+
positional parameter size.
|
182 |
+
"""
|
183 |
+
super().__init__()
|
184 |
+
self.norm1 = norm_layer(dim)
|
185 |
+
self.attn = Attention(
|
186 |
+
dim,
|
187 |
+
num_heads=num_heads,
|
188 |
+
qkv_bias=qkv_bias,
|
189 |
+
use_rel_pos=use_rel_pos,
|
190 |
+
rel_pos_zero_init=rel_pos_zero_init,
|
191 |
+
input_size=input_size if window_size == 0 else (window_size, window_size),
|
192 |
+
)
|
193 |
+
|
194 |
+
self.norm2 = norm_layer(dim)
|
195 |
+
self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
|
196 |
+
|
197 |
+
self.window_size = window_size
|
198 |
+
|
199 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
200 |
+
shortcut = x
|
201 |
+
x = self.norm1(x)
|
202 |
+
# Window partition
|
203 |
+
if self.window_size > 0:
|
204 |
+
H, W = x.shape[1], x.shape[2]
|
205 |
+
x, pad_hw = window_partition(x, self.window_size)
|
206 |
+
|
207 |
+
x = self.attn(x)
|
208 |
+
# Reverse window partition
|
209 |
+
if self.window_size > 0:
|
210 |
+
x = window_unpartition(x, self.window_size, pad_hw, (H, W))
|
211 |
+
|
212 |
+
x = shortcut + x
|
213 |
+
x = x + self.mlp(self.norm2(x))
|
214 |
+
|
215 |
+
return x
|
216 |
+
|
217 |
+
|
218 |
+
class Attention(nn.Module):
|
219 |
+
"""Multi-head Attention block with relative position embeddings."""
|
220 |
+
|
221 |
+
def __init__(
|
222 |
+
self,
|
223 |
+
dim: int,
|
224 |
+
num_heads: int = 8,
|
225 |
+
qkv_bias: bool = True,
|
226 |
+
use_rel_pos: bool = False,
|
227 |
+
rel_pos_zero_init: bool = True,
|
228 |
+
input_size: Optional[Tuple[int, int]] = None,
|
229 |
+
) -> None:
|
230 |
+
"""
|
231 |
+
Args:
|
232 |
+
dim (int): Number of input channels.
|
233 |
+
num_heads (int): Number of attention heads.
|
234 |
+
qkv_bias (bool): If True, add a learnable bias to query, key, value.
|
235 |
+
rel_pos (bool): If True, add relative positional embeddings to the attention map.
|
236 |
+
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
|
237 |
+
input_size (tuple(int, int) or None): Input resolution for calculating the relative
|
238 |
+
positional parameter size.
|
239 |
+
"""
|
240 |
+
super().__init__()
|
241 |
+
self.num_heads = num_heads
|
242 |
+
head_dim = dim // num_heads
|
243 |
+
self.scale = head_dim**-0.5
|
244 |
+
|
245 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
246 |
+
self.proj = nn.Linear(dim, dim)
|
247 |
+
|
248 |
+
self.use_rel_pos = use_rel_pos
|
249 |
+
if self.use_rel_pos:
|
250 |
+
assert (
|
251 |
+
input_size is not None
|
252 |
+
), "Input size must be provided if using relative positional encoding."
|
253 |
+
# initialize relative positional embeddings
|
254 |
+
self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
|
255 |
+
self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
|
256 |
+
|
257 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
258 |
+
B, H, W, _ = x.shape
|
259 |
+
# qkv with shape (3, B, nHead, H * W, C)
|
260 |
+
qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
|
261 |
+
# q, k, v with shape (B * nHead, H * W, C)
|
262 |
+
q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
|
263 |
+
|
264 |
+
attn = (q * self.scale) @ k.transpose(-2, -1)
|
265 |
+
|
266 |
+
if self.use_rel_pos:
|
267 |
+
attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
|
268 |
+
|
269 |
+
attn = attn.softmax(dim=-1)
|
270 |
+
x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
|
271 |
+
x = self.proj(x)
|
272 |
+
|
273 |
+
return x
|
274 |
+
|
275 |
+
|
276 |
+
def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
|
277 |
+
"""
|
278 |
+
Partition into non-overlapping windows with padding if needed.
|
279 |
+
Args:
|
280 |
+
x (tensor): input tokens with [B, H, W, C].
|
281 |
+
window_size (int): window size.
|
282 |
+
|
283 |
+
Returns:
|
284 |
+
windows: windows after partition with [B * num_windows, window_size, window_size, C].
|
285 |
+
(Hp, Wp): padded height and width before partition
|
286 |
+
"""
|
287 |
+
B, H, W, C = x.shape
|
288 |
+
|
289 |
+
pad_h = (window_size - H % window_size) % window_size
|
290 |
+
pad_w = (window_size - W % window_size) % window_size
|
291 |
+
if pad_h > 0 or pad_w > 0:
|
292 |
+
x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
|
293 |
+
Hp, Wp = H + pad_h, W + pad_w
|
294 |
+
|
295 |
+
x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
|
296 |
+
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
|
297 |
+
return windows, (Hp, Wp)
|
298 |
+
|
299 |
+
|
300 |
+
def window_unpartition(
|
301 |
+
windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
|
302 |
+
) -> torch.Tensor:
|
303 |
+
"""
|
304 |
+
Window unpartition into original sequences and removing padding.
|
305 |
+
Args:
|
306 |
+
windows (tensor): input tokens with [B * num_windows, window_size, window_size, C].
|
307 |
+
window_size (int): window size.
|
308 |
+
pad_hw (Tuple): padded height and width (Hp, Wp).
|
309 |
+
hw (Tuple): original height and width (H, W) before padding.
|
310 |
+
|
311 |
+
Returns:
|
312 |
+
x: unpartitioned sequences with [B, H, W, C].
|
313 |
+
"""
|
314 |
+
Hp, Wp = pad_hw
|
315 |
+
H, W = hw
|
316 |
+
B = windows.shape[0] // (Hp * Wp // window_size // window_size)
|
317 |
+
x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
|
318 |
+
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
|
319 |
+
|
320 |
+
if Hp > H or Wp > W:
|
321 |
+
x = x[:, :H, :W, :].contiguous()
|
322 |
+
return x
|
323 |
+
|
324 |
+
|
325 |
+
def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
|
326 |
+
"""
|
327 |
+
Get relative positional embeddings according to the relative positions of
|
328 |
+
query and key sizes.
|
329 |
+
Args:
|
330 |
+
q_size (int): size of query q.
|
331 |
+
k_size (int): size of key k.
|
332 |
+
rel_pos (Tensor): relative position embeddings (L, C).
|
333 |
+
|
334 |
+
Returns:
|
335 |
+
Extracted positional embeddings according to relative positions.
|
336 |
+
"""
|
337 |
+
max_rel_dist = int(2 * max(q_size, k_size) - 1)
|
338 |
+
# Interpolate rel pos if needed.
|
339 |
+
if rel_pos.shape[0] != max_rel_dist:
|
340 |
+
# Interpolate rel pos.
|
341 |
+
rel_pos_resized = F.interpolate(
|
342 |
+
rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
|
343 |
+
size=max_rel_dist,
|
344 |
+
mode="linear",
|
345 |
+
)
|
346 |
+
rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
|
347 |
+
else:
|
348 |
+
rel_pos_resized = rel_pos
|
349 |
+
|
350 |
+
# Scale the coords with short length if shapes for q and k are different.
|
351 |
+
q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
|
352 |
+
k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
|
353 |
+
relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
|
354 |
+
|
355 |
+
return rel_pos_resized[relative_coords.long()]
|
356 |
+
|
357 |
+
|
358 |
+
def add_decomposed_rel_pos(
|
359 |
+
attn: torch.Tensor,
|
360 |
+
q: torch.Tensor,
|
361 |
+
rel_pos_h: torch.Tensor,
|
362 |
+
rel_pos_w: torch.Tensor,
|
363 |
+
q_size: Tuple[int, int],
|
364 |
+
k_size: Tuple[int, int],
|
365 |
+
) -> torch.Tensor:
|
366 |
+
"""
|
367 |
+
Args:
|
368 |
+
attn (Tensor): attention map.
|
369 |
+
q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
|
370 |
+
rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
|
371 |
+
rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
|
372 |
+
q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
|
373 |
+
k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
|
374 |
+
|
375 |
+
Returns:
|
376 |
+
attn (Tensor): attention map with added relative positional embeddings.
|
377 |
+
"""
|
378 |
+
q_h, q_w = q_size
|
379 |
+
k_h, k_w = k_size
|
380 |
+
Rh = get_rel_pos(q_h, k_h, rel_pos_h)
|
381 |
+
Rw = get_rel_pos(q_w, k_w, rel_pos_w)
|
382 |
+
|
383 |
+
B, _, dim = q.shape
|
384 |
+
r_q = q.reshape(B, q_h, q_w, dim)
|
385 |
+
rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
|
386 |
+
rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
|
387 |
+
|
388 |
+
attn = (
|
389 |
+
attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
|
390 |
+
).view(B, q_h * q_w, k_h * k_w)
|
391 |
+
|
392 |
+
return attn
|
393 |
+
|
394 |
+
|
395 |
+
class PatchEmbed(nn.Module):
|
396 |
+
"""
|
397 |
+
Image to Patch Embedding.
|
398 |
+
"""
|
399 |
+
|
400 |
+
def __init__(
|
401 |
+
self,
|
402 |
+
kernel_size: Tuple[int, int] = (16, 16),
|
403 |
+
stride: Tuple[int, int] = (16, 16),
|
404 |
+
padding: Tuple[int, int] = (0, 0),
|
405 |
+
in_chans: int = 3,
|
406 |
+
embed_dim: int = 768,
|
407 |
+
) -> None:
|
408 |
+
"""
|
409 |
+
Args:
|
410 |
+
kernel_size (Tuple): kernel size of the projection layer.
|
411 |
+
stride (Tuple): stride of the projection layer.
|
412 |
+
padding (Tuple): padding size of the projection layer.
|
413 |
+
in_chans (int): Number of input image channels.
|
414 |
+
embed_dim (int): Patch embedding dimension.
|
415 |
+
"""
|
416 |
+
super().__init__()
|
417 |
+
|
418 |
+
self.proj = nn.Conv2d(
|
419 |
+
in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
|
420 |
+
)
|
421 |
+
|
422 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
423 |
+
x = self.proj(x)
|
424 |
+
# B C H W -> B H W C
|
425 |
+
x = x.permute(0, 2, 3, 1)
|
426 |
+
return x
|
427 |
+
|
428 |
+
|
429 |
+
|
430 |
+
def build_GOT_vit_b(checkpoint=None):
|
431 |
+
return _build_GOT_vision(
|
432 |
+
encoder_embed_dim=768,
|
433 |
+
encoder_depth=12,
|
434 |
+
encoder_num_heads=12,
|
435 |
+
encoder_global_attn_indexes=[2, 5, 8, 11],
|
436 |
+
checkpoint=checkpoint,
|
437 |
+
)
|
438 |
+
|
439 |
+
|
440 |
+
def _build_GOT_vision(
|
441 |
+
encoder_embed_dim,
|
442 |
+
encoder_depth,
|
443 |
+
encoder_num_heads,
|
444 |
+
encoder_global_attn_indexes,
|
445 |
+
checkpoint=None,
|
446 |
+
):
|
447 |
+
prompt_embed_dim = 256
|
448 |
+
image_size = 1024
|
449 |
+
vit_patch_size = 16
|
450 |
+
image_embedding_size = image_size // vit_patch_size
|
451 |
+
image_encoder=ImageEncoderViT(
|
452 |
+
depth=encoder_depth,
|
453 |
+
embed_dim=encoder_embed_dim,
|
454 |
+
img_size=image_size,
|
455 |
+
mlp_ratio=4,
|
456 |
+
norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
|
457 |
+
num_heads=encoder_num_heads,
|
458 |
+
patch_size=vit_patch_size,
|
459 |
+
qkv_bias=True,
|
460 |
+
use_rel_pos=True,
|
461 |
+
global_attn_indexes=encoder_global_attn_indexes,
|
462 |
+
window_size=14,
|
463 |
+
out_chans=prompt_embed_dim,
|
464 |
+
)
|
465 |
+
|
466 |
+
|
467 |
+
return image_encoder
|
468 |
+
|
modeling_GOT.py
ADDED
@@ -0,0 +1,659 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import Qwen2Config, Qwen2Model, Qwen2ForCausalLM, StoppingCriteria, TextStreamer
|
2 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
|
3 |
+
from typing import List, Optional, Tuple, Union
|
4 |
+
from transformers.cache_utils import Cache
|
5 |
+
import requests
|
6 |
+
from PIL import Image
|
7 |
+
from io import BytesIO
|
8 |
+
import torch
|
9 |
+
import torch.nn as nn
|
10 |
+
from torch.nn import CrossEntropyLoss
|
11 |
+
from .got_vision_b import build_GOT_vit_b
|
12 |
+
from torchvision import transforms
|
13 |
+
from torchvision.transforms.functional import InterpolationMode
|
14 |
+
import dataclasses
|
15 |
+
|
16 |
+
|
17 |
+
DEFAULT_IMAGE_TOKEN = "<image>"
|
18 |
+
DEFAULT_IMAGE_PATCH_TOKEN = '<imgpad>'
|
19 |
+
DEFAULT_IM_START_TOKEN = '<img>'
|
20 |
+
DEFAULT_IM_END_TOKEN = '</img>'
|
21 |
+
|
22 |
+
from enum import auto, Enum
|
23 |
+
class SeparatorStyle(Enum):
|
24 |
+
"""Different separator style."""
|
25 |
+
SINGLE = auto()
|
26 |
+
TWO = auto()
|
27 |
+
MPT = auto()
|
28 |
+
|
29 |
+
|
30 |
+
@dataclasses.dataclass
|
31 |
+
class Conversation:
|
32 |
+
"""A class that keeps all conversation history."""
|
33 |
+
system: str
|
34 |
+
roles: List[str]
|
35 |
+
messages: List[List[str]]
|
36 |
+
offset: int
|
37 |
+
sep_style: SeparatorStyle = SeparatorStyle.SINGLE
|
38 |
+
sep: str = "<|im_end|>"
|
39 |
+
sep2: str = None
|
40 |
+
version: str = "Unknown"
|
41 |
+
|
42 |
+
skip_next: bool = False
|
43 |
+
|
44 |
+
def get_prompt(self):
|
45 |
+
if self.sep_style == SeparatorStyle.SINGLE:
|
46 |
+
ret = self.system + self.sep + '\n'
|
47 |
+
for role, message in self.messages:
|
48 |
+
if message:
|
49 |
+
if type(message) is tuple:
|
50 |
+
message, _, _ = message
|
51 |
+
ret += role + ": " + message + self.sep
|
52 |
+
else:
|
53 |
+
ret += role + ":"
|
54 |
+
return ret
|
55 |
+
elif self.sep_style == SeparatorStyle.TWO:
|
56 |
+
seps = [self.sep, self.sep2]
|
57 |
+
ret = self.system + seps[0]
|
58 |
+
for i, (role, message) in enumerate(self.messages):
|
59 |
+
if message:
|
60 |
+
if type(message) is tuple:
|
61 |
+
message, _, _ = message
|
62 |
+
ret += role + ": " + message + seps[i % 2]
|
63 |
+
else:
|
64 |
+
ret += role + ":"
|
65 |
+
return ret
|
66 |
+
if self.sep_style == SeparatorStyle.MPT:
|
67 |
+
if self.system:
|
68 |
+
ret = self.system + self.sep
|
69 |
+
else:
|
70 |
+
ret = ''
|
71 |
+
for role, message in self.messages:
|
72 |
+
if message:
|
73 |
+
if type(message) is tuple:
|
74 |
+
message, _, _ = message
|
75 |
+
ret += role + message + self.sep
|
76 |
+
else:
|
77 |
+
ret += role
|
78 |
+
return ret
|
79 |
+
else:
|
80 |
+
raise ValueError(f"Invalid style: {self.sep_style}")
|
81 |
+
|
82 |
+
|
83 |
+
def append_message(self, role, message):
|
84 |
+
self.messages.append([role, message])
|
85 |
+
|
86 |
+
def copy(self):
|
87 |
+
return Conversation(
|
88 |
+
system=self.system,
|
89 |
+
roles=self.roles,
|
90 |
+
messages=[[x, y] for x, y in self.messages],
|
91 |
+
offset=self.offset,
|
92 |
+
sep_style=self.sep_style,
|
93 |
+
sep=self.sep,
|
94 |
+
sep2=self.sep2)
|
95 |
+
|
96 |
+
|
97 |
+
|
98 |
+
class KeywordsStoppingCriteria(StoppingCriteria):
|
99 |
+
def __init__(self, keywords, tokenizer, input_ids):
|
100 |
+
self.keywords = keywords
|
101 |
+
self.keyword_ids = [tokenizer(keyword).input_ids for keyword in keywords]
|
102 |
+
self.keyword_ids = [keyword_id[0] for keyword_id in self.keyword_ids if type(keyword_id) is list and len(keyword_id) == 1]
|
103 |
+
self.tokenizer = tokenizer
|
104 |
+
self.start_len = None
|
105 |
+
self.input_ids = input_ids
|
106 |
+
|
107 |
+
def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
108 |
+
if self.start_len is None:
|
109 |
+
self.start_len = self.input_ids.shape[1]
|
110 |
+
else:
|
111 |
+
for keyword_id in self.keyword_ids:
|
112 |
+
if output_ids[0, -1] == keyword_id:
|
113 |
+
return True
|
114 |
+
outputs = self.tokenizer.batch_decode(output_ids[:, self.start_len:], skip_special_tokens=True)[0]
|
115 |
+
for keyword in self.keywords:
|
116 |
+
if keyword in outputs:
|
117 |
+
return True
|
118 |
+
return False
|
119 |
+
|
120 |
+
|
121 |
+
class GOTImageEvalProcessor:
|
122 |
+
def __init__(self, image_size=384, mean=None, std=None):
|
123 |
+
if mean is None:
|
124 |
+
mean = (0.48145466, 0.4578275, 0.40821073)
|
125 |
+
if std is None:
|
126 |
+
std = (0.26862954, 0.26130258, 0.27577711)
|
127 |
+
|
128 |
+
self.normalize = transforms.Normalize(mean, std)
|
129 |
+
|
130 |
+
self.transform = transforms.Compose(
|
131 |
+
[
|
132 |
+
transforms.Resize(
|
133 |
+
(image_size, image_size), interpolation=InterpolationMode.BICUBIC
|
134 |
+
),
|
135 |
+
transforms.ToTensor(),
|
136 |
+
self.normalize,
|
137 |
+
]
|
138 |
+
)
|
139 |
+
def __call__(self, item):
|
140 |
+
return self.transform(item)
|
141 |
+
|
142 |
+
|
143 |
+
|
144 |
+
class GOTConfig(Qwen2Config):
|
145 |
+
model_type = "GOT"
|
146 |
+
|
147 |
+
|
148 |
+
class GOTQwenModel(Qwen2Model):
|
149 |
+
config_class = GOTConfig
|
150 |
+
|
151 |
+
def __init__(self, config: Qwen2Config):
|
152 |
+
super(GOTQwenModel, self).__init__(config)
|
153 |
+
|
154 |
+
self.vision_tower_high = build_GOT_vit_b()
|
155 |
+
|
156 |
+
self.mm_projector_vary = nn.Linear(1024, 1024)
|
157 |
+
|
158 |
+
|
159 |
+
def initialize_vision_modules(
|
160 |
+
self,
|
161 |
+
vision_tower,
|
162 |
+
pretrained_stage1_model=None,
|
163 |
+
freeze_vision_tower=False,
|
164 |
+
use_im_start_end=False,
|
165 |
+
vision_select_layer=-1,
|
166 |
+
dtype=torch.float16,
|
167 |
+
device="cuda"
|
168 |
+
):
|
169 |
+
|
170 |
+
|
171 |
+
image_processor_high = GOTImageEvalProcessor(image_size=1024)
|
172 |
+
|
173 |
+
self.vision_tower_high = self.vision_tower_high.to(dtype=dtype, device=device)
|
174 |
+
|
175 |
+
self.mm_projector_vary = self.mm_projector_vary.to(dtype=dtype, device=device)
|
176 |
+
|
177 |
+
|
178 |
+
image_token_len = 256
|
179 |
+
|
180 |
+
self.config.vision_tower = vision_tower
|
181 |
+
self.config.image_token_len = image_token_len
|
182 |
+
|
183 |
+
self.config.use_im_start_end = True
|
184 |
+
|
185 |
+
self.config.vision_select_layer = vision_select_layer
|
186 |
+
self.config.freeze_vision_tower = freeze_vision_tower
|
187 |
+
|
188 |
+
return dict(
|
189 |
+
image_processor_high=image_processor_high,
|
190 |
+
image_token_len=image_token_len,
|
191 |
+
)
|
192 |
+
|
193 |
+
|
194 |
+
def forward(
|
195 |
+
self,
|
196 |
+
input_ids: torch.LongTensor = None,
|
197 |
+
attention_mask: Optional[torch.Tensor] = None,
|
198 |
+
position_ids: Optional[torch.LongTensor] = None,
|
199 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
200 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
201 |
+
use_cache: Optional[bool] = None,
|
202 |
+
output_attentions: Optional[bool] = None,
|
203 |
+
output_hidden_states: Optional[bool] = None,
|
204 |
+
images: Optional[torch.FloatTensor] = None,
|
205 |
+
return_dict: Optional[bool] = None,
|
206 |
+
) -> Union[Tuple, BaseModelOutputWithPast]:
|
207 |
+
|
208 |
+
# HACK: replace back original embeddings for LLaVA pretraining
|
209 |
+
orig_embeds_params = getattr(self, 'orig_embeds_params', None)
|
210 |
+
if orig_embeds_params is not None:
|
211 |
+
with torch.no_grad():
|
212 |
+
self.get_input_embeddings().weight[:-self.num_new_tokens] = orig_embeds_params[:-self.num_new_tokens].data
|
213 |
+
|
214 |
+
if inputs_embeds is None:
|
215 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
216 |
+
|
217 |
+
|
218 |
+
vision_tower_high = getattr(self, 'vision_tower_high', None)
|
219 |
+
|
220 |
+
|
221 |
+
if vision_tower_high is not None and (input_ids.shape[1] != 1 or self.training) and images is not None:
|
222 |
+
use_im_start_end = getattr(self.config, "use_im_start_end", -1)
|
223 |
+
|
224 |
+
vision_select_layer = getattr(self.config, "vision_select_layer", -1)
|
225 |
+
im_patch_token = getattr(self.config, "im_patch_token", -1)
|
226 |
+
im_start_token = getattr(self.config, "im_start_token", -1)
|
227 |
+
im_end_token = getattr(self.config, "im_end_token", -1)
|
228 |
+
freeze_vision_tower = getattr(self.config, "freeze_vision_tower", False)
|
229 |
+
|
230 |
+
im_patch_token = 151859
|
231 |
+
|
232 |
+
im_start_token = 151857
|
233 |
+
|
234 |
+
im_end_token = 151858
|
235 |
+
|
236 |
+
image_features = []
|
237 |
+
|
238 |
+
|
239 |
+
for image in images:
|
240 |
+
P, C, H, W = image.shape
|
241 |
+
if P == 1:
|
242 |
+
with torch.set_grad_enabled(False):
|
243 |
+
cnn_feature = vision_tower_high(image)
|
244 |
+
cnn_feature = cnn_feature.flatten(2).permute(0, 2, 1) # 256*1024
|
245 |
+
image_feature = self.mm_projector_vary(cnn_feature)
|
246 |
+
image_features.append(image_feature)
|
247 |
+
|
248 |
+
else:
|
249 |
+
image_patches = torch.unbind(image)
|
250 |
+
image_patches_features = []
|
251 |
+
for image_patch in image_patches:
|
252 |
+
image_p = torch.stack([image_patch])
|
253 |
+
with torch.set_grad_enabled(False):
|
254 |
+
cnn_feature_p = vision_tower_high(image_p)
|
255 |
+
cnn_feature_p = cnn_feature_p.flatten(2).permute(0, 2, 1)
|
256 |
+
image_feature_p = self.mm_projector_vary(cnn_feature_p)
|
257 |
+
image_patches_features.append(image_feature_p)
|
258 |
+
image_feature = torch.cat(image_patches_features, dim=1)
|
259 |
+
image_features.append(image_feature)
|
260 |
+
|
261 |
+
|
262 |
+
dummy_image_features_2 = torch.zeros(256, 1024, device=inputs_embeds.device, dtype=inputs_embeds.dtype)
|
263 |
+
dummy_image_features = dummy_image_features_2
|
264 |
+
use_im_start_end = True
|
265 |
+
new_input_embeds = []
|
266 |
+
for cur_input_ids, cur_input_embeds, cur_image_features in zip(input_ids, inputs_embeds, image_features):
|
267 |
+
if (cur_input_ids == im_patch_token).sum() == 0:
|
268 |
+
cur_input_embeds = cur_input_embeds + (0. * dummy_image_features).sum()
|
269 |
+
new_input_embeds.append(cur_input_embeds)
|
270 |
+
continue
|
271 |
+
|
272 |
+
if use_im_start_end:
|
273 |
+
if (cur_input_ids == im_start_token).sum() != (cur_input_ids == im_end_token).sum():
|
274 |
+
raise ValueError("The number of image start tokens and image end tokens should be the same.")
|
275 |
+
|
276 |
+
image_start_tokens = torch.where(cur_input_ids == im_start_token)[0]
|
277 |
+
for image_start_token_pos, per_cur_image_features in zip(image_start_tokens, cur_image_features):
|
278 |
+
per_cur_image_features = per_cur_image_features.to(device=cur_input_embeds.device)
|
279 |
+
num_patches = per_cur_image_features.shape[0]
|
280 |
+
|
281 |
+
if cur_input_ids[image_start_token_pos + num_patches + 1] != im_end_token:
|
282 |
+
raise ValueError("The image end token should follow the image start token.")
|
283 |
+
|
284 |
+
cur_input_embeds = torch.cat(
|
285 |
+
(
|
286 |
+
cur_input_embeds[:image_start_token_pos+1],
|
287 |
+
per_cur_image_features,
|
288 |
+
cur_input_embeds[image_start_token_pos + num_patches + 1:]
|
289 |
+
),
|
290 |
+
dim=0
|
291 |
+
)
|
292 |
+
|
293 |
+
|
294 |
+
new_input_embeds.append(cur_input_embeds)
|
295 |
+
else:
|
296 |
+
raise NotImplementedError
|
297 |
+
|
298 |
+
inputs_embeds = torch.stack(new_input_embeds, dim=0)
|
299 |
+
|
300 |
+
return super(GOTQwenModel, self).forward(
|
301 |
+
input_ids=None, attention_mask=attention_mask, past_key_values=past_key_values,
|
302 |
+
inputs_embeds=inputs_embeds, use_cache=use_cache, position_ids = position_ids,
|
303 |
+
output_attentions=output_attentions, output_hidden_states=output_hidden_states,
|
304 |
+
return_dict=return_dict
|
305 |
+
)
|
306 |
+
|
307 |
+
|
308 |
+
|
309 |
+
class GOTQwenForCausalLM(Qwen2ForCausalLM):
|
310 |
+
config_class = GOTConfig
|
311 |
+
# supports_gradient_checkpointing = True
|
312 |
+
|
313 |
+
def __init__(self, config):
|
314 |
+
super(Qwen2ForCausalLM, self).__init__(config)
|
315 |
+
self.model = GOTQwenModel(config)
|
316 |
+
|
317 |
+
self.vocab_size = config.vocab_size
|
318 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
319 |
+
|
320 |
+
# Initialize weights and apply final processing
|
321 |
+
self.post_init()
|
322 |
+
|
323 |
+
def get_model(self):
|
324 |
+
return self.model
|
325 |
+
|
326 |
+
def forward(
|
327 |
+
self,
|
328 |
+
input_ids: torch.LongTensor = None,
|
329 |
+
attention_mask: Optional[torch.Tensor] = None,
|
330 |
+
position_ids: Optional[torch.LongTensor] = None,
|
331 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
332 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
333 |
+
labels: Optional[torch.LongTensor] = None,
|
334 |
+
use_cache: Optional[bool] = None,
|
335 |
+
output_attentions: Optional[bool] = None,
|
336 |
+
output_hidden_states: Optional[bool] = None,
|
337 |
+
images: Optional[torch.FloatTensor] = None,
|
338 |
+
return_dict: Optional[bool] = None,
|
339 |
+
|
340 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
341 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
342 |
+
output_hidden_states = (
|
343 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
344 |
+
)
|
345 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
346 |
+
|
347 |
+
outputs = self.model(
|
348 |
+
input_ids=input_ids,
|
349 |
+
past_key_values=past_key_values,
|
350 |
+
attention_mask=attention_mask,
|
351 |
+
position_ids=position_ids,
|
352 |
+
inputs_embeds=inputs_embeds,
|
353 |
+
use_cache=use_cache,
|
354 |
+
output_attentions=output_attentions,
|
355 |
+
output_hidden_states=output_hidden_states,
|
356 |
+
images=images,
|
357 |
+
return_dict=return_dict
|
358 |
+
|
359 |
+
)
|
360 |
+
|
361 |
+
hidden_states = outputs[0]
|
362 |
+
logits = self.lm_head(hidden_states)
|
363 |
+
logits = logits.float()
|
364 |
+
|
365 |
+
# logits
|
366 |
+
|
367 |
+
loss = None
|
368 |
+
if labels is not None:
|
369 |
+
# Shift so that tokens < n predict n
|
370 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
371 |
+
shift_labels = labels[..., 1:].contiguous()
|
372 |
+
# Flatten the tokens
|
373 |
+
loss_fct = CrossEntropyLoss()
|
374 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
375 |
+
shift_labels = shift_labels.view(-1)
|
376 |
+
# Enable model parallelism
|
377 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
378 |
+
loss = loss_fct(shift_logits, shift_labels)
|
379 |
+
|
380 |
+
if not return_dict:
|
381 |
+
output = (logits,) + outputs[1:]
|
382 |
+
return (loss,) + output if loss is not None else output
|
383 |
+
|
384 |
+
return CausalLMOutputWithPast(
|
385 |
+
loss=loss,
|
386 |
+
logits=logits,
|
387 |
+
past_key_values=outputs.past_key_values,
|
388 |
+
hidden_states=outputs.hidden_states,
|
389 |
+
attentions=outputs.attentions,
|
390 |
+
)
|
391 |
+
|
392 |
+
|
393 |
+
def prepare_inputs_for_generation(
|
394 |
+
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
395 |
+
):
|
396 |
+
# Omit tokens covered by past_key_values
|
397 |
+
if past_key_values is not None:
|
398 |
+
if isinstance(past_key_values, Cache):
|
399 |
+
cache_length = past_key_values.get_seq_length()
|
400 |
+
past_length = past_key_values.seen_tokens
|
401 |
+
max_cache_length = past_key_values.get_max_length()
|
402 |
+
else:
|
403 |
+
cache_length = past_length = past_key_values[0][0].shape[2]
|
404 |
+
max_cache_length = None
|
405 |
+
|
406 |
+
# Keep only the unprocessed tokens:
|
407 |
+
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
408 |
+
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
409 |
+
# input)
|
410 |
+
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
|
411 |
+
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
|
412 |
+
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
|
413 |
+
# input_ids based on the past_length.
|
414 |
+
elif past_length < input_ids.shape[1]:
|
415 |
+
input_ids = input_ids[:, past_length:]
|
416 |
+
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
|
417 |
+
|
418 |
+
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
|
419 |
+
if (
|
420 |
+
max_cache_length is not None
|
421 |
+
and attention_mask is not None
|
422 |
+
and cache_length + input_ids.shape[1] > max_cache_length
|
423 |
+
):
|
424 |
+
attention_mask = attention_mask[:, -max_cache_length:]
|
425 |
+
|
426 |
+
position_ids = kwargs.get("position_ids", None)
|
427 |
+
if attention_mask is not None and position_ids is None:
|
428 |
+
# create position_ids on the fly for batch generation
|
429 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
430 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
431 |
+
if past_key_values:
|
432 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
433 |
+
|
434 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
435 |
+
if inputs_embeds is not None and past_key_values is None:
|
436 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
437 |
+
else:
|
438 |
+
model_inputs = {"input_ids": input_ids}
|
439 |
+
|
440 |
+
model_inputs.update(
|
441 |
+
{
|
442 |
+
"position_ids": position_ids,
|
443 |
+
"past_key_values": past_key_values,
|
444 |
+
"use_cache": kwargs.get("use_cache"),
|
445 |
+
"attention_mask": attention_mask,
|
446 |
+
"images": kwargs.get("images", None),
|
447 |
+
}
|
448 |
+
)
|
449 |
+
return model_inputs
|
450 |
+
|
451 |
+
def initialize_vision_tokenizer(
|
452 |
+
self,
|
453 |
+
tokenizer,
|
454 |
+
freeze_lm_model=False,
|
455 |
+
pretrained_stage1_model=None,
|
456 |
+
device="cuda"
|
457 |
+
):
|
458 |
+
config = self.get_model().config
|
459 |
+
|
460 |
+
|
461 |
+
self.resize_token_embeddings(len(tokenizer))
|
462 |
+
|
463 |
+
config.im_patch_token = 151859
|
464 |
+
|
465 |
+
config.use_im_start_end = True
|
466 |
+
|
467 |
+
if config.use_im_start_end:
|
468 |
+
self.resize_token_embeddings(len(tokenizer))
|
469 |
+
config.im_start_token, config.im_end_token = 151857, 151858
|
470 |
+
|
471 |
+
def load_image(self, image_file):
|
472 |
+
if image_file.startswith('http') or image_file.startswith('https'):
|
473 |
+
response = requests.get(image_file)
|
474 |
+
image = Image.open(BytesIO(response.content)).convert('RGB')
|
475 |
+
else:
|
476 |
+
image = Image.open(image_file).convert('RGB')
|
477 |
+
return image
|
478 |
+
|
479 |
+
def disable_torch_init(self):
|
480 |
+
"""
|
481 |
+
Disable the redundant torch default initialization to accelerate model creation.
|
482 |
+
"""
|
483 |
+
import torch
|
484 |
+
setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
|
485 |
+
setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)
|
486 |
+
|
487 |
+
def chat(self, tokenizer, image_file, ocr_type, ocr_box='', ocr_color='', render=False):
|
488 |
+
|
489 |
+
self.disable_torch_init()
|
490 |
+
|
491 |
+
|
492 |
+
image_processor_high = GOTImageEvalProcessor(image_size=1024)
|
493 |
+
|
494 |
+
use_im_start_end = True
|
495 |
+
|
496 |
+
image_token_len = 256
|
497 |
+
|
498 |
+
image = self.load_image(image_file)
|
499 |
+
|
500 |
+
w, h = image.size
|
501 |
+
|
502 |
+
if ocr_type == 'format':
|
503 |
+
qs = 'OCR with format: '
|
504 |
+
else:
|
505 |
+
qs = 'OCR: '
|
506 |
+
|
507 |
+
if ocr_box:
|
508 |
+
bbox = eval(ocr_box)
|
509 |
+
if len(bbox) == 2:
|
510 |
+
bbox[0] = int(bbox[0]/w*1000)
|
511 |
+
bbox[1] = int(bbox[1]/h*1000)
|
512 |
+
if len(bbox) == 4:
|
513 |
+
bbox[0] = int(bbox[0]/w*1000)
|
514 |
+
bbox[1] = int(bbox[1]/h*1000)
|
515 |
+
bbox[2] = int(bbox[2]/w*1000)
|
516 |
+
bbox[3] = int(bbox[3]/h*1000)
|
517 |
+
if ocr_type == 'format':
|
518 |
+
qs = str(bbox) + ' ' + 'OCR with format: '
|
519 |
+
else:
|
520 |
+
qs = str(bbox) + ' ' + 'OCR: '
|
521 |
+
|
522 |
+
if ocr_color:
|
523 |
+
if ocr_type == 'format':
|
524 |
+
qs = '[' + ocr_color + ']' + ' ' + 'OCR with format: '
|
525 |
+
else:
|
526 |
+
qs = '[' + ocr_color + ']' + ' ' + 'OCR: '
|
527 |
+
|
528 |
+
if use_im_start_end:
|
529 |
+
qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_PATCH_TOKEN*image_token_len + DEFAULT_IM_END_TOKEN + '\n' + qs
|
530 |
+
else:
|
531 |
+
qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
|
532 |
+
|
533 |
+
|
534 |
+
conv_mpt = Conversation(
|
535 |
+
system="""<|im_start|>system
|
536 |
+
You should follow the instructions carefully and explain your answers in detail.""",
|
537 |
+
# system = None,
|
538 |
+
roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
|
539 |
+
version="mpt",
|
540 |
+
messages=(),
|
541 |
+
offset=0,
|
542 |
+
sep_style=SeparatorStyle.MPT,
|
543 |
+
sep="<|im_end|>",
|
544 |
+
)
|
545 |
+
|
546 |
+
conv = conv_mpt.copy()
|
547 |
+
conv.append_message(conv.roles[0], qs)
|
548 |
+
conv.append_message(conv.roles[1], None)
|
549 |
+
prompt = conv.get_prompt()
|
550 |
+
|
551 |
+
print(prompt)
|
552 |
+
|
553 |
+
inputs = tokenizer([prompt])
|
554 |
+
|
555 |
+
image_tensor_1 = image_processor_high(image)
|
556 |
+
|
557 |
+
input_ids = torch.as_tensor(inputs.input_ids).cuda()
|
558 |
+
|
559 |
+
stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
|
560 |
+
keywords = [stop_str]
|
561 |
+
stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
|
562 |
+
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
|
563 |
+
|
564 |
+
|
565 |
+
with torch.autocast("cuda", dtype=torch.bfloat16):
|
566 |
+
output_ids = self.generate(
|
567 |
+
input_ids,
|
568 |
+
images=[image_tensor_1.unsqueeze(0).half().cuda()],
|
569 |
+
do_sample=False,
|
570 |
+
num_beams = 1,
|
571 |
+
no_repeat_ngram_size = 20,
|
572 |
+
streamer=streamer,
|
573 |
+
max_new_tokens=4096,
|
574 |
+
stopping_criteria=[stopping_criteria]
|
575 |
+
)
|
576 |
+
|
577 |
+
|
578 |
+
# if render:
|
579 |
+
# print('==============rendering===============')
|
580 |
+
|
581 |
+
# outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
|
582 |
+
|
583 |
+
# if outputs.endswith(stop_str):
|
584 |
+
# outputs = outputs[:-len(stop_str)]
|
585 |
+
# outputs = outputs.strip()
|
586 |
+
|
587 |
+
# if '**kern' in outputs:
|
588 |
+
# import verovio
|
589 |
+
# from cairosvg import svg2png
|
590 |
+
# import cv2
|
591 |
+
# import numpy as np
|
592 |
+
# tk = verovio.toolkit()
|
593 |
+
# tk.loadData(outputs)
|
594 |
+
# tk.setOptions({"pageWidth": 2100, "footer": 'none',
|
595 |
+
# 'barLineWidth': 0.5, 'beamMaxSlope': 15,
|
596 |
+
# 'staffLineWidth': 0.2, 'spacingStaff': 6})
|
597 |
+
# tk.getPageCount()
|
598 |
+
# svg = tk.renderToSVG()
|
599 |
+
# svg = svg.replace("overflow=\"inherit\"", "overflow=\"visible\"")
|
600 |
+
|
601 |
+
# svg_to_html(svg, "./results/demo.html")
|
602 |
+
|
603 |
+
# if ocr_type == 'format' and '**kern' not in outputs:
|
604 |
+
|
605 |
+
|
606 |
+
# if '\\begin{tikzpicture}' not in outputs:
|
607 |
+
# html_path = "./render_tools/" + "/content-mmd-to-html.html"
|
608 |
+
# html_path_2 = "./results/demo.html"
|
609 |
+
# right_num = outputs.count('\\right')
|
610 |
+
# left_num = outputs.count('\left')
|
611 |
+
|
612 |
+
# if right_num != left_num:
|
613 |
+
# outputs = outputs.replace('\left(', '(').replace('\\right)', ')').replace('\left[', '[').replace('\\right]', ']').replace('\left{', '{').replace('\\right}', '}').replace('\left|', '|').replace('\\right|', '|').replace('\left.', '.').replace('\\right.', '.')
|
614 |
+
|
615 |
+
|
616 |
+
# outputs = outputs.replace('"', '``').replace('$', '')
|
617 |
+
|
618 |
+
# outputs_list = outputs.split('\n')
|
619 |
+
# gt= ''
|
620 |
+
# for out in outputs_list:
|
621 |
+
# gt += '"' + out.replace('\\', '\\\\') + r'\n' + '"' + '+' + '\n'
|
622 |
+
|
623 |
+
# gt = gt[:-2]
|
624 |
+
|
625 |
+
# with open(html_path, 'r') as web_f:
|
626 |
+
# lines = web_f.read()
|
627 |
+
# lines = lines.split("const text =")
|
628 |
+
# new_web = lines[0] + 'const text =' + gt + lines[1]
|
629 |
+
# else:
|
630 |
+
# html_path = "./render_tools/" + "/tikz.html"
|
631 |
+
# html_path_2 = "./results/demo.html"
|
632 |
+
# outputs = outputs.translate(translation_table)
|
633 |
+
# outputs_list = outputs.split('\n')
|
634 |
+
# gt= ''
|
635 |
+
# for out in outputs_list:
|
636 |
+
# if out:
|
637 |
+
# if '\\begin{tikzpicture}' not in out and '\\end{tikzpicture}' not in out:
|
638 |
+
# while out[-1] == ' ':
|
639 |
+
# out = out[:-1]
|
640 |
+
# if out is None:
|
641 |
+
# break
|
642 |
+
|
643 |
+
# if out:
|
644 |
+
# if out[-1] != ';':
|
645 |
+
# gt += out[:-1] + ';\n'
|
646 |
+
# else:
|
647 |
+
# gt += out + '\n'
|
648 |
+
# else:
|
649 |
+
# gt += out + '\n'
|
650 |
+
|
651 |
+
|
652 |
+
# with open(html_path, 'r') as web_f:
|
653 |
+
# lines = web_f.read()
|
654 |
+
# lines = lines.split("const text =")
|
655 |
+
# new_web = lines[0] + gt + lines[1]
|
656 |
+
|
657 |
+
# with open(html_path_2, 'w') as web_f_new:
|
658 |
+
# web_f_new.write(new_web)
|
659 |
+
|