Spaces:
Running
on
Zero
Running
on
Zero
Update model.py
Browse files
model.py
CHANGED
@@ -1,12 +1,13 @@
|
|
|
|
1 |
# Merge image encoder and fuse module to create an ID Encoder
|
2 |
-
#
|
3 |
|
4 |
import torch
|
5 |
import torch.nn as nn
|
6 |
from transformers.models.clip.modeling_clip import CLIPVisionModelWithProjection
|
7 |
from transformers.models.clip.configuration_clip import CLIPVisionConfig
|
8 |
-
from transformers import PretrainedConfig
|
9 |
|
|
|
10 |
VISION_CONFIG_DICT = {
|
11 |
"hidden_size": 1024,
|
12 |
"intermediate_size": 4096,
|
@@ -17,10 +18,11 @@ VISION_CONFIG_DICT = {
|
|
17 |
}
|
18 |
|
19 |
class MLP(nn.Module):
|
|
|
20 |
def __init__(self, in_dim, out_dim, hidden_dim, use_residual=True):
|
21 |
super().__init__()
|
22 |
if use_residual:
|
23 |
-
assert in_dim == out_dim
|
24 |
self.layernorm = nn.LayerNorm(in_dim)
|
25 |
self.fc1 = nn.Linear(in_dim, hidden_dim)
|
26 |
self.fc2 = nn.Linear(hidden_dim, out_dim)
|
@@ -34,11 +36,11 @@ class MLP(nn.Module):
|
|
34 |
x = self.act_fn(x)
|
35 |
x = self.fc2(x)
|
36 |
if self.use_residual:
|
37 |
-
x
|
38 |
return x
|
39 |
|
40 |
-
|
41 |
class FuseModule(nn.Module):
|
|
|
42 |
def __init__(self, embed_dim):
|
43 |
super().__init__()
|
44 |
self.mlp1 = MLP(embed_dim * 2, embed_dim, embed_dim, use_residual=False)
|
@@ -46,68 +48,85 @@ class FuseModule(nn.Module):
|
|
46 |
self.layer_norm = nn.LayerNorm(embed_dim)
|
47 |
|
48 |
def fuse_fn(self, prompt_embeds, id_embeds):
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
return
|
54 |
-
|
55 |
-
def forward(
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
|
|
|
|
|
|
62 |
id_embeds = id_embeds.to(prompt_embeds.dtype)
|
63 |
-
num_inputs = class_tokens_mask.sum().unsqueeze(0) # TODO: check for training case
|
64 |
batch_size, max_num_inputs = id_embeds.shape[:2]
|
65 |
-
# seq_length: 77
|
66 |
seq_length = prompt_embeds.shape[1]
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
)
|
71 |
-
|
72 |
-
|
73 |
-
torch.arange(max_num_inputs, device=flat_id_embeds.device)[None, :]
|
74 |
-
< num_inputs[:, None]
|
75 |
-
)
|
76 |
valid_id_embeds = flat_id_embeds[valid_id_mask.flatten()]
|
77 |
|
78 |
-
|
79 |
-
|
|
|
80 |
valid_id_embeds = valid_id_embeds.view(-1, valid_id_embeds.shape[-1])
|
81 |
-
|
82 |
-
image_token_embeds =
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
|
|
|
|
|
|
|
|
|
|
87 |
return updated_prompt_embeds
|
88 |
|
89 |
class PhotoMakerIDEncoder(CLIPVisionModelWithProjection):
|
|
|
90 |
def __init__(self):
|
91 |
super().__init__(CLIPVisionConfig(**VISION_CONFIG_DICT))
|
92 |
self.visual_projection_2 = nn.Linear(1024, 1280, bias=False)
|
93 |
-
self.fuse_module = FuseModule(2048)
|
94 |
|
95 |
def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
b, num_inputs, c, h, w = id_pixel_values.shape
|
97 |
id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
|
98 |
|
99 |
-
|
|
|
|
|
100 |
id_embeds = self.visual_projection(shared_id_embeds)
|
101 |
id_embeds_2 = self.visual_projection_2(shared_id_embeds)
|
102 |
|
103 |
id_embeds = id_embeds.view(b, num_inputs, 1, -1)
|
104 |
-
id_embeds_2 = id_embeds_2.view(b, num_inputs, 1, -1)
|
105 |
|
106 |
-
|
107 |
-
updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
|
108 |
|
109 |
-
|
110 |
|
|
|
111 |
|
112 |
if __name__ == "__main__":
|
113 |
-
PhotoMakerIDEncoder()
|
|
|
|
1 |
+
# model.py
|
2 |
# Merge image encoder and fuse module to create an ID Encoder
|
3 |
+
# Allows multiple ID images to update the text encoder with a stacked ID embedding.
|
4 |
|
5 |
import torch
|
6 |
import torch.nn as nn
|
7 |
from transformers.models.clip.modeling_clip import CLIPVisionModelWithProjection
|
8 |
from transformers.models.clip.configuration_clip import CLIPVisionConfig
|
|
|
9 |
|
10 |
+
# Vision backbone configuration for the CLIP-based encoder
|
11 |
VISION_CONFIG_DICT = {
|
12 |
"hidden_size": 1024,
|
13 |
"intermediate_size": 4096,
|
|
|
18 |
}
|
19 |
|
20 |
class MLP(nn.Module):
|
21 |
+
"""Simple MLP block with optional residual connection."""
|
22 |
def __init__(self, in_dim, out_dim, hidden_dim, use_residual=True):
|
23 |
super().__init__()
|
24 |
if use_residual:
|
25 |
+
assert in_dim == out_dim, "Input and output dimensions must match when using residual."
|
26 |
self.layernorm = nn.LayerNorm(in_dim)
|
27 |
self.fc1 = nn.Linear(in_dim, hidden_dim)
|
28 |
self.fc2 = nn.Linear(hidden_dim, out_dim)
|
|
|
36 |
x = self.act_fn(x)
|
37 |
x = self.fc2(x)
|
38 |
if self.use_residual:
|
39 |
+
x += residual
|
40 |
return x
|
41 |
|
|
|
42 |
class FuseModule(nn.Module):
|
43 |
+
"""Module that fuses prompt embeddings with ID embeddings."""
|
44 |
def __init__(self, embed_dim):
|
45 |
super().__init__()
|
46 |
self.mlp1 = MLP(embed_dim * 2, embed_dim, embed_dim, use_residual=False)
|
|
|
48 |
self.layer_norm = nn.LayerNorm(embed_dim)
|
49 |
|
50 |
def fuse_fn(self, prompt_embeds, id_embeds):
|
51 |
+
"""Performs two-step fusion of prompt and ID embeddings."""
|
52 |
+
stacked = torch.cat([prompt_embeds, id_embeds], dim=-1)
|
53 |
+
fused = self.mlp1(stacked) + prompt_embeds
|
54 |
+
fused = self.mlp2(fused)
|
55 |
+
return self.layer_norm(fused)
|
56 |
+
|
57 |
+
def forward(self, prompt_embeds, id_embeds, class_tokens_mask):
|
58 |
+
"""
|
59 |
+
Args:
|
60 |
+
prompt_embeds (Tensor): Text encoder embeddings [batch, seq_len, embed_dim]
|
61 |
+
id_embeds (Tensor): ID embeddings [batch, max_inputs, 1, embed_dim]
|
62 |
+
class_tokens_mask (Tensor): Mask indicating which tokens to replace [batch, seq_len]
|
63 |
+
|
64 |
+
Returns:
|
65 |
+
Tensor: Updated prompt embeddings.
|
66 |
+
"""
|
67 |
id_embeds = id_embeds.to(prompt_embeds.dtype)
|
|
|
68 |
batch_size, max_num_inputs = id_embeds.shape[:2]
|
|
|
69 |
seq_length = prompt_embeds.shape[1]
|
70 |
+
|
71 |
+
num_inputs = class_tokens_mask.sum(dim=1)
|
72 |
+
|
73 |
+
flat_id_embeds = id_embeds.view(-1, id_embeds.shape[-2], id_embeds.shape[-1])
|
74 |
+
valid_id_mask = (torch.arange(max_num_inputs, device=flat_id_embeds.device)[None, :] < num_inputs[:, None])
|
75 |
+
|
|
|
|
|
|
|
76 |
valid_id_embeds = flat_id_embeds[valid_id_mask.flatten()]
|
77 |
|
78 |
+
prompt_embeds_flat = prompt_embeds.view(-1, prompt_embeds.shape[-1])
|
79 |
+
class_tokens_mask_flat = class_tokens_mask.view(-1)
|
80 |
+
|
81 |
valid_id_embeds = valid_id_embeds.view(-1, valid_id_embeds.shape[-1])
|
82 |
+
|
83 |
+
image_token_embeds = prompt_embeds_flat[class_tokens_mask_flat]
|
84 |
+
stacked_embeds = self.fuse_fn(image_token_embeds, valid_id_embeds)
|
85 |
+
|
86 |
+
assert class_tokens_mask_flat.sum() == stacked_embeds.shape[0], (
|
87 |
+
f"Mismatch between mask sum and stacked embeds: {class_tokens_mask_flat.sum()} vs {stacked_embeds.shape[0]}"
|
88 |
+
)
|
89 |
+
|
90 |
+
prompt_embeds_flat.masked_scatter_(class_tokens_mask_flat[:, None], stacked_embeds.to(prompt_embeds.dtype))
|
91 |
+
updated_prompt_embeds = prompt_embeds_flat.view(batch_size, seq_length, -1)
|
92 |
+
|
93 |
return updated_prompt_embeds
|
94 |
|
95 |
class PhotoMakerIDEncoder(CLIPVisionModelWithProjection):
|
96 |
+
"""ID Encoder combining vision features and text prompts."""
|
97 |
def __init__(self):
|
98 |
super().__init__(CLIPVisionConfig(**VISION_CONFIG_DICT))
|
99 |
self.visual_projection_2 = nn.Linear(1024, 1280, bias=False)
|
100 |
+
self.fuse_module = FuseModule(embed_dim=2048)
|
101 |
|
102 |
def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask):
|
103 |
+
"""
|
104 |
+
Args:
|
105 |
+
id_pixel_values (Tensor): Images [batch, num_inputs, channels, height, width]
|
106 |
+
prompt_embeds (Tensor): Text embeddings [batch, seq_len, embed_dim]
|
107 |
+
class_tokens_mask (Tensor): Mask of class tokens to update
|
108 |
+
|
109 |
+
Returns:
|
110 |
+
Tensor: Updated text embeddings incorporating ID image features.
|
111 |
+
"""
|
112 |
b, num_inputs, c, h, w = id_pixel_values.shape
|
113 |
id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
|
114 |
|
115 |
+
vision_outputs = self.vision_model(id_pixel_values)
|
116 |
+
shared_id_embeds = vision_outputs[1] # Use pooled output
|
117 |
+
|
118 |
id_embeds = self.visual_projection(shared_id_embeds)
|
119 |
id_embeds_2 = self.visual_projection_2(shared_id_embeds)
|
120 |
|
121 |
id_embeds = id_embeds.view(b, num_inputs, 1, -1)
|
122 |
+
id_embeds_2 = id_embeds_2.view(b, num_inputs, 1, -1)
|
123 |
|
124 |
+
combined_id_embeds = torch.cat((id_embeds, id_embeds_2), dim=-1)
|
|
|
125 |
|
126 |
+
updated_prompt_embeds = self.fuse_module(prompt_embeds, combined_id_embeds, class_tokens_mask)
|
127 |
|
128 |
+
return updated_prompt_embeds
|
129 |
|
130 |
if __name__ == "__main__":
|
131 |
+
encoder = PhotoMakerIDEncoder()
|
132 |
+
print("PhotoMakerIDEncoder initialized successfully.")
|