Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -7,101 +7,7 @@ from huggingface_hub import hf_hub_download
|
|
7 |
import torch.nn as nn
|
8 |
|
9 |
class SpriteGenerator(nn.Module):
|
10 |
-
|
11 |
-
super(SpriteGenerator, self).__init__()
|
12 |
-
|
13 |
-
# Text encoder (T5 with lm_head)
|
14 |
-
self.text_encoder = T5ForConditionalGeneration.from_pretrained(text_encoder_name)
|
15 |
-
for param in self.text_encoder.parameters():
|
16 |
-
param.requires_grad = False
|
17 |
-
|
18 |
-
# Proiezione dal testo al latent space
|
19 |
-
self.text_projection = nn.Sequential(
|
20 |
-
nn.Linear(768, latent_dim),
|
21 |
-
nn.LeakyReLU(0.2),
|
22 |
-
nn.Linear(latent_dim, latent_dim)
|
23 |
-
)
|
24 |
-
|
25 |
-
# Generator
|
26 |
-
self.generator = nn.Sequential(
|
27 |
-
# Input: latent_dim x 1 x 1 -> 512 x 4 x 4
|
28 |
-
nn.ConvTranspose2d(latent_dim, 512, 4, 1, 0, bias=False),
|
29 |
-
nn.BatchNorm2d(512),
|
30 |
-
nn.ReLU(True),
|
31 |
-
|
32 |
-
# 512 x 4 x 4 -> 256 x 8 x 8
|
33 |
-
nn.ConvTranspose2d(512, 256, 4, 2, 1, bias=False),
|
34 |
-
nn.BatchNorm2d(256),
|
35 |
-
nn.ReLU(True),
|
36 |
-
|
37 |
-
# 256 x 8 x 8 -> 128 x 16 x 16
|
38 |
-
nn.ConvTranspose2d(256, 128, 4, 2, 1, bias=False),
|
39 |
-
nn.BatchNorm2d(128),
|
40 |
-
nn.ReLU(True),
|
41 |
-
|
42 |
-
# 128 x 16 x 16 -> 64 x 32 x 32
|
43 |
-
nn.ConvTranspose2d(128, 64, 4, 2, 1, bias=False),
|
44 |
-
nn.BatchNorm2d(64),
|
45 |
-
nn.ReLU(True),
|
46 |
-
|
47 |
-
# 64 x 32 x 32 -> 32 x 64 x 64
|
48 |
-
nn.ConvTranspose2d(64, 32, 4, 2, 1, bias=False),
|
49 |
-
nn.BatchNorm2d(32),
|
50 |
-
nn.ReLU(True),
|
51 |
-
|
52 |
-
# 32 x 64 x 64 -> 16 x 128 x 128
|
53 |
-
nn.ConvTranspose2d(32, 16, 4, 2, 1, bias=False),
|
54 |
-
nn.BatchNorm2d(16),
|
55 |
-
nn.ReLU(True),
|
56 |
-
|
57 |
-
# 16 x 128 x 128 -> 3 x 256 x 256
|
58 |
-
nn.ConvTranspose2d(16, 3, 4, 2, 1, bias=False),
|
59 |
-
)
|
60 |
-
|
61 |
-
# Frame interpolator
|
62 |
-
self.frame_interpolator = nn.Sequential(
|
63 |
-
nn.Linear(latent_dim + 1, latent_dim),
|
64 |
-
nn.LeakyReLU(0.2),
|
65 |
-
nn.Linear(latent_dim, latent_dim),
|
66 |
-
nn.LeakyReLU(0.2)
|
67 |
-
)
|
68 |
-
|
69 |
-
def forward(self, input_ids, attention_mask, num_frames=1):
|
70 |
-
batch_size = input_ids.shape[0]
|
71 |
-
|
72 |
-
# Encode text usando il T5 completo
|
73 |
-
text_outputs = self.text_encoder.encoder(
|
74 |
-
input_ids=input_ids,
|
75 |
-
attention_mask=attention_mask,
|
76 |
-
return_dict=True
|
77 |
-
)
|
78 |
-
|
79 |
-
# Get text features
|
80 |
-
text_features = text_outputs.last_hidden_state.mean(dim=1)
|
81 |
-
|
82 |
-
# Project to latent space
|
83 |
-
latent_vector = self.text_projection(text_features)
|
84 |
-
|
85 |
-
# Generate multiple frames if needed
|
86 |
-
all_frames = []
|
87 |
-
for frame_idx in range(max(num_frames.max().item(), 1)):
|
88 |
-
frame_info = torch.ones((batch_size, 1), device=latent_vector.device) * frame_idx / max(num_frames.max().item(), 1)
|
89 |
-
|
90 |
-
# Combine latent vector with frame info
|
91 |
-
frame_latent = self.frame_interpolator(
|
92 |
-
torch.cat([latent_vector, frame_info], dim=1)
|
93 |
-
)
|
94 |
-
|
95 |
-
# Generate frame
|
96 |
-
frame_latent_reshaped = frame_latent.unsqueeze(2).unsqueeze(3)
|
97 |
-
frame = self.generator(frame_latent_reshaped)
|
98 |
-
frame = torch.tanh(frame)
|
99 |
-
all_frames.append(frame)
|
100 |
-
|
101 |
-
# Stack all frames
|
102 |
-
sprites = torch.stack(all_frames, dim=1)
|
103 |
-
|
104 |
-
return sprites
|
105 |
|
106 |
def initialize_model():
|
107 |
print("Inizializzazione del modello...")
|
@@ -110,12 +16,19 @@ def initialize_model():
|
|
110 |
model = SpriteGenerator()
|
111 |
|
112 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
# Carica il modello
|
114 |
-
state_dict = torch.load(
|
115 |
model.load_state_dict(state_dict)
|
116 |
model = model.to(device)
|
117 |
model.eval()
|
118 |
-
print("Modello caricato con successo!")
|
119 |
return model, device
|
120 |
except Exception as e:
|
121 |
print(f"Errore nel caricamento del modello: {str(e)}")
|
|
|
7 |
import torch.nn as nn
|
8 |
|
9 |
class SpriteGenerator(nn.Module):
|
10 |
+
# ... (la classe SpriteGenerator rimane invariata) ...
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
def initialize_model():
|
13 |
print("Inizializzazione del modello...")
|
|
|
16 |
model = SpriteGenerator()
|
17 |
|
18 |
try:
|
19 |
+
# Scarica il modello da Hugging Face Hub
|
20 |
+
model_path = hf_hub_download(
|
21 |
+
repo_id="Lod34/Animator2D-v2",
|
22 |
+
filename="pytorch_model.bin",
|
23 |
+
repo_type="model"
|
24 |
+
)
|
25 |
+
|
26 |
# Carica il modello
|
27 |
+
state_dict = torch.load(model_path, map_location=device)
|
28 |
model.load_state_dict(state_dict)
|
29 |
model = model.to(device)
|
30 |
model.eval()
|
31 |
+
print("Modello caricato con successo da Hugging Face Hub!")
|
32 |
return model, device
|
33 |
except Exception as e:
|
34 |
print(f"Errore nel caricamento del modello: {str(e)}")
|