Update app.py
Browse files
app.py
CHANGED
@@ -1,13 +1,14 @@
|
|
|
|
1 |
import torch
|
2 |
import torch.nn as nn
|
3 |
from torchvision import transforms
|
4 |
from PIL import Image
|
5 |
from transformers import BertTokenizer, BertModel
|
6 |
-
import gradio as gr
|
7 |
import numpy as np
|
8 |
import os
|
9 |
import time
|
10 |
|
|
|
11 |
from train import CVAE, TextEncoder, LATENT_DIM, HIDDEN_DIM
|
12 |
|
13 |
# Initialize the BERT tokenizer
|
@@ -17,90 +18,81 @@ def clean_image(image, threshold=0.75):
|
|
17 |
np_image = np.array(image)
|
18 |
alpha_channel = np_image[:, :, 3]
|
19 |
alpha_channel[alpha_channel <= int(threshold * 255)] = 0
|
20 |
-
alpha_channel[alpha_channel > int(threshold * 255)] = 255
|
21 |
return Image.fromarray(np_image)
|
22 |
|
23 |
def generate_image(model, text_prompt, device, input_image=None, img_control=0.5):
|
24 |
encoded_input = tokenizer(text_prompt, padding=True, truncation=True, return_tensors="pt")
|
25 |
input_ids = encoded_input['input_ids'].to(device)
|
26 |
attention_mask = encoded_input['attention_mask'].to(device)
|
27 |
-
|
28 |
with torch.no_grad():
|
29 |
text_encoding = model.text_encoder(input_ids, attention_mask)
|
30 |
-
|
31 |
-
z = torch.randn(1, LATENT_DIM).to(device)
|
32 |
-
|
33 |
-
with torch.no_grad():
|
34 |
generated_image = model.decode(z, text_encoding)
|
35 |
-
|
36 |
if input_image is not None:
|
37 |
input_image = input_image.convert("RGBA").resize((16, 16), resample=Image.NEAREST)
|
38 |
input_image = transforms.ToTensor()(input_image).unsqueeze(0).to(device)
|
39 |
generated_image = img_control * input_image + (1 - img_control) * generated_image
|
40 |
-
|
41 |
generated_image = generated_image.squeeze(0).cpu()
|
42 |
-
generated_image = (generated_image + 1) / 2
|
43 |
generated_image = generated_image.clamp(0, 1)
|
44 |
generated_image = transforms.ToPILImage()(generated_image)
|
45 |
-
|
46 |
return generated_image
|
47 |
|
48 |
-
def
|
49 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
50 |
-
|
51 |
-
# Load input image if provided
|
52 |
-
input_image = Image.open(input_image).convert("RGBA") if input_image else None
|
53 |
-
|
54 |
-
# Initialize model
|
55 |
text_encoder = TextEncoder(hidden_size=HIDDEN_DIM, output_size=HIDDEN_DIM)
|
56 |
model = CVAE(text_encoder).to(device)
|
57 |
-
|
58 |
-
# Load the trained model
|
59 |
model.load_state_dict(torch.load(model_path, map_location=device))
|
60 |
model.eval()
|
|
|
|
|
|
|
|
|
|
|
61 |
|
62 |
start_time = time.time()
|
63 |
-
|
64 |
-
# Generate image from prompt
|
65 |
generated_image = generate_image(model, prompt, device, input_image, img_control)
|
66 |
-
|
67 |
end_time = time.time()
|
68 |
generation_time = end_time - start_time
|
69 |
-
|
70 |
-
|
71 |
-
if clean:
|
72 |
generated_image = clean_image(generated_image)
|
73 |
-
|
74 |
-
# Resize the generated image
|
75 |
generated_image = generated_image.resize((size, size), resample=Image.NEAREST)
|
76 |
-
|
77 |
-
# Save the generated image to the specified directory
|
78 |
-
model_name = os.path.splitext(os.path.basename(model_path))[0]
|
79 |
-
output_file = os.path.join(output_dir, f"{model_name}_{prompt}.png")
|
80 |
-
os.makedirs(output_dir, exist_ok=True)
|
81 |
-
generated_image.save(output_file)
|
82 |
-
|
83 |
-
print(f"Generated image saved as {output_file}")
|
84 |
-
print(f"Generation time: {generation_time:.10f} seconds")
|
85 |
-
|
86 |
-
return generated_image
|
87 |
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
gr.
|
94 |
-
|
95 |
-
gr.
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
104 |
|
105 |
if __name__ == "__main__":
|
106 |
-
|
|
|
|
1 |
+
import gradio as gr
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
from torchvision import transforms
|
5 |
from PIL import Image
|
6 |
from transformers import BertTokenizer, BertModel
|
|
|
7 |
import numpy as np
|
8 |
import os
|
9 |
import time
|
10 |
|
11 |
+
# Import the model architecture from train.py
|
12 |
from train import CVAE, TextEncoder, LATENT_DIM, HIDDEN_DIM
|
13 |
|
14 |
# Initialize the BERT tokenizer
|
|
|
18 |
np_image = np.array(image)
|
19 |
alpha_channel = np_image[:, :, 3]
|
20 |
alpha_channel[alpha_channel <= int(threshold * 255)] = 0
|
21 |
+
alpha_channel[alpha_channel > int(threshold * 255)] = 255
|
22 |
return Image.fromarray(np_image)
|
23 |
|
24 |
def generate_image(model, text_prompt, device, input_image=None, img_control=0.5):
|
25 |
encoded_input = tokenizer(text_prompt, padding=True, truncation=True, return_tensors="pt")
|
26 |
input_ids = encoded_input['input_ids'].to(device)
|
27 |
attention_mask = encoded_input['attention_mask'].to(device)
|
28 |
+
|
29 |
with torch.no_grad():
|
30 |
text_encoding = model.text_encoder(input_ids, attention_mask)
|
31 |
+
z = torch.randn(1, LATENT_DIM).to(device)
|
|
|
|
|
|
|
32 |
generated_image = model.decode(z, text_encoding)
|
33 |
+
|
34 |
if input_image is not None:
|
35 |
input_image = input_image.convert("RGBA").resize((16, 16), resample=Image.NEAREST)
|
36 |
input_image = transforms.ToTensor()(input_image).unsqueeze(0).to(device)
|
37 |
generated_image = img_control * input_image + (1 - img_control) * generated_image
|
38 |
+
|
39 |
generated_image = generated_image.squeeze(0).cpu()
|
40 |
+
generated_image = (generated_image + 1) / 2
|
41 |
generated_image = generated_image.clamp(0, 1)
|
42 |
generated_image = transforms.ToPILImage()(generated_image)
|
43 |
+
|
44 |
return generated_image
|
45 |
|
46 |
+
def load_model(model_path, device):
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
text_encoder = TextEncoder(hidden_size=HIDDEN_DIM, output_size=HIDDEN_DIM)
|
48 |
model = CVAE(text_encoder).to(device)
|
|
|
|
|
49 |
model.load_state_dict(torch.load(model_path, map_location=device))
|
50 |
model.eval()
|
51 |
+
return model
|
52 |
+
|
53 |
+
def generate_image_gradio(prompt, model_path, clean_image_flag, size, input_image=None, img_control=0.5):
|
54 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
55 |
+
model = load_model(model_path, device)
|
56 |
|
57 |
start_time = time.time()
|
|
|
|
|
58 |
generated_image = generate_image(model, prompt, device, input_image, img_control)
|
|
|
59 |
end_time = time.time()
|
60 |
generation_time = end_time - start_time
|
61 |
+
|
62 |
+
if clean_image_flag:
|
|
|
63 |
generated_image = clean_image(generated_image)
|
64 |
+
|
|
|
65 |
generated_image = generated_image.resize((size, size), resample=Image.NEAREST)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
|
67 |
+
return generated_image, f"Generation time: {generation_time:.4f} seconds"
|
68 |
+
|
69 |
+
# Gradio interface
|
70 |
+
def gradio_interface():
|
71 |
+
with gr.Blocks() as demo:
|
72 |
+
gr.Markdown("# Image Generator from Text Prompt")
|
73 |
+
|
74 |
+
with gr.Row():
|
75 |
+
with gr.Column():
|
76 |
+
prompt = gr.Textbox(label="Text Prompt")
|
77 |
+
model_path = gr.Textbox(label="Model Path", value="path/to/your/model.pth")
|
78 |
+
clean_image_flag = gr.Checkbox(label="Clean Image", value=False)
|
79 |
+
size = gr.Slider(minimum=16, maximum=512, step=16, label="Image Size", value=16)
|
80 |
+
img_control = gr.Slider(minimum=0, maximum=1, step=0.1, label="Image Control", value=0.5)
|
81 |
+
input_image = gr.Image(label="Input Image (optional)", type="pil")
|
82 |
+
generate_button = gr.Button("Generate Image")
|
83 |
+
|
84 |
+
with gr.Column():
|
85 |
+
output_image = gr.Image(label="Generated Image")
|
86 |
+
generation_time = gr.Textbox(label="Generation Time")
|
87 |
+
|
88 |
+
generate_button.click(
|
89 |
+
generate_image_gradio,
|
90 |
+
inputs=[prompt, model_path, clean_image_flag, size, input_image, img_control],
|
91 |
+
outputs=[output_image, generation_time]
|
92 |
+
)
|
93 |
+
|
94 |
+
return demo
|
95 |
|
96 |
if __name__ == "__main__":
|
97 |
+
demo = gradio_interface()
|
98 |
+
demo.launch()
|