Update app.py

app.py

@@ -1,3 +1,4 @@
+# -*- coding: utf-8 -*-
 import spaces
 import gradio as gr
 import os
@@ -46,56 +47,67 @@ model = model.cuda()
 # 번역 모델 로드
 translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")
 
-@spaces.GPU(duration=300, gpu_type="h100")
+@spaces.GPU(duration=300, gpu_type="l40s")
 def infer(image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
-    # 한글 입력 확인 및 번역
-    if any('\u3131' <= char <= '\u318E' or '\uAC00' <= char <= '\uD7A3' for char in prompt):
-        translated = translator(prompt, max_length=512)
-        prompt = translated[0]['translation_text']
-    
-    resolution = (576, 1024)
-    save_fps = 8
-    seed_everything(seed)
-    transform = transforms.Compose([
-        transforms.Resize(min(resolution)),
-        transforms.CenterCrop(resolution),
+    try:
+        # 한글 입력 확인 및 번역
+        if any('\u3131' <= char <= '\u318E' or '\uAC00' <= char <= '\uD7A3' for char in prompt):
+            translated = translator(prompt, max_length=512)
+            prompt = translated[0]['translation_text']
+        
+        resolution = (576, 1024)
+        save_fps = 8
+        seed_everything(seed)
+        transform = transforms.Compose([
+            transforms.Resize(min(resolution), antialias=True),
+            transforms.CenterCrop(resolution),
         ])
-    torch.cuda.empty_cache()
-    print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
-    start = time.time()
-    if steps > 60:
-        steps = 60 
+        
+        print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
+        start = time.time()
+        if steps > 60:
+            steps = 60 
 
-    batch_size = 1
-    channels = model.model.diffusion_model.out_channels
-    frames = model.temporal_length
-    h, w = resolution[0] // 8, resolution[1] // 8
-    noise_shape = [batch_size, channels, frames, h, w]
+        batch_size = 1
+        channels = model.model.diffusion_model.out_channels
+        frames = model.temporal_length
+        h, w = resolution[0] // 8, resolution[1] // 8
+        noise_shape = [batch_size, channels, frames, h, w]
 
-    with torch.no_grad(), torch.cuda.amp.autocast():
-        text_emb = model.get_learned_conditioning([prompt])
+        with torch.no_grad(), torch.cuda.amp.autocast():
+            text_emb = model.get_learned_conditioning([prompt])
+            
+            img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device)
+            img_tensor = (img_tensor / 255. - 0.5) * 2
+            image_tensor_resized = transform(img_tensor)
+            videos = image_tensor_resized.unsqueeze(0)
+            
+            z = get_latent_z(model, videos.unsqueeze(2))
+            img_tensor_repeat = repeat(z, 'b c t h w -> b c (repeat t) h w', repeat=frames)
         
-        img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device)
-        img_tensor = (img_tensor / 255. - 0.5) * 2
-        image_tensor_resized = transform(img_tensor)
-        videos = image_tensor_resized.unsqueeze(0)
+            cond_images = model.embedder(img_tensor.unsqueeze(0))
+            img_emb = model.image_proj_model(cond_images)
         
-        z = get_latent_z(model, videos.unsqueeze(2))
-        img_tensor_repeat = repeat(z, 'b c t h w -> b c (repeat t) h w', repeat=frames)
-    
-        cond_images = model.embedder(img_tensor.unsqueeze(0))
-        img_emb = model.image_proj_model(cond_images)
-    
-        imtext_cond = torch.cat([text_emb, img_emb], dim=1)
-    
-        fs = torch.tensor([fs], dtype=torch.long, device=model.device)
-        cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat]}
+            imtext_cond = torch.cat([text_emb, img_emb], dim=1)
         
-        batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale)
-    
-        video_path = './output.mp4'
-        save_videos(batch_samples, './', filenames=['output'], fps=save_fps)
-    return video_path
+            fs = torch.tensor([fs], dtype=torch.long, device=model.device)
+            cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat]}
+            
+            batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale)
+        
+            video_path = './output.mp4'
+            save_videos(batch_samples, './', filenames=['output'], fps=save_fps)
+        
+        # 메모리 정리
+        del text_emb, img_tensor, image_tensor_resized, videos, z, img_tensor_repeat, cond_images, img_emb, imtext_cond, cond, batch_samples
+        torch.cuda.empty_cache()
+        
+        return video_path
+    except Exception as e:
+        print(f"Error occurred: {e}")
+        return None
+    finally:
+        torch.cuda.empty_cache()
 
 i2v_examples = [
     ['prompts/1024/astronaut04.png', '우주인 복장으로 기타를 치는 남자', 30, 7.5, 1.0, 6, 123],
@@ -113,7 +125,7 @@ with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface:
                     with gr.Row():
                         i2v_input_image = gr.Image(label="Input Image",elem_id="input_img")
                     with gr.Row():
-                        i2v_input_text = gr.Text(label='Prompts (한글 입력 가능)')
+                        i2v_input_text = gr.Textbox(label='Prompts (한글 입력 가능)')
                     with gr.Row():
                         i2v_seed = gr.Slider(label='Random Seed', minimum=0, maximum=10000, step=1, value=123)
                         i2v_eta = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, label='ETA', value=1.0, elem_id="i2v_eta")
@@ -129,12 +141,11 @@ with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface:
                         inputs=[i2v_input_image, i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed],
                         outputs=[i2v_output_video],
                         fn = infer,
-                        cache_examples=True,
+                        cache_examples=False  # 이 부분을 False로 설정하여 캐시를 비활성화
             )
         i2v_end_btn.click(inputs=[i2v_input_image, i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed],
                         outputs=[i2v_output_video],
                         fn = infer
         )
 
-dynamicrafter_iface.launch()
-#dynamicrafter_iface.launch(server_port=7890, server_name="0.0.0.0", share=True)
+dynamicrafter_iface.launch(server_port=7890, server_name="0.0.0.0", share=True)