戻す

app.py

@@ -1,69 +1,34 @@
-import torch
-import numpy as np
 import gradio as gr
-from lavis.models import load_model_and_preprocess
+from transformers import CLIPSegProcessor, CLIPSegForImageSegmentation
+import torch
 from PIL import Image
-import matplotlib.pyplot as plt
-import matplotlib.cm as cm
-import torchvision
-
-def create_heatmap(activation_map):
-    # アクティベーションマップをnumpy配列に変換
-    activation_map_np = activation_map.squeeze().detach().cpu().numpy()
-
-    # アクティベーションマップの最小値と最大値を取得
-    min_value = np.min(activation_map_np)
-    max_value = np.max(activation_map_np)
-
-    # アクティベーションマップを0-1の範囲に正規化
-    normalized_map = (activation_map_np - min_value) / (max_value - min_value)
-
-    # 正規化されたアクティベーションマップをヒートマップに変換
-    heatmap = cm.jet(normalized_map)
-
-    # ヒートマップを [0, 255] の範囲にスケーリングし、uint8型に変換
-    heatmap = np.uint8(255 * heatmap)
-
-    return heatmap
+import numpy as np
 
 def process(input_image, prompt):
-    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-
-    model, vis_processors, txt_processors = load_model_and_preprocess(name="pnp_vqa", model_type="base", is_eval=True, device=device)
-    input_image = input_image.resize((256, 256))
-    image = vis_processors["eval"](input_image).unsqueeze(0).to(device)
-    text_input = txt_processors["eval"](prompt)
-    sample = {"image": image, "text_input": [text_input]}
-
-    output = model.forward_itm(samples=sample)
-    activation_map = output['gradcams'].reshape(24, 24)
-
-    relu = torch.nn.ReLU()
-    # ヒートマップを計算
-    heatmap = create_heatmap(activation_map)
-
-    heatmap = Image.fromarray(heatmap)
-    heatmap = torchvision.transforms.functional.to_tensor(heatmap)
-    heatmap = relu(heatmap)
-    heatmap = torchvision.transforms.functional.to_pil_image(heatmap)
-    heatmap = heatmap.resize((256, 256))
-    heatmap = np.array(heatmap)
-    heatmap = torch.sigmoid(torch.from_numpy(heatmap)).numpy()
-    preds = heatmap.reshape(256, 256, -1)
-    preds = Image.fromarray(preds.astype(np.uint8)).convert('L')
-    preds = np.array(preds)
-    preds = np.where(preds > 0.5, 255, 0)
-    
+    inputs = processor(text=prompt, images=input_image, padding="max_length", return_tensors="pt")
+    # predict
+    with torch.no_grad():
+        outputs = model(**inputs)
+    preds = torch.sigmoid(outputs.logits).squeeze().detach().cpu().numpy()
+    preds = np.where(preds > 0.3, 255, 0).astype(np.uint8)
+    preds = Image.fromarray(preds.astype(np.uint8))
+    preds = np.array(preds.resize((input_image.width, input_image.height)))
     return preds
-
+    
 if __name__ == '__main__':
+    processor = CLIPSegProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
+    model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
     input_image = gr.inputs.Image(label='image', type='pil')
     prompt = gr.Textbox(label='Prompt')
     ips = [
             input_image, prompt
         ]
     outputs = "image"
+    input_size = (256, 256)
+    output_size = (256, 256)
     iface = gr.Interface(fn=process,
                          inputs=ips,
-                         outputs=outputs)
+                         outputs=outputs,
+                         input_size=input_size, 
+                         output_size=output_size)
     iface.launch()