Implementation of evaluate.py

code/confs/dataset/video.yaml

@@ -20,7 +20,7 @@ valid:
     batch_size: 1
     drop_last: False
     shuffle: False
-    worker: 2
+    worker: 1
 
     num_sample : -1
     pixel_per_batch: 2048 

code/evaluate.py

@@ -2,31 +2,38 @@ import torch
 from v2a_model import V2AModel
 from skimage.metrics import structural_similarity as ssim
 from skimage.metrics import peak_signal_noise_ratio as psnr
+import glob
+from lib.datasets import create_dataset
 
-checkpoint_path = 'path/to/your/checkpoint.pth'  # Replace with the actual path
-checkpoint = torch.load(checkpoint_path)
+@hydra.main(config_path="confs", config_name="base")
+def main(opt):
+    checkpoint_path = sorted(glob.glob("checkpoints/*.ckpt"))[-1]  # Replace with the actual path (if not specified uses the last checkpoint)
+    checkpoint = torch.load(checkpoint_path)
 
-model = V2AModel(opt)  
+    model = V2AModel(opt)  
 
-model.load_state_dict(checkpoint['model_state_dict'])
+    model.load_state_dict(checkpoint['model_state_dict'])
 
-model.eval()
+    model.eval()
 
-# Prepare your input data (replace this with your actual data preparation)
-input_data = torch.randn(1, 3, 256, 256)  # Replace with your data
+    # Prepare your input data (replace this with your actual data preparation)
+    input_data = create_dataset(opt.dataset.metainfo, opt.dataset.test)  # Replace with your data
 
-# Pass the input through the model to get the outputs
-with torch.no_grad():
-    output = model(input_data)
+    # Pass the input through the model to get the outputs
+    with torch.no_grad():
+        output = model(input_data)
 
-# Assuming 'rgb_values' and 'fg_rgb_values' are the RGB images
-predicted_rgb_image = output['rgb_values'].cpu().numpy()  # Convert to NumPy array
-target_rgb_image = output['fg_rgb_values'].cpu().numpy()  # Convert to NumPy array
+    # Assuming 'rgb_values' and 'fg_rgb_values' are the RGB images
+    predicted_rgb_image = output['rgb_values'].cpu().numpy()  # Convert to NumPy array
+    target_rgb_image = output['fg_rgb_values'].cpu().numpy()  # Convert to NumPy array
 
-# Calculate SSIM and PSNR
-ssim_value = ssim(target_rgb_image, predicted_rgb_image, multichannel=True)
-psnr_value = psnr(target_rgb_image, predicted_rgb_image)
+    # Calculate SSIM and PSNR
+    ssim_value = ssim(target_rgb_image, predicted_rgb_image, multichannel=True)
+    psnr_value = psnr(target_rgb_image, predicted_rgb_image)
 
-# Log or print the SSIM and PSNR values
-print(f'SSIM: {ssim_value:.4f}')
-print(f'PSNR: {psnr_value:.4f}')
+    # Log or print the SSIM and PSNR values
+    print(f'SSIM: {ssim_value:.4f}')
+    print(f'PSNR: {psnr_value:.4f}')
+
+if __name__ == '__main__':
+    main()

visualization/vis.py

@@ -45,9 +45,9 @@ def vis_dynamic_canonical_train(args):
     faces = []
     vertex_normals = []
     deformed_mesh_paths = sorted(glob.glob(f'{args.path}/*.ply'))
-    print(deformed_mesh_paths)
     for deformed_mesh_path in deformed_mesh_paths:
         mesh = trimesh.load(deformed_mesh_path, process=False)
+        print(mesh)
         # center the human
         mesh.vertices = mesh.vertices - mesh.vertices.mean(axis=0)
         vertices.append(mesh.vertices)
@@ -67,6 +67,7 @@ def vis_dynamic_canonical_train(args):
     viewer.scene.origin.enabled = False
     viewer.scene.floor.enabled = True
     viewer.run()
+    
 if __name__ == '__main__':
     parser = argparse.ArgumentParser(description='3D Visualization')
     # static canonical mesh or dynamic sequence