vto conversion, some progress, not working properl

configs/direct_3.py

@@ -1,78 +0,0 @@
-from utils.convnext_utils import CNBlockConfig
-
-config = dict()
-
-config['gpu'] = '4'
-config['name'] = 'direct_3'
-config['notes'] = 'with curvature input and skip connection' 
-config['tags'] = ['torch', 'direct'] 
-config['script'] = 'scripts/direct_train.py'
-
-config['pretrained_model'] = ''
-config['use_curvature'] = True
-
-config['batch_size'] = 4
-config['train_dir'] = '/data-local/127/train'
-config['valid_dir'] = '/data-local/127/valid'
-config['test_dir'] = '/data-net/dataset/test_ready'
-config['train_frames'] = None
-config['valid_frames'] = None
-
-config['resolution'] = 512
-config['is_skirt'] = False
-config['dataset_in_memory'] = False
-config['is_cape'] = False
-config['remove_global_orientation'] = True
-config['device'] = 'cuda'
-config['input_features'] = 10
-
-config['smpl_res'] = 14475
-if config['is_cape']:
-  config['smpl_res'] = 3638
-
-config['latent_dim'] = 128
-
-config['bottleneck_scale'] = 2
-
-config['is_label'] = False
-config['kernel_size'] = 7
-config['bottleneck_size'] = 64 * config['bottleneck_scale']
-
-block_setting = [
-    CNBlockConfig(16, 32, 3),
-    CNBlockConfig(32, 128, 3),
-    CNBlockConfig(128, 256, 9),
-    CNBlockConfig(256, config['bottleneck_size'], 9),
-    CNBlockConfig(config['bottleneck_size'], None, 3),
-]
-
-block_setting_reverse = [
-    CNBlockConfig(config['bottleneck_size'], 256, 3),
-    CNBlockConfig(256*2, 128, 9),
-    CNBlockConfig(128*2, 32, 9),
-    CNBlockConfig(32*2, 16, 3),
-    CNBlockConfig(16*2, 8, 3),
-    CNBlockConfig(8, 3, 3),
-]
-
-config['encoder_block_settings'] = block_setting
-config['decoder_block_settings'] = block_setting_reverse
-config['activation'] = 'tanh'
-
-
-config['lr'] = 0.0001
-config['reg_loss_coeff'] = 0.1
-
-config['signature'] = { 'id',
-                        'frame_nr',
-                        'pose_shape',
-                        'uv_curvature',
-                        'mesh_mask',
-                        'mesh_body_posed',
-                        'mesh_posed',
-                        'uv_unposed',
-                        'mesh_unposed',
-                        'uv_static',
-                        'uv_mask',
-                        'uv_body_posed',
-                        }

data/convert_vto.py

@@ -0,0 +1,259 @@
+import numpy as np
+import os
+import sys
+import pickle
+import plotly.express as ex
+import plotly.graph_objects as go
+import torch
+import open3d as o3d
+import trimesh
+# from vedo import trimesh as trimesh_plot
+from scipy.spatial import KDTree
+from scipy.sparse import lil_matrix
+from scipy.optimize import minimize
+
+sys.path.append(os.path.dirname(os.path.abspath(__file__)) + '/..')
+
+from smpl.smpl_torch import SMPL
+
+# import mask faces
+from custom_utils.custom_utils import _maskFaces
+
+smpl = SMPL(path='smpl')
+
+def energy_function(params, source_points, target_points, indices):
+    num_points = source_points.shape[0]
+    num_dims = source_points.shape[1]
+    
+    # Reshape the parameters
+    weights = params[:num_points * num_dims].reshape((num_points, num_dims))
+    translation = params[num_points * num_dims:num_points * num_dims + num_dims]
+    control_points = params[num_points * num_dims + num_dims:].reshape((-1, num_dims))
+    
+    # Compute the transformed source points
+    transformed_points = source_points + np.dot(weights, control_points.T)
+    
+    # Compute the TPS energy
+    tps_energy = 0.5 * np.sum(np.square(target_points[indices] - transformed_points))
+    
+    return tps_energy
+
+def non_rigid(source_points, target_points):
+    tree = KDTree(target_points)
+    distances, indices = tree.query(source_points)
+
+    # Initialize the optimization parameters
+    num_points = source_points.shape[0]
+    num_dims = source_points.shape[1]
+    initial_params = np.zeros(num_points * num_dims + num_dims + num_points)
+
+    # Define the optimization bounds
+    bounds = [(None, None)] * (num_points * num_dims)
+    bounds += [(None, None)] * num_dims
+    bounds += [(None, None)] * (num_points * num_dims)
+
+    # Perform the optimization
+    result = minimize(energy_function, initial_params, method='L-BFGS-B',
+                    args=(source_points, target_points, indices), bounds=bounds)
+
+    # Get the optimized parameters
+    optimized_params = result.x
+
+    # Reshape the parameters
+    optimized_weights = optimized_params[:num_points * num_dims].reshape((num_points, num_dims))
+    optimized_translation = optimized_params[num_points * num_dims:num_points * num_dims + num_dims]
+    optimized_control_points = optimized_params[num_points * num_dims + num_dims:].reshape((-1, num_dims))
+
+    transformed_points = source_points + np.dot(optimized_weights, optimized_control_points.T)
+    return transformed_points
+
+
+def register_smpl():
+    # generate body with zero pose and shape
+    pose = torch.zeros((1, 72))
+    shape = torch.zeros((1, 11))
+    smpl_out = smpl(pose, shape)[0][0]
+
+    with open('/data-net/code/vto-dataset/tshirt/simulations/tshirt_shape00_55_27.pkl', "rb") as f:
+        data = pickle.load(f)
+        garment = data['vertices'][100]
+        faces = data['faces']
+
+    # rotate garment x degrees along y axis
+    # convert degrees to radians
+    degrees = 107
+    theta = degrees * np.pi / 180
+    rot = np.array([[np.cos(theta), 0, np.sin(theta)],
+                    [0, 1, 0],
+                    [-np.sin(theta), 0, np.cos(theta)]])
+    garment = np.dot(rot, garment.T).T
+
+    # rotate garment x degrees along x axis
+    # convert degrees to radians
+    degrees = 5
+    theta = degrees * np.pi / 180
+    rot = np.array([[1, 0, 0],
+                    [0, np.cos(theta), -np.sin(theta)],
+                    [0, np.sin(theta), np.cos(theta)]])
+    garment = np.dot(rot, garment.T).T
+
+    # rotate garment x degrees along z axis
+    # convert degrees to radians
+    degrees = -5
+    theta = degrees * np.pi / 180
+    rot = np.array([[np.cos(theta), -np.sin(theta), 0],
+                    [np.sin(theta), np.cos(theta), 0],
+                    [0, 0, 1]])
+    garment = np.dot(rot, garment.T).T
+
+
+    garment[:,0] -= 0.06
+    garment[:,1] -= 0.03
+    garment[:,2] -= 0.015
+    
+    # visualize the body and garment
+    # fig = go.Figure(data=[go.Mesh3d(x=smpl_out[:,0], y=smpl_out[:,1], z=smpl_out[:,2], i=smpl.f[:,0], j=smpl.f[:,1], k=smpl.f[:,2])])
+    # fig.add_trace(go.Mesh3d(x=garment[:,0], y=garment[:,1], z=garment[:,2], i=faces[:,0], j=faces[:,1], k=faces[:,2]))
+    # fig.update_layout(scene=dict(xaxis=dict(range=[-1,1],), yaxis=dict(range=[-1,1],), zaxis=dict(range=[-1,1],),),)
+    # fig.show()
+
+    # non_rigid(garment, smpl_out.numpy())
+
+    # Load the source and target meshes
+    source_mesh = o3d.geometry.TriangleMesh(o3d.utility.Vector3dVector(garment), o3d.utility.Vector3iVector(faces))
+    target_mesh = o3d.geometry.TriangleMesh(o3d.utility.Vector3dVector(smpl_out), o3d.utility.Vector3iVector(smpl.f))
+    # flip source and target mesh
+    source_mesh, target_mesh = target_mesh, source_mesh
+
+    source_mesh.compute_vertex_normals()
+    source_mesh = source_mesh.filter_smooth_laplacian(number_of_iterations=10)
+    # plot the smoothed mesh with plotly
+    # fig = go.Figure(data=[go.Mesh3d(x=np.asarray(source_mesh.vertices)[:,0], y=np.asarray(source_mesh.vertices)[:,1], z=np.asarray(source_mesh.vertices)[:,2], i=np.asarray(source_mesh.triangles)[:,0], j=np.asarray(source_mesh.triangles)[:,1], k=np.asarray(source_mesh.triangles)[:,2])])
+    # fig.show()
+
+    # Perform initial alignment (translation and/or scaling)
+    # You can modify this step based on your specific needs
+    # For example, you can manually set an initial transformation
+    init_transformation = [[1.0, 0.0, 0.0, 0.0],
+                        [0.0, 1.0, 0.0, 0.0],
+                        [0.0, 0.0, 1.0, 0.0],
+                        [0.0, 0.0, 0.0, 1.0]]
+
+    # Create the ICP object and set the parameters
+    icp = o3d.pipelines.registration.TransformationEstimationPointToPoint()
+    criteria = o3d.pipelines.registration.ICPConvergenceCriteria(max_iteration=2000)
+        
+    # Perform ICP registration
+    reg_result = o3d.pipelines.registration.registration_icp(o3d.geometry.PointCloud(source_mesh.vertices),
+                                                             o3d.geometry.PointCloud(target_mesh.vertices),
+                                                             0.075,
+                                                             init_transformation, icp, criteria)
+
+    # Get the aligned source mesh after registration
+    aligned_mesh = source_mesh.transform(reg_result.transformation)
+
+    # Visualize the aligned mesh and target mesh with plotly
+    fig = go.Figure(data=[go.Mesh3d(x=np.asarray(aligned_mesh.vertices)[:,0], y=np.asarray(aligned_mesh.vertices)[:,1], z=np.asarray(aligned_mesh.vertices)[:,2], i=np.asarray(aligned_mesh.triangles)[:,0], j=np.asarray(aligned_mesh.triangles)[:,1], k=np.asarray(aligned_mesh.triangles)[:,2])])
+    fig.add_trace(go.Mesh3d(x=np.asarray(target_mesh.vertices)[:,0], y=np.asarray(target_mesh.vertices)[:,1], z=np.asarray(target_mesh.vertices)[:,2], i=np.asarray(target_mesh.triangles)[:,0], j=np.asarray(target_mesh.triangles)[:,1], k=np.asarray(target_mesh.triangles)[:,2]))
+    fig.update_layout(scene=dict(xaxis=dict(range=[-1,1],), yaxis=dict(range=[-1,1],), zaxis=dict(range=[-1,1],),),)
+    fig.show()
+
+    # Get the vertices of the aligned source mesh and target mesh
+    aligned_source_vertices = np.asarray(aligned_mesh.vertices)
+    target_vertices = np.asarray(target_mesh.vertices)
+
+    # Perform nearest neighbor search to find corresponding vertex indices
+    kd_tree = o3d.geometry.KDTreeFlann(target_mesh)
+    correspondences = []
+    matched = set()
+    for i in range(len(aligned_source_vertices)):
+        k = 1
+        _, idx, _ = kd_tree.search_knn_vector_3d(aligned_source_vertices[i], k)
+        # while idx[-1] in matched:
+        #     k += 1
+        #     _, idx, _ = kd_tree.search_knn_vector_3d(aligned_source_vertices[i], k)
+        # matched.add(idx[-1])
+        correspondences.append((i, idx[-1]))
+
+    # correspondences is a list of tuples where each tuple contains the corresponding vertex indices
+    # For example, correspondences = [(0, 10), (1, 5), (2, 15), ...]
+
+    # Visualize the correspondences with plotly
+    # fig = go.Figure(data=[go.Scatter3d(x=aligned_source_vertices[:,0], y=aligned_source_vertices[:,1], z=aligned_source_vertices[:,2], mode='markers', marker=dict(color='red', size=1))])
+    # fig.add_trace(go.Scatter3d(x=target_vertices[:,0], y=target_vertices[:,1], z=target_vertices[:,2], mode='markers', marker=dict(color='blue', size=1)))
+    # for corr in correspondences:
+    #     fig.add_trace(go.Scatter3d(x=[aligned_source_vertices[corr[0],0], target_vertices[corr[1],0]], y=[aligned_source_vertices[corr[0],1], target_vertices[corr[1],1]], z=[aligned_source_vertices[corr[0],2], target_vertices[corr[1],2]], mode='lines', line=dict(color='green', width=1)))
+    # fig.update_layout(scene=dict(xaxis=dict(range=[-1,1],), yaxis=dict(range=[-1,1],), zaxis=dict(range=[-1,1],),),)
+    # fig.show()
+
+    registered = np.zeros((14475, 3))
+    mask = np.zeros((14475), dtype=bool)
+    for i in range(len(correspondences)):
+        registered[correspondences[i][1]] = aligned_source_vertices[correspondences[i][0]]
+        mask[correspondences[i][1]] = True
+
+    # F = np.load(f'smpl/smpl_faces.npy').astype(np.int32)
+
+    # registered_masked, masked_faces = _maskFaces(registered, np.asarray(target_mesh.triangles), mask)
+    registered_masked, masked_faces = registered, np.asarray(target_mesh.triangles)
+
+    # plot aligned source mesh with the target topology
+    fig = go.Figure(data=[go.Mesh3d(x=registered_masked[:,0], y=registered_masked[:,1], z=registered_masked[:,2], i=masked_faces[:,0], j=masked_faces[:,1], k=masked_faces[:,2])])
+    fig.update_layout(scene=dict(xaxis=dict(range=[-1,1],), yaxis=dict(range=[-1,1],), zaxis=dict(range=[-1,1],),),)
+    fig.show()
+
+    return smpl
+
+
+def convert_vto(folder, out_folder):
+    os.makedirs(out_folder, exist_ok=True)
+    os.makedirs(os.path.join(out_folder, 'mesh_posed_no_orient'), exist_ok=True)
+    os.makedirs(os.path.join(out_folder, 'mesh_posed_body_no_orient'), exist_ok=True)
+    os.makedirs(os.path.join(out_folder, 'mesh_mask'), exist_ok=True)
+    os.makedirs(os.path.join(out_folder, 'pose_shape'), exist_ok=True)
+    F = np.load(f'smpl/smpl_faces.npy').astype(np.int32)
+
+    files = os.listdir(folder)
+    # files = ['/data-net/code/vto-dataset/tshirt/simulations/tshirt_shape00_55_27.pkl']
+    for file in files:
+        # read pickle files and get vetices and faces from them
+        with open(os.path.join(folder, file), "rb") as f:
+            data = pickle.load(f)
+            vertices = data['vertices']
+            faces = data['faces']
+            pose = data['pose']
+            shape = data['shape']
+            trans = data['translation']
+            id = file.split('.')[0]
+            # pose[:, :3] = 0
+            id = id.replace('_', '-')
+            subject = id.split('-')[2]
+            sequence = '_'.join(id.split('-')[2:])
+            # cmu_data = np.load(f'/data-net/code/vto-garment-collisions/assets/CMU/{subject}/{sequence}_poses.npz')
+            # shape = np.tile(cmu_data['betas'], (shape.shape[0], 1))
+            # pose = cmu_data['poses']
+
+            body = smpl(torch.tensor(pose), torch.tensor(np.concatenate((shape, np.zeros((shape.shape[0],1),dtype=np.float32)),axis=1)))[0].numpy()
+            pose_shape = np.concatenate((pose, shape, np.zeros((shape.shape[0],1),dtype=np.float32)),axis=1)
+            
+            # for i in range(vertices.shape[0]):
+            #     np.savez_compressed(os.path.join(out_folder, 'mesh_posed_body_no_orient', f'{id}_{i}.npz'), mesh_posed_body_no_orient=body[i])
+            #     np.save(os.path.join(out_folder, 'pose_shape', f'{id}_{i}.npy'), pose_shape[i])
+
+
+            # plot mesh with plotly
+            frame = 0
+            fig = go.Figure(data=[go.Mesh3d(x=vertices[frame,:,0], y=vertices[frame,:,1], z=vertices[frame,:,2], i=faces[:,0], j=faces[:,1], k=faces[:,2])])
+            fig.add_trace(go.Mesh3d(x=body[frame,:,0], y=body[frame,:,1], z=body[frame,:,2], i=F[:,0], j=F[:,1], k=F[:,2]))
+            fig.update_layout(scene=dict(xaxis=dict(range=[-1,1],), yaxis=dict(range=[-1,1],), zaxis=dict(range=[-1,1],),),)
+            fig.show()
+
+            
+            pass
+    pass
+
+
+if __name__ == '__main__':
+    # register_smpl()
+    convert_vto('/data-net/code/vto-dataset/dress/simulations', '/data-net/code/vto-dataset/dress_converted')
+    pass

data/dataset_vto.py

@@ -0,0 +1,40 @@
+import os
+import sys
+
+import argparse
+parser = argparse.ArgumentParser(description='')
+parser.add_argument('--debug', action='store_true')
+args = vars(parser.parse_args())
+
+if args['debug']:
+    os.environ["CUDA_VISIBLE_DEVICES"] = '0'
+
+import math
+import time
+import torch
+import random
+import pytorch3d
+import torchvision as tv
+import numpy as np
+import scipy.io as sio
+import plotly.express as ex
+from torch.utils.data import Dataset
+
+sys.path.append(os.path.dirname(os.path.abspath(__file__)) + '/..')
+
+from data.utils import Frame
+
+class Dataset(Dataset):
+    def __init__(self, folder):
+        pass
+
+    def __len__(self):
+        return len(self.frames)
+
+
+    def __getitem__(self, idx):
+        frame : Frame = self.frames[idx]
+
+
+if __name__ == '__main__':
+    from torch.utils.data import DataLoader

models/base.py

@@ -53,13 +53,13 @@ class AEEncoder(nn.Module):
         else:
             x = inputs
 
-        x, features = self.encoder(x)
+        x = self.encoder(x)
         x = self.act(x)
 
         if len(inputs) == 2:
-            return x, features, inputs[1]
+            return x, inputs[1]
         else:
-            return x, features
+            return x
 
 
 class Decoder(nn.Module):
@@ -75,12 +75,12 @@ class Decoder(nn.Module):
             self.softmax = nn.Softmax(dim=1)
 
     def forward(self, inputs):
-        if len(inputs) == 3:
-            x, features, mask = inputs
+        if len(inputs) == 2:
+            x, mask = inputs
         else:
-            x, features = inputs
+            x = inputs
 
-        x = self.decoder([x, features])
+        x = self.decoder(x)
 
         if self.is_label:
             x = x * mask[:,None,:,:]

models/convnext_orig.py

@@ -81,6 +81,8 @@ class ConvNeXtBase(nn.Module):
         if norm_layer is None:
             norm_layer = partial(LayerNorm2d, eps=1e-6)
 
+        layers_list: List[nn.Module] = []
+
         # Stem
         firstconv_output_channels = block_setting[0].input_channels
         self.stem = Conv2dNormActivation(
@@ -93,8 +95,7 @@ class ConvNeXtBase(nn.Module):
                 activation_layer=None,
                 bias=True,
             ).to(device)
-        self.stages = []
-        self.downsamples = []
+        layers_list.append(self.stem)
 
         total_stage_blocks = sum(cnf.num_layers for cnf in block_setting)
         stage_block_id = 0
@@ -106,20 +107,17 @@ class ConvNeXtBase(nn.Module):
                 sd_prob = stochastic_depth_prob * stage_block_id / (total_stage_blocks - 1.0)
                 stage.append(block(cnf.input_channels, layer_scale, sd_prob, kernel_size))
                 stage_block_id += 1
-            self.stages.append(nn.Sequential(*stage).to(device))
+            layers_list.append(nn.Sequential(*stage).to(device))
             if cnf.out_channels is not None:
                 # Downsampling
-                self.downsamples.append(
+                layers_list.append(
                     nn.Sequential(
                         norm_layer(cnf.input_channels),
                         nn.Conv2d(cnf.input_channels, cnf.out_channels, kernel_size=2, stride=2),
                     ).to(device)
                 )
-        self.stages = nn.ModuleList(self.stages)
-        self.downsamples = nn.ModuleList(self.downsamples)
 
-        # self.layers = nn.Sequential(*layers_list).to(device)
-        # self.layers = nn.ModuleList(layers_list)
+        self.layers = nn.Sequential(*layers_list).to(device)
         self.act = nn.Tanh().to(device)
 
         # for m in self.modules():
@@ -129,16 +127,9 @@ class ConvNeXtBase(nn.Module):
         #             nn.init.zeros_(m.bias)
 
     def _forward_impl(self, x: Tensor) -> Tensor:
-        x = self.stem(x)
-        features = []
-        for i in range(len(self.stages)):
-            x = self.stages[i](x)
-            if i == len(self.stages) - 1:
-                continue
-            features.append(x)
-            x = self.downsamples[i](x)
+        x = self.layers(x)
         x = self.act(x)
-        return x, features
+        return x
 
     def forward(self, x: Tensor) -> Tensor:
         return self._forward_impl(x)
@@ -174,14 +165,13 @@ class ConvNeXtReverseBase(nn.Module):
             norm_layer = partial(LayerNorm2d, eps=1e-6)
 
         self.stem = nn.Sequential(
-                        nn.ConvTranspose2d(block_setting[0].input_channels, block_setting[0].out_channels, kernel_size=2, stride=2),
+                        nn.ConvTranspose2d(block_setting[0].input_channels, block_setting[0].out_channels, kernel_size=4, stride=4),
                         norm_layer(block_setting[0].out_channels),
                     ).to(device)
         
         total_stage_blocks = sum(cnf.num_layers for cnf in block_setting)
         stage_block_id = 0
-        self.stages = []
-        self.upsamples = []
+        layers_list: List[nn.Module] = []
 
 
         for i, cnf in enumerate(block_setting):
@@ -192,17 +182,17 @@ class ConvNeXtReverseBase(nn.Module):
                 sd_prob = stochastic_depth_prob * stage_block_id / (total_stage_blocks - 1.0)
                 stage.append(block(cnf.input_channels, layer_scale, sd_prob, kernel_size))
                 stage_block_id += 1
-            self.stages.append(nn.Sequential(*stage).to(device))
+            layers_list.append(nn.Sequential(*stage))
 
             if i == 0:
-                self.upsamples.append(self.stem)
+                layers_list.append(self.stem)
             else:
                 # Upsampling
-                self.upsamples.append(
+                layers_list.append(
                     nn.Sequential(
                         norm_layer(cnf.input_channels),
                         nn.ConvTranspose2d(cnf.input_channels, cnf.out_channels, kernel_size=2, stride=2),
-                    ).to(device)
+                    )
                 )
 
         self.last = nn.Sequential(
@@ -211,8 +201,7 @@ class ConvNeXtReverseBase(nn.Module):
                         Permute([0, 3, 1, 2]),
                     ).to(device)
         
-        self.stages = nn.ModuleList(self.stages)
-        self.upsamples = nn.ModuleList(self.upsamples)
+        self.layers = nn.Sequential(*layers_list).to(device)
         self.activation = nn.Tanh().to(device) if not is_label else nn.Identity().to(device)
 
 
@@ -224,15 +213,7 @@ class ConvNeXtReverseBase(nn.Module):
 
 
     def _forward_impl(self, x) -> Tensor:
-        x, features = x
-        for i in range(len(self.stages)-1):
-            if i == 0:
-                x = self.stages[i](x)
-            else:
-                x = self.stages[i](torch.cat([x, features[-i]], dim=1))
-            x = self.upsamples[i](x)
-        x = self.stages[-1](x)
-        x = self.upsamples[-1](x)
+        x = self.layers(x)
         x = self.last(x)
         x = self.activation(x)
         return x

scripts/direct_train.py

@@ -13,7 +13,7 @@ args = vars(parser.parse_args())
 config_module = importlib.import_module(args['config'])
 config = config_module.config
 if args['debug']:
-    from configs.direct_3 import config
+    from configs.direct_1 import config
     # from configs.ae_static_orig_2 import config
 os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
 os.environ["CUDA_VISIBLE_DEVICES"] = config['gpu']

smpl/smpl_torch.py

@@ -206,9 +206,9 @@ class SMPL(Module):
         )
       )
     
-    joints = torch.matmul(results, torch.transpose(torch.concat([J, torch.ones((J.shape[0], 24, 1), dtype=torch.float32, device=J.get_device())], axis=-1)[:,:,None,:],2,3))[:,:,:3,0]
+    joints = torch.matmul(results, torch.transpose(torch.concat([J, torch.ones((J.shape[0], 24, 1), dtype=torch.float32, device=J.device)], axis=-1)[:,:,None,:],2,3))[:,:,:3,0]
             
-    T = torch.stack([torch.tensordot(results[i], self.weights[torch.greater(g, 0.5).to(torch.long)], dims=([0], [1])).permute(2, 0, 1) for i, g in enumerate(betas[:,-1])])
+    T = torch.stack([torch.tensordot(results[i], self.weights_all[torch.greater(g, 0.5).to(torch.long)], dims=([0], [1])).permute(2, 0, 1) for i, g in enumerate(betas[:,-1])])
     rest_shape_h = torch.cat(
       (v_posed, torch.ones((batch_num, v_posed.shape[1], 1), dtype=torch.float32).to(self.device)), dim=2
     )
@@ -230,4 +230,4 @@ class SMPL(Module):
     if remove and verts is None:
       result = result[:,self.to_keep,:]
 
-    return result, joints, v_shaped, v_posed
+    return result, joints