Revert "try to train but wrong"

This reverts commit 47ff0fa.

SocketForUnity.py

@@ -237,8 +237,8 @@ def main():
     # save_pkl2json(testfilepath)
 
     ################################------npz数据文件读写测试
-    # testfilepath = os.path.join("data_split","CMU","01","01_01_poses.npz")
-    # save_npz2json(testfilepath)
+    testfilepath = os.path.join("data_split","CMU","01","01_01_poses.npz")
+    save_npz2json(testfilepath)
     return
 
 if __name__ == '__main__':

data/dataset_amass.py

@@ -53,7 +53,6 @@ def __getitem__(self, idx):
             data = pickle.load(f)
 
         if self.opt['phase'] == 'train':
-            #当文件长度小于窗口大小的时候，随机加载之前的文件
             while data['rotation_local_full_gt_list'].shape[0] <self.window_size:
                 idx = random.randint(0,idx)
                 filename = self.filename_list[idx]
@@ -67,21 +66,21 @@ def __getitem__(self, idx):
         head_global_trans_list = data['head_global_trans_list']
 
 
-        if self.opt['phase'] == 'train':  # train
-            #训练时   随机截取一段
-            frame = np.random.randint(hmd_position_global_full_gt_list.shape[0] - self.window_size)
-            input_hmd  = hmd_position_global_full_gt_list[frame:frame + self.window_size+1,...].reshape(self.window_size+1, -1)
-            output_gt = rotation_local_full_gt_list[frame : frame + self.window_size + 1,...]
+        if self.opt['phase'] == 'train':
+
+            frame = np.random.randint(hmd_position_global_full_gt_list.shape[0] - self.window_size + 1 - 1)
+            input_hmd  = hmd_position_global_full_gt_list[frame:frame + self.window_size+1,...].reshape(self.window_size+1, -1).float()
+            output_gt = rotation_local_full_gt_list[frame + self.window_size - 1 : frame + self.window_size - 1 + 1,...].float()
 
-            return {'L': input_hmd.float(),
-                    'H': output_gt.float(),
+            return {'L': input_hmd,
+                    'H': output_gt,
                     'P': 1,
                     'Head_trans_global':head_global_trans_list[frame + self.window_size - 1:frame + self.window_size - 1+1,...],
                     'pos_pelvis_gt':body_parms_list['trans'][frame + self.window_size - 1:frame + self.window_size - 1+1,...],
                     'vel_pelvis_gt':body_parms_list['trans'][frame + self.window_size - 1:frame + self.window_size - 1+1,...]-body_parms_list['trans'][frame + self.window_size - 2:frame + self.window_size - 2+1,...]
                     }
 
-        else: #  test
+        else:
 
             input_hmd  = hmd_position_global_full_gt_list.reshape(hmd_position_global_full_gt_list.shape[0], -1)[1:]
             output_gt = rotation_local_full_gt_list[1:]

main_train_avatarposer.py

@@ -103,7 +103,7 @@ def main(json_path='options/train_avatarposer.json'):
                                       shuffle=dataset_opt['dataloader_shuffle'],
                                       num_workers=dataset_opt['dataloader_num_workers'],
                                       drop_last=True,
-                                      pin_memory=False)
+                                      pin_memory=True)
         elif phase == 'test':
             test_set = define_Dataset(dataset_opt)
             test_loader = DataLoader(test_set, batch_size=dataset_opt['dataloader_batch_size'],

models/loss.py

@@ -1,4 +1,3 @@
-import math
 import torch
 import torch.nn as nn
 import torchvision

models/model_avatarposer.py

@@ -149,8 +149,8 @@ def feed_data(self, data, need_H=True, test=False):
         self.L = data['L'].to(self.device)
         self.P = data['P']
         self.Head_trans_global = data['Head_trans_global'].to(self.device)
-        self.H_global_orientation = data['H'][:,:6].to(self.device)
-        self.H_joint_rotation = data['H'][:,6:].to(self.device)
+        self.H_global_orientation = data['H'].squeeze()[:,:6].to(self.device)
+        self.H_joint_rotation = data['H'].squeeze()[:,6:].to(self.device)
 #        self.H = torch.cat([self.H_global_orientation, self.H_joint_rotation],dim=-1).to(self.device)
         self.H_joint_position = self.netG.module.fk_module(self.H_global_orientation, self.H_joint_rotation , self.bm)
 

options/train_avatarposer.json

@@ -18,8 +18,8 @@
       , "dataroot": "./data_fps60"// path of training dataset
 
       , "dataloader_shuffle": true
-      , "dataloader_num_workers": 1
-      , "dataloader_batch_size": 1    // batch size 1 | 16 | 32 | 48 | 64 | 128 | 256
+      , "dataloader_num_workers": 16
+      , "dataloader_batch_size": 256     // batch size 1 | 16 | 32 | 48 | 64 | 128 | 256
       , "num_input": 3
       , "window_size": 40
 

prepare_data.py

@@ -104,27 +104,28 @@
             rotation_global_matrot = local2global_pose(rotation_local_matrot, bm.kintree_table[0].long()) # rotation of joints relative to the origin
 
             head_rotation_global_matrot = rotation_global_matrot[:,[15],:,:]
-            #
+
             rotation_global_6d = utils_transform.matrot2sixd(rotation_global_matrot.reshape(-1,3,3)).reshape(rotation_global_matrot.shape[0],-1,6)
             input_rotation_global_6d = rotation_global_6d[1:,[15,20,21],:]
 
             rotation_velocity_global_matrot = torch.matmul(torch.inverse(rotation_global_matrot[:-1]),rotation_global_matrot[1:])
             rotation_velocity_global_6d = utils_transform.matrot2sixd(rotation_velocity_global_matrot.reshape(-1,3,3)).reshape(rotation_velocity_global_matrot.shape[0],-1,6)
             input_rotation_velocity_global_6d = rotation_velocity_global_6d[:,[15,20,21],:]
 
-            #
             position_global_full_gt_world = body_pose_world.Jtr[:,:22,:] # position of joints relative to the world origin
 
             position_head_world = position_global_full_gt_world[:,15,:] # world position of head
 
             head_global_trans = torch.eye(4).repeat(position_head_world.shape[0],1,1)
             head_global_trans[:,:3,:3] = head_rotation_global_matrot.squeeze()
             head_global_trans[:,:3,3] = position_global_full_gt_world[:,15,:]
+
+        #            embed()
             head_global_trans_list = head_global_trans[1:]
 
 
 
-            
+
             num_frames = position_global_full_gt_world.shape[0]-1