Onearth commited on
Commit
cb1ed64
·
verified ·
1 Parent(s): 4302d44

Upload vmodel.py

Browse files
Files changed (1) hide show
  1. vmodel.py +172 -0
vmodel.py ADDED
@@ -0,0 +1,172 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import torch
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+ from regress_module import VisProcess, VisTR, VisRes
6
+ from torchvision import transforms
7
+ import matplotlib.pyplot as plt
8
+ import sys
9
+ sys.path.append('/home/kejianshi/Desktop/Surgical_Robot/science_robotics/stateregress_back/utils')
10
+ from general_utils import AttrDict
11
+ # sys.path.append('/home/kejianshi/Desktop/Surgical_Robot/science_robotics/ar_surrol/surrol/tasks')
12
+ sys.path.append('/home/kejianshi/Desktop/Surgical_Robot/science_robotics/ar_surrol/surrol/tasks')
13
+ from depth_anything.dpt import DepthAnything
14
+ from depth_anything.util.transform import Resize, NormalizeImage, PrepareForNet
15
+
16
+
17
+ class vismodel(nn.Module):
18
+ def __init__(
19
+ self,
20
+ opts
21
+
22
+ ):
23
+ super().__init__()
24
+ self.opts=opts
25
+ self.device=opts.device
26
+
27
+ self.img_size=self.opts.img_size
28
+ self.obj_num=1
29
+ self.v_processor=VisRes()
30
+ if not self.opts.use_exist_depth:
31
+ self._load_dam()
32
+
33
+ def _load_dam(self):
34
+ encoder = 'vitb' # can also be 'vitb' or 'vitl'
35
+ self.depth_anything = DepthAnything.from_pretrained('LiheYoung/depth_anything_{:}14'.format(encoder)).eval()
36
+ #self.depth_anything.to(self.device)
37
+
38
+ def _get_depth_with_dam(self, img):
39
+ '''
40
+ input: rgb image 1xHxW
41
+ '''
42
+
43
+ #img=self.img_transform({'image': img})['image']
44
+ #img=torch.from_numpy(img).unsqueeze(0)
45
+ #img=transforms.Resize((518,518))(img)
46
+ with torch.no_grad():
47
+ depth = self.depth_anything(img)
48
+
49
+ #print(depth.shape)
50
+ depth = F.interpolate(depth[None], self.img_size, mode='bilinear', align_corners=False)[0]
51
+ depth_min = torch.amin(depth, dim=(1, 2), keepdim=True)
52
+ depth_max = torch.amax(depth, dim=(1, 2), keepdim=True)
53
+ depth = (depth - depth_min) / (depth_max - depth_min)
54
+
55
+ return depth
56
+
57
+
58
+ def normlize_angles(self, x):
59
+ return np.arctan2(np.sin(x),np.cos(x))
60
+
61
+ def get_action(self, state, noise=False, v_action=False):
62
+ #print("get_action")
63
+
64
+ if not v_action:
65
+ return super().get_action(state, noise)
66
+ #self.count+=1
67
+
68
+ self.v_processor.eval()
69
+
70
+ rgb=self.to_torch(state['depth']).unsqueeze(0)
71
+
72
+ depth=self._get_depth_with_dam(rgb)[0] #tensor
73
+ depth_norm=self.depth_norm.normalize(depth.reshape(-1,256*256),device=self.device).reshape(1,256,256)
74
+ seg=self.to_torch(transitions['seg'])
75
+ seg_d=torch.concat((seg,depth_norm))
76
+
77
+ inputs=seg_d.unsqueeze(0).float().to(self.device) # B 2 256 256
78
+ #print(inputs.shape)
79
+
80
+ with torch.no_grad():
81
+ v_output=self.v_processor(inputs).squeeze() # 9
82
+
83
+ #v_save=v_output.cpu().data.numpy()
84
+ #np.save('test_record/v_output.npy', v_save)
85
+
86
+ o, g = state['observation'], state['desired_goal']
87
+ g=self.g_norm.normalize(g)
88
+ #print("g: ",g)
89
+ g_norm=torch.tensor(g).float().to(self.device)
90
+ #print("g_norm: ",g_norm)
91
+
92
+ if not self.regress_rbt_staet:
93
+ robot_state=torch.tensor(o[:7]).to(self.device)
94
+ #pos=v_output[:3]
95
+ #rel_pos=pos-robot_state[:3]
96
+ rel_pos=v_output[:3*self.obj_num]
97
+ new_pos=robot_state[:3]+rel_pos[:3]
98
+
99
+ if self.obj_num>1:
100
+ for i in range(1, self.obj_num):
101
+ pos=robot_state[:3]+rel_pos[i*3:3*self.obj_num]
102
+ new_pos=torch.concat((new_pos,pos))
103
+
104
+ waypoint_pos_rot=v_output[3*self.obj_num:]
105
+
106
+
107
+ #o=torch.from_numpy(np.concatenate([o[:,7:10].copy(),o[:,13:19]].copy(),axis=1)).to(self.device)
108
+
109
+ #o_new=torch.concat((,o),dim=1) # B 19
110
+ o_new=torch.concat((robot_state, new_pos))
111
+ o_new=torch.concat((o_new, rel_pos))
112
+ o_new=torch.concat((o_new, waypoint_pos_rot))
113
+ o_norm=self.o_norm.normalize(o_new,device=self.device)
114
+ else:
115
+ o_norm=self.o_norm.normalize(v_output,device=self.device)
116
+ o_norm=torch.tensor(o_norm).float().to(self.device)
117
+
118
+ input_tensor=torch.concat((o_norm, g_norm), axis=0).to(torch.float32)
119
+ #save_input=input_tensor.cpu().data.numpy()
120
+ #np.save('test_record/actor_input.npy', save_input)
121
+ #exit()
122
+
123
+ action = self.actor(input_tensor).cpu().data.numpy().flatten()
124
+
125
+ self.v_processor.train()
126
+ return action
127
+
128
+
129
+ def forward(self, seg, v_gt):
130
+ # if self.opts.use_exist_depth:
131
+ # d=rgb
132
+ # else:
133
+ # d=self._get_depth_with_dam(rgb)
134
+
135
+ # seg_d=torch.concat((seg.unsqueeze(1),seg.unsqueeze(1)),dim=1)
136
+ seg_d = torch.unsqueeze(seg, 1)
137
+ # print(seg_d.shape)
138
+ output=self.v_processor(seg_d)
139
+
140
+
141
+ #print('output: ',type(output))
142
+ #print('v_gt: ',v_gt.shape)
143
+
144
+ pos_loss=F.mse_loss(output[:,:3*self.obj_num],v_gt[:,:3*self.obj_num])
145
+
146
+ # w_pos_loss=F.mse_loss(output[:,3*self.obj_num: 3*self.obj_num+3],v_gt[:,3*self.obj_num: 3*self.obj_num+3])
147
+ # v_loss+=w_pos_loss
148
+ # w_rot_loss=F.mse_loss(output[:,3*self.obj_num+3:],v_gt[:,3*self.obj_num+3: ])
149
+ # v_loss+=w_rot_loss
150
+ metrics = AttrDict(
151
+ v_pos=pos_loss.item(),
152
+ # w_pos_loss_loss=w_pos_loss.item(),
153
+ # w_rot_loss=w_rot_loss.item(),
154
+ # v_loss=v_loss.item()
155
+
156
+ )
157
+
158
+ return metrics, pos_loss
159
+
160
+ def get_obs(self, seg, rgb):
161
+ # if self.opts.use_exist_depth:
162
+ # d=rgb
163
+ # else:
164
+ # d=self._get_depth_with_dam(rgb)
165
+ #d=self._get_depth_with_dam(rgb)
166
+ #seg_d=torch.concat((seg.unsqueeze(1),d.unsqueeze(1)),dim=1)#.to(self.device)
167
+ # seg_d=torch.concat((seg.unsqueeze(1),d.unsqueeze(1)),dim=1)#.to(self.device)
168
+ seg_d = torch.unsqueeze(seg, 1)
169
+ output=self.v_processor(seg_d)
170
+
171
+ #print(output.shape)
172
+ return output