data_transfer / vmodel.py
Onearth's picture
Upload vmodel.py
cb1ed64 verified
raw
history blame
6.32 kB
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from regress_module import VisProcess, VisTR, VisRes
from torchvision import transforms
import matplotlib.pyplot as plt
import sys
sys.path.append('/home/kejianshi/Desktop/Surgical_Robot/science_robotics/stateregress_back/utils')
from general_utils import AttrDict
# sys.path.append('/home/kejianshi/Desktop/Surgical_Robot/science_robotics/ar_surrol/surrol/tasks')
sys.path.append('/home/kejianshi/Desktop/Surgical_Robot/science_robotics/ar_surrol/surrol/tasks')
from depth_anything.dpt import DepthAnything
from depth_anything.util.transform import Resize, NormalizeImage, PrepareForNet
class vismodel(nn.Module):
def __init__(
self,
opts
):
super().__init__()
self.opts=opts
self.device=opts.device
self.img_size=self.opts.img_size
self.obj_num=1
self.v_processor=VisRes()
if not self.opts.use_exist_depth:
self._load_dam()
def _load_dam(self):
encoder = 'vitb' # can also be 'vitb' or 'vitl'
self.depth_anything = DepthAnything.from_pretrained('LiheYoung/depth_anything_{:}14'.format(encoder)).eval()
#self.depth_anything.to(self.device)
def _get_depth_with_dam(self, img):
'''
input: rgb image 1xHxW
'''
#img=self.img_transform({'image': img})['image']
#img=torch.from_numpy(img).unsqueeze(0)
#img=transforms.Resize((518,518))(img)
with torch.no_grad():
depth = self.depth_anything(img)
#print(depth.shape)
depth = F.interpolate(depth[None], self.img_size, mode='bilinear', align_corners=False)[0]
depth_min = torch.amin(depth, dim=(1, 2), keepdim=True)
depth_max = torch.amax(depth, dim=(1, 2), keepdim=True)
depth = (depth - depth_min) / (depth_max - depth_min)
return depth
def normlize_angles(self, x):
return np.arctan2(np.sin(x),np.cos(x))
def get_action(self, state, noise=False, v_action=False):
#print("get_action")
if not v_action:
return super().get_action(state, noise)
#self.count+=1
self.v_processor.eval()
rgb=self.to_torch(state['depth']).unsqueeze(0)
depth=self._get_depth_with_dam(rgb)[0] #tensor
depth_norm=self.depth_norm.normalize(depth.reshape(-1,256*256),device=self.device).reshape(1,256,256)
seg=self.to_torch(transitions['seg'])
seg_d=torch.concat((seg,depth_norm))
inputs=seg_d.unsqueeze(0).float().to(self.device) # B 2 256 256
#print(inputs.shape)
with torch.no_grad():
v_output=self.v_processor(inputs).squeeze() # 9
#v_save=v_output.cpu().data.numpy()
#np.save('test_record/v_output.npy', v_save)
o, g = state['observation'], state['desired_goal']
g=self.g_norm.normalize(g)
#print("g: ",g)
g_norm=torch.tensor(g).float().to(self.device)
#print("g_norm: ",g_norm)
if not self.regress_rbt_staet:
robot_state=torch.tensor(o[:7]).to(self.device)
#pos=v_output[:3]
#rel_pos=pos-robot_state[:3]
rel_pos=v_output[:3*self.obj_num]
new_pos=robot_state[:3]+rel_pos[:3]
if self.obj_num>1:
for i in range(1, self.obj_num):
pos=robot_state[:3]+rel_pos[i*3:3*self.obj_num]
new_pos=torch.concat((new_pos,pos))
waypoint_pos_rot=v_output[3*self.obj_num:]
#o=torch.from_numpy(np.concatenate([o[:,7:10].copy(),o[:,13:19]].copy(),axis=1)).to(self.device)
#o_new=torch.concat((,o),dim=1) # B 19
o_new=torch.concat((robot_state, new_pos))
o_new=torch.concat((o_new, rel_pos))
o_new=torch.concat((o_new, waypoint_pos_rot))
o_norm=self.o_norm.normalize(o_new,device=self.device)
else:
o_norm=self.o_norm.normalize(v_output,device=self.device)
o_norm=torch.tensor(o_norm).float().to(self.device)
input_tensor=torch.concat((o_norm, g_norm), axis=0).to(torch.float32)
#save_input=input_tensor.cpu().data.numpy()
#np.save('test_record/actor_input.npy', save_input)
#exit()
action = self.actor(input_tensor).cpu().data.numpy().flatten()
self.v_processor.train()
return action
def forward(self, seg, v_gt):
# if self.opts.use_exist_depth:
# d=rgb
# else:
# d=self._get_depth_with_dam(rgb)
# seg_d=torch.concat((seg.unsqueeze(1),seg.unsqueeze(1)),dim=1)
seg_d = torch.unsqueeze(seg, 1)
# print(seg_d.shape)
output=self.v_processor(seg_d)
#print('output: ',type(output))
#print('v_gt: ',v_gt.shape)
pos_loss=F.mse_loss(output[:,:3*self.obj_num],v_gt[:,:3*self.obj_num])
# 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])
# v_loss+=w_pos_loss
# w_rot_loss=F.mse_loss(output[:,3*self.obj_num+3:],v_gt[:,3*self.obj_num+3: ])
# v_loss+=w_rot_loss
metrics = AttrDict(
v_pos=pos_loss.item(),
# w_pos_loss_loss=w_pos_loss.item(),
# w_rot_loss=w_rot_loss.item(),
# v_loss=v_loss.item()
)
return metrics, pos_loss
def get_obs(self, seg, rgb):
# if self.opts.use_exist_depth:
# d=rgb
# else:
# d=self._get_depth_with_dam(rgb)
#d=self._get_depth_with_dam(rgb)
#seg_d=torch.concat((seg.unsqueeze(1),d.unsqueeze(1)),dim=1)#.to(self.device)
# seg_d=torch.concat((seg.unsqueeze(1),d.unsqueeze(1)),dim=1)#.to(self.device)
seg_d = torch.unsqueeze(seg, 1)
output=self.v_processor(seg_d)
#print(output.shape)
return output