-
Notifications
You must be signed in to change notification settings - Fork 47
/
neat_agent.py
354 lines (289 loc) · 13.2 KB
/
neat_agent.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
import os
import json
import datetime
import pathlib
import time
import cv2
import carla
from collections import deque
import torch
import carla
import numpy as np
from PIL import Image, ImageDraw
from leaderboard.autoagents import autonomous_agent
from neat.architectures import AttentionField
from neat.config import GlobalConfig
from neat.utils import flow_to_color
from team_code.planner import RoutePlanner
SAVE_PATH = os.environ.get('SAVE_PATH', None)
def get_entry_point():
return 'MultiTaskAgent'
def scale_and_crop_image(image, scale=1, crop=256):
"""
Scale and crop a PIL image, returning a channels-first numpy array.
"""
# image = Image.open(filename)
(width, height) = (image.width // scale, image.height // scale)
im_resized = image.resize((width, height))
image = np.asarray(im_resized)
start_x = height//2 - crop//2
start_y = width//2 - crop//2
cropped_image = image[start_x:start_x+crop, start_y:start_y+crop]
cropped_image = np.transpose(cropped_image, (2,0,1))
return cropped_image
class MultiTaskAgent(autonomous_agent.AutonomousAgent):
def setup(self, path_to_conf_file):
self.track = autonomous_agent.Track.SENSORS
self.config_path = path_to_conf_file
self.step = -1
self.wall_start = time.time()
self.initialized = False
args_file = open(os.path.join(path_to_conf_file, 'args.txt'), 'r')
self.args = json.load(args_file)
args_file.close()
self.args['out_res'] = 100
self.input_buffer = {'rgb': deque(), 'rgb_left': deque(), 'rgb_right': deque()}
self.config = GlobalConfig()
self.net = AttentionField(self.config, 'cuda')
self.net.encoder.load_state_dict(torch.load(os.path.join(path_to_conf_file, 'best_encoder.pth')))
self.net.decoder.load_state_dict(torch.load(os.path.join(path_to_conf_file, 'best_decoder.pth')))
self.plan_grid = self.net.create_plan_grid(self.config.plan_scale, self.config.plan_points, 1)
self.light_grid = self.net.create_light_grid(self.config.light_x_steps, self.config.light_y_steps, 1)
self.net.cuda()
self.net.eval()
self.save_path = None
if SAVE_PATH is not None:
now = datetime.datetime.now()
string = pathlib.Path(os.environ['ROUTES']).stem + '_'
string += '_'.join(map(lambda x: '%02d' % x, (now.month, now.day, now.hour, now.minute, now.second)))
print (string)
self.save_path = pathlib.Path(os.environ['SAVE_PATH']) / string
self.save_path.mkdir(parents=True, exist_ok=False)
(self.save_path / 'rgb').mkdir(parents=True, exist_ok=False)
(self.save_path / 'bev').mkdir(parents=True, exist_ok=False)
(self.save_path / 'flow').mkdir(parents=True, exist_ok=False)
(self.save_path / 'out').mkdir(parents=True, exist_ok=False)
(self.save_path / 'img').mkdir(parents=True, exist_ok=False)
(self.save_path / 'meta').mkdir(parents=True, exist_ok=False)
def _init(self):
self._route_planner = RoutePlanner(4.0, 50.0)
self._route_planner.set_route(self._global_plan, True)
self.initialized = True
def _get_position(self, tick_data):
gps = tick_data['gps']
gps = (gps - self._route_planner.mean) * self._route_planner.scale
return gps
def sensors(self):
return [
{
'type': 'sensor.camera.rgb',
'x': 1.3, 'y': 0.0, 'z':2.3,
'roll': 0.0, 'pitch': 0.0, 'yaw': 0.0,
'width': 400, 'height': 300, 'fov': 100,
'id': 'rgb'
},
{
'type': 'sensor.camera.rgb',
'x': 1.3, 'y': 0.0, 'z':2.3,
'roll': 0.0, 'pitch': 0.0, 'yaw': -60.0,
'width': 400, 'height': 300, 'fov': 100,
'id': 'rgb_left'
},
{
'type': 'sensor.camera.rgb',
'x': 1.3, 'y': 0.0, 'z':2.3,
'roll': 0.0, 'pitch': 0.0, 'yaw': 60.0,
'width': 400, 'height': 300, 'fov': 100,
'id': 'rgb_right'
},
{
'type': 'sensor.camera.rgb',
'x': 1.3, 'y': 0.0, 'z':2.3,
'roll': 0.0, 'pitch': 0.0, 'yaw': 0.0,
'width': 800, 'height': 600, 'fov': 100,
'id': 'rgb_front'
},
{
'type': 'sensor.camera.rgb',
'x': 1.3, 'y': 0.0, 'z': 25,
'roll': 0.0, 'pitch': -90.0, 'yaw': 0.0,
'width': 800, 'height': 800, 'fov': 100,
'id': 'bev'
},
{
'type': 'sensor.other.imu',
'x': 0.0, 'y': 0.0, 'z': 0.0,
'roll': 0.0, 'pitch': 0.0, 'yaw': 0.0,
'sensor_tick': 0.05,
'id': 'imu'
},
{
'type': 'sensor.other.gnss',
'x': 0.0, 'y': 0.0, 'z': 0.0,
'roll': 0.0, 'pitch': 0.0, 'yaw': 0.0,
'sensor_tick': 0.01,
'id': 'gps'
},
{
'type': 'sensor.speedometer',
'reading_frequency': 20,
'id': 'speed'
}
]
def tick(self, input_data):
self.step += 1
rgb = cv2.cvtColor(input_data['rgb'][1][:, :, :3], cv2.COLOR_BGR2RGB)
rgb_left = cv2.cvtColor(input_data['rgb_left'][1][:, :, :3], cv2.COLOR_BGR2RGB)
rgb_right = cv2.cvtColor(input_data['rgb_right'][1][:, :, :3], cv2.COLOR_BGR2RGB)
rgb_front = cv2.cvtColor(input_data['rgb_front'][1][:, :, :3], cv2.COLOR_BGR2RGB)
bev = cv2.cvtColor(input_data['bev'][1][:, :, :3], cv2.COLOR_BGR2RGB)
gps = input_data['gps'][1][:2]
speed = input_data['speed'][1]['speed']
compass = input_data['imu'][1][-1]
result = {
'rgb': rgb,
'rgb_left': rgb_left,
'rgb_right': rgb_right,
'rgb_front': rgb_front,
'bev': bev,
'gps': gps,
'speed': speed,
'compass': compass,
}
pos = self._get_position(result)
next_wp, next_cmd = self._route_planner.run_step(pos)
result['next_command'] = next_cmd.value
theta = compass + np.pi/2
R = np.array([
[np.cos(theta), -np.sin(theta)],
[np.sin(theta), np.cos(theta)]
])
local_command_point = np.array([next_wp[0]-pos[0], next_wp[1]-pos[1]])
local_command_point = R.T.dot(local_command_point)
result['target_point'] = tuple(local_command_point)
return result
@torch.no_grad()
def run_step(self, input_data, timestamp):
if not self.initialized:
self._init()
tick_data = self.tick(input_data)
if self.step < self.config.seq_len:
rgb = torch.from_numpy(scale_and_crop_image(Image.fromarray(tick_data['rgb']))).unsqueeze(0)
self.input_buffer['rgb'].append(rgb.to('cuda', dtype=torch.float32))
rgb_left = torch.from_numpy(scale_and_crop_image(Image.fromarray(tick_data['rgb_left']))).unsqueeze(0)
self.input_buffer['rgb_left'].append(rgb_left.to('cuda', dtype=torch.float32))
rgb_right = torch.from_numpy(scale_and_crop_image(Image.fromarray(tick_data['rgb_right']))).unsqueeze(0)
self.input_buffer['rgb_right'].append(rgb_right.to('cuda', dtype=torch.float32))
control = carla.VehicleControl()
control.steer = 0.0
control.throttle = 0.0
control.brake = 0.0
return control
tick_data['target_point'] = [torch.FloatTensor([tick_data['target_point'][0]]),
torch.FloatTensor([tick_data['target_point'][1]])]
gt_velocity = torch.FloatTensor([tick_data['speed']]).to('cuda', dtype=torch.float32)
command = torch.FloatTensor([tick_data['next_command']]).to('cuda', dtype=torch.float32)
target_point = torch.stack(tick_data['target_point']).to('cuda', dtype=torch.float32)
self.target_point_model = target_point
rgb = torch.from_numpy(scale_and_crop_image(Image.fromarray(tick_data['rgb']))).unsqueeze(0)
self.input_buffer['rgb'].popleft()
self.input_buffer['rgb'].append(rgb.to('cuda', dtype=torch.float32))
rgb_left = torch.from_numpy(scale_and_crop_image(Image.fromarray(tick_data['rgb_left']))).unsqueeze(0)
self.input_buffer['rgb_left'].popleft()
self.input_buffer['rgb_left'].append(rgb_left.to('cuda', dtype=torch.float32))
rgb_right = torch.from_numpy(scale_and_crop_image(Image.fromarray(tick_data['rgb_right']))).unsqueeze(0)
self.input_buffer['rgb_right'].popleft()
self.input_buffer['rgb_right'].append(rgb_right.to('cuda', dtype=torch.float32))
images = []
for i in range(self.config.seq_len):
images.append(self.input_buffer['rgb'][i])
if self.config.num_camera == 3:
images.append(self.input_buffer['rgb_left'][i])
images.append(self.input_buffer['rgb_right'][i])
encoding = self.net.encoder(images, gt_velocity)
pred_waypoint_mean, red_light_occ = self.net.plan(target_point, encoding, self.plan_grid, self.light_grid, self.config.plan_points, self.config.plan_iters)
steer, throttle, brake, metadata = self.net.control_pid(pred_waypoint_mean[:, self.config.seq_len:], gt_velocity, target_point, red_light_occ)
self.encoding_model = encoding
self.pred_waypoint_mean_model = pred_waypoint_mean
self.pid_metadata = metadata
steer = float(steer)
throttle = float(throttle)
brake = float(brake)
if brake < 0.05: brake = 0.0
if throttle > brake: brake = 0.0
control = carla.VehicleControl()
control.steer = steer
control.throttle = throttle
control.brake = brake
if SAVE_PATH is not None and self.step % 10 == 0:
self.save(tick_data)
return control
def save(self, tick_data):
frame = self.step // 10
Image.fromarray(tick_data['rgb_front']).save(self.save_path / 'rgb' / ('%04d.png' % frame))
Image.fromarray(tick_data['bev']).save(self.save_path / 'bev' / ('%04d.png' % frame))
outfile = open(self.save_path / 'meta' / ('%04d.json' % frame), 'w')
json.dump(self.pid_metadata, outfile, indent=4)
outfile.close()
# grid used for visualizing occupancy and flow
linspace_x = torch.linspace(-self.config.axis/2, self.config.axis/2, steps=self.args['out_res'])
linspace_y = torch.linspace(-self.config.axis/2, self.config.axis/2, steps=self.args['out_res'])
linspace_t = torch.linspace(0, self.config.tot_len - 1, steps=self.config.tot_len)
for i in range(self.config.seq_len):
# save one sample per batch
front_numpy = (self.input_buffer['rgb'][i][0].data.cpu().numpy().transpose((1, 2, 0))).astype(np.uint8)
left_numpy = (self.input_buffer['rgb_left'][i][0].data.cpu().numpy().transpose((1, 2, 0))).astype(np.uint8)
right_numpy = (self.input_buffer['rgb_right'][i][0].data.cpu().numpy().transpose((1, 2, 0))).astype(np.uint8)
image_numpy = np.concatenate([left_numpy,front_numpy,right_numpy], axis=1)
image_display = Image.fromarray(image_numpy)
if not os.path.isdir(self.save_path / 'img' / str(frame).zfill(4)):
os.mkdir(self.save_path / 'img' / str(frame).zfill(4))
image_display.save(f"{self.save_path}/img/{str(frame).zfill(4)}/{str(i)}.png")
# target point in pixel coordinates
target_point_pixel = self.target_point_model.squeeze().cpu().numpy()
target_point_pixel[1] += self.config.offset * self.config.resolution
# hack for when actual target is outside image (axis/2 * resolution)
target_point_pixel = np.clip(target_point_pixel, -(self.config.axis/2 * self.config.resolution - 1), (self.config.axis/2 * self.config.resolution - 1))
target_point_pixel = (target_point_pixel*self.args['out_res']//50 + self.args['out_res']//2).astype(np.uint8)
for i in range(self.config.tot_len):
# predicted waypoint in pixel coordinates
pred_waypoint = self.pred_waypoint_mean_model[0,i].data.cpu().numpy()
pred_waypoint[1] += self.config.offset * self.config.resolution
pred_waypoint = np.clip(pred_waypoint, -(self.config.axis/2 * self.config.resolution - 1), (self.config.axis/2 * self.config.resolution - 1))
pred_waypoint = (pred_waypoint*self.args['out_res']//(self.config.axis * self.config.resolution) + self.args['out_res']//2).astype(np.uint8)
# visualization of occupancy and flow
img_rows = []
flow_rows = []
for row in range(self.args['out_res']):
grid_x, grid_y, grid_t = torch.meshgrid(linspace_x, linspace_y[row], linspace_t[i].unsqueeze(0))
grid_points = torch.stack((grid_x, grid_y, grid_t), dim=3).unsqueeze(0).repeat(1,1,1,1,1)
grid_points = grid_points.reshape(1,-1,3).to('cuda', dtype=torch.float32)
pred_img_pts, pred_img_offsets, _ = self.net.decode(grid_points, self.target_point_model, self.encoding_model)
pred_img_pts = torch.argmax(pred_img_pts[-1], dim=1)
pred_img = pred_img_pts.reshape(1,self.args['out_res'])
pred_flow = pred_img_offsets[-1].reshape(1,2,self.args['out_res'])
img_rows.append(pred_img)
flow_rows.append(pred_flow)
pred_img = torch.stack(img_rows, dim=-1)
pred_flow = torch.stack(flow_rows, dim=-1)
semantics = pred_img[0,:,:].transpose(1, 0).data.cpu().numpy().astype(np.uint8)
semantic_display = np.zeros((semantics.shape[0], semantics.shape[1], 3))
for key, value in self.config.classes.items():
semantic_display[np.where(semantics == key)] = value
semantic_display = semantic_display.astype(np.uint8)
semantic_display = Image.fromarray(semantic_display)
if not os.path.isdir(self.save_path / 'out' / str(frame).zfill(4)):
os.mkdir(self.save_path / 'out' / str(frame).zfill(4))
semantic_display.save(f"{self.save_path}/out/{str(frame).zfill(4)}/{str(i)}.png")
# flow image of predicted offsets
flow_uv = pred_flow[0,:,:,:].transpose(2,0).data.cpu().numpy()*self.args['out_res']/self.config.axis
flow_rgb = flow_to_color(flow_uv)
flow_display = Image.fromarray(flow_rgb)
draw = ImageDraw.Draw(flow_display)
draw.ellipse([tuple(target_point_pixel-1), tuple(target_point_pixel+1)], fill='Blue', outline='Blue')
draw.ellipse([tuple(pred_waypoint-1), tuple(pred_waypoint+1)], fill='Red', outline='Red')
if not os.path.isdir(self.save_path / 'flow' / str(frame).zfill(4)):
os.mkdir(self.save_path / 'flow' / str(frame).zfill(4))
flow_display.save(f"{self.save_path}/flow/{str(frame).zfill(4)}/{str(i)}.png")
def destroy(self):
del self.net