-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluation.py
275 lines (238 loc) · 11.3 KB
/
evaluation.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
from environment import *
from transformer import Transformer
from utils import get_device
import os
import time as t
from collections import deque
import copy
def evaluation(args):
# Determine if your system supports CUDA
cuda_available = torch.cuda.is_available()
device = get_device(cuda_available)
darp = Darp(args, mode='evaluate', device=device)
# Create a model
darp.model = Transformer(
device=device,
num_vehicles=darp.train_K,
input_seq_len=darp.train_N,
target_seq_len=darp.train_N + 2,
d_model=args.d_model,
num_layers=args.num_layers,
num_heads=args.num_heads,
d_k=args.d_k,
d_v=args.d_v,
d_ff=args.d_ff,
dropout=args.dropout)
# Load the trained model
model_name = darp.train_name + '-' + str(args.wait_time)
if args.model_type:
model = "rl"
print("Load the model trained by reinforcement learning.\n")
else:
model = "sl"
print("Load the model trained by supervised learning.\n")
checkpoint = torch.load('./model/' + model + '-' + model_name + '.model')
darp.model.load_state_dict(checkpoint['model_state_dict'])
darp.model.eval()
if cuda_available:
darp.model.cuda()
# Initialize the lists of metrics
eval_run_time = []
eval_time_penalty = []
eval_rist_cost = []
eval_pred_cost = []
eval_window = []
eval_ride_time = []
eval_not_same = []
eval_not_done = []
eval_rela_diff = []
# Set 'user_mask' and 'src_mask'
# user_mask = [1 for _ in range(9)] + \
# [1 for _ in range(darp.test_K)] + [0 for _ in range(darp.train_K - darp.test_K)]
src_mask = [1 for _ in range(darp.test_N)] + [0 for _ in range(darp.train_N - darp.test_N)]
# user_mask = torch.Tensor(user_mask).to(device)
src_mask = torch.Tensor(src_mask).to(device)
for num_instance in range(args.num_tt_instances):
print('--------Evaluation on Instance {}:--------'.format(num_instance + 1))
start = t.time()
true_cost = darp.reset(num_instance)
if args.beam:
darps = beam_search(darp, num_instance, src_mask, args.beam)
print('\n--------Beam results:--------')
for darp in darps:
print(round(darp[0].cost(), 2), round(abs(true_cost - darp[0].cost()) / true_cost * 100, 2),
len(darp[0].break_window), len(set(darp[0].break_ride_time)), round(darp[0].time_penalty, 2))
print()
darp, darp_cost = beam_choose(darps)
else:
darp_cost = greedy_evaluation(darp, num_instance, src_mask)
end = t.time()
# Evaluate the predicted results
for k in range(0, darp.test_K):
vehicle = darp.vehicles[k]
print('> Vehicle {}'.format(vehicle.id))
# Apply the node-user mapping
for r, node in enumerate(vehicle.route):
if 0 < node < 2 * darp.test_N + 1:
vehicle.route[r] = darp.node2user[node]
# Print the Rist's route of the vehicle
rist_route = zip(vehicle.route[1:-1], vehicle.schedule[1:-1])
print('Rist\'s route:', [term[0] for term in rist_route])
# Print the predicted route of the vehicle
pred_route = zip(vehicle.pred_route[1:-1], vehicle.pred_schedule[1:-1])
print('Predicted route:', [term[0] for term in pred_route])
run_time = end - start
# Append the lists of metrics
eval_rist_cost.append(true_cost)
eval_pred_cost.append(darp_cost)
eval_window.append(len(darp.break_window))
eval_ride_time.append(len(set(darp.break_ride_time)))
eval_not_same.append(len(darp.break_same))
eval_not_done.append(len(darp.break_done))
eval_rela_diff.append(abs(true_cost - darp_cost) / true_cost * 100)
eval_run_time.append(run_time)
eval_time_penalty.append(darp.time_penalty)
# Print the Rist's cost, the predicted cost, and the relative difference
print('> Objective')
print('Rist\'s cost: {:.4f}'.format(eval_rist_cost[-1]))
print('Predicted cost: {:.4f}'.format(eval_pred_cost[-1]))
print('Relative difference (%): {:.2f}'.format(eval_rela_diff[-1]))
# Print the number of broken constraints
print('> Constraint')
print('# broken time window: {}'.format(eval_window[-1]))
print('# broken ride time: {}\n'.format(eval_ride_time[-1]))
# Print the metrics on one standard instance
print('--------Metrics on one standard instance:--------')
print('Cost (Rist 2021): {:.2f}'.format(eval_rist_cost[0]))
print('Cost (predicted): {:.2f}'.format(eval_pred_cost[0]))
print('Diff. (%): {:.2f}'.format(eval_rela_diff[0]))
print('# Time Window: {}'.format(eval_window[0]))
print('# Ride Time: {}'.format(eval_ride_time[0]))
print('Time penalty: {:.2f}'.format(eval_time_penalty[0]))
print('Run time: {:.2f}\n'.format(eval_run_time[0]))
# Print the metrics on multiple random instances
print('--------Average metrics on {} random instances:--------'.format(args.num_tt_instances))
print('Aver. Cost (Rist 2021): {:.2f}'.format(sum(eval_rist_cost) / len(eval_rist_cost)))
print('Aver. Cost (predicted): {:.2f}'.format(sum(eval_pred_cost) / len(eval_pred_cost)))
print('Aver. Diff. (%): {:.2f}'.format(sum(eval_rela_diff) / len(eval_rela_diff)))
print('Aver. # Time Window: {:.2f}'.format(sum(eval_window) / len(eval_window)))
print('Aver. # Ride Time: {:.2f}'.format(sum(eval_ride_time) / len(eval_ride_time)))
print('Aver. Time penalty: {:.2f}'.format(sum(eval_time_penalty) / len(eval_time_penalty)))
print('Aver. Run time: {:.2f}'.format(sum(eval_run_time) / len(eval_run_time)))
# Print the number of problematic requests
print('# Not Same: {}'.format(np.sum(np.asarray(eval_not_same) > 0)))
print('# Not Done: {}'.format(np.sum(np.asarray(eval_not_done) > 0)))
path_result = './result/'
os.makedirs(path_result, exist_ok=True)
with open(path_result + 'evaluation.txt', 'a+') as output:
# Dump the parameters of training instances
output.write('Training instances -> ')
json.dump({
'Type': darp.train_type, 'K': darp.train_K, 'N': darp.train_N,
'T': darp.train_T, 'Q': darp.train_Q, 'L': darp.train_L,
}, output)
output.write('\n')
# Dump the parameters of test instances
output.write('Test instances -> ')
json.dump({
'Type': darp.test_type, 'K': darp.test_K, 'N': darp.test_N,
'T': darp.test_T, 'Q': darp.test_Q, 'L': darp.test_L,
}, output)
output.write('\n')
# Dump the metrics on one standard instance
json.dump({
'Cost (Rist 2021)': round(eval_rist_cost[0], 2),
'Cost (predicted)': round(eval_pred_cost[0], 2),
'Diff. (%)': round(eval_rela_diff[0], 2),
'# Time Window': eval_window[0],
'# Ride Time': eval_ride_time[0],
'Time penalty': round(eval_time_penalty[0], 2),
'Run time': round(eval_run_time[0], 2),
}, output)
output.write('\n')
# Dump the metrics on multiple random instances
json.dump({
'Aver. Cost (Rist 2021)': round(sum(eval_rist_cost) / len(eval_rist_cost), 2),
'Aver. Cost (predicted)': round(sum(eval_pred_cost) / len(eval_pred_cost), 2),
'Aver. Diff. (%)': round(sum(eval_rela_diff) / len(eval_rela_diff), 2),
'Aver. # Time Window': round(sum(eval_window) / len(eval_window), 2),
'Aver. # Ride Time': round(sum(eval_ride_time) / len(eval_ride_time), 2),
'Aver. Time penalty': round(sum(eval_time_penalty) / len(eval_time_penalty), 2),
'Aver. Run time': round(sum(eval_run_time) / len(eval_run_time), 2),
}, output)
output.write('\n')
# Dump the number of problematic requests
json.dump({
'# Not Same': int(np.sum(np.asarray(eval_not_same) > 0)),
'# Not Done': int(np.sum(np.asarray(eval_not_done) > 0)),
}, output)
output.write('\n')
def greedy_evaluation(darp, num_instance, src_mask=None, logs=True):
# Run the simulator
darp.log_probs = []
while darp.finish():
free_times = [vehicle.free_time for vehicle in darp.vehicles]
time = np.min(free_times)
indices = np.argwhere(free_times == time)
indices = indices.flatten().tolist()
for _, k in enumerate(indices):
if darp.vehicles[k].free_time == 1440:
continue
darp.beta(k)
state = darp.state(k, time)
action, probs = darp.predict(state, user_mask=None, src_mask=src_mask)
darp.log_probs.append(torch.log(probs.squeeze(0)[action]))
darp.evaluate_step(k, action)
return darp.cost()
def beam_search(darp, num_instance, src_mask, beam_width):
"""
Beam search algorithm for the DARP problem. Maintain
the best beam_width solutions at each time step and expand them to the next time step.
WARNING: Increases running time exponentially with beam_width.
a2-16 greedy takes ~10s while beam search with beam_width=10 takes ~1000s.
"""
# TODO: transpositions are possible, they need to be detected and removed from the beam
darp.load_from_file(num_instance)
k_best = [(darp, False, 0.0)] # (darp, finish, sumlogprob)
# Run the simulator
while sum([done for (env, done, score) in k_best]) < beam_width:
k_best_new = []
envs = {}
for i, (env, done, score) in enumerate(k_best):
if not done:
waiting = True
while waiting:
if not env.indices:
free_times = [vehicle.free_time for vehicle in env.vehicles]
time = np.min(free_times)
indices = np.argwhere(free_times == time)
env.indices = deque(indices.flatten().tolist())
env.time = time
k = env.indices.popleft()
if env.vehicles[k].free_time == 1440:
continue
env.beta(k)
state = env.state(k, env.time)
action, outputs = env.predict(state, user_mask=None, src_mask=src_mask)
if action == env.train_N + 1:
env.evaluate_step(k, action)
else:
waiting = False
log_probs, actions = torch.topk(torch.log(outputs.squeeze(0)[:-1]), min(beam_width, darp.train_N))
envs[i] = env
for log_prob, action in zip(log_probs, actions):
# expand each current candidate
k_best_new.append((i, score - log_prob.item(), k, action.item()))
# order by score, select k best
k_best_new = sorted(k_best_new, key=lambda x: x[1])[:beam_width]
# step the env in potential envs
k_best = []
for (i, score, k, action) in k_best_new:
env = copy.deepcopy(envs[i])
env.evaluate_step(k, action)
k_best.append((env, not env.finish(), score))
return k_best
def beam_choose(darps):
idx = np.argmin([darp[0].time_penalty for darp in darps])
darp = darps[idx]
return darp[0], darp[0].cost()