route-explainer / utils /data_utils /pctsptw_dataset.py
daisuke.kikuta
first commit
719d0db
raw
history blame
15 kB
import os
import random
import numpy as np
import torch
from utils.utils import load_dataset, save_dataset
from scipy.spatial.distance import cdist
from utils.data_utils.dataset_base import DatasetBase, DataLoaderBase
from utils.data_utils.pctsp_dataset import get_total_prizes, get_total_penalty
from utils.data_utils.tsptw_dataset import get_tw_mask2
from models.classifiers.ground_truth.ground_truth_base import get_visited_mask
from models.solvers.general_solver import GeneralSolver
from models.classifiers.ground_truth.ground_truth_pctsptw import GroundTruthPCTSPTW
class PCTSPTWDataset(DatasetBase):
def __init__(self, coord_dim, num_samples, num_nodes, solver="ortools", classifier="ortools", annotation=True, parallel=True, random_seed=1234, num_cpus=os.cpu_count(),
penalty_factor=3.):
"""
Parameters
----------
num_samples: int
number of samples(instances)
num_nodes: int
number of nodes
grid_size: int or float32
x-pos/y-pos of cities will be in the range [0, grid_size]
max_tw_gap:
maximum time windows gap allowed between the cities constituing the feasible tour
max_tw_size:
time windows of cities will be in the range [0, max_tw_size]
is_integer_instance: bool
True if we want the distances and time widows to have integer values
seed: int
seed used for generating the instance. -1 means no seed (instance is random)
"""
super().__init__(coord_dim, num_samples, num_nodes, annotation, parallel, random_seed, num_cpus)
self.penalty_factor = penalty_factor
MAX_LENGTHS = {
20: 2.,
50: 3.,
100: 4.
}
self.max_length = MAX_LENGTHS[num_nodes]
solver_type = solver
classifier_type = classifier
problem = "pctsptw"
distribution="da_silva"
max_tw_gap=10
MAX_TW_COEFF = {
20: 1,
50: 5,
100: 10
}
self.da_silva_style = distribution == "da_silva"
self.max_tw_size = MAX_TW_COEFF[num_nodes] * 1000 if self.da_silva_style else 100
self.max_tw_gap = max_tw_gap
self.pctsptw_solver = GeneralSolver(problem=problem, solver_type=solver_type)
self.classifier = GroundTruthPCTSPTW(solver_type=classifier_type)
def generate_instance(self, seed):
"""
Minor change of https://github.com/wouterkool/attention-learn-to-route/blob/master/problems/pctsp/problem_pctsp.py
"""
if seed is not None:
np.random.seed(seed)
rand = random.Random()
rand.seed(seed)
#-----------------------------
# generate locations of nodes
#-----------------------------
coords = np.random.uniform(size=(self.num_nodes+1, self.coord_dim))
# For the penalty to make sense it should be not too large (in which case all nodes will be visited) nor too small
# so we want the objective term to be approximately equal to the length of the tour, which we estimate with half
# of the nodes by half of the tour length (which is very rough but similar to op)
# This means that the sum of penalties for all nodes will be approximately equal to the tour length (on average)
# The expected total (uniform) penalty of half of the nodes (since approx half will be visited by the constraint)
# is (n / 2) / 2 = n / 4 so divide by this means multiply by 4 / n,
# However instead of 4 we use penalty_factor (3 works well) so we can make them larger or smaller
penalty_max = self.max_length * (self.penalty_factor) / float(self.num_nodes)
penalties = np.random.uniform(size=(self.num_nodes+1, )) * penalty_max
# Take uniform prizes
# Now expectation is 0.5 so expected total prize is n / 2, we want to force to visit approximately half of the nodes
# so the constraint will be that total prize >= (n / 2) / 2 = n / 4
# equivalently, we divide all prizes by n / 4 and the total prize should be >= 1
deterministic_prizes = np.random.uniform(size=(self.num_nodes+1, )) * 4 / float(self.num_nodes)
deterministic_prizes[0] = 0.0 # Prize at the depot is zero
#-------------
# time window
#-------------
# dist = np.sqrt(((coords[0:1] - coords) ** 2).sum(-1)) * 100
# # define sampling horizon
# a0 = 0; b0 = 1000
# a_sample = np.floor(dist) + 1
# b_sample = b0 - a_sample - 10
# # sample horizon of each node
# a = np.random.uniform(size=(self.num_nodes+1,))
# a = (a * (b_sample - a_sample) + a_sample).astype("int")
# eps = np.maximum(np.abs(np.random.normal(0, 1, (self.num_nodes+1,))), 0.01)
# b = np.minimum(np.ceil(a + 300 * eps), b_sample)
# a[0] = a0; b[0] = b0
# a = a / 100
# b = b / 100
# time_window = np.concatenate((a[:, None], b[:, None]), -1)
self.grid_size = 100
random_solution = list(range(1, self.num_nodes+1))
rand.shuffle(random_solution)
random_solution = [0] + random_solution # add the depot (node_id=0)
travel_time = cdist(coords, coords) * self.grid_size # [num_nodes x num_nodes]
time_windows = np.zeros((self.num_nodes+1, 2))
time_windows[0, :] = [0, 1000 * self.grid_size] # time window for the depot
total_dist = 0
for i in range(1, self.num_nodes+1):
prev_node_id = random_solution[i-1]
cur_node_id = random_solution[i]
cur_dist = travel_time[prev_node_id][cur_node_id]
tw_lb_min = time_windows[prev_node_id, 0] + cur_dist
total_dist += cur_dist
if self.da_silva_style:
# Style by Da Silva and Urrutia, 2010, "A VNS Heuristic for TSPTW"
rand_tw_lb = rand.uniform(total_dist - self.max_tw_size / 2, total_dist)
rand_tw_ub = rand.uniform(total_dist, total_dist + self.max_tw_size / 2)
else:
# Cappart et al. style 'propagates' the time windows resulting in little overlap / easier instances
rand_tw_lb = rand.uniform(tw_lb_min, tw_lb_min + self.max_tw_gap)
rand_tw_ub = rand.uniform(rand_tw_lb, rand_tw_lb + self.max_tw_size)
time_windows[cur_node_id, :] = [rand_tw_lb, rand_tw_ub] # [num_nodes x 2(start, end)]
return {
"coords": coords,
"penalties": penalties,
"prizes": deterministic_prizes,
"time_window": time_windows / self.grid_size,
"min_prize": np.min([np.sum(deterministic_prizes), 1.0]),
"grid_size": np.array([1.0])
}
def annotate(self, instance):
# solve PCTSPTW
node_feats = instance
pctsptw_tour = self.pctsptw_solver.solve(node_feats)
# print(pctsptw_tour)
if pctsptw_tour is None:
return
# annotate each path
inputs = self.classifier.get_inputs(pctsptw_tour, 0, node_feats)
labels = self.classifier(inputs, annotation=True)
if labels is None:
return
instance.update({"tour": pctsptw_tour, "labels": labels})
return instance
class PCTSPTWDataloader(DataLoaderBase):
# @override
def load_randomly(self, instance, fname=None):
data = []
coords = torch.FloatTensor(instance["coords"]) # [num_nodes x coord_dim]
prizes = torch.FloatTensor(instance["prizes"]) # [num_nodes x 1]
penalties = torch.FloatTensor(instance["penalties"]) # [num_nodes x 1]
raw_time_window = torch.FloatTensor(instance["time_window"]).clamp(0.0)
time_window = torch.FloatTensor(instance["time_window"]).clamp(0.0) # [num_nodes x 2]
time_window = (time_window - time_window[1:].min()) / (time_window[1:].max() - time_window[1:].min()) # min-max normalization
node_feats = torch.cat((coords, prizes[:, None], penalties[:, None], time_window), -1) # [num_nodes x (coord_dim + 4)]
tours = instance["tour"]
labels = instance["labels"]
for vehicle_id in range(len(labels)):
for step, label in labels[vehicle_id]:
visited = get_visited_mask(tours[vehicle_id], step, instance)
curr_prize = get_total_prizes(tours[vehicle_id], step, instance)
curr_penalty = get_total_penalty(visited, instance)
not_exceed_tw, curr_time = get_tw_mask2(tours[vehicle_id], step, instance)
curr_time = ((curr_time - raw_time_window[1:].min()) / (raw_time_window[1:].max() - raw_time_window[1:].min())).item()
mask = torch.from_numpy((~visited) & not_exceed_tw)
mask[0] = True # depot is always feasible
data.append({
"node_feats": node_feats,
"curr_node_id": torch.tensor(tours[vehicle_id][step-1]).to(torch.long),
"next_node_id": torch.tensor(tours[vehicle_id][step]).to(torch.long),
"mask": mask,
"state": torch.FloatTensor([curr_prize, curr_penalty, curr_time]),
"labels": torch.tensor(label).to(torch.long)
})
if fname is not None:
save_dataset(data, fname, display=False)
return fname
else:
return data
# @override
def load_sequentially(self, instance, fname=None):
data = []
coords = torch.FloatTensor(instance["coords"]) # [num_nodes x coord_dim]
prizes = torch.FloatTensor(instance["prizes"]) # [num_nodes x 1]
penalties = torch.FloatTensor(instance["penalties"]) # [num_nodes x 1]
raw_time_window = torch.FloatTensor(instance["time_window"]).clamp(0.0)
time_window = torch.FloatTensor(instance["time_window"]).clamp(0.0) # [num_nodes x 2]
time_window = (time_window - time_window[1:].min()) / (time_window[1:].max() - time_window[1:].min()) # min-max normalization
node_feats = torch.cat((coords, prizes[:, None], penalties[:, None], time_window), -1) # [num_nodes x (coord_dim + 4)]
tours = instance["tour"]
labels = instance["labels"]
num_nodes, node_dim = node_feats.size()
for vehicle_id in range(len(labels)):
seq_length = len(labels[vehicle_id])
curr_node_id_list = []; next_node_id_list = []
mask_list = []; state_list = []; label_list = []
for step, label in labels[vehicle_id]:
visited = get_visited_mask(tours[vehicle_id], step, instance)
curr_prize = get_total_prizes(tours[vehicle_id], step, instance)
curr_penalty = get_total_penalty(visited, instance)
not_exceed_tw, curr_time = get_tw_mask2(tours[vehicle_id], step, instance)
curr_time = ((curr_time - raw_time_window[1:].min()) / (raw_time_window[1:].max() - raw_time_window[1:].min())).item()
mask = torch.from_numpy((~visited) & not_exceed_tw)
mask[0] = True # depot is always feasible
# add values to the lists
curr_node_id_list.append(tours[vehicle_id][step-1])
next_node_id_list.append(tours[vehicle_id][step])
mask_list.append(mask)
state_list.append([curr_prize, curr_penalty, curr_time])
label_list.append(label)
data.append({
"node_feats": node_feats.unsqueeze(0).expand(seq_length, num_nodes, node_dim), # [seq_length x num_nodes x node_feats]
"curr_node_id": torch.LongTensor(curr_node_id_list), # [seq_length]
"next_node_id": torch.LongTensor(next_node_id_list), # [seq_length]
"mask": torch.stack(mask_list, 0), # [seq_length x num_nodes]
"state": torch.FloatTensor(state_list), # [seq_length x state_dim(1)]
"labels": torch.LongTensor(label_list) # [seq_length]
})
if fname is not None:
save_dataset(data, fname, display=False)
return fname
else:
return data
def load_pctsptw_sequentially(instance, fname=None):
data = []
coords = torch.FloatTensor(instance["coords"]) # [num_nodes x coord_dim]
prizes = torch.FloatTensor(instance["prizes"]) # [num_nodes x 1]
penalties = torch.FloatTensor(instance["penalties"]) # [num_nodes x 1]
raw_time_window = torch.FloatTensor(instance["time_window"]).clamp(0.0)
time_window = torch.FloatTensor(instance["time_window"]).clamp(0.0) # [num_nodes x 2]
time_window = (time_window - time_window[1:].min()) / (time_window[1:].max() - time_window[1:].min()) # min-max normalization
node_feats = torch.cat((coords, prizes[:, None], penalties[:, None], time_window), -1) # [num_nodes x (coord_dim + 4)]
tours = instance["tour"]
labels = instance["labels"]
num_nodes, node_dim = node_feats.size()
for vehicle_id in range(len(labels)):
seq_length = len(tours[vehicle_id])
curr_node_id_list = []; next_node_id_list = []
mask_list = []; state_list = []
for step in range(1, len(tours[vehicle_id])):
visited = get_visited_mask(tours[vehicle_id], step, instance)
curr_prize = get_total_prizes(tours[vehicle_id], step, instance)
curr_penalty = get_total_penalty(visited, instance)
not_exceed_tw, curr_time = get_tw_mask2(tours[vehicle_id], step, instance)
curr_time = ((curr_time - raw_time_window[1:].min()) / (raw_time_window[1:].max() - raw_time_window[1:].min())).item()
mask = torch.from_numpy((~visited) & not_exceed_tw)
mask[0] = True # depot is always feasible
# add values to the lists
curr_node_id_list.append(tours[vehicle_id][step-1])
next_node_id_list.append(tours[vehicle_id][step])
mask_list.append(mask)
state_list.append([curr_prize, curr_penalty, curr_time])
data.append({
"node_feats": node_feats.unsqueeze(0).expand(seq_length, num_nodes, node_dim), # [seq_length x num_nodes x node_feats]
"curr_node_id": torch.LongTensor(curr_node_id_list), # [seq_length]
"next_node_id": torch.LongTensor(next_node_id_list), # [seq_length]
"mask": torch.stack(mask_list, 0), # [seq_length x num_nodes]
"state": torch.FloatTensor(state_list), # [seq_length x state_dim(1)]
})
if fname is not None:
save_dataset(data, fname, display=False)
return fname
else:
return data