Spaces:
Runtime error
Runtime error
File size: 6,729 Bytes
f75943a 4121f92 0888538 4121f92 f75943a 0888538 f75943a 0888538 f75943a 0888538 f75943a 0888538 f75943a 0888538 f75943a 0888538 f75943a 4e231e2 |
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 |
import os
import sys
import torch
import torch.nn as nn
from datetime import datetime
from torch.nn.utils import clip_grad_norm_
import torch.optim as optim
from torch.utils.data import DataLoader
import json
dir_path = os.path.dirname(os.path.realpath(__file__))
sys.path.append(os.path.join(dir_path, '..'))
from nets.model import Model
from Actor.actor import Actor
from dataloader import VRP_Dataset
from google_solver.google_model import evaluate_google_model
# Load params
with open('params.json', 'r') as f:
params = json.load(f)
# Save params into a local file for tracking
with open('params_saved.json', 'w') as f:
json.dump(params, f)
# Settings
device = params['device']
run_tests = params['run_tests']
save_results = params['save_results']
dataset_path = params['dataset_path']
# Dataset sizes
train_dataset_size = params['train_dataset_size']
validation_dataset_size = params['validation_dataset_size']
baseline_dataset_size = params['baseline_dataset_size']
# Problem config
num_nodes = params['num_nodes']
num_depots = params['num_depots']
embedding_size = params['embedding_size']
sample_size = params['sample_size']
gradient_clipping = params['gradient_clipping']
num_neighbors_encoder = params['num_neighbors_encoder']
num_neighbors_action = params['num_neighbors_action']
num_movers = params['num_movers']
learning_rate = params['learning_rate']
batch_size = params['batch_size']
test_batch_size = params['test_batch_size']
baseline_update_period = params['baseline_update_period']
# Datasets
validation_dataset = VRP_Dataset(validation_dataset_size, num_nodes, num_depots, dataset_path, device)
baseline_dataset = VRP_Dataset(train_dataset_size, num_nodes, num_depots, dataset_path, device)
if params['overfit_test']:
train_dataset = VRP_Dataset(train_dataset_size, num_nodes, num_depots, dataset_path, device)
baseline_dataset = train_dataset
validation_dataset = train_dataset
# Evaluate Google solver
google_scores = evaluate_google_model(validation_dataset)
tot_google_scores = google_scores.sum().item()
input_size = validation_dataset.model_input_length()
# Models
model = Model(input_size=input_size, embedding_size=embedding_size, decoder_input_size=params["decoder_input_size"])
actor = Actor(model=model, num_movers=num_movers, num_neighbors_encoder=num_neighbors_encoder,
num_neighbors_action=num_neighbors_action, device=device, normalize=False)
actor.train_mode()
baseline_model = Model(input_size=input_size, embedding_size=embedding_size, decoder_input_size=params["decoder_input_size"])
baseline_actor = Actor(model=baseline_model, num_movers=num_movers, num_neighbors_encoder=num_neighbors_encoder,
num_neighbors_action=num_neighbors_action, device=device, normalize=False)
baseline_actor.greedy_search()
baseline_actor.load_state_dict(actor.state_dict())
nn_actor = Actor(model=None, num_movers=1, num_neighbors_action=1, device=device)
nn_actor.nearest_neighbors()
optimizer = optim.Adam(params=actor.parameters(), lr=learning_rate)
train_batch_record = 100
validation_record = 100
baseline_record = None
# Training loop
for epoch in range(params['num_epochs']):
if not params['overfit_test']:
train_dataset = VRP_Dataset(train_dataset_size, num_nodes, num_depots, dataset_path, device)
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, collate_fn=train_dataset.collate)
for i, batch in enumerate(train_dataloader):
with torch.no_grad():
nn_actor.nearest_neighbors()
nn_output = nn_actor(batch)
tot_nn_cost = nn_output['total_time'].sum().item()
baseline_actor.greedy_search()
baseline_cost = baseline_actor(batch)['total_time']
actor.train_mode()
actor_output = actor(batch)
actor_cost, log_probs = actor_output['total_time'], actor_output['log_probs']
loss = ((actor_cost - baseline_cost).detach() * log_probs).mean()
optimizer.zero_grad()
loss.backward()
if gradient_clipping:
for group in optimizer.param_groups:
clip_grad_norm_(group['params'], 1, norm_type=2)
optimizer.step()
tot_actor_cost = actor_cost.sum().item()
tot_baseline_cost = baseline_cost.sum().item()
actor_nn_ratio = tot_actor_cost / tot_nn_cost
actor_baseline_ratio = tot_actor_cost / tot_baseline_cost
train_batch_record = min(train_batch_record, actor_nn_ratio)
result = f"{epoch}, {i}, {actor_nn_ratio:.4f}, {actor_baseline_ratio:.4f}, {train_batch_record:.4f}"
print(result, flush=True)
if save_results:
with open('train_results.txt', 'a') as f:
f.write(result + '\n')
del batch
# Validation
if epoch % 5 == 0:
baseline_dataloader = DataLoader(baseline_dataset, batch_size=batch_size, collate_fn=baseline_dataset.collate)
tot_cost = []
for batch in baseline_dataloader:
with torch.no_grad():
actor.greedy_search()
actor_output = actor(batch)
cost = actor_output['total_time']
tot_cost.append(cost)
del batch
tot_cost = torch.cat(tot_cost, dim=0)
if baseline_record is None or (tot_cost < baseline_record).float().mean().item() > 0.9:
baseline_record = tot_cost
baseline_actor.load_state_dict(actor.state_dict())
print('\nNew baseline record\n')
# Test every 10 epochs
if (epoch % 10 == 0) and run_tests:
b = max(int(batch_size // sample_size**2), 1)
validation_dataloader = DataLoader(validation_dataset, batch_size=b, collate_fn=validation_dataset.collate)
tot_cost = 0
tot_nn_cost = 0
for batch in validation_dataloader:
with torch.no_grad():
actor.beam_search(sample_size)
actor_output = actor(batch)
cost = actor_output['total_time']
nn_actor.nearest_neighbors()
nn_output = nn_actor(batch)
nn_cost = nn_output['total_time']
tot_cost += cost.sum().item()
tot_nn_cost += nn_cost.sum().item()
ratio = tot_cost / tot_nn_cost
validation_record = min(validation_record, ratio)
actor_google_ratio = tot_cost / tot_google_scores
print(f"\nTest results:\nActor/Google: {actor_google_ratio:.4f}, Actor/NN: {ratio:.4f}, Best NN Ratio: {validation_record:.4f}\n")
if save_results:
with open('test_results.txt', 'a') as f:
f.write(f"{epoch}, {actor_google_ratio:.4f}, {ratio:.4f}, {validation_record:.4f}\n")
print("End")
|