iicd/xautodl
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update REA, REINFORCE, and RANDOM

Browse Source
This commit is contained in:
D-X-Y
2020-07-13 11:35:13 +00:00
parent 6dc494be08
commit ebad9197f7
5 changed files with 38 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
exps/algos-v2/random_wo_share.py
View File

@@ -4,6 +4,8 @@
# Random Search for Hyper-Parameter Optimization, JMLR 2012 ################## # Random Search for Hyper-Parameter Optimization, JMLR 2012 ##################
############################################################################## ##############################################################################
# python ./exps/algos-v2/random_wo_share.py --dataset cifar10 --search_space tss # python ./exps/algos-v2/random_wo_share.py --dataset cifar10 --search_space tss
# python ./exps/algos-v2/random_wo_share.py --dataset cifar100 --search_space tss
# python ./exps/algos-v2/random_wo_share.py --dataset ImageNet16-120 --search_space tss
############################################################################## ##############################################################################
import os, sys, time, glob, random, argparse import os, sys, time, glob, random, argparse
import numpy as np, collections import numpy as np, collections
@@ -20,7 +22,7 @@ from utils import get_model_infos, obtain_accuracy
from log_utils import AverageMeter, time_string, convert_secs2time from log_utils import AverageMeter, time_string, convert_secs2time
from models import get_search_spaces from models import get_search_spaces
from nas_201_api import NASBench201API, NASBench301API from nas_201_api import NASBench201API, NASBench301API
from .regularized_ea import random_topology_func, random_size_func from regularized_ea import random_topology_func, random_size_func
def main(xargs, api): def main(xargs, api):
@@ -28,16 +30,18 @@ def main(xargs, api):
prepare_seed(xargs.rand_seed) prepare_seed(xargs.rand_seed)
logger = prepare_logger(args) logger = prepare_logger(args)
logger.log('{:} use api : {:}'.format(time_string(), api))
api.reset_time()
search_space = get_search_spaces(xargs.search_space, 'nas-bench-301') search_space = get_search_spaces(xargs.search_space, 'nas-bench-301')
if xargs.search_space == 'tss': if xargs.search_space == 'tss':
random_arch = random_topology_func(search_space) random_arch = random_topology_func(search_space)
else: else:
random_arch = random_size_func(search_space) random_arch = random_size_func(search_space)
x_start_time = time.time()
logger.log('{:} use nas_bench : {:}'.format(time_string(), nas_bench))
best_arch, best_acc, total_time_cost, history = None, -1, [], [] best_arch, best_acc, total_time_cost, history = None, -1, [], []
while total_time_cost[-1] < xargs.time_budget: current_best_index = []
while len(total_time_cost) == 0 or total_time_cost[-1] < xargs.time_budget:
arch = random_arch() arch = random_arch()
accuracy, _, _, total_cost = api.simulate_train_eval(arch, xargs.dataset, '12') accuracy, _, _, total_cost = api.simulate_train_eval(arch, xargs.dataset, '12')
total_time_cost.append(total_cost) total_time_cost.append(total_cost)
@@ -45,13 +49,14 @@ def main(xargs, api):
if best_arch is None or best_acc < accuracy: if best_arch is None or best_acc < accuracy:
best_acc, best_arch = accuracy, arch best_acc, best_arch = accuracy, arch
logger.log('[{:03d}] : {:} : accuracy = {:.2f}%'.format(len(history), arch, accuracy)) logger.log('[{:03d}] : {:} : accuracy = {:.2f}%'.format(len(history), arch, accuracy))
logger.log('{:} best arch is {:}, accuracy = {:.2f}%, visit {:} archs with {:.1f} s (real-cost = {:.3f} s).'.format(time_string(), best_arch, best_acc, len(history), total_time_cost, time.time()-x_start_time)) current_best_index.append(api.query_index_by_arch(best_arch))
logger.log('{:} best arch is {:}, accuracy = {:.2f}%, visit {:} archs with {:.1f} s.'.format(time_string(), best_arch, best_acc, len(history), total_time_cost[-1]))
info = api.query_info_str_by_arch(best_arch, '200' if xargs.search_space == 'tss' else '90') info = api.query_info_str_by_arch(best_arch, '200' if xargs.search_space == 'tss' else '90')
logger.log('{:}'.format(info)) logger.log('{:}'.format(info))
logger.log('-'*100) logger.log('-'*100)
logger.close() logger.close()
return logger.log_dir, total_time_cost, history return logger.log_dir, current_best_index, total_time_cost
if __name__ == '__main__': if __name__ == '__main__':
@@ -62,7 +67,7 @@ if __name__ == '__main__':
parser.add_argument('--time_budget', type=int, default=20000, help='The total time cost budge for searching (in seconds).') parser.add_argument('--time_budget', type=int, default=20000, help='The total time cost budge for searching (in seconds).')
parser.add_argument('--loops_if_rand', type=int, default=500, help='The total runs for evaluation.') parser.add_argument('--loops_if_rand', type=int, default=500, help='The total runs for evaluation.')
# log # log
parser.add_argument('--save_dir', type=str, help='Folder to save checkpoints and log.') parser.add_argument('--save_dir', type=str, default='./output/search', help='Folder to save checkpoints and log.')
parser.add_argument('--rand_seed', type=int, default=-1, help='manual seed') parser.add_argument('--rand_seed', type=int, default=-1, help='manual seed')
args = parser.parse_args() args = parser.parse_args()
@@ -77,7 +82,7 @@ if __name__ == '__main__':
print('save-dir : {:}'.format(args.save_dir)) print('save-dir : {:}'.format(args.save_dir))
if args.rand_seed < 0: if args.rand_seed < 0:
save_dir, all_info = None, {} save_dir, all_info = None, collections.OrderedDict()
for i in range(args.loops_if_rand): for i in range(args.loops_if_rand):
print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand)) print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand))
args.rand_seed = random.randint(1, 100000) args.rand_seed = random.randint(1, 100000)

18
exps/algos-v2/regularized_ea.py
View File

@@ -155,7 +155,7 @@ def regularized_evolution(cycles, population_size, sample_size, time_budget, ran
population = collections.deque() population = collections.deque()
api.reset_time() api.reset_time()
history, total_time_cost = [], [] # Not used by the algorithm, only used to report results. history, total_time_cost = [], [] # Not used by the algorithm, only used to report results.
current_best_index = []
# Initialize the population with random models. # Initialize the population with random models.
while len(population) < population_size: while len(population) < population_size:
model = Model() model = Model()
@@ -163,8 +163,9 @@ def regularized_evolution(cycles, population_size, sample_size, time_budget, ran
model.accuracy, _, _, total_cost = api.simulate_train_eval(model.arch, dataset, '12') model.accuracy, _, _, total_cost = api.simulate_train_eval(model.arch, dataset, '12')
# Append the info # Append the info
population.append(model) population.append(model)
history.append(model) history.append((model.accuracy, model.arch))
total_time_cost.append(total_cost) total_time_cost.append(total_cost)
current_best_index.append(api.query_index_by_arch(max(history, key=lambda x: x[0])[1]))
# Carry out evolution in cycles. Each cycle produces a model and removes another. # Carry out evolution in cycles. Each cycle produces a model and removes another.
while total_time_cost[-1] < time_budget: while total_time_cost[-1] < time_budget:
@@ -183,15 +184,16 @@ def regularized_evolution(cycles, population_size, sample_size, time_budget, ran
# Create the child model and store it. # Create the child model and store it.
child = Model() child = Model()
child.arch = mutate_arch(parent.arch) child.arch = mutate_arch(parent.arch)
child.accuracy, _, _, total_cost = api.simulate_train_eval(model.arch, dataset, '12') child.accuracy, _, _, total_cost = api.simulate_train_eval(child.arch, dataset, '12')
# Append the info # Append the info
population.append(child) population.append(child)
history.append(child) history.append((child.accuracy, child.arch))
current_best_index.append(api.query_index_by_arch(max(history, key=lambda x: x[0])[1]))
total_time_cost.append(total_cost) total_time_cost.append(total_cost)
# Remove the oldest model. # Remove the oldest model.
population.popleft() population.popleft()
return history, total_time_cost return history, current_best_index, total_time_cost
def main(xargs, api): def main(xargs, api):
@@ -210,7 +212,7 @@ def main(xargs, api):
x_start_time = time.time() x_start_time = time.time()
logger.log('{:} use api : {:}'.format(time_string(), api)) logger.log('{:} use api : {:}'.format(time_string(), api))
logger.log('-'*30 + ' start searching with the time budget of {:} s'.format(xargs.time_budget)) logger.log('-'*30 + ' start searching with the time budget of {:} s'.format(xargs.time_budget))
history, total_times = regularized_evolution(xargs.ea_cycles, xargs.ea_population, xargs.ea_sample_size, xargs.time_budget, random_arch, mutate_arch, api, xargs.dataset) history, current_best_index, total_times = regularized_evolution(xargs.ea_cycles, xargs.ea_population, xargs.ea_sample_size, xargs.time_budget, random_arch, mutate_arch, api, xargs.dataset)
logger.log('{:} regularized_evolution finish with history of {:} arch with {:.1f} s (real-cost={:.2f} s).'.format(time_string(), len(history), total_times[-1], time.time()-x_start_time)) logger.log('{:} regularized_evolution finish with history of {:} arch with {:.1f} s (real-cost={:.2f} s).'.format(time_string(), len(history), total_times[-1], time.time()-x_start_time))
best_arch = max(history, key=lambda i: i.accuracy) best_arch = max(history, key=lambda i: i.accuracy)
best_arch = best_arch.arch best_arch = best_arch.arch
@@ -220,7 +222,7 @@ def main(xargs, api):
logger.log('{:}'.format(info)) logger.log('{:}'.format(info))
logger.log('-'*100) logger.log('-'*100)
logger.close() logger.close()
return logger.log_dir, [api.query_index_by_arch(x.arch) for x in history], total_times return logger.log_dir, current_best_index, total_times
if __name__ == '__main__': if __name__ == '__main__':
@@ -249,7 +251,7 @@ if __name__ == '__main__':
print('save-dir : {:}'.format(args.save_dir)) print('save-dir : {:}'.format(args.save_dir))
if args.rand_seed < 0: if args.rand_seed < 0:
save_dir, all_info = None, {} save_dir, all_info = None, collections.OrderedDict()
for i in range(args.loops_if_rand): for i in range(args.loops_if_rand):
print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand)) print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand))
args.rand_seed = random.randint(1, 100000) args.rand_seed = random.randint(1, 100000)

10
exps/algos-v2/reinforce.py
View File

@@ -145,6 +145,7 @@ def main(xargs, api):
x_start_time = time.time() x_start_time = time.time()
logger.log('Will start searching with time budget of {:} s.'.format(xargs.time_budget)) logger.log('Will start searching with time budget of {:} s.'.format(xargs.time_budget))
total_steps, total_costs, trace = 0, [], [] total_steps, total_costs, trace = 0, [], []
current_best_index = []
while len(total_costs) == 0 or total_costs[-1] < xargs.time_budget: while len(total_costs) == 0 or total_costs[-1] < xargs.time_budget:
start_time = time.time() start_time = time.time()
log_prob, action = select_action( policy ) log_prob, action = select_action( policy )
@@ -162,9 +163,8 @@ def main(xargs, api):
# accumulate time # accumulate time
total_steps += 1 total_steps += 1
logger.log('step [{:3d}] : average-reward={:.3f} : policy_loss={:.4f} : {:}'.format(total_steps, baseline.value(), policy_loss.item(), policy.genotype())) logger.log('step [{:3d}] : average-reward={:.3f} : policy_loss={:.4f} : {:}'.format(total_steps, baseline.value(), policy_loss.item(), policy.genotype()))
#logger.log('----> {:}'.format(policy.arch_parameters)) # to analyze
#logger.log('') current_best_index.append(api.query_index_by_arch(max(trace, key=lambda x: x[0])[1]))
# best_arch = policy.genotype() # first version # best_arch = policy.genotype() # first version
best_arch = max(trace, key=lambda x: x[0])[1] best_arch = max(trace, key=lambda x: x[0])[1]
logger.log('REINFORCE finish with {:} steps and {:.1f} s (real cost={:.3f}).'.format(total_steps, total_costs[-1], time.time()-x_start_time)) logger.log('REINFORCE finish with {:} steps and {:.1f} s (real cost={:.3f}).'.format(total_steps, total_costs[-1], time.time()-x_start_time))
@@ -173,7 +173,7 @@ def main(xargs, api):
logger.log('-'*100) logger.log('-'*100)
logger.close() logger.close()
return logger.log_dir, [api.query_index_by_arch(x[1]) for x in trace], total_costs return logger.log_dir, current_best_index, total_costs
if __name__ == '__main__': if __name__ == '__main__':
@@ -203,7 +203,7 @@ if __name__ == '__main__':
print('save-dir : {:}'.format(args.save_dir)) print('save-dir : {:}'.format(args.save_dir))
if args.rand_seed < 0: if args.rand_seed < 0:
save_dir, all_info = None, {} save_dir, all_info = None, collections.OrderedDict()
for i in range(args.loops_if_rand): for i in range(args.loops_if_rand):
print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand)) print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand))
args.rand_seed = random.randint(1, 100000) args.rand_seed = random.randint(1, 100000)

1
exps/algos-v2/run-all.sh
View File

@@ -13,5 +13,6 @@ do
do do
python ./exps/algos-v2/reinforce.py --dataset ${dataset} --search_space ${search_space} --learning_rate 0.001 python ./exps/algos-v2/reinforce.py --dataset ${dataset} --search_space ${search_space} --learning_rate 0.001
python ./exps/algos-v2/regularized_ea.py --dataset ${dataset} --search_space ${search_space} --ea_cycles 200 --ea_population 10 --ea_sample_size 3 python ./exps/algos-v2/regularized_ea.py --dataset ${dataset} --search_space ${search_space} --ea_cycles 200 --ea_population 10 --ea_sample_size 3
python ./exps/algos-v2/random_wo_share.py --dataset ${dataset} --search_space ${search_space}
done done
done done

14
exps/experimental/vis-bench-algos.py
View File

@@ -3,7 +3,7 @@
############################################################### ###############################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.06 # # Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.06 #
############################################################### ###############################################################
# Usage: python exps/experimental/vis-bench-algos.py # Usage: python exps/experimental/vis-bench-algos.py #
############################################################### ###############################################################
import os, sys, time, torch, argparse import os, sys, time, torch, argparse
import numpy as np import numpy as np
@@ -30,6 +30,7 @@ def fetch_data(root_dir='./output/search', search_space='tss', dataset=None):
alg2name, alg2path = OrderedDict(), OrderedDict() alg2name, alg2path = OrderedDict(), OrderedDict()
alg2name['REA'] = 'R-EA-SS3' alg2name['REA'] = 'R-EA-SS3'
alg2name['REINFORCE'] = 'REINFORCE-0.001' alg2name['REINFORCE'] = 'REINFORCE-0.001'
# alg2name['RANDOM'] = 'RANDOM'
for alg, name in alg2name.items(): for alg, name in alg2name.items():
alg2path[alg] = os.path.join(ss_dir, dataset, name, 'results.pth') alg2path[alg] = os.path.join(ss_dir, dataset, name, 'results.pth')
assert os.path.isfile(alg2path[alg]) assert os.path.isfile(alg2path[alg])
@@ -62,14 +63,15 @@ def visualize_curve(api, vis_save_dir, search_space, max_time):
vis_save_dir = vis_save_dir.resolve() vis_save_dir = vis_save_dir.resolve()
vis_save_dir.mkdir(parents=True, exist_ok=True) vis_save_dir.mkdir(parents=True, exist_ok=True)
dpi, width, height = 250, 4700, 1500 dpi, width, height = 250, 5100, 1500
figsize = width / float(dpi), height / float(dpi) figsize = width / float(dpi), height / float(dpi)
LabelSize, LegendFontsize = 14, 14 LabelSize, LegendFontsize = 14, 14
def sub_plot_fn(ax, dataset): def sub_plot_fn(ax, dataset):
alg2data = fetch_data(search_space=search_space, dataset=dataset) alg2data = fetch_data(search_space=search_space, dataset=dataset)
alg2accuracies = OrderedDict() alg2accuracies = OrderedDict()
time_tickets = [float(i) / 100 * max_time for i in range(100)] total_tickets = 150
time_tickets = [float(i) / total_tickets * max_time for i in range(total_tickets)]
colors = ['b', 'g', 'c', 'm', 'y'] colors = ['b', 'g', 'c', 'm', 'y']
for idx, (alg, data) in enumerate(alg2data.items()): for idx, (alg, data) in enumerate(alg2data.items()):
print('plot alg : {:}'.format(alg)) print('plot alg : {:}'.format(alg))
@@ -78,7 +80,10 @@ def visualize_curve(api, vis_save_dir, search_space, max_time):
accuracy = query_performance(api, data, dataset, ticket) accuracy = query_performance(api, data, dataset, ticket)
accuracies.append(accuracy) accuracies.append(accuracy)
alg2accuracies[alg] = accuracies alg2accuracies[alg] = accuracies
ax.plot(time_tickets, accuracies, c=colors[idx], label='{:}'.format(alg)) ax.plot([x/100 for x in time_tickets], accuracies, c=colors[idx], label='{:}'.format(alg))
ax.set_xlabel('Estimated wall-clock time (1e2 seconds)', fontsize=LabelSize)
ax.set_ylabel('Test accuracy on {:}'.format(dataset), fontsize=LabelSize)
ax.set_title('Searching results on {:}'.format(dataset), fontsize=LabelSize+4)
ax.legend(loc=4, fontsize=LegendFontsize) ax.legend(loc=4, fontsize=LegendFontsize)
fig, axs = plt.subplots(1, 3, figsize=figsize) fig, axs = plt.subplots(1, 3, figsize=figsize)
@@ -104,4 +109,3 @@ if __name__ == '__main__':
visualize_curve(api201, save_dir, 'tss', args.max_time) visualize_curve(api201, save_dir, 'tss', args.max_time)
api301 = NASBench301API(verbose=False) api301 = NASBench301API(verbose=False)
visualize_curve(api301, save_dir, 'sss', args.max_time) visualize_curve(api301, save_dir, 'sss', args.max_time)