Skip to content

Commit df1c1ff

Browse files
CanyonWindCanyonWind
authored
Add -1 constraint and debug population is None
1 parent 0b79748 commit df1c1ff

1 file changed

Lines changed: 30 additions & 23 deletions

File tree

‎search_supernet.py‎

Lines changed: 30 additions & 23 deletions
Original file line numberDiff line numberDiff line change
@@ -16,11 +16,12 @@
1616
from oneshot_nas_blocks import NasBatchNorm
1717

1818
FLOP_MAX = 1
19-
PARAM_MAX = 1
19+
PARAM_MAX = -1
2020
SCORE_ACC_RATIO = 1 # flop_param_score_weight/acc_weight for fitness
2121
BLOCK_CHOICE = None # [0, 0, 3, 1, 1, 1, 0, 0, 2, 0, 2, 1, 1, 0, 2, 0, 2, 1, 3, 2]
2222
CHANNEL_CHOICE = None # [6, 5, 3, 5, 2, 6, 3, 4, 2, 5, 7, 5, 4, 6, 7, 4, 4, 5, 4, 3]
2323

24+
2425
def parse_args():
2526
parser = argparse.ArgumentParser(description='Train a model for image classification.')
2627

@@ -236,22 +237,22 @@ def update_log(elem, logger=None):
236237
"""
237238
if logger:
238239
logger.info('-' * 40)
239-
logger.info("Overall score: {}".format(elem[0]))
240-
logger.info("Val accuracy: {}".format(elem[1]))
241-
logger.info("Model normalized score: {}.".format(elem[2]))
242-
logger.info('Flops: {} MFLOPS'.format(elem[3]))
243-
logger.info('# parameters: {} M'.format(elem[4]))
244-
logger.info("Block choices: {}".format(elem[5]))
245-
logger.info("Channel choices: {}".format(elem[6]))
240+
logger.info("Acc/computation balanced score: {}".format(elem[0]))
241+
logger.info("Val accuracy: {}".format(elem[1]))
242+
logger.info("Model normalized score: {}.".format(elem[2]))
243+
logger.info('Flops: {} MFLOPS'.format(elem[3]))
244+
logger.info('# parameters: {} M'.format(elem[4]))
245+
logger.info("Block choices: {}".format(elem[5]))
246+
logger.info("Channel choices: {}".format(elem[6]))
246247
else:
247248
print('-' * 40)
248-
print("Overall score: {}".format(elem[0]))
249-
print("Val accuracy: {}".format(elem[1]))
250-
print("Model normalized score: {}.".format(elem[2]))
251-
print('Flops: {} MFLOPS'.format(elem[3]))
252-
print('# parameters: {} M'.format(elem[4]))
253-
print("Block choices: {}".format(elem[5]))
254-
print("Channel choices: {}".format(elem[6]))
249+
print("Acc/computation balanced score: {}".format(elem[0]))
250+
print("Val accuracy: {}".format(elem[1]))
251+
print("Model normalized score: {}.".format(elem[2]))
252+
print('Flops: {} MFLOPS'.format(elem[3]))
253+
print('# parameters: {} M'.format(elem[4]))
254+
print("Block choices: {}".format(elem[5]))
255+
print("Channel choices: {}".format(elem[6]))
255256

256257

257258
class TopKHeap(object):
@@ -325,21 +326,30 @@ def create_population(self):
325326
get_flop_param_score(block_choices, channel_choices, comparison_model='SinglePathOneShot')
326327

327328
combined_score = 0.5 * flop_score + 0.5 * model_size_score
328-
if flop_score > FLOP_MAX:
329+
if FLOP_MAX != -1 and flop_score > FLOP_MAX:
329330
print("[SKIPPED] Current model normalized score: {}.".format(combined_score))
330331
print("[SKIPPED] Block choices: {}".format(block_choices.asnumpy()))
331332
print("[SKIPPED] Channel choices: {}".format(channel_choices))
332333
print('[SKIPPED] Flops: {} MFLOPS'.format(flops))
333334
print('[SKIPPED] # parameters: {} M'.format(model_size))
334335
continue
335-
if model_size_score > PARAM_MAX:
336+
if PARAM_MAX != -1 and model_size_score > PARAM_MAX:
336337
print("[SKIPPED] Current model normalized score: {}.".format(combined_score))
337338
print("[SKIPPED] Block choices: {}".format(block_choices.asnumpy()))
338339
print("[SKIPPED] Channel choices: {}".format(channel_choices))
339340
print('[SKIPPED] Flops: {} MFLOPS'.format(flops))
340341
print('[SKIPPED] # parameters: {} M'.format(model_size))
341342
continue
342-
print("Population size + 1, total {}, with normalized score: {}".format(len(population) + 1, combined_score))
343+
if combined_score > 1:
344+
print("[SKIPPED] Current model normalized score: {}.".format(combined_score))
345+
print("[SKIPPED] Block choices: {}".format(block_choices.asnumpy()))
346+
print("[SKIPPED] Channel choices: {}".format(channel_choices))
347+
print('[SKIPPED] Flops: {} MFLOPS'.format(flops))
348+
print('[SKIPPED] # parameters: {} M'.format(model_size))
349+
continue
350+
351+
print("Population size + 1, total {}, with normalized score: {}, flop score: {}, param score: {}"
352+
.format(len(population) + 1, combined_score, flop_score, model_size_score))
343353
# Add the network to our population.
344354
instance['flops'] = flops
345355
instance['model_size'] = model_size
@@ -390,9 +400,7 @@ def breed(self, mother, father):
390400
"""
391401
children = []
392402
for _ in range(2):
393-
394403
child = {}
395-
396404
# Crossover: loop through the parameters and pick params for the kid.
397405
# for param_name in self.param_dict.keys():
398406
# child[param_name] = [0] * len(father[param_name])
@@ -444,8 +452,7 @@ def evolve(self, population, topk_items, logger=None):
444452
copy.deepcopy(person['block']), copy.deepcopy(person['channel']))
445453
topk_items.push(net_obj)
446454
update_log(net_obj, logger)
447-
population = population.sort(key=lambda x: -SCORE_ACC_RATIO * x['score'] + x['acc'], reverse=True)
448-
455+
population.sort(key=lambda x: -SCORE_ACC_RATIO * x['score'] + x['acc'], reverse=True)
449456
# The parents are every network we want to keep.
450457
parents = population[:self.retain_length]
451458

@@ -579,7 +586,7 @@ def genetic_search(net, dtype='float32', logger=None, ctx=[mx.cpu()], comparison
579586
topk_nets = TopKHeap(topk) # a list of tuple (acc, score, flops, model_size, block_choices, channel_choices)
580587

581588
# set channel and block value list
582-
param_dict = {'channel': [2, 3, 4, 5, 6, 7, 8, 9],
589+
param_dict = {'channel': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
583590
'block': [0, 1, 2, 3]}
584591

585592
# evolution

0 commit comments

Comments
 (0)