def convert():
architecture = [0, 0, 0, 1, 0, 0, 1, 0, 3, 2, 0, 1, 2, 2, 1, 2, 0, 0, 2, 0]
scale_ids = [8, 7, 6, 8, 5, 7, 3, 4, 2, 4, 2, 3, 4, 5, 6, 6, 3, 3, 4, 6]
net = get_shufflenas_oneshot(architecture=architecture, scale_ids=scale_ids, use_se=True,
last_conv_after_pooling=True, shuffle_by_conv=True)
# load params
net._initialize(force_reinit=True, dtype='float32')
net.cast('float16')
net.load_parameters('../models/oneshot-s+model-0000.params', allow_missing=True)
net.cast('float32')
# save both gluon model and symbols
test_data = nd.ones([5, 3, 224, 224], dtype='float32')
_ = net(test_data)
net.summary(test_data)
net.hybridize()
if not os.path.exists('./symbols'):
os.makedirs('./symbols')
if not os.path.exists('./params'):
os.makedirs('./params')
net.cast('float16')
net.load_parameters('../models/oneshot-s+model-0000.params', allow_missing=True)
net.cast('float32')
net.hybridize()
net(test_data)
net.save_parameters('./params/ShuffleNas_fixArch_shuffleByConv-0000.params')
net.export("./symbols/ShuffleNas_fixArch_shuffleByConv", epoch=0)
flops, model_size = get_flops(symbol_path='./symbols/ShuffleNas_fixArch_shuffleByConv-symbol.json')
print("Last conv after pooling: {}, use se: {}".format(True, True))
print("FLOPS: {}M, # parameters: {}M".format(flops, model_size))
def get_flop_param_score(block_choices,
channel_choices,
comparison_model='SinglePathOneShot',
use_se=False,
last_conv_after_pooling=False,
channels_layout='OneShot'):
""" Return the flops and num of params """
# build fix_arch network and calculate flop
fixarch_net = get_shufflenas_oneshot(
block_choices,
channel_choices,
use_se=use_se,
last_conv_after_pooling=last_conv_after_pooling,
channels_layout=channels_layout)
fixarch_net._initialize()
if not os.path.exists('./symbols'):
os.makedirs('./symbols')
fixarch_net.hybridize()
# calculate flops and num of params
dummy_data = nd.ones([1, 3, 224, 224])
fixarch_net(dummy_data)
fixarch_net.export("./symbols/ShuffleNas_fixArch", epoch=1)
flops, model_size = get_flops(
symbol_path="./symbols/ShuffleNas_fixArch-symbol.json"
) # both in Millions
# proves ShuffleNet series calculate == google paper's
if comparison_model == 'MobileNetV3_large':
flops_constraint = 217
parameter_number_constraint = 5.4
# proves MicroNet challenge doubles what google paper claimed
elif comparison_model == 'MobileNetV2_1.4':
flops_constraint = 585
parameter_number_constraint = 6.9
elif comparison_model == 'SinglePathOneShot':
flops_constraint = 328
parameter_number_constraint = 3.4
# proves mine calculation == ShuffleNet series' == google paper's
elif comparison_model == 'ShuffleNetV2+_medium':
flops_constraint = 222
parameter_number_constraint = 5.6
else:
raise ValueError(
"Unrecognized comparison model: {}".format(comparison_model))
flop_score = flops / flops_constraint
model_size_score = model_size / parameter_number_constraint
return flops, model_size, flop_score, model_size_score
def main():
if FIX_ARCH:
# architecture = [0, 0, 3, 1, 1, 1, 0, 0, 2, 0, 2, 1, 1, 0, 2, 0, 2, 1, 3, 2]
# scale_ids = [6, 5, 3, 5, 2, 6, 3, 4, 2, 5, 7, 5, 4, 6, 7, 4, 4, 5, 4, 3]
architecture = [0, 0, 0, 1, 0, 0, 1, 0, 3, 2, 0, 1, 2, 2, 1, 2, 0, 0, 2, 0]
scale_ids = [8, 7, 6, 8, 5, 7, 3, 4, 2, 4, 2, 3, 4, 5, 6, 6, 3, 3, 4, 6]
net = get_shufflenas_oneshot(architecture=architecture, scale_ids=scale_ids, channels_layout=CHANNELS_LAYOUT,
use_se=USE_SE, last_conv_after_pooling=LAST_CONV_AFTER_POOLING)
else:
net = get_shufflenas_oneshot(use_se=USE_SE, last_conv_after_pooling=LAST_CONV_AFTER_POOLING,
channels_layout=CHANNELS_LAYOUT)
""" Test customized initialization """
net._initialize(force_reinit=True)
print(net)
""" Test ShuffleNasOneShot """
test_data = nd.ones([5, 3, 224, 224])
for step in range(1):
if FIX_ARCH:
test_outputs = net(test_data)
net.summary(test_data)
net.hybridize()
else:
block_choices = net.random_block_choices(select_predefined_block=False, dtype='float32')
full_channel_mask, _ = net.random_channel_mask(select_all_channels=False, dtype='float32')
test_outputs = net(test_data, block_choices, full_channel_mask)
net.summary(test_data, block_choices, full_channel_mask)
if FIX_ARCH:
if not os.path.exists('./symbols'):
os.makedirs('./symbols')
net(test_data)
net.export("./symbols/ShuffleNas_fixArch", epoch=0)
flops, model_size = get_flops()
print("Last conv after pooling: {}, use se: {}".format(LAST_CONV_AFTER_POOLING, USE_SE))
print("FLOPS: {}M, # parameters: {}M".format(flops, model_size))
else:
if not os.path.exists('./params'):
os.makedirs('./params')
net.save_parameters('./params/ShuffleNasOneshot-imagenet-supernet.params')
""" Test generating random channels """
epoch_start_cs = 30
use_all_channels = True if epoch_start_cs != -1 else False
dtype = 'float16'
for epoch in range(120):
if epoch == epoch_start_cs:
use_all_channels = False
for batch in range(1):
full_channel_mask, channel_choices = net.random_channel_mask(select_all_channels=use_all_channels,
epoch_after_cs=epoch - epoch_start_cs,
dtype=dtype,
ignore_first_two_cs=True)
print("Epoch {}: {}".format(epoch, channel_choices))
def search_supernet(net, search_iters=2000, bn_iters=50000, num_gpus=0):
# TODO: use a heapq here to store top-5 models
train_data, val_data, batch_fn = get_data(num_gpus=num_gpus)
best_acc = 0
best_block_choices = None
best_channel_choices = None
context = [mx.gpu(i)
for i in range(num_gpus)] if num_gpus > 0 else [mx.cpu()]
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
for _ in range(search_iters):
block_choices = net.random_block_choices(select_predefined_block=False,
dtype='float32')
full_channel_mask, channel_choices = net.random_channel_mask(
select_all_channels=False, dtype='float32')
# Update BN
# for _ in range(bn_iters):
# data, _ = generate_random_data_label()
# net(data, block_choices, full_channel_mask)
# Get validation accuracy
val_acc = get_accuracy(net,
val_data,
batch_fn,
block_choices,
full_channel_mask,
acc_top1=acc_top1,
acc_top5=acc_top5,
ctx=context)
if val_acc > best_acc:
best_acc = val_acc
best_block_choices = copy.deepcopy(block_choices)
best_channel_choices = copy.deepcopy(channel_choices)
# build fix_arch network and calculate flop
fixarch_net = get_shufflenas_oneshot(block_choices.asnumpy(),
channel_choices)
fixarch_net.initialize()
if not os.path.exists('./symbols'):
os.makedirs('./symbols')
fixarch_net.hybridize()
dummy_data = nd.ones([1, 3, 224, 224])
fixarch_net(dummy_data)
fixarch_net.export("./symbols/ShuffleNas_fixArch", epoch=1)
flops, model_size = get_flops()
print('-' * 40)
print("Val accuracy: {}".format(val_acc))
print('flops: ', str(flops), ' MFLOPS')
print('model size: ', str(model_size), ' MB')
print('-' * 40)
print("Best val accuracy: {}".format(best_acc))
print("Best block choices: {}".format(best_block_choices.asnumpy()))
print("Best channel choices: {}".format(best_channel_choices))
def get_block_flop(block, input_data):
fixarch_net = block
fixarch_net.initialize()
if not os.path.exists('./symbols'):
os.makedirs('./symbols')
fixarch_net.hybridize()
# calculate flops and num of params
output_data = fixarch_net(input_data)
fixarch_net.export("./symbols/ShuffleNas_block", epoch=1)
data_shapes = [('data', input_data.shape)]
flops, model_size = get_flops(symbol_path="./symbols/ShuffleNas_block-symbol.json", data_shapes=data_shapes)
return flops, model_size, output_data
def main():
if FIX_ARCH:
architecture = [
0, 0, 3, 1, 1, 1, 0, 0, 2, 0, 2, 1, 1, 0, 2, 0, 2, 1, 3, 2
]
scale_ids = [
6, 5, 3, 5, 2, 6, 3, 4, 2, 5, 7, 5, 4, 6, 7, 4, 4, 5, 4, 3
]
net = get_shufflenas_oneshot(
architecture=architecture,
scale_ids=scale_ids,
use_se=USE_SE,
last_conv_after_pooling=LAST_CONV_AFTER_POOLING)
else:
net = get_shufflenas_oneshot(
use_se=USE_SE, last_conv_after_pooling=LAST_CONV_AFTER_POOLING)
""" Test customized initialization """
net._initialize(force_reinit=True)
print(net)
""" Test ShuffleNasOneShot """
test_data = nd.ones([5, 3, 224, 224])
for step in range(1):
if FIX_ARCH:
test_outputs = net(test_data)
net.summary(test_data)
net.hybridize()
else:
block_choices = net.random_block_choices(
select_predefined_block=False, dtype='float32')
full_channel_mask, _ = net.random_channel_mask(
select_all_channels=False, dtype='float32')
test_outputs = net(test_data, block_choices, full_channel_mask)
net.summary(test_data, block_choices, full_channel_mask)
if FIX_ARCH:
if not os.path.exists('./symbols'):
os.makedirs('./symbols')
net(test_data)
net.export("./symbols/ShuffleNas_fixArch", epoch=0)
flops, model_size = get_flops()
print("Last conv after pooling: {}, use se: {}".format(
LAST_CONV_AFTER_POOLING, USE_SE))
print("FLOPS: {}M, # parameters: {}M".format(flops, model_size))
else:
if not os.path.exists('./params'):
os.makedirs('./params')
net.save_parameters(
'./params/ShuffleNasOneshot-imagenet-supernet.params')
print(test_outputs.shape)
def main():
# Save a toy SE SuperNet for playing with the search codes
supernet = get_shufflenas_oneshot(use_se=True,
last_conv_after_pooling=True,
channels_layout='OneShot')
supernet._initialize(force_reinit=True)
if not os.path.exists('./params'):
os.makedirs('./params')
test_data = nd.ones([5, 3, 224, 224])
for step in range(1):
block_choices = supernet.random_block_choices(
select_predefined_block=False, dtype='float32')
full_channel_mask, _ = supernet.random_channel_mask(
select_all_channels=False, dtype='float32')
_ = supernet(test_data, block_choices, full_channel_mask)
if not os.path.exists('./params'):
os.makedirs('./params')
supernet.save_parameters(
'./params/ShuffleNasOneshot-imagenet-supernet.params')
# Profiling the OneShot-s+
architecture = [0, 0, 0, 1, 0, 0, 1, 0, 3, 2, 0, 1, 2, 2, 1, 2, 0, 0, 2, 0]
scale_ids = [8, 7, 6, 8, 5, 7, 3, 4, 2, 4, 2, 3, 4, 5, 6, 6, 3, 3, 4, 6]
net = get_shufflenas_oneshot(architecture=architecture,
scale_ids=scale_ids,
channels_layout='OneShot',
use_se=True,
last_conv_after_pooling=True)
net._initialize(force_reinit=True)
print(net)
test_data = nd.ones([5, 3, 224, 224])
for step in range(1):
_ = net(test_data)
net.summary(test_data)
net.hybridize()
if not os.path.exists('./symbols'):
os.makedirs('./symbols')
net(test_data)
net.export("./symbols/ShuffleNas_fixArch", epoch=0)
flops, model_size = get_flops()
print("Last conv after pooling: {}, use se: {}".format(True, True))
print("FLOPS: {}M, # parameters: {}M".format(flops, model_size))
