def main():
if opt.builtin_profiler > 0:
profiler.set_config(profile_all=True, aggregate_stats=True)
profiler.set_state('run')
if opt.mode == 'symbolic':
data = mx.sym.var('data')
if opt.dtype == 'float16':
data = mx.sym.Cast(data=data, dtype=np.float16)
out = net(data)
if opt.dtype == 'float16':
out = mx.sym.Cast(data=out, dtype=np.float32)
softmax = mx.sym.SoftmaxOutput(out, name='softmax')
mod = mx.mod.Module(softmax, context=context)
train_data, val_data = get_data_iters(dataset, batch_size, opt)
mod.fit(train_data,
eval_data=val_data,
num_epoch=opt.epochs,
kvstore=kv,
batch_end_callback = mx.callback.Speedometer(batch_size, max(1, opt.log_interval)),
epoch_end_callback = mx.callback.do_checkpoint('image-classifier-%s'% opt.model),
optimizer = 'sgd',
optimizer_params = {'learning_rate': opt.lr, 'wd': opt.wd, 'momentum': opt.momentum, 'multi_precision': True},
initializer = mx.init.Xavier(magnitude=2))
mod.save_parameters('image-classifier-%s-%d-final.params'%(opt.model, opt.epochs))
else:
if opt.mode == 'hybrid':
net.hybridize()
train(opt, context)
if opt.builtin_profiler > 0:
profiler.set_state('stop')
print(profiler.dumps())
def test_profile_create_domain_dept():
profiler.set_config(profile_symbolic=True, filename='test_profile_create_domain_dept.json')
profiler.set_state('run')
domain = profiler.Domain(name='PythonDomain')
print("Domain created: {}".format(str(domain)))
profiler.dump_profile()
profiler.set_state('stop')
def __init__(self, opts, ctx):
self._opts = opts
self._epochs = opts.epochs
self._batch_size = opts.batch_size
self._ctx = ctx
self._chkpt_interval = opts.chkpt_interval
self._log_interval = opts.log_interval
self._weight_interval = opts.weight_interval
self._profile = opts.profile
self._epoch_tick = 0
self._batch_tick = 0
self._networks = []
self._overwrite = opts.overwrite
self._outdir = opts.outdir or os.path.join(os.getcwd(), '{}-{}e-{}'.format(self.model_name(), self._epochs, datetime.now().strftime('%y_%m_%d-%H_%M')))
self._outdir = os.path.expanduser(self._outdir)
self._outlogs = os.path.join(self._outdir, 'logs')
self._outchkpts = os.path.join(self._outdir, 'checkpoints')
self._outsounds = os.path.join(self._outdir, 'sounds')
self._prepare_outdir()
if self._profile:
self._outprofile = os.path.join(self._outdir, 'profile.json')
profiler.set_config(profile_all=True, aggregate_stats=True, filename=self._outprofile)
logging.basicConfig()
self._logger = logging.getLogger()
self._logger.setLevel(logging.INFO)
def test_profile_create_domain_dept():
profiler.set_config(profile_symbolic=True,
filename='test_profile_create_domain_dept.json')
profiler.set_state('run')
domain = profiler.Domain(name='PythonDomain')
profiler.dump()
profiler.set_state('stop')
def main():
if opt.builtin_profiler > 0:
profiler.set_config(profile_all=True, aggregate_stats=True)
profiler.set_state('run')
if opt.mode == 'symbolic':
data = mx.sym.var('data')
out = net(data)
softmax = mx.sym.SoftmaxOutput(out, name='softmax')
mod = mx.mod.Module(softmax, context=[mx.gpu(i) for i in range(num_gpus)] if num_gpus > 0 else [mx.cpu()])
kv = mx.kv.create(opt.kvstore)
train_data, val_data = get_data_iters(dataset, batch_size, kv.num_workers, kv.rank)
mod.fit(train_data,
eval_data = val_data,
num_epoch=opt.epochs,
kvstore=kv,
batch_end_callback = mx.callback.Speedometer(batch_size, max(1, opt.log_interval)),
epoch_end_callback = mx.callback.do_checkpoint('image-classifier-%s'% opt.model),
optimizer = 'sgd',
optimizer_params = {'learning_rate': opt.lr, 'wd': opt.wd, 'momentum': opt.momentum, 'multi_precision': True},
initializer = mx.init.Xavier(magnitude=2))
mod.save_params('image-classifier-%s-%d-final.params'%(opt.model, opt.epochs))
else:
if opt.mode == 'hybrid':
net.hybridize()
train(opt, context)
if opt.builtin_profiler > 0:
profiler.set_state('stop')
print(profiler.dumps())
def test_profile_create_domain_dept():
profiler.set_config(profile_symbolic=True,
filename='test_profile_create_domain_dept.json')
profiler.set_state('run')
domain = profiler.Domain(name='PythonDomain')
print("Domain created: {}".format(str(domain)))
profiler.dump_profile()
profiler.set_state('stop')
def main():
if opt.builtin_profiler > 0:
profiler.set_config(profile_all=True, aggregate_stats=True)
profiler.set_state('run')
if opt.mode == 'hybrid':
net.hybridize()
train(opt, device)
if opt.builtin_profiler > 0:
profiler.set_state('stop')
print(profiler.dumps())
def main():
if opt.builtin_profiler > 0:
profiler.set_config(profile_all=True, aggregate_stats=True)
profiler.set_state('run')
if opt.mode == 'hybrid':
net.hybridize(static_alloc = True, static_shape = True)
train(opt, context)
if opt.builtin_profiler > 0:
profiler.set_state('stop')
print(profiler.dumps())
def enable_profiler(profile_filename, run=True, continuous_dump=False, aggregate_stats=False):
profiler.set_config(profile_symbolic=True,
profile_imperative=True,
profile_memory=True,
profile_api=True,
filename=profile_filename,
continuous_dump=continuous_dump,
aggregate_stats=aggregate_stats)
if run is True:
profiler.set_state('run')
def main():
if opt.builtin_profiler > 0:
profiler.set_config(profile_all=True, aggregate_stats=True)
profiler.set_state('run')
if opt.mode == 'symbolic':
train_symbolic(opt, context)
else:
train(opt, context)
if opt.builtin_profiler > 0:
profiler.set_state('stop')
print(profiler.dumps())
def enable_profiler(profile_filename, run=True, continuous_dump=False, aggregate_stats=False):
profiler.set_config(profile_symbolic=True,
profile_imperative=True,
profile_memory=True,
profile_api=True,
filename=profile_filename,
continuous_dump=continuous_dump,
aggregate_stats=aggregate_stats
)
print('profile file save to {}'.format(profile_filename))
if run is True:
profiler.set_state('run')
def enable_profiler(run=True, continuous_dump=False, aggregate_stats=False):
profile_filename = 'test_profile.json'
profiler.set_config(profile_symbolic=True,
profile_imperative=True,
profile_memory=True,
profile_api=True,
filename=profile_filename,
continuous_dump=continuous_dump,
aggregate_stats=aggregate_stats)
print('profile file save to {}'.format(profile_filename))
if run is True:
profiler.set_state('run')
def main():
if opt.builtin_profiler > 0:
profiler.set_config(profile_all=True, aggregate_stats=True)
profiler.set_state('run')
if opt.mode == 'symbolic':
data = mx.sym.var('data')
out = net(data)
softmax = mx.sym.SoftmaxOutput(out, name='softmax')
mod = mx.mod.Module(softmax, context=context)
kv = mx.kv.create(opt.kvstore)
eval_metric = []
eval_metric.append(mx.metric.create('acc'))
eval_metric.append(mx.metric.create('top_k_accuracy', top_k=5))
eval_metric.append(mx.metric.create('ce'))
train_data, val_data = get_data_iters(dataset, batch_size,
kv.num_workers, kv.rank)
mod.fit(train_data,
eval_data=val_data,
eval_metric=eval_metric,
num_epoch=opt.epochs,
kvstore=kv,
batch_end_callback=mx.callback.Speedometer(
batch_size, max(1, opt.log_interval)),
epoch_end_callback=mx.callback.do_checkpoint(
'image-classifier-%s' % opt.model),
optimizer='sgd',
optimizer_params={
'learning_rate': opt.lr,
'wd': opt.wd,
'momentum': opt.momentum,
'multi_precision': True
},
initializer=mx.init.Xavier(magnitude=2))
mod.save_params('image-classifier-%s-%d-final.params' %
(opt.model, opt.epochs))
else:
if opt.mode == 'hybrid':
net.hybridize()
train(opt, context)
if opt.builtin_profiler > 0:
profiler.set_state('stop')
print(profiler.dumps())
def main():
if opt.builtin_profiler > 0:
profiler.set_config(profile_all=True, aggregate_stats=True)
profiler.set_state("run")
if opt.mode == "symbolic":
data = mx.sym.var("data")
if opt.dtype == "float16":
data = mx.sym.Cast(data=data, dtype=np.float16)
out = net(data)
if opt.dtype == "float16":
out = mx.sym.Cast(data=out, dtype=np.float32)
softmax = mx.sym.SoftmaxOutput(out, name="softmax")
mod = mx.mod.Module(softmax, context=context)
train_data, val_data = get_data_iters(dataset, batch_size, opt)
mod.fit(
train_data,
eval_data=val_data,
num_epoch=opt.epochs,
kvstore=kv,
batch_end_callback=mx.callback.Speedometer(
batch_size, max(1, opt.log_interval)),
epoch_end_callback=mx.callback.do_checkpoint(
"image-classifier-%s" % opt.model),
optimizer="sgd",
optimizer_params={
"learning_rate": opt.lr,
"wd": opt.wd,
"momentum": opt.momentum,
"multi_precision": True,
},
initializer=mx.init.Xavier(magnitude=2),
)
mod.save_parameters("image-classifier-%s-%d-final.params" %
(opt.model, opt.epochs))
else:
if opt.mode == "hybrid":
net.hybridize()
train(opt, context)
if opt.builtin_profiler > 0:
profiler.set_state("stop")
print(profiler.dumps())
def cpp_profile_it(*args, **kwargs):
# Profile the operation
profiler.set_config(profile_all=True, aggregate_stats=True)
profiler.set_state('run')
res = func(*args, **kwargs)
profiler.set_state('stop')
# Prepare the results
profiler_dump = profiler.dumps(reset=True)
# args[0] is assumed to be operator name, if not found check for block name.
# NOTE: This parameter should be removed when we get away from parsing
# profiler output and start using new profiler APIs - get_summary(), reset()
if len(args) > 0:
operator_name = args[0].__name__
elif 'block' in kwargs:
operator_name = kwargs['block']._op_name
else:
raise ValueError("Unable to identify operator name to extract profiler output!")
# Get the MXNet profile output
profiler_output = parse_profiler_dump(operator_name, profiler_dump)
return res, profiler_output
# Horovod: fetch and broadcast parameters
params = model.collect_params()
if params is not None:
hvd.broadcast_parameters(params, root_rank=0)
# Horovod: create DistributedTrainer, a subclass of gluon.Trainer
trainer = hvd.DistributedTrainer(params, opt)
# Create loss function and train metric
loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
metric = mx.metric.Accuracy()
# Set profiler
profiler.set_config(profile_all=True,
aggregate_stats=True,
filename="profile_mx_mnist.json")
# Train model
for epoch in range(args.epochs):
tic = time.time()
train_data.reset()
metric.reset()
for nbatch, batch in enumerate(train_data, start=1):
# Start and pause profiling
if nbatch == 100:
if epoch == 0:
profiler.set_state('run')
else:
profiler.resume()
elif nbatch == 200:
import sys
import os
import math
import mxnet as mx
import time
import psutil
import gc
from mxnet import profiler
from util import *
profiler.set_config(profile_all=True,
aggregate_stats=True,
continuous_dump=True,
filename='profile_output.json')
def cpuStats():
# print(sys.version)
# print(psutil.cpu_percent())
# print(psutil.virtual_memory()) # physical memory usage
pid = os.getpid()
py = psutil.Process(pid)
memoryUse = py.memory_info()[0] / 2. ** 30 # memory use in GB...I think
# print('memory GB:', memoryUse)
return memoryUse
jitter_param = 0.4
lighting_param = 0.1
mean_rgb = [123.68, 116.779, 103.939]
import mxnet as mx
from mxnet import nd
from mxnet import profiler
profiler.set_config(profile_all=True,
aggregate_stats=True,
filename='cpu_gpu_data_copy_profiler_output.json')
# Create a large Tensor on CPU
data1 = nd.random.uniform(shape=(10000, 10000), ctx=mx.cpu())
data2 = nd.random.uniform(shape=(10000, 10000), ctx=mx.cpu())
nd.waitall()
# Profiler copying data and operation only
profiler.set_state('run')
# Copy data to GPU
data1.as_in_context(context=mx.gpu(0))
data2.as_in_context(context=mx.gpu(0))
# Do couple of operations on GPU
res = data1 + data2
#res = nd.mean(res)
# Copy result back to CPU
res_cpu = res.as_in_context(context=mx.cpu())
nd.waitall()
profiler.set_state('stop')
print(profiler.dumps())
profiler.dump()
def train_net(net, config, check_flag, logger, sig_state, sig_pgbar, sig_table):
print(config)
# config = Configs()
# matplotlib.use('Agg')
# import matplotlib.pyplot as plt
sig_pgbar.emit(-1)
mx.random.seed(1)
matplotlib.use('Agg')
import matplotlib.pyplot as plt
classes = 10
num_epochs = config.train_cfg.epoch
batch_size = config.train_cfg.batchsize
optimizer = config.lr_cfg.optimizer
lr = config.lr_cfg.lr
num_gpus = config.train_cfg.gpu
batch_size *= max(1, num_gpus)
context = [mx.gpu(i)
for i in range(num_gpus)] if num_gpus > 0 else [mx.cpu()]
num_workers = config.data_cfg.worker
warmup = config.lr_cfg.warmup
if config.lr_cfg.decay == 'cosine':
lr_sch = lr_scheduler.CosineScheduler((50000//batch_size)*num_epochs,
base_lr=lr,
warmup_steps=warmup *
(50000//batch_size),
final_lr=1e-5)
else:
lr_sch = lr_scheduler.FactorScheduler((50000//batch_size)*config.lr_cfg.factor_epoch,
factor=config.lr_cfg.factor,
base_lr=lr,
warmup_steps=warmup*(50000//batch_size))
model_name = config.net_cfg.name
if config.data_cfg.mixup:
model_name += '_mixup'
if config.train_cfg.amp:
model_name += '_amp'
base_dir = './'+model_name
if os.path.exists(base_dir):
base_dir = base_dir + '-' + \
time.strftime("%m-%d-%H.%M.%S", time.localtime())
makedirs(base_dir)
if config.save_cfg.tensorboard:
logdir = base_dir+'/tb/'+model_name
if os.path.exists(logdir):
logdir = logdir + '-' + \
time.strftime("%m-%d-%H.%M.%S", time.localtime())
sw = SummaryWriter(logdir=logdir, flush_secs=5, verbose=False)
cmd_file = open(base_dir+'/tb.bat', mode='w')
cmd_file.write('tensorboard --logdir=./')
cmd_file.close()
save_period = 10
save_dir = base_dir+'/'+'params'
makedirs(save_dir)
plot_name = base_dir+'/'+'plot'
makedirs(plot_name)
stat_name = base_dir+'/'+'stat.txt'
csv_name = base_dir+'/'+'data.csv'
if os.path.exists(csv_name):
csv_name = base_dir+'/'+'data-' + \
time.strftime("%m-%d-%H.%M.%S", time.localtime())+'.csv'
csv_file = open(csv_name, mode='w', newline='')
csv_writer = csv.writer(csv_file)
csv_writer.writerow(['Epoch', 'train_loss', 'train_acc',
'valid_loss', 'valid_acc', 'lr', 'time'])
logging_handlers = [logging.StreamHandler(), logger]
logging_handlers.append(logging.FileHandler(
'%s/train_cifar10_%s.log' % (model_name, model_name)))
logging.basicConfig(level=logging.INFO, handlers=logging_handlers)
logging.info(config)
if config.train_cfg.amp:
amp.init()
if config.save_cfg.profiler:
profiler.set_config(profile_all=True,
aggregate_stats=True,
continuous_dump=True,
filename=base_dir+'/%s_profile.json' % model_name)
is_profiler_run = False
trans_list = []
imgsize = config.data_cfg.size
if config.data_cfg.crop:
trans_list.append(gcv_transforms.RandomCrop(
32, pad=config.data_cfg.crop_pad))
if config.data_cfg.cutout:
trans_list.append(CutOut(config.data_cfg.cutout_size))
if config.data_cfg.flip:
trans_list.append(transforms.RandomFlipLeftRight())
if config.data_cfg.erase:
trans_list.append(gcv_transforms.block.RandomErasing(s_max=0.25))
trans_list.append(transforms.Resize(imgsize))
trans_list.append(transforms.ToTensor())
trans_list.append(transforms.Normalize([0.4914, 0.4822, 0.4465],
[0.2023, 0.1994, 0.2010]))
transform_train = transforms.Compose(trans_list)
transform_test = transforms.Compose([
transforms.Resize(imgsize),
transforms.ToTensor(),
transforms.Normalize([0.4914, 0.4822, 0.4465],
[0.2023, 0.1994, 0.2010])
])
def label_transform(label, classes):
ind = label.astype('int')
res = nd.zeros((ind.shape[0], classes), ctx=label.context)
res[nd.arange(ind.shape[0], ctx=label.context), ind] = 1
return res
def test(ctx, val_data):
metric = mx.metric.Accuracy()
loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
num_batch = len(val_data)
test_loss = 0
for i, batch in enumerate(val_data):
data = gluon.utils.split_and_load(
batch[0], ctx_list=ctx, batch_axis=0)
label = gluon.utils.split_and_load(
batch[1], ctx_list=ctx, batch_axis=0)
outputs = [net(X) for X in data]
loss = [loss_fn(yhat, y) for yhat, y in zip(outputs, label)]
metric.update(label, outputs)
test_loss += sum([l.sum().asscalar() for l in loss])
test_loss /= batch_size * num_batch
name, val_acc = metric.get()
return name, val_acc, test_loss
def train(epochs, ctx):
if isinstance(ctx, mx.Context):
ctx = [ctx]
if config.train_cfg.param_init:
init_func = getattr(mx.init, config.train_cfg.init)
net.initialize(init_func(), ctx=ctx, force_reinit=True)
else:
net.load_parameters(config.train_cfg.param_file, ctx=ctx)
summary(net, stat_name, nd.uniform(
shape=(1, 3, imgsize, imgsize), ctx=ctx[0]))
# net = nn.HybridBlock()
net.hybridize()
root = config.dir_cfg.dataset
train_data = gluon.data.DataLoader(
gluon.data.vision.CIFAR10(
root=root, train=True).transform_first(transform_train),
batch_size=batch_size, shuffle=True, last_batch='discard', num_workers=num_workers)
val_data = gluon.data.DataLoader(
gluon.data.vision.CIFAR10(
root=root, train=False).transform_first(transform_test),
batch_size=batch_size, shuffle=False, num_workers=num_workers)
trainer_arg = {'learning_rate': config.lr_cfg.lr,
'wd': config.lr_cfg.wd, 'lr_scheduler': lr_sch}
extra_arg = eval(config.lr_cfg.extra_arg)
trainer_arg.update(extra_arg)
trainer = gluon.Trainer(net.collect_params(), optimizer, trainer_arg)
if config.train_cfg.amp:
amp.init_trainer(trainer)
metric = mx.metric.Accuracy()
train_metric = mx.metric.RMSE()
loss_fn = gluon.loss.SoftmaxCrossEntropyLoss(
sparse_label=False if config.data_cfg.mixup else True)
train_history = TrainingHistory(['training-error', 'validation-error'])
# acc_history = TrainingHistory(['training-acc', 'validation-acc'])
loss_history = TrainingHistory(['training-loss', 'validation-loss'])
iteration = 0
best_val_score = 0
# print('start training')
sig_state.emit(1)
sig_pgbar.emit(0)
# signal.emit('Training')
for epoch in range(epochs):
tic = time.time()
train_metric.reset()
metric.reset()
train_loss = 0
num_batch = len(train_data)
alpha = 1
for i, batch in enumerate(train_data):
if epoch == 0 and iteration == 1 and config.save_cfg.profiler:
profiler.set_state('run')
is_profiler_run = True
if epoch == 0 and iteration == 1 and config.save_cfg.tensorboard:
sw.add_graph(net)
lam = np.random.beta(alpha, alpha)
if epoch >= epochs - 20 or not config.data_cfg.mixup:
lam = 1
data_1 = gluon.utils.split_and_load(
batch[0], ctx_list=ctx, batch_axis=0)
label_1 = gluon.utils.split_and_load(
batch[1], ctx_list=ctx, batch_axis=0)
if not config.data_cfg.mixup:
data = data_1
label = label_1
else:
data = [lam*X + (1-lam)*X[::-1] for X in data_1]
label = []
for Y in label_1:
y1 = label_transform(Y, classes)
y2 = label_transform(Y[::-1], classes)
label.append(lam*y1 + (1-lam)*y2)
with ag.record():
output = [net(X) for X in data]
loss = [loss_fn(yhat, y) for yhat, y in zip(output, label)]
if config.train_cfg.amp:
with ag.record():
with amp.scale_loss(loss, trainer) as scaled_loss:
ag.backward(scaled_loss)
# scaled_loss.backward()
else:
for l in loss:
l.backward()
trainer.step(batch_size)
train_loss += sum([l.sum().asscalar() for l in loss])
output_softmax = [nd.SoftmaxActivation(out) for out in output]
train_metric.update(label, output_softmax)
metric.update(label_1, output_softmax)
name, acc = train_metric.get()
if config.save_cfg.tensorboard:
sw.add_scalar(tag='lr', value=trainer.learning_rate,
global_step=iteration)
if epoch == 0 and iteration == 1 and config.save_cfg.profiler:
nd.waitall()
profiler.set_state('stop')
profiler.dump()
iteration += 1
sig_pgbar.emit(iteration)
if check_flag()[0]:
sig_state.emit(2)
while(check_flag()[0] or check_flag()[1]):
if check_flag()[1]:
print('stop')
return
else:
time.sleep(5)
print('pausing')
epoch_time = time.time() - tic
train_loss /= batch_size * num_batch
name, acc = train_metric.get()
_, train_acc = metric.get()
name, val_acc, _ = test(ctx, val_data)
# if config.data_cfg.mixup:
# train_history.update([acc, 1-val_acc])
# plt.cla()
# train_history.plot(save_path='%s/%s_history.png' %
# (plot_name, model_name))
# else:
train_history.update([1-train_acc, 1-val_acc])
plt.cla()
train_history.plot(save_path='%s/%s_history.png' %
(plot_name, model_name))
if val_acc > best_val_score:
best_val_score = val_acc
net.save_parameters('%s/%.4f-cifar-%s-%d-best.params' %
(save_dir, best_val_score, model_name, epoch))
current_lr = trainer.learning_rate
name, val_acc, val_loss = test(ctx, val_data)
logging.info('[Epoch %d] loss=%f train_acc=%f train_RMSE=%f\n val_acc=%f val_loss=%f lr=%f time: %f' %
(epoch, train_loss, train_acc, acc, val_acc, val_loss, current_lr, epoch_time))
loss_history.update([train_loss, val_loss])
plt.cla()
loss_history.plot(save_path='%s/%s_loss.png' %
(plot_name, model_name), y_lim=(0, 2), legend_loc='best')
if config.save_cfg.tensorboard:
sw._add_scalars(tag='Acc',
scalar_dict={'train_acc': train_acc, 'test_acc': val_acc}, global_step=epoch)
sw._add_scalars(tag='Loss',
scalar_dict={'train_loss': train_loss, 'test_loss': val_loss}, global_step=epoch)
sig_table.emit([epoch, train_loss, train_acc,
val_loss, val_acc, current_lr, epoch_time])
csv_writer.writerow([epoch, train_loss, train_acc,
val_loss, val_acc, current_lr, epoch_time])
csv_file.flush()
if save_period and save_dir and (epoch + 1) % save_period == 0:
net.save_parameters('%s/cifar10-%s-%d.params' %
(save_dir, model_name, epoch))
if save_period and save_dir:
net.save_parameters('%s/cifar10-%s-%d.params' %
(save_dir, model_name, epochs-1))
train(num_epochs, context)
if config.save_cfg.tensorboard:
sw.close()
for ctx in context:
ctx.empty_cache()
csv_file.close()
logging.shutdown()
reload(logging)
sig_state.emit(0)
model_prefix, load_epoch)
# initialize the module
mod = mx.module.Module(symbol=sym,
context=ctx,
data_names=['user', 'item'],
label_names=['score'])
mod.bind(data_shapes=train_iter.provide_data,
label_shapes=train_iter.provide_label)
# get the sparse weight parameter
mod.set_params(arg_params=arg_params, aux_params=aux_params)
# profile
profiler.set_config(profile_all=True,
aggregate_stats=True,
filename='profile_neumf.json')
profiler.set_state('run')
if benchmark:
logging.info('Evaluating...')
(hits, ndcgs) = evaluate_model(mod, testRatings, testNegatives, topK,
evaluation_threads)
hr, ndcg = np.array(hits).mean(), np.array(ndcgs).mean()
print('HR = %.4f, NDCG = %.4f' % (hr, ndcg))
logging.info('Evaluating completed')
profiler.set_state('stop')
else:
logging.info('Inference started ...')
nbatch = 0
tic = time()
import mxnet as mx
from mxnet import autograd
from mxnet import profiler
#################### Set Profiler Config ######################
profiler.set_config(profile_all=True,
aggregate_stats=True,
filename='cpu_mnist_cnn_profile_output.json')
###############################################################
# Build Network
from mxnet import gluon
net = gluon.nn.HybridSequential()
with net.name_scope():
net.add(gluon.nn.Conv2D(channels=20, kernel_size=5, activation='relu'))
net.add(gluon.nn.MaxPool2D(pool_size=2, strides=2))
net.add(gluon.nn.Conv2D(channels=50, kernel_size=5, activation='relu'))
net.add(gluon.nn.MaxPool2D(pool_size=2, strides=2))
net.add(gluon.nn.Flatten())
net.add(gluon.nn.Dense(512, activation="relu"))
net.add(gluon.nn.Dense(10))
from mxnet.gluon.data.vision import transforms
train_data = gluon.data.DataLoader(
gluon.data.vision.MNIST(train=True).transform_first(transforms.ToTensor()),
batch_size=64,
shuffle=True)
# Set Context
ctx = mx.cpu()
def train(net, train_data, val_data, eval_metric, batch_size, ctx, args):
"""Training pipeline"""
print("rank:{}, training...".format(
kv.rank)) if "perseus" in args.kv_store else None
if args.profiler == "1":
# profiler config
profiler.set_config(profile_all=True,
aggregate_stats=True,
continuous_dump=True,
filename='profile_output_{}.json'.format(
kv.rank if "perseus" in args.kv_store else 0))
net.collect_params().setattr('grad_req', 'null')
net.collect_train_params().setattr('grad_req', 'write')
optimizer_params = {
'learning_rate': args.lr,
'wd': args.wd,
'momentum': args.momentum
}
if args.amp:
optimizer_params['multi_precision'] = True
if args.horovod:
hvd.broadcast_parameters(net.collect_params(), root_rank=0)
trainer = hvd.DistributedTrainer(
net.collect_train_params(
), # fix batchnorm, fix first stage, etc...
'sgd',
optimizer_params)
else:
trainer = gluon.Trainer(
net.collect_train_params(
), # fix batchnorm, fix first stage, etc...
'sgd',
optimizer_params,
update_on_kvstore=None,
kvstore=kv) #(False if args.amp else None), kvstore=kv)
if args.amp:
amp.init_trainer(trainer)
# lr decay policy
lr_decay = float(args.lr_decay)
lr_steps = sorted(
[float(ls) for ls in args.lr_decay_epoch.split(',') if ls.strip()])
lr_warmup = float(args.lr_warmup) # avoid int division
# TODO(zhreshold) losses?
rpn_cls_loss = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss(
from_sigmoid=False)
rpn_box_loss = mx.gluon.loss.HuberLoss(
rho=args.rpn_smoothl1_rho) # == smoothl1
rcnn_cls_loss = mx.gluon.loss.SoftmaxCrossEntropyLoss()
rcnn_box_loss = mx.gluon.loss.HuberLoss(
rho=args.rcnn_smoothl1_rho) # == smoothl1
metrics = [
mx.metric.Loss('RPN_Conf'),
mx.metric.Loss('RPN_SmoothL1'),
mx.metric.Loss('RCNN_CrossEntropy'),
mx.metric.Loss('RCNN_SmoothL1'),
]
rpn_acc_metric = RPNAccMetric()
rpn_bbox_metric = RPNL1LossMetric()
rcnn_acc_metric = RCNNAccMetric()
rcnn_bbox_metric = RCNNL1LossMetric()
metrics2 = [
rpn_acc_metric, rpn_bbox_metric, rcnn_acc_metric, rcnn_bbox_metric
]
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
log_file_path = args.save_prefix + '_train.log'
log_dir = os.path.dirname(log_file_path)
if log_dir and not os.path.exists(log_dir):
os.makedirs(log_dir)
fh = logging.FileHandler(log_file_path)
logger.addHandler(fh)
if args.custom_model:
logger.info(
'Custom model enabled. Expert Only!! Currently non-FPN model is not supported!!'
' Default setting is for MS-COCO.')
logger.info(args)
if args.verbose:
logger.info('Trainable parameters:')
logger.info(net.collect_train_params().keys())
logger.info('Start training from [Epoch {}]'.format(args.start_epoch))
best_map = [0]
for epoch in range(args.start_epoch, args.epochs):
mix_ratio = 1.0
if not args.disable_hybridization:
net.hybridize(static_alloc=args.static_alloc)
rcnn_task = ForwardBackwardTask(net,
trainer,
rpn_cls_loss,
rpn_box_loss,
rcnn_cls_loss,
rcnn_box_loss,
mix_ratio=1.0)
if "perseus" in args.kv_store:
args.executor_threads = 1
executor = Parallel(args.executor_threads, rcnn_task) if (
not args.horovod and "perseus" not in args.kv_store) else None
if args.mixup:
# TODO(zhreshold) only support evenly mixup now, target generator needs to be modified otherwise
train_data._dataset._data.set_mixup(np.random.uniform, 0.5, 0.5)
mix_ratio = 0.5
if epoch >= args.epochs - args.no_mixup_epochs:
train_data._dataset._data.set_mixup(None)
mix_ratio = 1.0
while lr_steps and epoch >= lr_steps[0]:
new_lr = trainer.learning_rate * lr_decay
lr_steps.pop(0)
trainer.set_learning_rate(new_lr)
logger.info("[Epoch {}] Set learning rate to {}".format(
epoch, new_lr))
for metric in metrics:
metric.reset()
tic = time.time()
btic = time.time()
base_lr = trainer.learning_rate
rcnn_task.mix_ratio = mix_ratio
if args.profiler == "1":
# profiler 1
profiler.set_state('run')
for i, batch in enumerate(train_data):
if epoch == 0 and i <= lr_warmup:
# adjust based on real percentage
new_lr = base_lr * get_lr_at_iter(i / lr_warmup,
args.lr_warmup_factor)
if new_lr != trainer.learning_rate:
if i % args.log_interval == 0:
logger.info(
'[Epoch 0 Iteration {}] Set learning rate to {}'.
format(i, new_lr))
trainer.set_learning_rate(new_lr)
batch = split_and_load(batch, ctx_list=ctx)
metric_losses = [[] for _ in metrics]
add_losses = [[] for _ in metrics2]
if executor is not None:
for data in zip(*batch):
executor.put(data)
for j in range(len(ctx)):
if executor is not None:
result = executor.get()
else:
result = rcnn_task.forward_backward(list(zip(*batch))[0])
if (not args.horovod) or hvd.rank() == 0:
for k in range(len(metric_losses)):
metric_losses[k].append(result[k])
for k in range(len(add_losses)):
add_losses[k].append(result[len(metric_losses) + k])
for metric, record in zip(metrics, metric_losses):
metric.update(0, record)
for metric, records in zip(metrics2, add_losses):
for pred in records:
metric.update(pred[0], pred[1])
trainer.step(batch_size)
# update metrics
if ((not args.horovod) or hvd.rank() == 0) and args.log_interval \
and not (i + 1) % args.log_interval:
msg = ','.join([
'{}={:.3f}'.format(*metric.get())
for metric in metrics + metrics2
])
cur_rank = kv.rank if "perseus" in args.kv_store else 0
logger.info(
'rank:{}, [Epoch {}][Batch {}], Speed: {:.3f} samples/sec, {}'
.format(
cur_rank, epoch, i,
args.log_interval * batch_size / (time.time() - btic),
msg))
btic = time.time()
if i >= 100 and args.profiler == "1":
profiler.set_state('stop')
print(profiler.dumps())
break
if ((not args.horovod) and ("perseus" not in args.kv_store)) or (
args.horovod and hvd.rank() == 0) or (
("perseus" in args.kv_store) and kv.rank == 0): # perseus
#if (not args.horovod) or hvd.rank() == 0:
msg = ','.join(
['{}={:.3f}'.format(*metric.get()) for metric in metrics])
logger.info('[Epoch {}] Training cost: {:.3f}, {}'.format(
epoch, (time.time() - tic), msg))
if not (epoch + 1) % args.val_interval:
# consider reduce the frequency of validation to save time
map_name, mean_ap = validate(net, val_data, ctx, eval_metric,
args)
val_msg = '\n'.join(
['{}={}'.format(k, v) for k, v in zip(map_name, mean_ap)])
logger.info('[Epoch {}] Validation: \n{}'.format(
epoch, val_msg))
current_map = float(mean_ap[-1])
else:
current_map = 0.
save_params(net, logger, best_map, current_map, epoch,
args.save_interval, args.save_prefix)
mx.nd.waitall()
def train(net, train_data, val_data, eval_metric, ctx, args):
"""Training pipeline"""
net.collect_params().setattr('grad_req', 'null')
net.collect_train_params().setattr('grad_req', 'write')
rescale_factor = float(
cfg.GENERAL.FP16_RESCALE_FACTOR) if cfg.GENERAL.FP16 else None
trainer = gluon.Trainer(
net.collect_train_params(), # fix batchnorm, fix first stage, etc...
'sgd',
{
'learning_rate':
cfg.TRAIN.BASE_LR,
'wd':
cfg.TRAIN.WEIGHT_DECAY,
'momentum':
cfg.TRAIN.MOMENTUM,
'clip_gradient':
5,
'multi_precision':
cfg.GENERAL.FP16,
'rescale_grad':
1.0 / cfg.GENERAL.FP16_RESCALE_FACTOR if cfg.GENERAL.FP16 else 1.0
})
# lr decay policy
lr_steps = cfg.AUTO.LR_DECAY_EPOCH
lr_warmup = float(cfg.TRAIN.LR_WARMUP) # avoid int division
# TODO(zhreshold) losses?
rpn_cls_loss = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss(
from_sigmoid=False, weight=rescale_factor)
rpn_box_loss = mx.gluon.loss.HuberLoss(
rho=1 / 9., weight=rescale_factor) # i.e. smoothl1
rcnn_cls_loss = mx.gluon.loss.SoftmaxCrossEntropyLoss(
weight=rescale_factor)
rcnn_box_loss = mx.gluon.loss.HuberLoss(
weight=rescale_factor) # i.e. smoothl1
metrics = [
mx.metric.Loss('RPN_Conf'),
mx.metric.Loss('RPN_SmoothL1'),
mx.metric.Loss('RCNN_CrossEntropy'),
mx.metric.Loss('RCNN_SmoothL1'),
]
metrics2 = [
RPNAccMetric(),
RPNL1LossMetric(),
RCNNAccMetric(),
RCNNL1LossMetric()
]
logger.info("Trainable parameters: ------------------------------------------\n" + \
pprint.pformat(net.collect_train_params().keys(), indent=1, width=100, compact=True))
logger.info('LR Schedule [Epochs {} - {}].'.format(
cfg.AUTO.LR_DECAY_EPOCH, [
cfg.TRAIN.BASE_LR * cfg.TRAIN.LR_DECAY_FACTOR**i
for i in range(len(cfg.AUTO.LR_DECAY_EPOCH))
]))
logger.info('Start training from [Epoch {}] to [Epoch {}].'.format(
cfg.TRAIN.START_EPOCH, cfg.AUTO.END_EPOCH))
best_map = [0]
steps_per_epoch = cfg.TRAIN.STEPS_PER_EPOCH if cfg.TRAIN.STEPS_PER_EPOCH else len(
train_data)
for epoch in range(cfg.TRAIN.START_EPOCH, cfg.AUTO.END_EPOCH + 1):
mix_ratio = 1.0
if cfg.TRAIN.MODE_MIXUP:
# TODO(zhreshold) only support evenly mixup now, target generator needs to be modified otherwise
train_data._dataset.set_mixup(np.random.uniform, 0.5, 0.5)
mix_ratio = 0.5
if epoch >= (cfg.AUTO.END_EPOCH + 1) - cfg.AUTO.NO_MIXUP_EPOCH:
train_data._dataset.set_mixup(None)
mix_ratio = 1.0
if lr_steps and epoch >= lr_steps[0]:
while lr_steps and epoch >= lr_steps[0]:
new_lr = trainer.learning_rate * cfg.TRAIN.LR_DECAY_FACTOR
lr_steps.pop(0)
trainer.set_learning_rate(new_lr)
logger.info("[Epoch {}] Set learning rate to {}".format(
epoch, new_lr))
for metric in metrics:
metric.reset()
tic = time.time()
btic = time.time()
if epoch == cfg.TRAIN.START_EPOCH or (
epoch - 1) % cfg.TRAIN.EVAL_INTERVAL == 0:
net.hybridize(static_alloc=True)
base_lr = trainer.learning_rate
tbar = tqdm(train_data, total=steps_per_epoch)
tbar.set_description_str("[ TRAIN ]")
for i, batch in enumerate(tbar):
i += 1
total_iter = (epoch - 1) * steps_per_epoch + i
if total_iter <= lr_warmup:
# adjust based on real percentage
new_lr = base_lr * get_lr_at_iter(total_iter / lr_warmup)
if new_lr != trainer.learning_rate:
if total_iter % cfg.GENERAL.LOG_INTERVAL == 0:
tqdm.write(
'[Warm Up] Set learning rate to {}'.format(new_lr))
trainer.set_learning_rate(new_lr)
batch = split_and_load(
batch, ctx_list=ctx) # Split data to 1 batch each device.
batch_size = len(batch[0])
losses = []
metric_losses = [[] for _ in metrics]
add_losses = [[] for _ in metrics2]
if args.profile and i == 10:
profiler.set_config(profile_all=True,
aggregate_stats=True,
filename='profile_output.json')
profiler.set_state('run')
with autograd.record():
for data, label, rpn_cls_targets, rpn_box_targets, rpn_box_masks in zip(
*batch):
gt_label = label[:, :, 4:5]
gt_box = label[:, :, :4]
cls_pred, box_pred, roi, samples, matches, rpn_score, rpn_box, anchors = net(
data, gt_box)
# losses of rpn
if cfg.GENERAL.FP16:
rpn_score = rpn_score.astype('float32')
rpn_box = rpn_box.astype('float32')
rpn_cls_targets = rpn_cls_targets.astype('float32')
rpn_box_targets = rpn_box_targets.astype('float32')
rpn_box_masks = rpn_box_masks.astype('float32')
rpn_score = rpn_score.squeeze(axis=-1)
num_rpn_pos = (rpn_cls_targets >= 0).sum()
rpn_loss1 = rpn_cls_loss(
rpn_score, rpn_cls_targets, rpn_cls_targets >=
0) * rpn_cls_targets.size / num_rpn_pos
rpn_loss2 = rpn_box_loss(
rpn_box, rpn_box_targets,
rpn_box_masks) * rpn_box.size / num_rpn_pos
# rpn overall loss, use sum rather than average
rpn_loss = rpn_loss1 + rpn_loss2
# generate targets for rcnn
cls_targets, box_targets, box_masks = net.target_generator(
roi, samples, matches, gt_label, gt_box)
# losses of rcnn
if cfg.GENERAL.FP16:
cls_pred = cls_pred.astype('float32')
box_pred = box_pred.astype('float32')
cls_targets = cls_targets.astype('float32')
box_targets = box_targets.astype('float32')
box_masks = box_masks.astype('float32')
num_rcnn_pos = (cls_targets >= 0).sum()
rcnn_loss1 = rcnn_cls_loss(
cls_pred, cls_targets, cls_targets >= 0
) * cls_targets.size / cls_targets.shape[0] / num_rcnn_pos
rcnn_loss2 = rcnn_box_loss(
box_pred, box_targets, box_masks
) * box_pred.size / box_pred.shape[0] / num_rcnn_pos
rcnn_loss = rcnn_loss1 + rcnn_loss2
# overall losses
losses.append(rpn_loss.sum() * mix_ratio +
rcnn_loss.sum() * mix_ratio)
metric_losses[0].append(rpn_loss1.sum() * mix_ratio)
metric_losses[1].append(rpn_loss2.sum() * mix_ratio)
metric_losses[2].append(rcnn_loss1.sum() * mix_ratio)
metric_losses[3].append(rcnn_loss2.sum() * mix_ratio)
add_losses[0].append(
[[rpn_cls_targets, rpn_cls_targets >= 0], [rpn_score]])
add_losses[1].append([[rpn_box_targets, rpn_box_masks],
[rpn_box]])
add_losses[2].append([[cls_targets], [cls_pred]])
add_losses[3].append([[box_targets, box_masks],
[box_pred]])
autograd.backward(losses)
for metric, record in zip(metrics, metric_losses):
metric.update(0, record)
for metric, records in zip(metrics2, add_losses):
for pred in records:
metric.update(pred[0], pred[1])
trainer.step(batch_size)
if args.profile:
mx.nd.waitall()
profiler.set_state('stop')
# update metrics
if cfg.GENERAL.LOG_INTERVAL and total_iter % cfg.GENERAL.LOG_INTERVAL == 0:
msg = ','.join([
'{}={:.3f}'.format(*metric.get())
for metric in metrics + metrics2
])
total_speed = cfg.GENERAL.LOG_INTERVAL * batch_size / (
time.time() - btic)
speed = total_speed / batch_size # batch size rely on the gpu num.
epoch_time_left = (steps_per_epoch - i + 1) / speed
total_time_left = (
(cfg.AUTO.END_EPOCH - epoch) * steps_per_epoch - i +
1) / speed
epoch_tl_h, epoch_tl_m, epoch_tl_s = sec_to_time(
epoch_time_left)
total_tl_h, total_tl_m, _ = sec_to_time(total_time_left)
tqdm.write(
'[Epoch {}][Batch {}], {:.3f}/{:0>2}h{:0>2}m{:0>2}s/{:0>2}h{:0>2}m, {}'
.format(epoch, total_iter, total_speed, epoch_tl_h,
epoch_tl_m, epoch_tl_s, total_tl_h, total_tl_m,
msg))
btic = time.time()
if cfg.TRAIN.STEPS_PER_EPOCH and i >= cfg.TRAIN.STEPS_PER_EPOCH:
break
tbar.close()
msg = ','.join(
['{}={:.3f}'.format(*metric.get()) for metric in metrics])
logger.info('[Epoch {}] Training cost: {:.3f}s, {}'.format(
epoch, (time.time() - tic), msg))
if epoch % cfg.TRAIN.EVAL_INTERVAL == 0:
# consider reduce the frequency of validation to save time
map_name, mean_ap = validate(net, val_data, ctx, eval_metric)
val_msg = '\n'.join(
['{}={}'.format(k, v) for k, v in zip(map_name, mean_ap)])
logger.info('[Epoch {}] Validation: \n{}'.format(epoch, val_msg))
current_map = float(mean_ap[-1])
else:
current_map = 0.
save_params(net, logger, best_map, current_map, epoch,
cfg.TRAIN.SAVE_INTERVAL, args.logdir)
import mxnet as mx
from mxnet import nd, gluon, init, autograd
from mxnet.gluon import nn
from mxnet.gluon.data.vision import datasets, transforms
import time
from mxnet import profiler
ctx = mx.gpu(0)
profiler.set_config(profile_all=True, aggregate_stats=True, filename='gpu_fashion_mnist_profile_output.json')
# Get Data
mnist_train = datasets.FashionMNIST(train=True)
text_labels = ['t-shirt', 'trouser', 'pullover', 'dress', 'coat',
'sandal', 'shirt', 'sneaker', 'bag', 'ankle boot']
# Data preprocessing
transformer = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(0.13, 0.31)])
mnist_train = mnist_train.transform_first(transformer)
batch_size = 256
train_data = gluon.data.DataLoader(
mnist_train, batch_size=batch_size, shuffle=True, num_workers=4)
mnist_valid = gluon.data.vision.FashionMNIST(train=False)
valid_data = gluon.data.DataLoader(
mnist_valid.transform_first(transformer),
def main():
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
opt = parse_args()
batch_size = opt.batch_size
classes = 10
num_gpus = opt.num_gpus
batch_size *= max(1, num_gpus)
context = [mx.gpu(i)
for i in range(num_gpus)] if num_gpus > 0 else [mx.cpu()]
num_workers = opt.num_workers
lr_sch = lr_scheduler.CosineScheduler((50000//batch_size)*opt.num_epochs,
base_lr=opt.lr,
warmup_steps=5*(50000//batch_size),
final_lr=1e-5)
# lr_sch = lr_scheduler.FactorScheduler((50000//batch_size)*20,
# factor=0.2, base_lr=opt.lr,
# warmup_steps=5*(50000//batch_size))
# lr_sch = LRScheduler('cosine',opt.lr, niters=(50000//batch_size)*opt.num_epochs,)
model_name = opt.model
net = SKT_Lite()
# if model_name.startswith('cifar_wideresnet'):
# kwargs = {'classes': classes,
# 'drop_rate': opt.drop_rate}
# else:
# kwargs = {'classes': classes}
# net = get_model(model_name, **kwargs)
if opt.mixup:
model_name += '_mixup'
if opt.amp:
model_name += '_amp'
makedirs('./'+model_name)
os.chdir('./'+model_name)
sw = SummaryWriter(
logdir='.\\tb\\'+model_name, flush_secs=5, verbose=False)
makedirs(opt.save_plot_dir)
if opt.resume_from:
net.load_parameters(opt.resume_from, ctx=context)
optimizer = 'nag'
save_period = opt.save_period
if opt.save_dir and save_period:
save_dir = opt.save_dir
makedirs(save_dir)
else:
save_dir = ''
save_period = 0
plot_name = opt.save_plot_dir
logging_handlers = [logging.StreamHandler()]
if opt.logging_dir:
logging_dir = opt.logging_dir
makedirs(logging_dir)
logging_handlers.append(logging.FileHandler(
'%s/train_cifar10_%s.log' % (logging_dir, model_name)))
logging.basicConfig(level=logging.INFO, handlers=logging_handlers)
logging.info(opt)
if opt.amp:
amp.init()
if opt.profile_mode:
profiler.set_config(profile_all=True,
aggregate_stats=True,
continuous_dump=True,
filename='%s_profile.json' % model_name)
transform_train = transforms.Compose([
gcv_transforms.RandomCrop(32, pad=4),
CutOut(8),
# gcv_transforms.block.RandomErasing(s_max=0.25),
transforms.RandomFlipLeftRight(),
# transforms.RandomFlipTopBottom(),
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize([0.4914, 0.4822, 0.4465],
[0.2023, 0.1994, 0.2010])
])
transform_test = transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize([0.4914, 0.4822, 0.4465],
[0.2023, 0.1994, 0.2010])
])
def label_transform(label, classes):
ind = label.astype('int')
res = nd.zeros((ind.shape[0], classes), ctx=label.context)
res[nd.arange(ind.shape[0], ctx=label.context), ind] = 1
return res
def test(ctx, val_data):
metric = mx.metric.Accuracy()
loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
num_batch = len(val_data)
test_loss = 0
for i, batch in enumerate(val_data):
data = gluon.utils.split_and_load(
batch[0], ctx_list=ctx, batch_axis=0)
label = gluon.utils.split_and_load(
batch[1], ctx_list=ctx, batch_axis=0)
outputs = [net(X) for X in data]
loss = [loss_fn(yhat, y) for yhat, y in zip(outputs, label)]
metric.update(label, outputs)
test_loss += sum([l.sum().asscalar() for l in loss])
test_loss /= batch_size * num_batch
name, val_acc = metric.get()
return name, val_acc, test_loss
def train(epochs, ctx):
if isinstance(ctx, mx.Context):
ctx = [ctx]
net.initialize(mx.init.MSRAPrelu(), ctx=ctx)
root = os.path.join('..', 'datasets', 'cifar-10')
train_data = gluon.data.DataLoader(
gluon.data.vision.CIFAR10(
root=root, train=True).transform_first(transform_train),
batch_size=batch_size, shuffle=True, last_batch='discard', num_workers=num_workers)
val_data = gluon.data.DataLoader(
gluon.data.vision.CIFAR10(
root=root, train=False).transform_first(transform_test),
batch_size=batch_size, shuffle=False, num_workers=num_workers)
trainer = gluon.Trainer(net.collect_params(), optimizer,
{'learning_rate': opt.lr, 'wd': opt.wd,
'momentum': opt.momentum, 'lr_scheduler': lr_sch})
if opt.amp:
amp.init_trainer(trainer)
metric = mx.metric.Accuracy()
train_metric = mx.metric.RMSE()
loss_fn = gluon.loss.SoftmaxCrossEntropyLoss(
sparse_label=False if opt.mixup else True)
train_history = TrainingHistory(['training-error', 'validation-error'])
# acc_history = TrainingHistory(['training-acc', 'validation-acc'])
loss_history = TrainingHistory(['training-loss', 'validation-loss'])
iteration = 0
best_val_score = 0
for epoch in range(epochs):
tic = time.time()
train_metric.reset()
metric.reset()
train_loss = 0
num_batch = len(train_data)
alpha = 1
for i, batch in enumerate(train_data):
if epoch == 0 and iteration == 1 and opt.profile_mode:
profiler.set_state('run')
lam = np.random.beta(alpha, alpha)
if epoch >= epochs - 20 or not opt.mixup:
lam = 1
data_1 = gluon.utils.split_and_load(
batch[0], ctx_list=ctx, batch_axis=0)
label_1 = gluon.utils.split_and_load(
batch[1], ctx_list=ctx, batch_axis=0)
if not opt.mixup:
data = data_1
label = label_1
else:
data = [lam*X + (1-lam)*X[::-1] for X in data_1]
label = []
for Y in label_1:
y1 = label_transform(Y, classes)
y2 = label_transform(Y[::-1], classes)
label.append(lam*y1 + (1-lam)*y2)
with ag.record():
output = [net(X) for X in data]
loss = [loss_fn(yhat, y) for yhat, y in zip(output, label)]
if opt.amp:
with ag.record():
with amp.scale_loss(loss, trainer) as scaled_loss:
ag.backward(scaled_loss)
# scaled_loss.backward()
else:
for l in loss:
l.backward()
trainer.step(batch_size)
train_loss += sum([l.sum().asscalar() for l in loss])
output_softmax = [nd.SoftmaxActivation(out) for out in output]
train_metric.update(label, output_softmax)
metric.update(label_1, output_softmax)
name, acc = train_metric.get()
sw.add_scalar(tag='lr', value=trainer.learning_rate,
global_step=iteration)
if epoch == 0 and iteration == 1 and opt.profile_mode:
nd.waitall()
profiler.set_state('stop')
iteration += 1
train_loss /= batch_size * num_batch
name, acc = train_metric.get()
_, train_acc = metric.get()
name, val_acc, _ = test(ctx, val_data)
if opt.mixup:
train_history.update([acc, 1-val_acc])
plt.cla()
train_history.plot(save_path='%s/%s_history.png' %
(plot_name, model_name))
else:
train_history.update([1-train_acc, 1-val_acc])
plt.cla()
train_history.plot(save_path='%s/%s_history.png' %
(plot_name, model_name))
# acc_history.update([train_acc, val_acc])
# plt.cla()
# acc_history.plot(save_path='%s/%s_acc.png' %
# (plot_name, model_name), legend_loc='best')
if val_acc > best_val_score:
best_val_score = val_acc
net.save_parameters('%s/%.4f-cifar-%s-%d-best.params' %
(save_dir, best_val_score, model_name, epoch))
current_lr = trainer.learning_rate
name, val_acc, val_loss = test(ctx, val_data)
loss_history.update([train_loss, val_loss])
plt.cla()
loss_history.plot(save_path='%s/%s_loss.png' %
(plot_name, model_name), y_lim=(0, 2), legend_loc='best')
logging.info('[Epoch %d] loss=%f train_acc=%f train_RMSE=%f\n val_acc=%f val_loss=%f lr=%f time: %f' %
(epoch, train_loss, train_acc, acc, val_acc, val_loss, current_lr, time.time()-tic))
sw._add_scalars(tag='Acc',
scalar_dict={'train_acc': train_acc, 'test_acc': val_acc}, global_step=epoch)
sw._add_scalars(tag='Loss',
scalar_dict={'train_loss': train_loss, 'test_loss': val_loss}, global_step=epoch)
if save_period and save_dir and (epoch + 1) % save_period == 0:
net.save_parameters('%s/cifar10-%s-%d.params' %
(save_dir, model_name, epoch))
if save_period and save_dir:
net.save_parameters('%s/cifar10-%s-%d.params' %
(save_dir, model_name, epochs-1))
if opt.mode == 'hybrid':
net.hybridize()
train(opt.num_epochs, context)
if opt.profile_mode:
profiler.dump(finished=False)
sw.close()
def run(self):
# Helper methods
def get_random_lot(data_loader):
return next(iter(data_loader))
# Data importing, pre-processing, and loading
num_training_examples, num_testing_examples, train_data_lot_iterator, train_data_eval_iterator, test_data = self._load_data(
)
# parameters calculated from loaded data
self._num_training_examples = num_training_examples
self._num_testing_examples = num_testing_examples
self._hyperparams[
'sample_fraction'] = self._lot_size / num_training_examples
rounds_per_epoch = round(num_training_examples / self._lot_size)
# Set up privacy accountant
accountant = rdp_acct.anaRDPacct() # dpacct.anaCGFAcct()
eps_sequence = []
# Network structure creation
self._create_network_params()
# Loss function
loss_func = self._get_loss_func()
# Optimization procedure
trainer = self._optimizer(self._hyperparams, self._net, self._params,
loss_func, self._model_ctx, accountant)
# begin profiling if enabled
if self._enable_mxnet_profiling:
from mxnet import profiler
profiler.set_config(profile_all=True,
aggregate_stats=True,
filename='profile_output.json')
profiler.set_state('run')
# Training sequence
rounds = round(self._epochs * rounds_per_epoch)
loss_sequence = []
current_epoch_loss = mx.nd.zeros(1, ctx=self._model_ctx)
for t in range(1, rounds + 1):
if self._verbose and self._print_epoch_status:
# show current epoch progress
epoch_number = 1 + (t - 1) // rounds_per_epoch
epoch_progress = 1 + (t - 1) % rounds_per_epoch
printProgressBar(
epoch_progress,
rounds_per_epoch,
prefix='Epoch {} progress:'.format(epoch_number),
length=50)
if self._run_training:
# prepare random lot of data for DPSGD step
data, labels = get_random_lot(train_data_lot_iterator)
data = data.as_in_context(self._model_ctx).reshape(
(-1, 1, self._input_layer))
labels = labels.as_in_context(self._model_ctx)
else:
data, labels = [], []
# perform DPSGD step
lot_mean_loss = trainer.step(
data,
labels,
accumulate_privacy=self._accumulate_privacy,
run_training=self._run_training)
loss_sequence.append(lot_mean_loss)
current_epoch_loss += lot_mean_loss
# no need to continue running training if NaNs are present
if not np.isfinite(lot_mean_loss):
self._run_training = False
if self._verbose: print("NaN loss on round {}.".format(t))
if self._params_not_finite():
self._run_training = False
if self._verbose:
print("Non-finite parameters on round {}.".format(t))
if self._accumulate_privacy and self._debugging:
eps_sequence.append(accountant.get_eps(self._fixed_delta))
# print some stats after an "epoch"
if t % rounds_per_epoch == 0:
if self._verbose:
print("Epoch {} (round {}) complete.".format(
t / rounds_per_epoch, t))
if self._run_training:
print("mean epoch loss: {}".format(
current_epoch_loss.asscalar() * self._lot_size /
self._num_training_examples))
if self._compute_epoch_accuracy:
print("training accuracy: {}".format(
self._evaluate_accuracy(
train_data_eval_iterator)))
print("testing accuracy: {}".format(
self._evaluate_accuracy(test_data)))
if self._accumulate_privacy and self._debugging:
print("eps used: {}\n".format(eps_sequence[-1]))
print()
current_epoch_loss = mx.nd.zeros(1, ctx=self._model_ctx)
# end profiling if enabled
if self._enable_mxnet_profiling:
mx.nd.waitall()
profiler.set_state('stop')
print(profiler.dumps())
# Make sure we don't report a bogus number
if self._accumulate_privacy:
final_eps = accountant.get_eps(self._fixed_delta)
else:
final_eps = -1
test_accuracy = self._evaluate_accuracy(test_data)
if self._save_plots or self._debugging:
self._create_and_save_plots(t, eps_sequence, loss_sequence,
final_eps, test_accuracy)
return final_eps, test_accuracy
# %% [markdown]
# # Train
# %%
trainer1 = {
k: gluon.Trainer(v.collect_params(), 'adagrad', {'clip_gradient': 1.25})
for (k, v) in net1.items()
}
trainer2 = gluon.Trainer(net2.collect_params(), 'adagrad',
{'clip_gradient': 1.25})
loss = gluon.loss.L2Loss()
# %%
profiler.set_config(profile_all=True,
profile_imperative=True,
aggregate_stats=True,
continuous_dump=True,
filename='profile.json')
# %%
def train_model(dataiter, epoch):
train_loss = 0
total_size = 0
for i, batch in enumerate(dataiter):
with mx.autograd.record():
# iterate over the left and right question
embs = []
data_lists = []
for k in range(2):
embedding = [
def main():
data_p = Path('/storage/data/').resolve()
checkpoint_p = Path('./checkpoints/').resolve()
checkpoint_p.mkdir(parents=True, exist_ok=True)
logs_p = Path('./logs/').resolve()
shutil.rmtree(logs_p, ignore_errors=True)
encoder = SevenPlaneEncoder((19, 19))
builder = SGFDatasetBuilder(data_p, encoder=encoder)
builder.download_and_prepare()
train_itr = builder.train_dataset(batch_size=BATCH_SIZE,
max_worker=cpu_count(),
factor=FACTOR)
test_itr = builder.test_dataset(batch_size=BATCH_SIZE,
max_worker=cpu_count(),
factor=FACTOR)
# build model
betago = Model()
# convert to half-presicion floating point FP16
# NOTE: all NVIDIA GPUs with compute capability 6.1 have a low-rate FP16 performance == FFP16 is not the fast path on these GPUs
# data passed to split_and_load() must be float16 too
#betago.cast('float16')
# hybridize for speed
betago.hybridize(static_alloc=True, static_shape=True)
# print graph
shape = (1, ) + encoder.shape()
mx.viz.print_summary(betago(mx.sym.var('data')), shape={'data': shape})
# pin GPUs
ctx = [mx.gpu(i) for i in range(GPU_COUNT)]
# optimizer
opt_params = {
'learning_rate': 0.001,
'beta1': 0.9,
'beta2': 0.999,
'epsilon': 1e-08
}
opt = mx.optimizer.create('adam', **opt_params)
# initialize parameters
# MXNet initializes the weight matrices uniformly by drawing from [−0.07,0.07], bias parameters are all set to 0
# 'Xavier': initializer is designed to keep the scale of gradients roughly the same in all layers
betago.initialize(mx.init.Xavier(magnitude=2.3),
ctx=ctx,
force_reinit=True)
# fetch and broadcast parameters
params = betago.collect_params()
# trainer
trainer = Trainer(params=params, optimizer=opt, kvstore='device')
# loss function
loss_fn = SoftmaxCrossEntropyLoss()
# use accuracy as the evaluation metric
metric = Accuracy()
with mxb.SummaryWriter(logdir='./logs') as sw:
# add graph to MXBoard
#betago.forward(mx.nd.ones(shape, ctx=ctx[0]))
#betago.forward(mx.nd.ones(shape, ctx=ctx[1]))
#sw.add_graph(betago)
profiler.set_config(profile_all=True,
aggregate_stats=True,
continuous_dump=True,
filename='profile_output.json')
start = time.perf_counter()
# train
for e in range(EPOCHS):
if 0 == e:
profiler.set_state('run')
tick = time.time()
# reset the train data iterator.
train_itr.reset()
# loop over the train data iterator
for i, batch in enumerate(train_itr):
if 0 == i:
tick_0 = time.time()
# splits train data into multiple slices along batch_axis
# copy each slice into a context
data = split_and_load(batch.data[0],
ctx_list=ctx,
batch_axis=0,
even_split=False)
# splits train label into multiple slices along batch_axis
# copy each slice into a context
label = split_and_load(batch.label[0],
ctx_list=ctx,
batch_axis=0,
even_split=False)
outputs = []
losses = []
# inside training scope
with ag.record():
for x, y in zip(data, label):
z = betago(x)
# computes softmax cross entropy loss
l = loss_fn(z, y)
outputs.append(z)
losses.append(l)
# backpropagate the error for one iteration
for l in losses:
l.backward()
# make one step of parameter update.
# trainer needs to know the batch size of data
# to normalize the gradient by 1/batch_size
trainer.step(BATCH_SIZE)
# updates internal evaluation
metric.update(label, outputs)
# Print batch metrics
if 0 == i % PRINT_N and 0 < i:
# checkpointing
betago.save_parameters(
str(checkpoint_p.joinpath(
'betago-{}.params'.format(e))))
sw.add_scalar(tag='Accuracy',
value={'naive': metric.get()[1]},
global_step=i - PRINT_N)
sw.add_scalar(tag='Speed',
value={
'naive':
BATCH_SIZE * (PRINT_N) /
(time.time() - tick)
},
global_step=i - PRINT_N)
print(
'epoch[{}] batch [{}], accuracy {:.4f}, samples/sec: {:.4f}'
.format(e, i,
metric.get()[1],
BATCH_SIZE * (PRINT_N) / (time.time() - tick)))
tick = time.time()
if 0 == e:
profiler.set_state('stop')
profiler.dump()
# gets the evaluation result
print('epoch [{}], accuracy {:.4f}, samples/sec: {:.4f}'.format(
e,
metric.get()[1],
BATCH_SIZE * (i + 1) / (time.time() - tick_0)))
# reset evaluation result to initial state
metric.reset()
elapsed = time.perf_counter() - start
print('elapsed: {:0.3f}'.format(elapsed))
# use Accuracy as the evaluation metric
metric = Accuracy()
for batch in test_itr:
data = split_and_load(batch.data[0], ctx_list=ctx, batch_axis=0)
label = split_and_load(batch.label[0], ctx_list=ctx, batch_axis=0)
outputs = []
for x in data:
outputs.append(betago(x))
metric.update(label, outputs)
print('validation %s=%f' % metric.get())
def train(
args,
model,
train_sampler,
valid_samplers=None,
rank=0,
rel_parts=None,
barrier=None,
):
assert args.num_proc <= 1, "MXNet KGE does not support multi-process now"
assert (args.rel_part == False
), "No need for relation partition in single process for MXNet KGE"
logs = []
for arg in vars(args):
logging.info("{:20}:{}".format(arg, getattr(args, arg)))
if len(args.gpu) > 0:
gpu_id = (args.gpu[rank % len(args.gpu)]
if args.mix_cpu_gpu and args.num_proc > 1 else args.gpu[0])
else:
gpu_id = -1
if args.strict_rel_part:
model.prepare_relation(mx.gpu(gpu_id))
if mxprofiler:
from mxnet import profiler
profiler.set_config(
profile_all=True,
aggregate_stats=True,
continuous_dump=True,
filename="profile_output.json",
)
start = time.time()
for step in range(0, args.max_step):
pos_g, neg_g = next(train_sampler)
args.step = step
if step == 1 and mxprofiler:
profiler.set_state("run")
with mx.autograd.record():
loss, log = model.forward(pos_g, neg_g, gpu_id)
loss.backward()
logs.append(log)
model.update(gpu_id)
if step % args.log_interval == 0:
for k in logs[0].keys():
v = sum(l[k] for l in logs) / len(logs)
print("[Train]({}/{}) average {}: {}".format(
step, args.max_step, k, v))
logs = []
print(time.time() - start)
start = time.time()
if (args.valid and step % args.eval_interval == 0 and step > 1
and valid_samplers is not None):
start = time.time()
test(args, model, valid_samplers, mode="Valid")
print("test:", time.time() - start)
if args.strict_rel_part:
model.writeback_relation(rank, rel_parts)
if mxprofiler:
nd.waitall()
profiler.set_state("stop")
profiler.dump()
print(profiler.dumps())
# clear cache
logs = []
from mxnet import profiler
import Graph
from learners.IterativeLearner import learn_iterative
from common import data_ctx, measure_time
from feature_transformations import FeatureTransformation
from feature_transformations.FeatureScalingTransformation import FeatureScalingTransformation
from feature_transformations.KernelTransformation import KernelTransformation
from feature_transformations.LinearConvolutionTransformation import LinearConvolutionTransformation
from feature_transformations.PcaTransformation import PcaTransformation
############## PARAMETERS ###################
from feature_transformations.RealFeatureScalingTransformation import RealFeatureScalingTransformation
from learners.IterativeLogisticLearner import learn_iterative_logistic
profiler.set_config(aggregate_stats=True, filename='profile_output.json')
mx.random.seed(101)
random.seed(101)
def learn(graph, net_type, training_set, test_set, iterations_per_epoch,
batch_size):
def try_standard_approach(approach):
def get_all_vertices(data_loader: DataLoader):
res_X = []
for X, y in data_loader:
for x in X:
res_X.append(round(float(x.asscalar())))
return res_X
import logging
import argparse
import os
import sys
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import mxnet as mx
from models.E3DNet import create_m3d
from lib.data import ClipBatchIter
from mxnet import profiler
train_list = ["fc", "comp_17", "comp_16", "comp_15", "comp_14", "softmax"]
tmp_pool_list = ["final_fc", "softmax_label"]
profiler.set_config(profile_all=True,
aggregate_stats=True,
filename='profile_output_m3d.json')
def plot_schedule(schedule_fn, iterations=1500):
# Iteration count starting at 1
iterations = [i + 1 for i in range(iterations)]
lrs = [schedule_fn(i) for i in iterations]
plt.scatter(iterations, lrs)
plt.xlabel("Iteration")
plt.ylabel("Learning Rate")
#plt.savefig('learning_rate.png')
def train(args):
gpus = [int(i) for i in args.gpus.split(',')]