def insert2db(rcu, qcu, ic_obj, created, db):
# making a new row, "testlab", it will be recorded to the db anyway
# regardless of the usage of other accounts, i.e. even when all of
# them are 0, it's very important since we need a datapoint at this
# snapshot!
# rcu, qcu could be {}, {}, meaning cluster is down, else they can
# be rcu, qcu could only have a bunch of {key:0}s
#### Here is a flaw in the design because when the cluster is
#### down, the 0 core usages ({},{}) cannot be distinguished from the real
#### 0 core usages ({u1:0, u2:0}, {u1:0, u2:0})
for dd in [rcu, qcu]:
dd.update(dict(testlab=sum(dd.values())))
# remove items where its value is 0,
rcu = util.prune(rcu, preserved_keys=['testlab'])
qcu = util.prune(qcu, preserved_keys=['testlab'])
# uccm: users consuming cores at the moment (realnames)
# using uccm instead of usermap.values() is trying to eliminate usage items
# where the both of user's runningcores notrunning cores are zero
uccm = set(rcu.keys() + qcu.keys())
# must use rcu.get(realname, 0), qcu.get(realname, 0) instead of
# rcu[realname], or qcu[realname] since the realname could be in either rcu
# or qcu, or both.
for realname in sorted(uccm):
usage = Usage(ic_obj.clustername, realname, rcu.get(realname, 0),
qcu.get(realname, 0), created)
db.session.add(usage)
db.session.commit()
def insert2db(rcu, qcu, ic_obj, created, db):
# making a new row, "testlab", it will be recorded to the db anyway
# regardless of the usage of other accounts, i.e. even when all of
# them are 0, it's very important since we need a datapoint at this
# snapshot!
# rcu, qcu could be {}, {}, meaning cluster is down, else they can
# be rcu, qcu could only have a bunch of {key:0}s
#### Here is a flaw in the design because when the cluster is
#### down, the 0 core usages ({},{}) cannot be distinguished from the real
#### 0 core usages ({u1:0, u2:0}, {u1:0, u2:0})
for dd in [rcu, qcu]:
dd.update(dict(testlab=sum(dd.values())))
# remove items where its value is 0,
rcu = util.prune(rcu, preserved_keys=['testlab'])
qcu = util.prune(qcu, preserved_keys=['testlab'])
# uccm: users consuming cores at the moment (realnames)
# using uccm instead of usermap.values() is trying to eliminate usage items
# where the both of user's runningcores notrunning cores are zero
uccm = set(rcu.keys() + qcu.keys())
# must use rcu.get(realname, 0), qcu.get(realname, 0) instead of
# rcu[realname], or qcu[realname] since the realname could be in either rcu
# or qcu, or both.
for realname in sorted(uccm):
usage = Usage(
ic_obj.clustername,
realname,
rcu.get(realname, 0),
qcu.get(realname, 0),
created)
db.session.add(usage)
db.session.commit()
def gen_report(self, rcu, qcu, usermap, created=None):
# first, title of the report
title = "{0}|{1}|{2}".format(self.clustername, self.quota,
self.cores_per_node)
# second, created datetime
created = created if created else datetime.datetime.now()
report_content = []
# datetime, an addtional "\n" is just for making the reportprettier
report_content.append("{0}\n".format(util.format_datetime(created)))
if not rcu and not qcu:
report_content.append("data not available at the moment")
else:
rcu, qcu = util.prune(rcu), util.prune(qcu)
total_usage = {}
for realname in set(usermap.values()):
total_usage[realname] = sum(
dd.get(realname, 0) for dd in [rcu, qcu])
total_usage = util.prune(total_usage)
# from this step on, basically it's about print data from 3 dicts
# in a pretty way: rcu, qcu, total_usage
# 1. print headers
report_content.append("{0:13s} {1:8s} {2:8s} {3:8s}".format(
'USERNAME', 'Running', 'NotRunning', 'TOTAL'))
# 2. sort the order of key by total_usage
sorted_keys = reversed(sorted(total_usage, key=total_usage.get))
if not len(total_usage) == 0: # not the usage of everyone is zero
report_content.append('=' * 44)
# 3. print the table
for k in sorted_keys:
# full name is too long, so last name is used since
# firstname is confusing
name = k.split()[0]
report_content.append(
"{0:13s} {1:<8d} {2:<8d} {3:<8d}".format(
name, rcu.get(k, 0), qcu.get(k, 0),
total_usage.get(k, 0)))
# 4. print the footer sum
report_content.append('=' * 44)
report_content.append("{0:13s} {1:<8d} {2:<8d} {3:<8d}".format(
'SUM', sum(rcu.values()), sum(qcu.values()),
sum(total_usage.values())))
report_content.append('=' * 44)
# 5. join the final work
report_content = '\n'.join(report_content)
# 6. since it's displayed on line, need such replacements
report_content = report_content.replace("\n",
"<br>").replace(" ", " ")
return Report(self, report_content, created)
def gen_report(self, rcu, qcu, usermap, created=None):
# first, title of the report
title = "{0}|{1}|{2}".format(self.clustername, self.quota, self.cores_per_node)
# second, created datetime
created = created if created else datetime.datetime.now()
report_content = []
# datetime, an addtional "\n" is just for making the reportprettier
report_content.append("{0}\n".format(util.format_datetime(created)))
if not rcu and not qcu:
report_content.append("data not available at the moment")
else:
rcu, qcu = util.prune(rcu), util.prune(qcu)
total_usage = {}
for realname in set(usermap.values()):
total_usage[realname] = sum(dd.get(realname, 0) for dd in [rcu, qcu])
total_usage = util.prune(total_usage)
# from this step on, basically it's about print data from 3 dicts
# in a pretty way: rcu, qcu, total_usage
# 1. print headers
report_content.append("{0:13s} {1:8s} {2:8s} {3:8s}".format(
'USERNAME', 'Running', 'NotRunning', 'TOTAL'))
# 2. sort the order of key by total_usage
sorted_keys = reversed(sorted(total_usage, key=total_usage.get))
if not len(total_usage) == 0: # not the usage of everyone is zero
report_content.append('=' * 44)
# 3. print the table
for k in sorted_keys:
# full name is too long, so last name is used since
# firstname is confusing
name = k.split()[0]
report_content.append("{0:13s} {1:<8d} {2:<8d} {3:<8d}".format(
name, rcu.get(k, 0), qcu.get(k, 0), total_usage.get(k, 0)))
# 4. print the footer sum
report_content.append('=' * 44)
report_content.append("{0:13s} {1:<8d} {2:<8d} {3:<8d}".format(
'SUM', sum(rcu.values()), sum(qcu.values()),
sum(total_usage.values())))
report_content.append('=' * 44)
# 5. join the final work
report_content = '\n'.join(report_content)
# 6. since it's displayed on line, need such replacements
report_content = report_content.replace("\n", "<br>").replace(" ", " ")
return Report(self, report_content, created)
def configure(save=False, machine=None):
"""
Read configuration files
"""
cwd = os.getcwd()
path = os.path.realpath(os.path.dirname(__file__))
os.chdir(path)
conf = {}
exec open('conf/conf.py') in conf
if not machine and os.path.isfile('machine'):
machine = open('machine').read().strip()
if machine:
machine = os.path.basename(machine)
path = os.path.join('conf', machine, 'conf.py')
exec open(path) in conf
conf['machine'] = machine
if save:
open('machine', 'w').write(machine)
util.prune(conf, pattern='(^_)|(^.$)')
os.chdir(cwd)
return conf
def train_model(wrapped_model, model, train_loss_f, model_path, train_loader,
test_loader, init_lr, epochs, args):
val_loss_f = nn.CrossEntropyLoss()
best_model_path = '.'.join(model_path.split('.')[:-1]) + '.best.pth'
# tracking stats
if not hasattr(model, 'stats'):
model.stats = {
'train_loss': [],
'test_acc': [],
'test_loss': [],
'weight': [],
'lr': [],
'macs': [],
'efficiency': []
}
start_epoch = 1
else:
start_epoch = len(model.stats['test_loss'])
curr_weights, _ = util.num_nonzeros(model)
if hasattr(model, 'packed_layer_size'):
macs = np.sum([x * y for x, y in model.packed_layer_size])
else:
macs = curr_weights
# optimizer
optimizer = optim.RMSprop(util.group_weight(model),
lr=init_lr,
momentum=0.9,
alpha=0.9,
weight_decay=4e-5,
eps=1.0)
print("Optimizer:")
print(optimizer)
best_acc = 0
prune_epoch = 0
max_prune_rate = 0.8
final_prune_epoch = int(0.9 * args.epochs)
num_prune_epochs = 10
prune_rates = [
max_prune_rate * (1 - (1 - (i / num_prune_epochs))**3)
for i in range(num_prune_epochs)
]
prune_rates[-1] = max_prune_rate
prune_epochs = np.linspace(0, final_prune_epoch,
num_prune_epochs).astype('i').tolist()
prune_rate = 0.1
prune_total = 0.0
prune_cycle = 8
max_prune = 0.7
# pruning stage
for epoch in range(start_epoch, epochs + 1):
print('[Epoch {}]'.format(epoch))
for g in optimizer.param_groups:
lr = g['lr']
break
if epoch % prune_cycle == 0 and prune_total < max_prune:
prune_total += prune_rate
print('Prune Total: {:2.2f}'.format(100. * prune_total))
util.prune(model, prune_total)
packing.pack_model(model, args.gamma)
macs = np.sum([x * y for x, y in model.packed_layer_size])
curr_weights, num_weights = util.num_nonzeros(model)
train_loss = util.train(train_loader, wrapped_model, train_loss_f,
optimizer, epoch - 1, args)
test_loss, test_acc = util.validate(test_loader, model, val_loss_f,
epoch - 1, args)
print('LR :: {}'.format(lr))
print('Train Loss:: {}'.format(train_loss))
print('Test Loss:: {}'.format(test_loss))
print('Test Acc.:: {}'.format(test_acc))
print('Nonzeros :: {}'.format(curr_weights))
print('')
print('')
model.stats['lr'].append(lr)
model.optimizer = optimizer.state_dict()
model.cpu()
torch.save(model, model_path)
if test_acc > best_acc and prune_total >= max_prune:
print('New best model found')
torch.save(model, best_model_path)
best_acc = test_acc
model.cuda()
def stage(inputs):
"""
Setup, and optionally launch, a SORD job.
"""
import glob, time, getopt, shutil
import setup
# Save start time
starttime = time.asctime()
print('SORD setup')
# Read defaults
pm = {}
f = os.path.join(os.path.dirname(__file__), 'parameters.py')
exec open(f) in pm
if 'machine' in inputs:
cf = configure.configure(machine=inputs['machine'])
else:
cf = configure.configure()
# Merge inputs
inputs = inputs.copy()
util.prune(inputs)
util.prune(pm)
util.prune(cf, pattern='(^_)|(^.$)')
for k, v in inputs.iteritems():
if k in cf:
cf[k] = v
elif k in pm:
pm[k] = v
else:
sys.exit('Unknown parameter: %s = %r' % (k, v))
cf = util.namespace(cf)
cf.rundir = os.path.expanduser(cf.rundir)
pm = prepare_param(util.namespace(pm), cf.itbuff)
# Command line options
opts = [
'n',
'dryrun',
's',
'serial',
'm',
'mpi',
'i',
'interactive',
'q',
'queue',
'd',
'debug',
'g',
'debugging',
't',
'testing',
'p',
'profiling',
'O',
'optimized',
'f',
'force',
]
options = ''.join(opts[::2])
long_options = opts[1::2]
opts = getopt.getopt(sys.argv[1:], options, long_options)[0]
for o, v in opts:
if o in ('-n', '--dry-run'):
cf.prepare = False
elif o in ('-s', '--serial'):
cf.mode = 's'
elif o in ('-m', '--mpi'):
cf.mode = 'm'
elif o in ('-i', '--interactive'):
cf.run = 'i'
elif o in ('-q', '--queue'):
cf.run = 'q'
elif o in ('-d', '--debug'):
cf.optimize = 'g'
cf.run = 'g'
elif o in ('-g', '--debugging'):
cf.optimize = 'g'
elif o in ('-t', '--testing'):
cf.optimize = 't'
elif o in ('-p', '--profiling'):
cf.optimize = 'p'
elif o in ('-O', '--optimized'):
cf.optimize = 'O'
elif o in ('-f', '--force'):
if os.path.isdir(cf.rundir):
shutil.rmtree(cf.rundir)
else:
sys.exit('Error: unknown option: ' + o)
if not cf.prepare:
cf.run = False
# Partition for parallelization
pm.nn = tuple(int(i) for i in pm.nn)
maxtotalcores = cf.maxnodes * cf.maxcores
if not cf.mode and maxtotalcores == 1:
cf.mode = 's'
np3 = pm.np3[:]
if cf.mode == 's':
np3 = [1, 1, 1]
nl = [(pm.nn[i] - 1) / np3[i] + 1 for i in range(3)]
i = abs(pm.faultnormal) - 1
if i >= 0:
nl[i] = max(nl[i], 2)
pm.np3 = tuple((pm.nn[i] - 1) / nl[i] + 1 for i in range(3))
cf.np = pm.np3[0] * pm.np3[1] * pm.np3[2]
if not cf.mode:
cf.mode = 's'
if cf.np > 1:
cf.mode = 'm'
# Resources
if cf.maxcores:
cf.nodes = min(cf.maxnodes, (cf.np - 1) / cf.maxcores + 1)
cf.ppn = (cf.np - 1) / cf.nodes + 1
cf.cores = min(cf.maxcores, cf.ppn)
cf.totalcores = cf.nodes * cf.maxcores
else:
cf.nodes = 1
cf.ppn = cf.np
cf.cores = cf.np
cf.totalcores = cf.np
# RAM and Wall time usage
if pm.oplevel in (1, 2):
nvars = 20
elif pm.oplevel in (3, 4, 5):
nvars = 23
else:
nvars = 44
nm = (nl[0] + 2) * (nl[1] + 2) * (nl[2] + 2)
cf.pmem = 32 + int(1.2 * nm * nvars * int(cf.dtype[-1]) / 1024 / 1024)
cf.ram = cf.pmem * cf.ppn
ss = (pm.nt + 10) * cf.ppn * nm / cf.cores / cf.rate
sus = int(ss / 3600 * cf.totalcores + 1)
mm = ss / 60 * 3.0 + 10
if cf.maxtime:
mm = min(mm, 60 * cf.maxtime[0] + cf.maxtime[1])
mm = mm * 3
hh = mm / 60
mm = mm % 60
cf.walltime = '%d:%02d:00' % (hh, mm) #v1.1.01 2 times of orignal hours
cf.walltime = '1:00:00' # used in v1.1
print('Machine: ' + cf.machine)
print('Cores: %s of %s' % (cf.np, maxtotalcores))
print('Nodes: %s of %s' % (cf.nodes, cf.maxnodes))
print('RAM: %sMb of %sMb per node' % (cf.ram, cf.maxram))
print('Time limit: ' + cf.walltime)
print('SUs: %s' % sus)
if cf.maxcores and cf.ppn > cf.maxcores:
print('Warning: exceding available cores per node (%s)' % cf.maxcores)
if cf.ram and cf.ram > cf.maxram:
print('Warning: exceding available RAM per node (%sMb)' % cf.maxram)
# Compile code
if not cf.prepare:
return cf
setup.build(cf.mode, cf.optimize)
# Create run directory
print('Run directory: ' + cf.rundir)
try:
os.makedirs(cf.rundir)
except (OSError):
sys.exit('%r exists or cannot be created. Use --force to overwrite.' %
cf.rundir)
for f in 'in', 'out', 'prof', 'stats', 'debug', 'checkpoint':
os.mkdir(os.path.join(cf.rundir, f))
# Copy files to run directory
cwd = os.path.realpath(os.getcwd())
cf.rundate = time.asctime()
cf.name = os.path.basename(cf.rundir)
cf.rundir = os.path.realpath(cf.rundir)
os.chdir(os.path.realpath(os.path.dirname(__file__)))
cf.bin = os.path.join('.', 'sord-' + cf.mode + cf.optimize)
path = os.path.join('bin', 'sord-' + cf.mode + cf.optimize)
shutil.copy(path, cf.rundir)
if os.path.isfile('sord.tgz'):
shutil.copy('sord.tgz', cf.rundir)
if cf.optimize == 'g':
for f in glob.glob(os.path.join('src', '*.f90')):
shutil.copy(f, cf.rundir)
f = os.path.join('conf', cf.machine, 'templates')
if not os.path.isdir(f):
f = os.path.join('conf', 'default', 'templates')
for d in os.path.join('conf', 'common', 'templates'), f:
for f in glob.glob(os.path.join(d, '*')):
ff = os.path.join(cf.rundir, os.path.basename(f))
out = open(f).read() % cf.__dict__
open(ff, 'w').write(out)
shutil.copymode(f, ff)
# Combine metadata
meta = util.namespace(pm.__dict__)
for k in 'name', 'rundate', 'rundir', 'user', 'os_', 'dtype':
setattr(meta, k, getattr(cf, k))
meta.indices = {}
meta.xi = {}
for f in meta.fieldio:
op, filename = f[0], f[8]
if filename != '-':
meta.indices[filename] = f[7]
if 'wi' in op:
meta.xi[filename] = f[4]
meta.shape = {}
for k in meta.indices:
nn = [(i[1] - i[0]) / i[2] + 1 for i in meta.indices[k]]
nn = [n for n in nn if n > 1]
if nn == []:
nn = [1]
meta.shape[k] = nn
# Write files
os.chdir(cf.rundir)
log = open('log', 'w')
log.write(starttime + ': setup started\n')
util.save('conf.py', cf, prune_pattern='(^_)|(^.$)')
util.save('parameters.py', pm, expand=['fieldio'])
util.save('meta.py', meta, expand=['shape', 'xi', 'indices', 'fieldio'])
# Return to initial directory
os.chdir(cwd)
return cf
def train(model, train_loader, val_loader, args):
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs)
prune_epoch = 0
max_prune_rate = 0.85
max_prune_rate = 0.8
final_prune_epoch = int(0.5 * args.epochs)
num_prune_epochs = 10
prune_rates = [
max_prune_rate * (1 - (1 - (i / num_prune_epochs))**3)
for i in range(num_prune_epochs)
]
prune_rates[-1] = max_prune_rate
prune_epochs = np.linspace(0, final_prune_epoch,
num_prune_epochs).astype('i').tolist()
print("Pruning Epochs: {}".format(prune_epochs))
print("Pruning Rates: {}".format(prune_rates))
curr_weights, num_weights = util.num_nonzeros(model)
macs = curr_weights
model.stats = {
'train_loss': [],
'test_acc': [],
'test_loss': [],
'weight': [],
'lr': [],
'macs': [],
'efficiency': []
}
best_path = args.save_path.split('.pth')[0] + '.best.pth'
best_test_acc = 0
for epoch in range(1, args.epochs + 1):
scheduler.step()
for g in optimizer.param_groups:
lr = g['lr']
break
# prune smallest weights up to a set prune_rate
if epoch in prune_epochs:
util.prune(model, prune_rates[prune_epoch])
curr_weights, num_weights = util.num_nonzeros(model)
packing.pack_model(model, args.gamma)
macs = np.sum([x * y for x, y in model.packed_layer_size])
curr_weights, num_weights = util.num_nonzeros(model)
prune_epoch += 1
if epoch == prune_epochs[-1]:
# disable l1 penalty, as target sparsity is reached
args.l1_penalty = 0
print(' :: [{}]\tLR {:.4f}\tNonzeros ({}/{})'.format(
epoch, lr, curr_weights, num_weights))
train_loss = util.train(train_loader, model, criterion, optimizer,
epoch, args)
test_loss, test_acc = util.validate(val_loader, model, criterion,
epoch, args)
is_best = test_acc > best_test_acc
best_test_acc = max(test_acc, best_test_acc)
model.stats['lr'].append(lr)
model.stats['macs'].append(macs)
model.stats['weight'].append(curr_weights)
model.stats['efficiency'].append(100.0 * (curr_weights / macs))
model.optimizer = optimizer.state_dict()
model.epoch = epoch
model.cpu()
torch.save(model, args.save_path)
if is_best:
torch.save(model, best_path)
model.cuda()
