def compress(fn):
u"压缩合并后的文件，需要网络支持"
print "compressing..."
path, filename = os.path.split(fn)
compressed_fn = os.path.join(path, filename.replace("-pkg.js", "-pkg-min.js"))
compressor(compressed_fn, fn)
print "compressed!"
def test_set_suffix2(self):
''' Test case than time suffix admin dont want to compressed file
'''
now = datetime.strptime("2019-01-01", "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
self.clear_environment(compressor1.conf)
compressor1.conf["add_date_suffix"] = False
compressor1.set_suffix()
self.assertEqual(compressor1.suffix, '', 'bad suffix settings')
def test_set_suffix1(self):
''' Tests that if configured by admin that suffix will be append to compressed file
'''
now = datetime.strptime("2019-01-01", "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
self.clear_environment(compressor1.conf)
compressor1.conf["add_date_suffix"] = True
compressor1.set_suffix()
self.assertEqual(compressor1.suffix, '2019-01-01',
'bad suffix settings')
def test_not_recursive_searching(self):
''' Test case that recusive searching in dir tree is not allowed.
'''
now = datetime.strptime("2019-12-01", "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
self.clear_environment(compressor1.conf)
compressor1.conf["recursive"] = False
self.create_subdir_tree(compressor1.conf["direcrory"])
compressor1.find_files(compressor1.conf["direcrory"])
#print(self.subdirs)
for subd in self.subdirs:
self.assertNotIn(subd, compressor1.searched_directories,
'subdir wasnt searched')
def test_maturation_is_lower(self):
''' In days before plan compression must not be done.
Tests 1 day before plan.
'''
tested_data = ("2019-01-01", "2019-01-02", "2019-10-25", "2019-11-30",
"2019-12-01", "2019-12-31")
for x in tested_data:
now = datetime.strptime(x, "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
self.clear_environment(compressor1.conf)
diff = timedelta(days=compressor1.conf["job_days_step"] - 1)
last_date = now - diff
compressor1.set_timestamp(last_date)
result = compressor1.is_job_maturate()
self.assertEqual(result, False, 'it run a job before maturation')
def test_maturation_is_higher(self):
''' Tests situation that server was shutdown in time when job was planed. May be plan was 30th day.
After repearing sever cron will start program. It is higher than 30th, but compression should be done
'''
tested_data = ("2019-01-01", "2019-01-02", "2019-10-25", "2019-11-30",
"2019-12-01", "2019-12-31")
for x in tested_data:
now = datetime.strptime(x, "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
self.clear_environment(compressor1.conf)
diff = timedelta(days=compressor1.conf["job_days_step"] + 1)
last_date = now - diff
compressor1.set_timestamp(last_date)
result = compressor1.is_job_maturate()
self.assertEqual(result, True,
'it didnt run a job after enough days')
def test_maturation_is_precisely(self):
''' Tests if compression task will start precisely N-th day after last job.
Default in conf file is 30
'''
tested_data = ("2019-01-01", "2019-01-02", "2019-10-25", "2019-11-30",
"2019-12-01", "2019-12-31")
for x in tested_data:
now = datetime.strptime(x, "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
self.clear_environment(compressor1.conf)
diff = timedelta(days=compressor1.conf["job_days_step"])
last_date = now - diff
compressor1.set_timestamp(last_date)
result = compressor1.is_job_maturate()
self.assertEqual(result, True,
'it didnt run a job after enough days')
def test_first_time_run_job(self):
''' In first time job will start file with timestamp dosnt exist. In such case compression must start.
'''
tested_data = ("2019-01-01", "2019-01-02", "2019-10-25", "2019-11-30",
"2019-12-01", "2019-12-31")
for x in tested_data:
now = datetime.strptime(x, "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
#before first run dosnt exist file with timestamp, it is prepared in clear_environment too
self.clear_environment(compressor1.conf)
#but for case that clear_environment will be modified I explicitly check it
timestamp_file_name = compressor1.conf[
"job_name"] + "_timestamp.log"
if (os.path.exists(timestamp_file_name)):
os.remove(timestamp_file_name)
result = compressor1.is_job_maturate()
self.assertEqual(result, True, 'first time run wasnt started')
def test_searching_files(self):
''' test if all files for compression will be find
'''
now = datetime.strptime("2019-12-01", "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
self.clear_environment(compressor1.conf)
compressor1.conf["recursive"] = True
self.create_subdir_tree(compressor1.conf["direcrory"])
self.create_all_random_files(compressor1.conf["exclude_ext"],
compressor1.conf["direcrory"])
compressor1.find_files(compressor1.conf["direcrory"])
#test that files with extension for exclusion are really excluded
for afile in self.files_exclude:
self.assertNotIn(afile, compressor1.files_to_compress,
'file should be excluded and wasnt')
#test that all files for compression are really find
#print(compressor1.files_to_compress)
for afile in self.files_compress:
self.assertIn(afile, compressor1.files_to_compress,
'file should be compress and isnt')
def test_compression(self):
''' Tests if files which are specified for compression are really compressed and
after that are original files deleted
'''
now = datetime.strptime("2019-12-01", "%Y-%m-%d")
compressor1 = compressor.compressor("job1.conf", now)
self.clear_environment(compressor1.conf)
compressor1.conf["recursive"] = True
self.create_subdir_tree(compressor1.conf["direcrory"])
self.create_all_random_files(compressor1.conf["exclude_ext"],
compressor1.conf["direcrory"])
compressor1.find_files(compressor1.conf["direcrory"])
compressor1.set_suffix()
for file_to_compress in compressor1.files_to_compress:
new_compressed_file = compressor1.get_compressed_file_name(
file_to_compress)
compressor1.compress(file_to_compress)
self.assertTrue(os.path.exists(new_compressed_file),
"new compressed file wasnt created")
self.assertFalse(os.path.exists(file_to_compress),
"old file after compression wasnt deleted")
def __init__(self,
params,
lr=0.01,
momentum=0.9,
weight_decay=0,
compression_buffer=False,
all_reduce=False,
local_rank=0,
gpus_per_machine=1,
single_worker=False,
**kwargs):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= momentum:
raise ValueError("Invalid momentum value: {}".format(momentum))
if not 0.0 <= weight_decay:
raise ValueError(
"Invalid weight_decay value: {}".format(weight_decay))
defaults = dict(lr=lr, momentum=momentum, weight_decay=weight_decay)
super(SGD_distribute, self).__init__(params, defaults)
#custom code
self.compression_buffer = compression_buffer
self.all_reduce = all_reduce
self.single_worker = single_worker
self.gpus_per_machine = gpus_per_machine
#gpus_per_machine = torch.cuda.device_count()
print('The number of GPUs is', gpus_per_machine)
print('Single worker', single_worker)
self.MB = 1024 * 1024
self.bucket_size = 200 * self.MB
if self.compression_buffer and not self.single_worker:
self.compressor = compressor.compressor(using_cuda=True,
local_rank=1)
self.local_rank = local_rank
self.global_rank = dist.get_rank()
self.local_dst_in_global = self.global_rank - self.local_rank
self.inter_node_group = []
self.nodes = dist.get_world_size() // gpus_per_machine
self.intra_node_group_list = []
for index in range(self.nodes):
# set inter_node_group
self.inter_node_group.append(0 + index * gpus_per_machine)
# set all intra_node_group
intra_node_group_temp = []
for intra_index in range(gpus_per_machine):
intra_node_group_temp.append(intra_index +
index * gpus_per_machine)
intra_node_group_temp = dist.new_group(intra_node_group_temp)
self.intra_node_group_list.append(intra_node_group_temp)
if self.local_dst_in_global == 0 + index * gpus_per_machine:
self.nodes_rank = index
#self.intra_node_list = self.intra_node_group
self.inter_node_list = self.inter_node_group
self.inter_node_group_list = []
for index in range(len(self.inter_node_list)):
if index is not 0:
temp = dist.new_group(
[self.inter_node_list[0], self.inter_node_list[index]])
self.inter_node_group_list.append(temp)
self.all_gpu = dist.new_group()
self.all_inter_node_group = dist.new_group(self.inter_node_list)
if dist.get_rank() == 0 or dist.get_rank() == 8:
print('nodes', self.nodes)
print('intra_node_group_list', self.intra_node_group_list)
print('inter_node_group', self.inter_node_group_list)
print('all_inter_node_group', self.inter_node_list)
#add time record
self.all_time = Time_recorder()
self.all_reduce_time = Time_recorder()
self.compression_time = Time_recorder()
self.uncompression_time = Time_recorder()
self.broadcast_time = Time_recorder()
self.majority_vote_time = Time_recorder()
self.compress_all_time = Time_recorder()
self.gather_all_time = Time_recorder()
self.calculate_all_time = Time_recorder()
self.broadcast_all_time = Time_recorder()
self.update_para_time = Time_recorder()
self.bucketing_time = Time_recorder()
self.debucketing_time = Time_recorder()
x_test = x_test.astype("float32") / 255
x_test = np.expand_dims(x_test, -1)
num_classes = 10
y_test = keras.utils.to_categorical(y_test, num_classes)
print("Original Model:")
model.evaluate(x_test, y_test)
k = 16
bits = 2
uniform = False #uniform or non uniform quantization of model weights
block_size = 0
c = compressor(model, k, bits, uniform, block_size)
compressedModel = c.compressedModel
parameters = c.parameters
print("Compressed:")
#h = huffmanCompressor(compressedModel, bits, k)
#print("huffman Compression Ratio:", h.compressionRatio)
d = decompressor(compressedModel, parameters, block_size)
decompressedModel = d.decompressedModel
decompressedModel.evaluate(x_test, y_test)
model = keras.models.load_model(modelPath)
q = quantizer(model, bits, uniform)
quantizedModel = q.quantizedModel
daughterSigma=float(args.clustered[1]))
else:
#data['blockage'], data['gamma'] = blockage(density, data['scenario'], data['baseStation'], data['userEquipment'])
data['blockers'], data['blockage'] = blockage(density,
data['scenario'],
data['baseStation'],
data['userEquipment'],
tolerance=2 * bs_radius)
else:
data['blockage'] = []
for j in range(n_BS):
data['blockage'].append([])
data['blockage'][j].append([])
for i in range(data['scenario']['simTime']):
data['blockage'][j][0].append(1)
#if ((i < data['scenario']['simTime']/4)
# or (i > data['scenario']['simTime']/2 and i <= 3*data['scenario']['simTime']/4)):
#else:
# data['blockage'][j][0][i] = 1
#f.close()
compressor(data)
#outname = str(n_BS)+'-'+str(n_UE)+'-'+str(density)+'-'+str(args.vx)+'-'+str(args.vy)+'-'+str(args.seed) #'mobility.json'
#with open(outname, 'w') as outfile:
# json.dump(data, outfile, indent=4)
y = json.dumps(data, indent=4)
print(y)
from intro import intro from input_parser import input_parser from compressor import compressor from printer import printer from robo_browser import robo_browser from misc_data import ListAlbum # This is what runs the program. It calls multiple methods from different files to do so. user_object = intro() album_list = ListAlbum() input_parser(user_object, album_list) album_object = compressor(album_list) printer(album_object) robo_browser(album_object, user_object)
import compressor
import torch
import sys
import time
torch.cuda.set_device(0)
a = torch.randn(22, 100, 224, 224)
a = a.cuda()
#print('before sending')
#print(a)
compressor = compressor.compressor(using_cuda=True)
# warmups to amortize allocation costs
c, size = compressor.compress(a)
d = compressor.uncompress(c, size)
del c, size, d
c, size = compressor.compress(a)
d = compressor.uncompress(c, size)
del c, size, d
# benchmark
torch.cuda.synchronize()
start = time.time()
c, size = compressor.compress(a)
torch.cuda.synchronize()
end = time.time()
print('Compression time cost')
print(str(end - start))
def __init__(self,
params,
lr=0.01,
momentum=0.9,
weight_decay=0,
compression_buffer=False,
all_reduce=False,
local_rank=0,
gpus_per_machine=1,
args=None,
**kwargs):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= momentum:
raise ValueError("Invalid momentum value: {}".format(momentum))
if not 0.0 <= weight_decay:
raise ValueError(
"Invalid weight_decay value: {}".format(weight_decay))
defaults = dict(lr=lr, momentum=momentum, weight_decay=weight_decay)
super(SGD_distribute, self).__init__(params, defaults)
#custom code
self.compression_buffer = compression_buffer
self.all_reduce = all_reduce
self.use_majority_vote = not args.disable_majority_vote
self.enable_krum = args.enable_krum
self.krum_f = args.krum_f
self.enable_adversary = args.enable_adversary
self.adversary_num = args.adversary_num
self.enable_minus_adversary = args.enable_minus_adversary
if self.enable_krum:
self.compression_buffer = True
self.all_reduce = False
self.use_majority_vote = True
self.enable_minus_adversary = False
self.MB = 1024 * 1024
self.bucket_size = 100 * self.MB
if self.compression_buffer:
self.compressor = compressor.compressor(using_cuda=True,
local_rank=local_rank)
self.local_rank = local_rank
self.global_rank = dist.get_rank()
self.local_dst_in_global = self.global_rank - self.local_rank
self.inter_node_group = []
self.nodes = dist.get_world_size() // gpus_per_machine
self.intra_node_group_list = []
for index in range(self.nodes):
# set inter_node_group
self.inter_node_group.append(0 + index * gpus_per_machine)
# set all intra_node_group
intra_node_group_temp = []
for intra_index in range(gpus_per_machine):
intra_node_group_temp.append(intra_index +
index * gpus_per_machine)
intra_node_group_temp = dist.new_group(intra_node_group_temp)
self.intra_node_group_list.append(intra_node_group_temp)
if self.local_dst_in_global == 0 + index * gpus_per_machine:
self.nodes_rank = index
#self.intra_node_list = self.intra_node_group
self.inter_node_list = self.inter_node_group
self.inter_node_group_list = []
for index in range(len(self.inter_node_list)):
if index is not 0:
temp = dist.new_group(
[self.inter_node_list[0], self.inter_node_list[index]])
self.inter_node_group_list.append(temp)
self.all_gpu = dist.new_group()
self.all_inter_node_group = dist.new_group(self.inter_node_list)
if dist.get_rank() == 0 or dist.get_rank() == 8:
print('nodes', self.nodes)
print('intra_node_group_list', self.intra_node_group_list)
print('inter_node_group', self.inter_node_group_list)
print('all_inter_node_group', self.inter_node_list)
import distortion
import compressor
overdrive = distortion.overdrive_sigmoid(2)
distor1 = distortion.distortion_threshold(2)
distor2 = distortion.distortion_amplification(2)
compressor1 = compressor.compressor()
compressor2 = compressor.compressor_attack()
# overdrive .simulate()
# distor1.simulate()
# distor2.simulate()
compressor1.simulate()
compressor2.simulate()
try:
input("Press Enter to exit...")
except:
pass
def __init__(self, params, args, log_writer, **kwargs):
lr = args.lr
momentum = args.momentum
weight_decay = args.weight_decay
compression_buffer = args.compress
all_reduce = args.all_reduce
local_rank = args.local_rank
gpus_per_machine = args.gpus_per_machine
self.compression_buffer = compression_buffer
self.all_reduce = all_reduce
self.signum = args.signum
self.log_writer = log_writer
self.args = args
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= momentum:
raise ValueError("Invalid momentum value: {}".format(momentum))
if not 0.0 <= weight_decay:
raise ValueError(
"Invalid weight_decay value: {}".format(weight_decay))
defaults = dict(lr=lr, momentum=momentum, weight_decay=weight_decay)
super(SGD_distribute, self).__init__(params, defaults)
self.MB = 1024 * 1024
self.bucket_size = 100 * self.MB
if self.compression_buffer:
import compressor
self.compressor = compressor.compressor(
using_cuda=True,
local_rank=local_rank,
cpp_extend_load=args.cpp_extend_load)
self.local_rank = local_rank
self.global_rank = dist.get_rank()
self.local_dst_in_global = self.global_rank - self.local_rank
self.inter_node_group = []
self.nodes = dist.get_world_size() // gpus_per_machine
self.intra_node_group_list = []
for index in range(self.nodes):
# set inter_node_group
self.inter_node_group.append(0 + index * gpus_per_machine)
# set all intra_node_group
intra_node_group_temp = []
for intra_index in range(gpus_per_machine):
intra_node_group_temp.append(intra_index +
index * gpus_per_machine)
intra_node_group_temp = dist.new_group(intra_node_group_temp)
self.intra_node_group_list.append(intra_node_group_temp)
if self.local_dst_in_global == 0 + index * gpus_per_machine:
self.nodes_rank = index
#self.intra_node_list = self.intra_node_group
self.inter_node_list = self.inter_node_group
self.inter_node_group_list = []
for index in range(len(self.inter_node_list)):
if index is not 0:
temp = dist.new_group(
[self.inter_node_list[0], self.inter_node_list[index]])
self.inter_node_group_list.append(temp)
self.all_gpu = dist.new_group()
self.all_inter_node_group = dist.new_group(self.inter_node_list)
if dist.get_rank() == 0 or dist.get_rank() == 8:
print('nodes', self.nodes)
print('intra_node_group_list', self.intra_node_group_list)
print('inter_node_group', self.inter_node_group_list)
print('all_inter_node_group', self.inter_node_list)
lead = np.append(lead, zeros_vec)
zz += 4
lead = np.array(lead)
# get lead-vector that has same length as signal, because we are calculating
# a little bit more than that
lead = lead[0:signal_length]
# divide signal into three filter-bands
lower_border_freq = 250
upper_border_freq = 4000
filt = filterbank.filterbank(fs, lower_border_freq, upper_border_freq)
low_sig, mid_sig, high_sig = filt.filter(signal)
# compress signal in filter-bands depending on where speech was detected
comp = compressor.compressor(fs=fs, ratio=2, mu_auto=True)
gain = 1/4 # -12 dB
compressed_low = comp.compress_mono(low_sig, lead)
compressed_mid = comp.compress_mono(mid_sig, lead)
compressed_high = comp.compress_mono(high_sig, lead)
signal_out = gain * compressed_low + compressed_mid + gain * compressed_high
# write signal into file so that you can (hopefully) hear a difference
sf.write('test_signal_compressed.wav', data=signal_out, samplerate=fs)
# plotting classifier output (lead) and the original audio signal
time_vec = np.linspace(0, signal_length/fs, signal_length )
fig, ax1 = plt.subplots()
color = 'tab:blue'
