def init_exp_logger(repo=None, experiment_name=None, flush_frequency=1):
if repo is None:
repo = os.path.join(log_path(),"./.aim")
if not os.path.exists(repo):
logger.info('{} dir is not exist, create {}',repo, repo)
os.system(str("cd " + os.path.join(repo,"../") + "&& aim init"))
aim_logger = aim.Session(repo = repo, experiment=experiment_name, flush_frequency=flush_frequency)
aim_logger.experiment_name = experiment_name
return aim_logger
def __init__(
self,
name = 'default',
results_dir = 'results',
models_dir = 'models',
base_dir = './',
image_size = 128,
network_capacity = 16,
fmap_max = 512,
transparent = False,
batch_size = 4,
mixed_prob = 0.9,
gradient_accumulate_every=1,
lr = 2e-4,
lr_mlp = 1.,
ttur_mult = 2,
rel_disc_loss = False,
num_workers = None,
save_every = 1000,
evaluate_every = 1000,
num_image_tiles = 8,
trunc_psi = 0.6,
fp16 = False,
cl_reg = False,
fq_layers = [],
fq_dict_size = 256,
attn_layers = [],
no_const = False,
aug_prob = 0.,
aug_types = ['translation', 'cutout'],
top_k_training = False,
generator_top_k_gamma = 0.99,
generator_top_k_frac = 0.5,
dataset_aug_prob = 0.,
calculate_fid_every = None,
is_ddp = False,
rank = 0,
world_size = 1,
log = False,
*args,
**kwargs
):
self.GAN_params = [args, kwargs]
self.GAN = None
self.name = name
base_dir = Path(base_dir)
self.base_dir = base_dir
self.results_dir = base_dir / results_dir
self.models_dir = base_dir / models_dir
print(self.results_dir, self.models_dir)
self.config_path = self.models_dir / name / '.config.json'
assert log2(image_size).is_integer(), 'image size must be a power of 2 (64, 128, 256, 512, 1024)'
self.image_size = image_size
self.network_capacity = network_capacity
self.fmap_max = fmap_max
self.transparent = transparent
self.fq_layers = cast_list(fq_layers)
self.fq_dict_size = fq_dict_size
self.has_fq = len(self.fq_layers) > 0
self.attn_layers = cast_list(attn_layers)
self.no_const = no_const
self.aug_prob = aug_prob
self.aug_types = aug_types
self.lr = lr
self.lr_mlp = lr_mlp
self.ttur_mult = ttur_mult
self.rel_disc_loss = rel_disc_loss
self.batch_size = batch_size
self.num_workers = num_workers
self.mixed_prob = mixed_prob
self.num_image_tiles = num_image_tiles
self.evaluate_every = evaluate_every
self.save_every = save_every
self.steps = 0
self.av = None
self.trunc_psi = trunc_psi
self.pl_mean = None
self.gradient_accumulate_every = gradient_accumulate_every
assert not fp16 or fp16 and APEX_AVAILABLE, 'Apex is not available for you to use mixed precision training'
self.fp16 = fp16
self.cl_reg = cl_reg
self.d_loss = 0
self.g_loss = 0
self.q_loss = None
self.last_gp_loss = None
self.last_cr_loss = None
self.last_fid = None
self.pl_length_ma = EMA(0.99)
self.init_folders()
self.loader = None
self.dataset_aug_prob = dataset_aug_prob
self.calculate_fid_every = calculate_fid_every
self.top_k_training = top_k_training
self.generator_top_k_gamma = generator_top_k_gamma
self.generator_top_k_frac = generator_top_k_frac
assert not (is_ddp and cl_reg), 'Contrastive loss regularization does not work well with multi GPUs yet'
self.is_ddp = is_ddp
self.is_main = rank == 0
self.rank = rank
self.world_size = world_size
self.logger = aim.Session(experiment=name) if log else None
import aim
import random
import math
epochs = 4
steps = 30
c_step = 100
sess = aim.Session(experiment='test_epoch_alignment_x', flush_frequency=10)
sess.set_params({
'name': 'Dataset name',
'version': 'Dataset version',
},
name='dataset')
sess.set_params({
'epochs': epochs,
'steps': steps,
'c_step': c_step,
},
name='hparmas')
sess.set_params({
'foo': random.random() * 100,
'bar': random.random() * 100,
'baz': random.random() * 100,
'cluster': int(random.random() * 3),
'nested': {
'arr': ['aa', 'bb', 'cc'],
'obj': {
import aim
import math
sess = aim.Session(experiment='test_params')
sess.set_params({
'num_epochs': 10,
'fc_units': 128,
}, name='hparams')
sess.set_params({
'name': 'Dataset name',
'version': 'Dataset version',
},
name='dataset')
sess.set_params({
'foo': 'bar',
})
sess.set_params({
'inf': float('inf'),
'inf_in_nested_obj': (1, 2, 3, {
'inf': math.inf,
}),
})
import random
import os
import aim
exp1_run1 = aim.Session(experiment='test_metrics', flush_frequency=10)
exp1_run2 = aim.Session(experiment='test_metrics', flush_frequency=10)
exp2 = aim.Session(experiment='test_metrics_2', flush_frequency=10)
print(os.getpid(), exp1_run1.run_hash, exp1_run2.run_hash, exp2.run_hash)
exp1_run1.set_params({
'foo': random.random() * 100,
'bar': random.random() * 100,
'baz': random.random() * 100,
})
exp1_run2.set_params({
'foo': random.random() * 100,
'bar': random.random() * 100,
'baz': random.random() * 100,
})
exp2.set_params({
'foo': random.random() * 100,
'bar': random.random() * 100,
'baz': random.random() * 100,
})
for i in range(100, 200):
# Experiment 1, run 1
exp1_run1.track(i, name='metric_2')
import aim
import random
import os
exp1_run1 = aim.Session(experiment='test_const', flush_frequency=10)
exp1_run1.set_params({
'foo': random.random() * 100,
'bar': random.random() * 100,
'baz': random.random() * 100,
})
for i in range(100, 200):
exp1_run1.track(3.45, name='const3')
import random
import time
import os
import aim
sess = aim.Session(experiment='TEST_METRICS')
sess.set_params({'key': random.random()})
print(os.getpid())
for i in range(100, 4000):
print(i, i * 2, i * 3)
sess.track(i, name='metric')
sess.track(i * 2, name='metric', subset='train', sub=12)
sess.track(i * 3, name='metric', subset='test', aab=22)
time.sleep(0.1)
import random
import time
import os
import aim
sess = aim.Session(experiment='TEST_METRICS', flush_frequency=10)
sess2 = aim.Session(experiment='TEST_METRICS_2', flush_frequency=30)
sess.set_params({'key': random.random()})
print(os.getpid())
for i in range(100, 200):
sess.track(i, name='metric')
sess.track(i * 2, name='metric', subset='train', foo='baz')
sess.track(i * 3, name='metric', subset='test', bar='baz')
sess2.track(i, name='metric', subset='train', bar='baz')
time.sleep(1)
import aim
import math
sess = aim.Session(experiment='test_floats')
sess.set_params({
'num_epochs': 10,
'fc_units': 128,
}, name='hparams')
sess.set_params({
'name': 'Dataset name',
'version': 'Dataset version',
},
name='dataset')
sess.set_params({
'foo': 'bar',
})
sess.set_params({
'nan_x': 'NaN',
'inf_x': 'Infinity',
'nan': float('nan'),
'inf': float('inf'),
'inf_in_nested_obj': (1, 2, 3, {
'inf': math.inf,
}),
'nan_in_nested_obj': (1, 2, 3, {
'nan': math.nan,
}),
import random
import aim
sess = aim.Session(experiment='test_2_metrics')
# sess.set_params({'key': random.random()})
for _ in range(100):
sess.track(random.random(), name='metric')
import aim
sess = aim.Session(experiment='test_metrics_context', flush_frequency=10)
sess.set_params({
'num_epochs': 5,
'lr': 10,
}, name='hparams')
for e in range(5):
for i in range(50):
sess.track(i, name='loss', epoch=e, subset='train', subtask='lm')
sess.track(i, name='acc', epoch=e, subset='train', subtask='lm')
if i % 10 == 0:
sess.track(i, name='loss', epoch=e, subset='val', subtask='lm')
sess.track(i, name='acc', epoch=e, subset='val', subtask='lm')
for e in range(5):
for i in range(50):
sess.track(i, name='loss', epoch=e, subset='train', subtask='nmt')
sess.track(i, name='acc', epoch=e, subset='train', subtask='nmt')
if i % 10 == 0:
sess.track(i, name='loss', epoch=e, subset='val', subtask='nmt')
sess.track(i, name='acc', epoch=e, subset='val', subtask='nmt')
import aim
import random
import math
import time
foo = 1
bar = 1
seed = 1003
epochs = 4
steps = 30
k = 2.3
sess = aim.Session(experiment='test_system', system_tracking_interval=2)
sess.set_params(
{
'epochs': epochs,
'steps': steps,
'k': k,
'foo': foo,
'bar': bar,
'seed': seed,
},
name='hparmas')
#
# for e in range(epochs):
# for i in range(steps):
# sess.track(k, name='agg_metric', epoch=e, subset='train')
# sess.track(k, name='agg_metric', epoch=e, subset='val')
