class RunConfiguration(zconf.RunConfig):
val_json = zconf.attr(help='train_json path')
casm_path = zconf.attr(help='model_checkpoint')
output_path = zconf.attr(help='output_path')
layer_depth = zconf.attr(default=2, type=int)
plot_img_i = zconf.attr(default=0, type=int)
num_batches = zconf.attr(default=1, type=int)
do_plot = zconf.attr(default=1, type=int)
workers = zconf.attr(default=4, type=int, help='number of data loading workers (default: 4)')
batch_size = zconf.attr(default=128, type=int, help='mini-batch size (default: 256)')
break_ratio = zconf.attr(action='store_true', help='break original aspect ratio when resizing')
class RunConfiguration(zconf.RunConfig):
train_path = zconf.attr(type=str)
val_path = zconf.attr(type=str)
val_annotation_path = zconf.attr(type=str)
output_base_path = zconf.attr(type=str)
extended_annot_base_path = zconf.attr(type=str, default=None)
num_per_class_in_train_val = zconf.attr(type=int, default=50)
seed = zconf.attr(type=int, default=1234)
class RunConfiguration(zconf.RunConfig):
# Experiment Setup
train_json = zconf.attr(help='train_json path')
val_json = zconf.attr(help='val_json path')
output_path = zconf.attr(help='output_path')
name = zconf.attr(
default='random',
help=
'name used to build a path where the models and log are saved (default: random)'
)
log_buffer = zconf.attr(default=10, type=int, help='log buffer')
workers = zconf.attr(default=4,
type=int,
help='number of data loading workers (default: 4)')
epochs = zconf.attr(default=60,
type=int,
help='number of total epochs to run')
batch_size = zconf.attr(default=128,
type=int,
help='mini-batch size (default: 128)')
perc_of_training = zconf.attr(
default=0.2,
type=float,
help='percent of training set seen in each epoch')
do_val = zconf.attr(action="store_true")
lr = zconf.attr(default=0.001,
type=float,
help='initial learning rate for classifier')
lr_casme = zconf.attr(default=0.001,
type=float,
help='initial learning rate for casme')
lrde = zconf.attr(default=20,
type=int,
help='how often is the learning rate decayed')
momentum = zconf.attr(default=0.9,
type=float,
help='momentum for classifier')
weight_decay = zconf.attr(
default=1e-4,
type=float,
help='weight decay for both classifier and casme (default: 1e-4)')
upsample = zconf.attr(
default='nearest',
help='mode for final upsample layer in the decoder (default: nearest)')
fixed_classifier = zconf.attr(action='store_true', help='train classifier')
prob_historic = zconf.attr(
default=0.5,
type=float,
help='probability for evaluating historic model')
save_freq = zconf.attr(
default=1000,
type=int,
help='frequency of model saving to history (in batches)')
actual_save_freq = zconf.attr(default=1, type=int)
f_size = zconf.attr(
default=30,
type=int,
help=
'size of F set - maximal number of previous classifier iterations stored'
)
resnet_path = zconf.attr(
default=None,
type=str,
help="If none, defaults to loading from full ResNet-50")
lambda_r = zconf.attr(default=None, type=float)
lambda_tv = zconf.attr(default=None, type=float)
masker_use_layers = zconf.attr(default="0,1,2,3,4", type=str)
mask_in_criterion = zconf.attr(default="none",
type=str,
help='crossentropy|kldivergence|none')
mask_in_criterion_config = zconf.attr(default="", type=str, help='etc')
mask_in_objective_direction = zconf.attr(default="maximize",
help="maximize|minimize")
mask_in_objective_type = zconf.attr(default="entropy",
help="entropy|classification")
mask_in_weight = zconf.attr(default=1.0, type=float)
mask_in_lambda_r = zconf.attr(default=10, type=float)
mask_in_lambda_tv = zconf.attr(default=None, type=float)
mask_out_criterion = zconf.attr(default="none",
type=str,
help='crossentropy|kldivergence|none')
mask_out_criterion_config = zconf.attr(default="", type=str, help='etc')
mask_out_objective_direction = zconf.attr(default="maximize",
help="maximize|minimize")
mask_out_objective_type = zconf.attr(default="entropy",
help="entropy|classification")
mask_out_weight = zconf.attr(default=1.0, type=float)
mask_out_lambda_r = zconf.attr(default=10, type=float)
mask_out_lambda_tv = zconf.attr(default=None, type=float)
reproduce = zconf.attr(
default='', help='reproducing paper results (F|L|FL|L100|L1000)')
add_prob_layers = zconf.attr(action='store_true')
prob_sample_low = zconf.attr(default=0.25, type=float)
prob_sample_high = zconf.attr(default=0.75, type=float)
prob_loss_func = zconf.attr(default="l1")
add_class_ids = zconf.attr(action='store_true')
apply_gumbel = zconf.attr(action='store_true')
apply_gumbel_tau = zconf.attr(default=0.1, type=float)
gumbel_output_mode = zconf.attr(default="hard", type=str)
# Placeholders
casms_path = zconf.attr(default='')
log_path = zconf.attr(default='')
infiller_model = zconf.attr(default=None, type=str)
do_infill_for_mask_in = zconf.attr(default=0, type=int)
do_infill_for_mask_out = zconf.attr(default=0, type=int)
def _post_init(self):
randomhash = ''.join(str(time.time()).split('.'))
self.name = self.name + "___" + randomhash
self.need_infiller = self.do_infill_for_mask_in or self.do_infill_for_mask_out
if self.lambda_r is not None:
assert self.mask_in_lambda_r == 10
assert self.mask_out_lambda_r == 10
self.mask_in_lambda_r = self.lambda_r
self.mask_out_lambda_r = self.lambda_r
if self.lambda_tv is not None:
assert self.mask_in_lambda_tv is None
assert self.mask_out_lambda_tv is None
self.mask_in_lambda_tv = self.lambda_tv
self.mask_out_lambda_tv = self.lambda_tv
if self.resnet_path == "none":
self.resnet_path = None
set_args(self)
class RunConfiguration(zconf.RunConfig):
input_base_path = zconf.attr(type=str)
output_base_path = zconf.attr(type=str)
class RunConfiguration(zconf.RunConfig):
val_json = zconf.attr(help='train_json path')
mode = zconf.attr(type=str)
bboxes_path = zconf.attr(help='path to bboxes_json')
casm_path = zconf.attr(help='model_checkpoint')
classifier_load_mode = zconf.attr(default="pickled")
output_path = zconf.attr(help='output_path')
record_bboxes = zconf.attr(type=str, default=None)
use_p = zconf.attr(type=float, default=None)
workers = zconf.attr(default=4,
type=int,
help='number of data loading workers (default: 4)')
batch_size = zconf.attr(default=128,
type=int,
help='mini-batch size (default: 256)')
print_freq = zconf.attr(default=10,
type=int,
help='print frequency (default: 10)')
break_ratio = zconf.attr(action='store_true',
help='break original aspect ratio when resizing')
not_normalize = zconf.attr(action='store_true',
help='prevents normalization')
pot = zconf.attr(default=1,
type=float,
help='percent of validation set seen')
class RunConfiguration(zconf.RunConfig):
casm_path1 = zconf.attr(default="best")
casm_path2 = zconf.attr(default="best")
val_json = zconf.attr(type=str)
output_path = zconf.attr(default=None)
class RunConfiguration(zconf.RunConfig):
# === RunConfig Parameters === #
jiant_task_container_path = zconf.attr(type=str, default=None)
# === Required Parameters === #
supertask = zconf.attr(type=str, default=None)
output_dir = zconf.attr(type=str, required=True)
# === Optional Parameters === #
skip_if_done = zconf.attr(action="store_true")
bucc_val_metrics_path = zconf.attr(
type=str,
default=None,
help=
"Path to val_metrics.json for bucc2018. Contains the optimal threshold,"
" to be used for generating test predictions.",
)
# === Model parameters === #
model_type = zconf.attr(type=str, required=True)
model_path = zconf.attr(type=str, required=True)
model_config_path = zconf.attr(default=None, type=str)
model_tokenizer_path = zconf.attr(default=None, type=str)
model_load_mode = zconf.attr(default="from_ptt", type=str)
# === Nuisance Parameters === #
# Required for quickly setting up runner
# Remove/refactor with config refactor (issue #1176)
learning_rate = zconf.attr(default=1e-5, type=float)
adam_epsilon = zconf.attr(default=1e-8, type=float)
max_grad_norm = zconf.attr(default=1.0, type=float)
optimizer_type = zconf.attr(default="adam", type=str)
# Specialized config
no_cuda = zconf.attr(action="store_true")
fp16 = zconf.attr(action="store_true")
fp16_opt_level = zconf.attr(default="O1", type=str)
local_rank = zconf.attr(default=-1, type=int)
server_ip = zconf.attr(default="", type=str)
server_port = zconf.attr(default="", type=str)
force_overwrite = zconf.attr(action="store_true")
seed = zconf.attr(type=int, default=-1)
class RunConfiguration(zconf.RunConfig):
# === Required parameters === #
task_config_path = zconf.attr(type=str, required=True)
unsup_task_config_path = zconf.attr(type=str, required=True)
output_dir = zconf.attr(type=str, required=True)
# === Model parameters === #
model_type = zconf.attr(type=str, required=True)
model_path = zconf.attr(type=str, required=True)
model_config_path = zconf.attr(default=None, type=str)
model_tokenizer_path = zconf.attr(default=None, type=str)
model_load_mode = zconf.attr(default="safe", type=str)
model_save_mode = zconf.attr(default="all", type=str)
max_seq_length = zconf.attr(default=128, type=int)
# === Running Setup === #
# cache_dir
do_train = zconf.attr(action='store_true')
do_val = zconf.attr(action='store_true')
do_test = zconf.attr(action='store_true')
do_save = zconf.attr(action="store_true")
eval_every_steps = zconf.attr(type=int, default=0)
save_every_steps = zconf.attr(type=int, default=0)
partial_eval_number = zconf.attr(type=int, default=1000)
train_batch_size = zconf.attr(default=8, type=int) # per gpu
eval_batch_size = zconf.attr(default=8, type=int) # per gpu
force_overwrite = zconf.attr(action="store_true")
# overwrite_cache = zconf.attr(action="store_true")
seed = zconf.attr(type=int, default=-1)
# === Training Learning Parameters === #
learning_rate = zconf.attr(default=1e-5, type=float)
num_train_epochs = zconf.attr(default=3, type=int)
max_steps = zconf.attr(default=-1, type=int) ## Change to None
adam_epsilon = zconf.attr(default=1e-8, type=float)
max_grad_norm = zconf.attr(default=1.0, type=float)
warmup_steps = zconf.attr(default=None, type=int)
warmup_proportion = zconf.attr(default=0.1, type=float)
optimizer_type = zconf.attr(default="adam", type=str)
# Specialized config
gradient_accumulation_steps = zconf.attr(default=1, type=int)
no_cuda = zconf.attr(action='store_true')
fp16 = zconf.attr(action='store_true')
fp16_opt_level = zconf.attr(default='O1', type=str)
local_rank = zconf.attr(default=-1, type=int)
server_ip = zconf.attr(default='', type=str)
server_port = zconf.attr(default='', type=str)
# === UDA === #
unsup_ratio = zconf.attr(type=int, default=3)
no_tsa = zconf.attr(action="store_true")
tsa_schedule = zconf.attr(type=str, default="linear_schedule")
uda_softmax_temp = zconf.attr(type=float, default=-1)
uda_confidence_thresh = zconf.attr(type=float, default=-1)
uda_coeff = zconf.attr(type=float, default=1.)
class RunConfiguration(zconf.RunConfig):
# === Required parameters === #
uda_task_config_path = zconf.attr(type=str, required=True)
output_dir = zconf.attr(type=str, required=True)
# === Model parameters === #
model_type = zconf.attr(type=str, required=True)
model_path = zconf.attr(type=str, required=True)
model_config_path = zconf.attr(default=None, type=str)
model_tokenizer_path = zconf.attr(default=None, type=str)
model_load_mode = zconf.attr(type=str, required=True)
model_save_mode = zconf.attr(default="all", type=str)
max_seq_length = zconf.attr(default=128, type=int)
# === Running Setup === #
# cache_dir
do_train = zconf.attr(action='store_true')
do_val = zconf.attr(action='store_true')
do_test = zconf.attr(action='store_true')
do_save = zconf.attr(action="store_true")
do_val_history = zconf.attr(action='store_true')
train_save_every = zconf.attr(type=int, default=None)
train_save_every_epoch = zconf.attr(action="store_true")
eval_every_epoch = zconf.attr(action="store_true")
eval_every = zconf.attr(type=int, default=None)
train_batch_size = zconf.attr(default=8, type=int) # per gpu
eval_batch_size = zconf.attr(default=8, type=int) # per gpu
force_overwrite = zconf.attr(action="store_true")
# overwrite_cache = zconf.attr(action="store_true")
seed = zconf.attr(type=int, default=-1)
# === Training Learning Parameters === #
learning_rate = zconf.attr(default=1e-5, type=float)
num_train_epochs = zconf.attr(default=3, type=int)
max_steps = zconf.attr(default=-1, type=int) ## Change to None
adam_epsilon = zconf.attr(default=1e-8, type=float)
max_grad_norm = zconf.attr(default=1.0, type=float)
warmup_steps = zconf.attr(default=None, type=int)
warmup_proportion = zconf.attr(default=0.1, type=float)
optimizer_type = zconf.attr(default="adam", type=str)
# Specialized config
gradient_accumulation_steps = zconf.attr(default=1, type=int)
no_cuda = zconf.attr(action='store_true')
fp16 = zconf.attr(action='store_true')
fp16_opt_level = zconf.attr(default='O1', type=str)
local_rank = zconf.attr(default=-1, type=int)
server_ip = zconf.attr(default='', type=str)
server_port = zconf.attr(default='', type=str)
# LLP hyperparams
llp_embedding_dim = zconf.attr(type=int, default=128)
llp_const_k = zconf.attr(type=int, default=10)
llp_const_t = zconf.attr(type=int, default=25)
llp_const_tau = zconf.attr(type=float, default=0.07)
llp_prop_chunk_size = zconf.attr(type=int, default=500)
llp_mem_bank_t = zconf.attr(type=float, default=0.5)
llp_rep_global_agg_loss_lambda = zconf.attr(type=float, default=1.)
llp_embedding_norm_loss = zconf.attr(type=float, default=0.01)
llp_compute_global_agg_loss_mode = zconf.attr(type=str, default="v1")
llp_load_override = zconf.attr(type=str, default=None)
unlabeled_train_examples_number = zconf.attr(type=int, default=None)
unlabeled_train_examples_fraction = zconf.attr(type=float, default=None)
# UDA LLP
uda_coeff = zconf.attr(type=float, default=1.0)
unsup_ratio = zconf.attr(type=int, default=1)
class RunConfiguration(zconf.RunConfig):
cam_loader = zconf.attr(type=str, required=True)
casm_base_path = zconf.attr(type=str, default=None)
output_base_path = zconf.attr(type=str, default=None)
dataset = zconf.attr(type=str)
dataset_split = zconf.attr(type=str)
dataset_path = zconf.attr(type=str)
metadata_path = zconf.attr(type=str)
cam_curve_interval = zconf.attr(type=float, default=0.01)
box_v2_metric = zconf.attr(action="store_true")
classifier_load_mode = zconf.attr(default="pickled")
workers = zconf.attr(default=4,
type=int,
help='number of data loading workers (default: 4)')
batch_size = zconf.attr(default=128,
type=int,
help='mini-batch size (default: 256)')
break_ratio = True
# === Dataset-specific === #
# Used for ILSVRC/test
imagenet_val_path = zconf.attr(type=str, default=None)
wsoleval_dataset_path = zconf.attr(type=str, default=None)
# === Method-specific === #
torchray_method = zconf.attr(default=None)
casme_load_mode = zconf.attr(type=str, default="best")
class RunConfiguration(zconf.RunConfig):
# === Required parameters === #
task_config_path = zconf.attr(type=str, required=True)
output_dir = zconf.attr(type=str, required=True)
# === Model parameters === #
model_type = zconf.attr(type=str, required=True)
model_path = zconf.attr(type=str, required=True)
model_config_path = zconf.attr(default=None, type=str)
model_tokenizer_path = zconf.attr(default=None, type=str)
model_load_mode = zconf.attr(default="safe", type=str)
model_save_mode = zconf.attr(default="all", type=str)
max_seq_length = zconf.attr(default=128, type=int)
# === Running Setup === #
# cache_dir
do_train = zconf.attr(action='store_true')
do_val = zconf.attr(action='store_true')
do_test = zconf.attr(action='store_true')
do_save = zconf.attr(action="store_true")
eval_every_steps = zconf.attr(type=int, default=0)
save_every_steps = zconf.attr(type=int, default=0)
partial_eval_number = zconf.attr(type=int, default=1000)
train_batch_size = zconf.attr(default=8, type=int) # per gpu
eval_batch_size = zconf.attr(default=8, type=int) # per gpu
force_overwrite = zconf.attr(action="store_true")
seed = zconf.attr(type=int, default=-1)
train_examples_number = zconf.attr(type=int, default=None)
train_examples_fraction = zconf.attr(type=float, default=None)
# === Training Learning Parameters === #
learning_rate = zconf.attr(default=1e-5, type=float)
num_train_epochs = zconf.attr(default=3, type=int)
max_steps = zconf.attr(default=None, type=int)
adam_epsilon = zconf.attr(default=1e-8, type=float)
max_grad_norm = zconf.attr(default=1.0, type=float)
warmup_steps = zconf.attr(default=None, type=int)
warmup_proportion = zconf.attr(default=0.1, type=float)
optimizer_type = zconf.attr(default="adam", type=str)
# Specialized config
gradient_accumulation_steps = zconf.attr(default=1, type=int)
no_cuda = zconf.attr(action='store_true')
fp16 = zconf.attr(action='store_true')
fp16_opt_level = zconf.attr(default='O1', type=str)
local_rank = zconf.attr(default=-1, type=int)
server_ip = zconf.attr(default='', type=str)
server_port = zconf.attr(default='', type=str)
# Multi Adapters
adapter_weights_path = zconf.attr(type=str, required=True)
adapter_exclude = zconf.attr(type=str, default="")
adapter_num_weight_sets = zconf.attr(type=int, default=1)
adapter_ft_mode = zconf.attr(type=str, default="weights")
adapter_use_optimized = zconf.attr(type=int, default=0)
adapter_include_base = zconf.attr(default=None)
adapter_include_flex = zconf.attr(default=None)
def _post_init(self):
if self.adapter_ft_mode == "base":
self.adapter_include_base = True
self.adapter_include_flex = False
elif self.adapter_ft_mode == "flex":
self.adapter_include_base = False
self.adapter_include_flex = True
elif self.adapter_ft_mode == "base_ft":
self.adapter_include_base = True
self.adapter_include_flex = False
elif self.adapter_ft_mode == "full_ft":
self.adapter_include_base = True
self.adapter_include_flex = False
else:
raise KeyError(self.adapter_ft_mode)
class RunConfiguration(zconf.RunConfig):
# === Required parameters === #
task_config_path = zconf.attr(type=str, required=True)
extra_train_paths = zconf.attr(type=str, action='append')
output_dir = zconf.attr(type=str, required=True)
# === Model parameters === #
model_type = zconf.attr(type=str, required=True)
model_path = zconf.attr(type=str, required=True)
model_config_path = zconf.attr(default=None, type=str)
model_tokenizer_path = zconf.attr(default=None, type=str)
#model_load_mode = zconf.attr(type=str, required=True)
model_save_mode = zconf.attr(default="all", type=str)
max_seq_length = zconf.attr(default=128, type=int)
# === Running Setup === #
# cache_dir
do_train = zconf.attr(action='store_true')
do_val = zconf.attr(action='store_true')
do_test = zconf.attr(action='store_true')
do_save = zconf.attr(action="store_true")
do_val_history = zconf.attr(action='store_true')
train_save_every = zconf.attr(type=int, default=None)
train_save_every_epoch = zconf.attr(action="store_true")
eval_every_epoch = zconf.attr(action="store_true")
eval_every = zconf.attr(type=int, default=None)
train_batch_size = zconf.attr(default=8, type=int) # per gpu
eval_batch_size = zconf.attr(default=8, type=int) # per gpu
force_overwrite = zconf.attr(action="store_true")
# overwrite_cache = zconf.attr(action="store_true")
seed = zconf.attr(type=int, default=-1)
use_tensorboard = zconf.attr(action="store_true")
# === Training Learning Parameters === #
learning_rate = zconf.attr(default=1e-5, type=float)
num_train_epochs = zconf.attr(default=3, type=int)
max_steps = zconf.attr(default=-1, type=int) ## Change to None
adam_epsilon = zconf.attr(default=1e-8, type=float)
max_grad_norm = zconf.attr(default=1.0, type=float)
warmup_steps = zconf.attr(default=None, type=int)
warmup_proportion = zconf.attr(default=0.1, type=float)
optimizer_type = zconf.attr(default="adam", type=str)
# Specialized config
gradient_accumulation_steps = zconf.attr(default=1, type=int)
no_cuda = zconf.attr(action='store_true')
fp16 = zconf.attr(action='store_true')
fp16_opt_level = zconf.attr(default='O1', type=str)
local_rank = zconf.attr(default=-1, type=int)
server_ip = zconf.attr(default='', type=str)
server_port = zconf.attr(default='', type=str)
class RunConfiguration(zconf.RunConfig):
model_config_base_path = zconf.attr(type=str, required=True)
task_config_base_path = zconf.attr(type=str, default=None)
output_base_path = zconf.attr(type=str, required=True)
reference_base_path = zconf.attr(type=str, default=None)