def setUp(self):
HOCON = """
lr = 123.456
pretrain_data_fraction = .123
target_train_data_fraction = .1234
mnli = {
lr = 4.56,
batch_size = 123
max_epochs = 456
training_data_fraction = .456
}
qqp = {
max_epochs = 789
}
"""
DEFAULTS_PATH = "config/defaults.conf" # To get other required values.
params = params_from_file(DEFAULTS_PATH, HOCON)
self.processed_pretrain_params = build_trainer_params(
params, ["mnli", "qqp"], phase="pretrain"
)
self.processed_mnli_target_params = build_trainer_params(
params, ["mnli"], phase="target_train"
)
self.processed_qqp_target_params = build_trainer_params(
params, ["qqp"], phase="target_train"
)
def setUp(self):
HOCON = """
lr = 123.456
pretrain_data_fraction = .123
target_train_data_fraction = .1234
mnli = {
lr = 4.56,
batch_size = 123
max_epochs = 456
training_data_fraction = .456
}
qqp = {
max_epochs = 789
}
"""
DEFAULTS_PATH = resource_filename(
"jiant",
"config/defaults.conf") # To get other required replacers.
params = params_from_file(DEFAULTS_PATH, HOCON)
cuda_device = -1
self.processed_pretrain_params = build_trainer_params(params,
cuda_device,
["mnli", "qqp"],
phase="pretrain")
self.processed_mnli_target_params = build_trainer_params(
params, cuda_device, ["mnli"], phase="target_train")
self.processed_qqp_target_params = build_trainer_params(
params, cuda_device, ["qqp"], phase="target_train")
def setUp(self):
HOCON = """
lr = 123.456
pretrain_data_fraction = .123
target_train_data_fraction = .1234
mnli = {
lr = 4.56,
batch_size = 123
max_epochs = 456
training_data_fraction = .456
}
qqp = {
max_epochs = 789
}
"""
DEFAULTS_PATH = resource_filename(
"jiant", "config/defaults.conf") # To get other required values.
params = params_from_file(DEFAULTS_PATH, HOCON)
cuda_device = -1
self.processed_pretrain_params = build_trainer_params(params,
cuda_device,
["mnli", "qqp"],
phase="pretrain")
self.processed_mnli_target_params = build_trainer_params(
params, cuda_device, ["mnli"], phase="target_train")
self.processed_qqp_target_params = build_trainer_params(
params, cuda_device, ["qqp"], phase="target_train")
self.pretrain_tasks = []
pretrain_task_registry = {
"sst": REGISTRY["sst"],
"winograd-coreference": REGISTRY["winograd-coreference"],
"commitbank": REGISTRY["commitbank"],
}
for name, (cls, _, kw) in pretrain_task_registry.items():
task = cls("dummy_path",
max_seq_len=1,
name=name,
tokenizer_name="dummy_tokenizer_name",
**kw)
self.pretrain_tasks.append(task)