for i, (env, env_weights) in enumerate(in_splits)
if i not in args.test_envs
]
uda_loaders = [
InfiniteDataLoader(dataset=env,
weights=env_weights,
batch_size=hparams['batch_size'],
num_workers=dataset.N_WORKERS)
for i, (env, env_weights) in enumerate(uda_splits)
if i in args.test_envs
]
eval_loaders = [
FastDataLoader(dataset=env,
batch_size=64,
num_workers=dataset.N_WORKERS)
for env, _ in (in_splits + out_splits + uda_splits)
]
eval_weights = [
None for _, weights in (in_splits + out_splits + uda_splits)
]
eval_loader_names = ['env{}_in'.format(i) for i in range(len(in_splits))]
eval_loader_names += [
'env{}_out'.format(i) for i in range(len(out_splits))
]
eval_loader_names += [
'env{}_uda'.format(i) for i in range(len(uda_splits))
]
algorithm_class = algorithms.get_algorithm_class(args.algorithm)
for i, (env, env_weights) in enumerate(in_splits)
if i not in args.test_envs
]
uda_loaders = [
InfiniteDataLoader(dataset=env,
weights=env_weights,
batch_size=hparams['batch_size'],
num_workers=dataset.N_WORKERS)
for i, (env, env_weights) in enumerate(uda_splits)
if i in args.test_envs
]
eval_loaders = [
FastDataLoader(dataset=env,
batch_size=hparams['batch_size'],
num_workers=dataset.N_WORKERS)
for env, _ in (in_splits + out_splits + uda_splits)
]
if split_indata:
invenio_mata_out_splits = in_val_splits
else:
invenio_mata_out_splits = out_splits
val_loaders_invenio = [
InfiniteDataLoader(dataset=env,
weights=env_weights,
batch_size=hparams['batch_size'],
num_workers=dataset.N_WORKERS)
for i, (env, env_weights) in enumerate(invenio_mata_out_splits)
if i not in args.test_envs
]
in_weights, out_weights = None, None
in_splits.append((in_, in_weights))
out_splits.append((out, out_weights))
hparams['batch_size'] = 24
hparams['weight_decay'] = 1e-6
train_loaders = [
InfiniteDataLoader(dataset=env,
weights=env_weights,
batch_size=hparams['batch_size'],
num_workers=train_dataset.N_WORKERS)
for i, (env, env_weights) in enumerate(in_splits)
if i not in args.test_envs
]
eval_loaders = [
FastDataLoader(dataset=env,
batch_size=hparams['batch_size'] * 4,
num_workers=val_dataset.N_WORKERS)
for env, _ in (out_splits)
]
eval_weights = [None for _, weights in (out_splits)]
eval_loader_names = ['env{}_out'.format(i) for i in range(len(out_splits))]
train_eval_loaders = [
FastDataLoader(dataset=env,
batch_size=hparams['batch_size'] * 4,
num_workers=train_dataset.N_WORKERS)
for env, _ in (in_splits)
]
train_eval_weights = [None for _, weights in (in_splits)]
train_eval_loader_names = [
'env{}_in'.format(i) for i in range(len(in_splits))
]
for env_i, env in enumerate(dataset):
out, in_ = misc.split_dataset(env,
int(len(env) * args.holdout_fraction),
misc.seed_hash(args.trial_seed, env_i))
if hparams['class_balanced']:
in_weights = misc.make_weights_for_balanced_classes(in_)
out_weights = misc.make_weights_for_balanced_classes(out)
else:
in_weights, out_weights = None, None
in_splits.append((in_, in_weights))
out_splits.append((out, out_weights))
train_loaders = [
FastDataLoader(dataset=env,
weights=env_weights,
batch_size=hparams['batch_size'],
num_workers=dataset.N_WORKERS,
length=FastDataLoader.INFINITE)
for i, (env, env_weights) in enumerate(in_splits)
if i not in args.test_envs
]
eval_loaders = [
FastDataLoader(dataset=env,
weights=None,
batch_size=64,
num_workers=dataset.N_WORKERS,
length=FastDataLoader.EPOCH)
for env, _ in (in_splits + out_splits)
]
eval_weights = [None for _, weights in (in_splits + out_splits)]
else:
in_weights, out_weights = None, None
if env_i in args.target_envs:
target_envs.append(len(in_splits))
in_splits.append((in_, in_weights))
out_splits.append((out, out_weights))
train_loaders.append(
InfiniteDataLoader(dataset=in_,
weights=in_weights,
batch_size=hparams['batch_size'],
num_workers=dataset.N_WORKERS))
#for i, (env, env_weights) in enumerate(in_splits)
#if i not in args.test_envs]
eval_loaders += [
FastDataLoader(dataset=in_,
batch_size=64,
num_workers=dataset.N_WORKERS),
FastDataLoader(dataset=out,
batch_size=64,
num_workers=dataset.N_WORKERS)
]
eval_loader_names += [
'source_tr{0}'.format(len(in_splits)),
'source_te{0}'.format(len(in_splits))
]
eval_weights += [None, None]
else:
out_splits.append((env, None))
eval_loaders.append(
FastDataLoader(dataset=env,
batch_size=64,