def main():
parser = argparse.ArgumentParser(
"Train Viggo transformer model (ls=inc freq, da=base+phrases)")
parser.add_argument("output_dir", type=Path, help="save directory")
parser.add_argument("--seed", type=int, default=234222452)
parser.add_argument("--layers", default=2, type=int, help="num layers")
parser.add_argument("--ls", default=0.1, type=float,
help="label smoothing")
parser.add_argument("--wd", default=0.0, type=float,
help="weight decay")
parser.add_argument("--opt", choices=["adam", "sgd"], default="sgd",
help='optimizer')
parser.add_argument("--lr", default=0.5, type=float,
help="learning rate")
parser.add_argument("--tie-embeddings", action="store_true",
help="share decoder input/output embeddings")
parser.add_argument("--attn", choices=['bahdanau', 'luong-general'],
default='luong-general', help='attention type')
parser.add_argument("--max-epochs", default=100, type=int,
help="max training epochs")
parser.add_argument("--gpu", default=-1, type=int)
parser.add_argument("--n-procs", default=2, type=int,
help="num data loader processes")
parser.add_argument("--tr-batch-size", default=128, type=int,
help="training batch size")
parser.add_argument("--va-batch-size", default=1, type=int,
help="valid batch size")
args = parser.parse_args()
seeds(args.seed)
assert os.getenv("MRT_EVAL_SCRIPT") is not None
eval_script = os.getenv("MRT_EVAL_SCRIPT")
dataset = "E2E"
lin_strat = "inc_freq"
is_delex = False
mr_vcb, utt_vcb = setup_vocab(dataset, lin_strat, is_delex)
tr_ds = setup_training_data(dataset, lin_strat, is_delex, mr_vcb, utt_vcb,
include_phrases=True)
va_ds = setup_validation_data(dataset, lin_strat, is_delex,
mr_vcb, utt_vcb)
tr_batches = make_batches(tr_ds, args.tr_batch_size, args.n_procs,
lin_strat, is_delex)
va_batches = make_batches(va_ds, args.va_batch_size, args.n_procs,
lin_strat, is_delex)
model = setup_model("gru", "bi", args.layers, args.attn,
args.tie_embeddings, mr_vcb, utt_vcb, beam_size=8)
trainer = setup_trainer(model, args.opt, args.lr, args.wd, args.ls,
tr_batches, va_batches, args.max_epochs,
utt_vcb, eval_script, mr_utils,
lambda: f"E2E/bi-gru/if/base+phrases/{args.seed}")
env = {'proj_dir': args.output_dir, "gpu": args.gpu}
trainer.run(env, verbose=True)
def main():
parser = argparse.ArgumentParser(
"Train Viggo transformer model (ls=inc freq, da=base+phrases)")
parser.add_argument("output_dir", type=Path, help="save directory")
parser.add_argument("--seed", type=int, default=234222452)
parser.add_argument("--layers", default=2, type=int, help="num layers")
parser.add_argument("--ls", default=0.1, type=float,
help="label smoothing")
parser.add_argument("--tie-embeddings", action="store_true",
help="share decoder input/output embeddings")
parser.add_argument("--max-epochs", default=100, type=int,
help="max training epochs")
parser.add_argument("--gpu", default=-1, type=int)
parser.add_argument("--n-procs", default=2, type=int,
help="num data loader processes")
parser.add_argument("--tr-batch-size", default=128, type=int,
help="training batch size")
parser.add_argument("--va-batch-size", default=1, type=int,
help="valid batch size")
args = parser.parse_args()
seeds(args.seed)
assert os.getenv("MRT_EVAL_SCRIPT") is not None
eval_script = os.getenv("MRT_EVAL_SCRIPT")
dataset = "Viggo"
lin_strat = "inc_freq"
is_delex = True
mr_vcb, utt_vcb = setup_vocab(dataset, lin_strat, is_delex)
tr_ds = setup_training_data(dataset, lin_strat, is_delex, mr_vcb, utt_vcb,
include_phrases=True)
va_ds = setup_validation_data(dataset, lin_strat, is_delex,
mr_vcb, utt_vcb)
tr_batches = make_batches(tr_ds, args.tr_batch_size, args.n_procs,
lin_strat, is_delex)
va_batches = make_batches(va_ds, args.va_batch_size, args.n_procs,
lin_strat, is_delex)
model = setup_model("transformer", None, args.layers, None,
args.tie_embeddings, mr_vcb, utt_vcb, beam_size=8)
trainer = setup_trainer(model, "adamtri", None, None, args.ls, tr_batches,
va_batches, args.max_epochs, utt_vcb, eval_script,
mr_utils,
lambda: f"Viggo/transformer/if/base+phrases/{args.seed}")
env = {'proj_dir': args.output_dir, "gpu": args.gpu}
trainer.run(env, verbose=True)
va_ds = setup_validation_data(dataset, lin_strat, is_delex, mr_vcb, utt_vcb)
tr_batches = make_batches(tr_ds, 128, WKRS, lin_strat, is_delex)
tr_phrase_batches = make_batches(tr_phrase_ds, 128, WKRS, lin_strat, is_delex)
tr_template_batches = make_batches(tr_templ_ds, 128, WKRS, lin_strat, is_delex)
tr_phrase_template_batches = make_batches(tr_phrase_templ_ds, 128, WKRS,
lin_strat, is_delex)
va_batches = make_batches(va_ds, 1 if model_type == 'transformer' else 64,
WKRS, lin_strat, is_delex)
model = setup_model(model_type,
rnn_dir,
L,
rnn_attn,
tie_dec_emb,
mr_vcb,
utt_vcb,
beam_size=8)
if dataset == 'Viggo':
mr_utils = mrt.viggo.mr_utils
elif dataset == 'E2E':
mr_utils = mrt.e2e.mr_utils
else:
raise Exception(f"Bad dataset: {dataset}")
def make_template():
if model_type in ['gru', 'lstm']:
TEMPLATE = "rnn/{lin_strat}_{delex}/{arch}_{dir}_L={layers}_attn={attn}_tied={tied}_{opt}_lr={lr}_wd={wd}_ls={ls}"
WKRS = HP('WKRS', 2, description='data loader processes')
T = HP('T', 500, description='max training epochs')
LR = HP('LR', 0.1, description='learning rate')
WD = HP('WD', 0.0, description='weight decay')
LS = HP('LS', 0.1, description='label smoothing')
L = HP('L', 2, description='num layers')
mr_vcb, utt_vcb = setup_vocab(dataset, lin_strat, is_delex)
tr_ds = setup_training_data(dataset, lin_strat, is_delex, mr_vcb, utt_vcb)
va_ds = setup_validation_data(dataset, lin_strat, is_delex, mr_vcb, utt_vcb)
tr_batches = make_batches(tr_ds, 128, WKRS, lin_strat, is_delex)
va_batches = make_batches(va_ds, 128, WKRS, lin_strat, is_delex)
model = setup_model(model_type, rnn_dir, L, rnn_attn, tie_dec_emb,
mr_vcb, utt_vcb, no_posemb=no_posemb,
no_encoder=no_encoder)
if dataset == 'Viggo':
mr_utils = mrt.viggo.mr_utils
elif dataset == 'E2E':
mr_utils = mrt.e2e.mr_utils
else:
raise Exception(f"Bad dataset: {dataset}")
def make_template():
if model_type in ['gru', 'lstm']:
TEMPLATE = "{no_enc}rnn/{lin_strat}_{delex}/{arch}_{dir}_L={layers}_attn={attn}_tied={tied}_{opt}_lr={lr}_wd={wd}_ls={ls}"
return TEMPLATE.format(
lin_strat=lin_strat, delex='delex' if is_delex else 'lex',
dir=rnn_dir,
