def debug():
parser = ArgumentParser()
parser = PosDM.add_data_specific_args(parser)
parser.add_argument(
"--datasets",
nargs="*",
type=str,
default=["taskmaster", "multiwoz", "metalwoz", "coached"],
)
datasets = parser.parse_args().datasets
parser = add_builder_specific_args(parser,
datasets) # add for all builders
args = parser.parse_args()
dm = PosDM(args)
dm.prepare_data()
dm.setup("fit")
pos_model = GramTrainer()
for data in tqdm(dm.train_dset, desc=f"Train Pos"):
pos_model.process(data["pos"])
stats = pos_model.finalize()
sp_uni_prob = round(100 * stats["uni"]["SPEAKER"], 2)
print(f"Speaker: {sp_uni_prob}%")
# bigram
max_prob = -1
all_sequences = []
all_probs = []
print("=" * 70)
for w1, sp_prob in stats["bi"].items():
print(f"{w1} -> {round(100*sp_prob, 2)}%")
all_sequences.append(w1)
all_probs.append(sp_prob)
print("-" * 70)
all_probs, perm_idx = torch.tensor(all_probs).sort(descending=True)
all_seq = []
for i in perm_idx:
all_seq.append(all_sequences[i])
n = 50
for seq, prob in zip(all_seq[:n], all_probs[:n]):
print(seq, "->", prob.item())
# trigram
all_sequences = []
all_probs = []
print("=" * 70)
for w1, w2s in stats["tri"].items():
for w2, sp_prob in w2s.items():
all_sequences.append((w1, w2))
all_probs.append(sp_prob)
all_probs, perm_idx = torch.tensor(all_probs).sort(descending=True)
all_seq = []
for i in perm_idx:
all_seq.append(all_sequences[i])
n = 50
for seq, prob in zip(all_seq[:n], all_probs[:n]):
print(seq, prob.item())
def load():
parser = ArgumentParser()
parser.add_argument("--chunk_size", default=512, type=int)
parser = TurnGPTEval.add_eval_specific_args(parser)
# Data
parser = TurnGPTDM.add_data_specific_args(parser)
parser.add_argument(
"--datasets",
nargs="*",
type=str,
default=["coached"],
)
parser.add_argument(
"--split",
type=str,
default="train", # val, test
)
temp_args, _ = parser.parse_known_args()
# Add all datasets
datasets = temp_args.datasets
parser = add_builder_specific_args(parser, datasets) # add for all builders
args = parser.parse_args()
# Tokenizer
tokenizer = torch.load(args.tokenizer)
# Data
dm = TurnGPTDM(args, tokenizer)
dm.prepare_data()
# Choose Model
if args.model == "mini":
from turngpt.models.gpt_mini import TurnGPTMini
model = TurnGPTMini.load_from_checkpoint(checkpoint_path=args.checkpoint)
elif args.model == "pretrained":
from turngpt.models.pretrained import TurnGPTPretrained
model = TurnGPTPretrained.load_from_checkpoint(checkpoint_path=args.checkpoint)
return model, dm, args
elif temp_args.model == "sparse":
from models.gpt_sparse import TurnGPTSparse
parser = TurnGPTSparse.add_model_specific_args(parser)
elif temp_args.model == "pretrained":
from models.pretrained import TurnGPTPretrained
parser = TurnGPTPretrained.add_model_specific_args(parser)
elif temp_args.model == "rnn":
from gpt_rnn import TurnGPTRnn
parser = TurnGPTRnn.add_model_specific_args(parser)
else:
raise NotImplementedError(
'The model argument must be one of "mini", "sparse" or "pretrained"'
)
# Add all datasets
datasets = temp_args.datasets
parser = add_builder_specific_args(parser, datasets) # add for all builders
args = parser.parse_args()
# Where to save the training
print()
print("-" * 70)
args.save_dir = get_run_dir(__file__)
print(args.save_dir)
print()
main(args)