def dcase_predict(self,
experiment_path: str,
feature_file: str,
predict_scp: str,
output: str = "predition.csv",
**kwargs):
"""kwargs: {'max_length': int, 'method': str, 'beam_size': int}"""
dump = torch.load(os.path.join(experiment_path, "saved.pth"),
map_location="cpu")
# Load previous training config
config = dump["config"]
vocabulary = torch.load(config["vocab_file"])
model = self._get_model(config, len(vocabulary))
model.load_state_dict(dump["model"])
# Some scaler (sklearn standardscaler)
scaler = dump["scaler"]
zh = config["zh"]
model = model.to(self.device)
dataset = SJTUDatasetEval(feature=feature_file,
eval_scp=predict_scp,
transform=scaler.transform)
dataloader = torch.utils.data.DataLoader(dataset,
shuffle=False,
collate_fn=collate_fn((1, )),
batch_size=32,
num_workers=0)
pbar = ProgressBar(persist=False, ascii=True)
predictions = []
def _sample(engine, batch):
# batch: [keys, feats, feat_lens]
with torch.no_grad():
model.eval()
keys = batch[0]
output = self._forward(model, batch, mode="sample", **kwargs)
seqs = output["seqs"].cpu().numpy()
for idx, seq in enumerate(seqs):
caption = self._convert_idx2sentence(seq,
vocabulary,
zh=zh)
predictions.append({
"file_name": keys[idx] + ".wav",
"caption_predicted": caption
})
sample_engine = Engine(_sample)
pbar.attach(sample_engine)
sample_engine.run(dataloader)
pred_df = pd.DataFrame(predictions)
pred_df.to_csv(os.path.join(experiment_path, output), index=False)
def _ensemble(self,
path1: str,
path2: str,
kaldi_stream: str,
kaldi_scp: str,
max_length: int = None):
dump1 = torch.load(path1, map_location="cpu")
dump2 = torch.load(path2, map_location="cpu")
model1 = dump1["model"].to(device)
model2 = dump2["model"].to(device)
scaler = dump1["scaler"]
config = dump1["config"]
vocabulary = torch.load(config["vocab_file"])
dataset = SJTUDatasetEval(kaldi_stream=kaldi_stream,
kaldi_scp=kaldi_scp,
transform=scaler.transform)
# dataset[i]: key, feature
dataloader = torch.utils.data.DataLoader(dataset,
shuffle=False,
collate_fn=collate_fn((1, )),
batch_size=32,
num_workers=0)
if max_length is None:
max_length = model1.max_length
pbar = ProgressBar(persist=False, ascii=True)
key2pred = {}
def _sample(engine, batch):
# batch: [ids, feats, feat_lens]
ids = batch[0]
feats = batch[1]
feat_lens = batch[-1]
seqs = self._sample_batch(model1, model2, feats, feat_lens,
max_length)
seqs = seqs.cpu().numpy()
for idx, seq in enumerate(seqs):
caption = []
for word_id in seq:
word = vocabulary.idx2word[word_id]
if word == "<start>":
continue
elif word == "<end>":
break
else:
caption.append(word)
caption = " ".join(caption)
key2pred[ids[idx]] = [
caption,
]
sampler = Engine(_sample)
pbar.attach(sampler)
sampler.run(dataloader)
return key2pred
def _get_dataloaders(config, vocabulary):
scaler = getattr(pre, config["scaler"])(**config["scaler_args"])
inputdim = -1
caption_df = pd.read_json(config["caption_file"], dtype={"key": str})
for batch in tqdm(torch.utils.data.DataLoader(
SJTUDataset(
feature=config["feature_file"],
caption_df=caption_df,
vocabulary=vocabulary,
),
collate_fn=collate_fn([0, 1]),
**config["dataloader_args"]),
ascii=True):
feat = batch[0]
feat_lens = batch[-2]
packed_feat = torch.nn.utils.rnn.pack_padded_sequence(
feat, feat_lens, batch_first=True, enforce_sorted=False).data
scaler.partial_fit(packed_feat)
inputdim = feat.shape[-1]
assert inputdim > 0, "Reading input feature failed"
augments = train_util.parse_augments(config["augments"])
train_keys = np.random.choice(caption_df["key"].unique(),
int(
len(caption_df["key"].unique()) *
(config["train_percent"] / 100.)),
replace=False)
train_df = caption_df[caption_df["key"].apply(
lambda x: x in train_keys)]
val_df = caption_df[~caption_df.index.isin(train_df.index)]
train_loader = torch.utils.data.DataLoader(
SJTUDataset(feature=config["feature_file"],
caption_df=train_df,
vocabulary=vocabulary,
transform=[scaler.transform, augments]),
shuffle=True,
collate_fn=collate_fn([0, 1]),
**config["dataloader_args"])
if config["zh"]:
train_key2refs = train_df.groupby("key")["tokens"].apply(
list).to_dict()
val_key2refs = val_df.groupby("key")["tokens"].apply(
list).to_dict()
else:
train_key2refs = train_df.groupby("key")["caption"].apply(
list).to_dict()
val_key2refs = val_df.groupby("key")["caption"].apply(
list).to_dict()
val_loader = torch.utils.data.DataLoader(SJTUDataset(
feature=config["feature_file"],
caption_df=val_df,
vocabulary=vocabulary,
transform=scaler.transform),
shuffle=False,
collate_fn=collate_fn([0, 1]),
**config["dataloader_args"])
return train_loader, val_loader, {
"scaler": scaler,
"inputdim": inputdim,
"train_key2refs": train_key2refs,
"val_key2refs": val_key2refs
}
def evaluate(self,
experiment_path: str,
feature_file: str,
feature_scp: str,
caption_file: str,
caption_output: str = "eval_output.json",
score_output: str = "scores.txt",
**kwargs):
"""kwargs: {'max_length': int, 'method': str, 'beam_size': int}"""
dump = torch.load(os.path.join(experiment_path, "saved.pth"),
map_location="cpu")
# Load previous training config
config = dump["config"]
vocabulary = torch.load(config["vocab_file"])
model = self._get_model(config, vocabulary)
model.load_state_dict(dump["model"])
# Some scaler (sklearn standardscaler)
scaler = dump["scaler"]
zh = config["zh"]
model = model.to(self.device)
dataset = SJTUDatasetEval(feature=feature_file,
eval_scp=feature_scp,
transform=scaler.transform)
dataloader = torch.utils.data.DataLoader(dataset,
shuffle=False,
collate_fn=collate_fn((1, )),
batch_size=32,
num_workers=0)
caption_df = pd.read_json(caption_file, dtype={"key": str})
if zh:
key2refs = caption_df.groupby("key")["tokens"].apply(
list).to_dict()
else:
key2refs = caption_df.groupby("key")["caption"].apply(
list).to_dict()
model.eval()
key2pred = {}
def _sample(engine, batch):
with torch.no_grad():
model.eval()
keys = batch[0]
output = self._forward(model, batch, mode="sample", **kwargs)
seqs = output["seqs"].cpu().numpy()
for idx, seq in enumerate(seqs):
caption = self._convert_idx2sentence(seq, vocabulary, zh)
key2pred[keys[idx]] = [
caption,
]
pbar = ProgressBar(persist=False, ascii=True)
sampler = Engine(_sample)
pbar.attach(sampler)
sampler.run(dataloader)
pred_df = []
for key, pred in key2pred.items():
pred_df.append({
"filename": key + ".wav",
"caption": "".join(pred[0]) if zh else pred[0],
"tokens": pred[0] if zh else pred[0].split()
})
pred_df = pd.DataFrame(pred_df)
pred_df.to_json(os.path.join(experiment_path, caption_output))
from pycocoevalcap.bleu.bleu import Bleu
from pycocoevalcap.rouge.rouge import Rouge
from pycocoevalcap.cider.cider import Cider
from pycocoevalcap.meteor.meteor import Meteor
from pycocoevalcap.spice.spice import Spice
f = open(os.path.join(experiment_path, score_output), "w")
scorer = Bleu(n=4, zh=zh)
score, scores = scorer.compute_score(key2refs, key2pred)
for n in range(4):
f.write("Bleu-{}: {:6.3f}\n".format(n + 1, score[n]))
scorer = Rouge(zh=zh)
score, scores = scorer.compute_score(key2refs, key2pred)
f.write("ROUGE: {:6.3f}\n".format(score))
scorer = Cider(zh=zh)
score, scores = scorer.compute_score(key2refs, key2pred)
f.write("CIDEr: {:6.3f}\n".format(score))
if not zh:
scorer = Meteor()
score, scores = scorer.compute_score(key2refs, key2pred)
f.write("Meteor: {:6.3f}\n".format(score))
scorer = Spice()
score, scores = scorer.compute_score(key2refs, key2pred)
f.write("Spice: {:6.3f}\n".format(score))
f.close()
def sample(self,
experiment_path: str,
feature_file: str,
feature_scp: str,
output: str = "output_word.txt",
**kwargs):
"""Generate captions given experiment model"""
"""kwargs: {'max_length': int, 'method': str, 'beam_size': int}"""
import tableprint as tp
dump = torch.load(os.path.join(experiment_path, "saved.pth"),
map_location="cpu")
# Load previous training config
config = dump["config"]
vocab_size = len(torch.load(config["vocab_file"]))
model = self._get_model(config, vocab_size)
model.load_state_dict(dump["model"])
# Some scaler (sklearn standardscaler)
scaler = dump["scaler"]
vocabulary = torch.load(config["vocab_file"])
zh = config["zh"]
model = model.to(self.device)
dataset = SJTUDatasetEval(feature=feature_file,
eval_scp=feature_scp,
transform=scaler.transform)
dataloader = torch.utils.data.DataLoader(dataset,
shuffle=False,
collate_fn=collate_fn((1, )),
batch_size=16,
num_workers=0)
width_length = 80
pbar = ProgressBar(persist=False, ascii=True)
writer = open(os.path.join(experiment_path, output), "w")
writer.write(
tp.header(["InputUtterance", "Output Sentence"],
width=[len("InputUtterance"), width_length]))
writer.write('\n')
sentences = []
def _sample(engine, batch):
# batch: [keys, feats, feat_lens]
with torch.no_grad():
model.eval()
keys = batch[0]
output = self._forward(model, batch, mode="sample", **kwargs)
seqs = output["seqs"].cpu().numpy()
for idx, seq in enumerate(seqs):
caption = self._convert_idx2sentence(seq,
vocabulary,
zh=zh)
if zh:
sentence = " ".join(caption)
else:
sentence = caption
writer.write(
tp.row([keys[idx], sentence],
width=[len("InputUtterance"), width_length]) +
"\n")
sentences.append(sentence)
sample_engine = Engine(_sample)
pbar.attach(sample_engine)
sample_engine.run(dataloader)
writer.write(
tp.bottom(2, width=[len("InputUtterance"), width_length]) + "\n")
writer.write("Unique sentence number: {}\n".format(len(
set(sentences))))
writer.close()
def _get_dataloaders(config, vocabulary):
scaler = getattr(pre, config["scaler"])(**config["scaler_args"])
inputdim = -1
caption_df = pd.read_json(config["caption_file"], dtype={"key": str})
for batch in tqdm(torch.utils.data.DataLoader(
SJTUDataset(
feature_file=config["feature_file"],
caption_df=caption_df,
vocabulary=vocabulary,
),
collate_fn=collate_fn([0, 1]),
**config["dataloader_args"]),
ascii=True):
feat = batch[0]
feat = feat.reshape(-1, feat.shape[-1])
scaler.partial_fit(feat)
inputdim = feat.shape[-1]
assert inputdim > 0, "Reading inputstream failed"
augments = train_util.parse_augments(config["augments"])
cv_keys = np.random.choice(caption_df["key"].unique(),
int(
len(caption_df["key"].unique()) *
(1 - config["train_percent"] / 100.)),
replace=False)
cv_df = caption_df[caption_df["key"].apply(lambda x: x in cv_keys)]
train_df = caption_df[~caption_df.index.isin(cv_df.index)]
trainloader = torch.utils.data.DataLoader(
SJTUDataset(feature_file=config["feature_file"],
caption_df=train_df,
vocabulary=vocabulary,
transform=[scaler.transform, augments]),
shuffle=True,
collate_fn=collate_fn([0, 1]),
**config["dataloader_args"])
if config["zh"]:
train_key2refs = train_df.groupby("key")["tokens"].apply(
list).to_dict()
cv_key2refs = cv_df.groupby("key")["tokens"].apply(list).to_dict()
else:
train_key2refs = train_df.groupby("key")["caption"].apply(
list).to_dict()
cv_key2refs = cv_df.groupby("key")["caption"].apply(list).to_dict()
cvloader = torch.utils.data.DataLoader(SJTUDataset(
feature_file=config["feature_file"],
caption_df=cv_df,
vocabulary=vocabulary,
transform=[scaler.transform]),
shuffle=False,
collate_fn=collate_fn([0, 1]),
**config["dataloader_args"])
return trainloader, cvloader, {
"scaler": scaler,
"inputdim": inputdim,
"train_key2refs": train_key2refs,
"cv_key2refs": cv_key2refs
}
