def filter_symbols(data: MergedDataset, symbols: SymbolIdDict,
accent_ids: AccentsDict, speakers: SpeakersDict,
allowed_symbol_ids: Set[int],
logger: Logger) -> MergedDatasetContainer:
# maybe check all symbol ids are valid before
allowed_symbols = [symbols.get_symbol(x) for x in allowed_symbol_ids]
not_allowed_symbols = [
symbols.get_symbol(x) for x in symbols.get_all_symbol_ids()
if x not in allowed_symbol_ids
]
logger.info(
f"Keep utterances with these symbols: {' '.join(allowed_symbols)}")
logger.info(
f"Remove utterances with these symbols: {' '.join(not_allowed_symbols)}"
)
logger.info("Statistics before filtering:")
log_stats(data, symbols, accent_ids, speakers, logger)
result = MergedDataset([
x for x in data.items() if contains_only_allowed_symbols(
deserialize_list(x.serialized_symbol_ids), allowed_symbol_ids)
])
if len(result) > 0:
logger.info(
f"Removed {len(data) - len(result)} from {len(data)} total entries and got {len(result)} entries ({len(result)/len(data)*100:.2f}%)."
)
else:
logger.info("Removed all utterances!")
new_symbol_ids = update_symbols(result, symbols)
new_accent_ids = update_accents(result, accent_ids)
new_speaker_ids = update_speakers(result, speakers)
logger.info("Statistics after filtering:")
log_stats(result, new_symbol_ids, new_accent_ids, new_speaker_ids, logger)
res = MergedDatasetContainer(
name=None,
data=result,
accent_ids=new_accent_ids,
speaker_ids=new_speaker_ids,
symbol_ids=new_symbol_ids,
)
return res
def sims_to_csv(sims: Dict[int, List[Tuple[int, float]]],
symbols: SymbolIdDict) -> pd.DataFrame:
lines = []
assert len(sims) == len(symbols)
for symbol_id, similarities in sims.items():
sims = [f"{symbols.get_symbol(symbol_id)}", "<=>"]
for other_symbol_id, similarity in similarities:
sims.append(symbols.get_symbol(other_symbol_id))
sims.append(f"{similarity:.2f}")
lines.append(sims)
df = pd.DataFrame(lines)
return df
def plot_embeddings(
symbols: SymbolIdDict, emb: torch.Tensor,
logger: Logger) -> Tuple[pd.DataFrame, go.Figure, go.Figure]:
assert emb.shape[0] == len(symbols)
logger.info(f"Emb size {emb.shape}")
logger.info(f"Sym len {len(symbols)}")
sims = get_similarities(emb.numpy())
df = sims_to_csv(sims, symbols)
all_symbols_sorted = [symbols.get_symbol(x) for x in range(len(symbols))]
emb_normed = norm2emb(emb)
fig_2d = emb_plot_2d(emb_normed, all_symbols_sorted)
fig_3d = emb_plot_3d(emb_normed, all_symbols_sorted)
return df, fig_2d, fig_3d
