def _load_state_vars(state, gtruth_df, median_win=None):
pred_df = gtruth_df.copy()
# Define dataloader
many_hot_encoder = ManyHotEncoder.load_state_dict(
state["many_hot_encoder"])
scaler = _load_scaler(state)
crnn = _load_crnn(state)
transforms_valid = get_transforms(cfg.max_frames,
scaler=scaler,
add_axis=0)
strong_dataload = DataLoadDf(pred_df,
many_hot_encoder.encode_strong_df,
transforms_valid,
return_indexes=True)
strong_dataloader_ind = DataLoader(strong_dataload,
batch_size=cfg.batch_size,
drop_last=False)
pooling_time_ratio = state["pooling_time_ratio"]
many_hot_encoder = ManyHotEncoder.load_state_dict(
state["many_hot_encoder"])
if median_win is None:
median_win = state["median_window"]
return {
"model": crnn,
"dataloader": strong_dataloader_ind,
"pooling_time_ratio": pooling_time_ratio,
"many_hot_encoder": many_hot_encoder,
"median_window": median_win
}
def audio_tagging_results(reference, estimated):
classes = []
if "event_label" in reference.columns:
classes.extend(reference.event_label.dropna().unique())
classes.extend(estimated.event_label.dropna().unique())
classes = list(set(classes))
mhe = ManyHotEncoder(classes)
reference = format_df(reference, mhe)
estimated = format_df(estimated, mhe)
else:
classes.extend(
reference.event_labels.str.split(
',', expand=True).unstack().dropna().unique())
classes.extend(
estimated.event_labels.str.split(
',', expand=True).unstack().dropna().unique())
classes = list(set(classes))
mhe = ManyHotEncoder(classes)
matching = reference.merge(estimated,
how='outer',
on="filename",
suffixes=["_ref", "_pred"])
def na_values(val):
if type(val) is np.ndarray:
return val
if pd.isna(val):
return np.zeros(len(classes))
return val
if not estimated.empty:
matching.event_label_pred = matching.event_label_pred.apply(na_values)
matching.event_label_ref = matching.event_label_ref.apply(na_values)
tp, fp, fn, tn = intermediate_at_measures(
np.array(matching.event_label_ref.tolist()),
np.array(matching.event_label_pred.tolist()))
macro_res = macro_f_measure(tp, fp, fn)
else:
macro_res = np.zeros(len(classes))
results_serie = pd.DataFrame(macro_res, index=mhe.labels)
return results_serie[0]
def _load_state_vars(state, gtruth_df, median_win=None):
pred_df = gtruth_df.copy()
# Define dataloader
many_hot_encoder = ManyHotEncoder.load_state_dict(state["many_hot_encoder"])
scaler = _load_scaler(state)
crnn = _load_crnn(state)
# Note, need to unsqueeze axis 1
transforms_valid = get_transforms(cfg.max_frames, scaler=scaler, add_axis=1)
# Note, no dataloader here
strong_dataload = DataLoadDf(pred_df, many_hot_encoder.encode_strong_df, transforms_valid, return_indexes=True)
pooling_time_ratio = state["pooling_time_ratio"]
many_hot_encoder = ManyHotEncoder.load_state_dict(state["many_hot_encoder"])
if median_win is None:
median_win = state["median_window"]
return {
"model": crnn,
"dataload": strong_dataload,
"pooling_time_ratio": pooling_time_ratio,
"many_hot_encoder": many_hot_encoder,
"median_window": median_win
}
out_nb_frames_1s = cfg.sample_rate / cfg.hop_size / pooling_time_ratio
median_window = max(int(cfg.median_window_s * out_nb_frames_1s), 1)
logger.debug(f"median_window: {median_window}")
# ##############
# DATA
# ##############
dataset = DESED(base_feature_dir=os.path.join(cfg.workspace, "dataset",
"features"),
compute_log=False)
dfs = get_dfs(dataset, reduced_number_of_data)
# Meta path for psds
durations_synth = get_durations_df(cfg.synthetic)
many_hot_encoder = ManyHotEncoder(cfg.classes,
n_frames=cfg.max_frames //
pooling_time_ratio)
encod_func = many_hot_encoder.encode_strong_df
# Normalisation per audio or on the full dataset
if cfg.scaler_type == "dataset":
transforms = get_transforms(cfg.max_frames, add_axis=add_axis_conv)
weak_data = DataLoadDf(dfs["weak"], encod_func, transforms)
unlabel_data = DataLoadDf(dfs["unlabel"], encod_func, transforms)
train_synth_data = DataLoadDf(dfs["train_synthetic"], encod_func,
transforms)
scaler_args = []
scaler = Scaler()
# # Only on real data since that's our final goal and test data are real
scaler.calculate_scaler(
ConcatDataset([weak_data, unlabel_data, train_synth_data]))