def test_model(state,
reference_tsv_path,
reduced_number_of_data=None,
strore_predicitions_fname=None):
dataset = DatasetDcase2019Task4(os.path.join(cfg.workspace),
base_feature_dir=os.path.join(
cfg.workspace, "dataset", "features"),
save_log_feature=False)
crnn_kwargs = state["model"]["kwargs"]
crnn = CRNN(**crnn_kwargs)
crnn.load(parameters=state["model"]["state_dict"])
LOG.info("Model loaded at epoch: {}".format(state["epoch"]))
pooling_time_ratio = state["pooling_time_ratio"]
crnn.load(parameters=state["model"]["state_dict"])
scaler = Scaler()
scaler.load_state_dict(state["scaler"])
classes = cfg.classes
many_hot_encoder = ManyHotEncoder.load_state_dict(
state["many_hot_encoder"])
crnn = crnn.eval()
[crnn] = to_cuda_if_available([crnn])
transforms_valid = get_transforms(cfg.max_frames, scaler=scaler)
LOG.info(reference_tsv_path)
df = dataset.initialize_and_get_df(reference_tsv_path,
reduced_number_of_data)
strong_dataload = DataLoadDf(df,
dataset.get_feature_file,
many_hot_encoder.encode_strong_df,
transform=transforms_valid)
predictions = get_predictions(crnn,
strong_dataload,
many_hot_encoder.decode_strong,
pooling_time_ratio,
save_predictions=strore_predicitions_fname)
compute_strong_metrics(predictions, df)
weak_dataload = DataLoadDf(df,
dataset.get_feature_file,
many_hot_encoder.encode_weak,
transform=transforms_valid)
weak_metric = get_f_measure_by_class(
crnn, len(classes), DataLoader(weak_dataload,
batch_size=cfg.batch_size))
LOG.info("Weak F1-score per class: \n {}".format(
pd.DataFrame(weak_metric * 100, many_hot_encoder.labels)))
LOG.info("Weak F1-score macro averaged: {}".format(np.mean(weak_metric)))
def set_df_list(self, train):
dataset = DatasetDcase2019Task4(cfg.workspace,
base_feature_dir=os.path.join(cfg.workspace, "dataset", "features"),
save_log_feature=False)
transforms = get_transforms(cfg.max_frames)
weak_df = dataset.initialize_and_get_df(cfg.weak)
load_weak = DataLoadDf(weak_df, dataset.get_feature_file, None, transform=transforms)
if train ==True:
self.list_dataset = [load_weak]
else:
synthetic_df = dataset.initialize_and_get_df(cfg.synthetic, download=False)
synthetic_df.onset = synthetic_df.onset * cfg.sample_rate // cfg.hop_length
synthetic_df.offset = synthetic_df.offset * cfg.sample_rate // cfg.hop_length
validation_df = dataset.initialize_and_get_df(cfg.validation)
validation_df.onset = validation_df.onset * cfg.sample_rate // cfg.hop_length
validation_df.offset = validation_df.offset * cfg.sample_rate // cfg.hop_length
eval_desed_df = dataset.initialize_and_get_df(cfg.eval_desed)
eval_desed_df.onset = eval_desed_df.onset * cfg.sample_rate // cfg.hop_length
eval_desed_df.offset = eval_desed_df.offset * cfg.sample_rate // cfg.hop_length
# many_hot_encoder = ManyHotEncoder(classes, n_frames=cfg.max_frames // pooling_time_ratio)
load_synthetic = DataLoadDf(synthetic_df, dataset.get_feature_file, None, transform=transforms)
load_validation = DataLoadDf(validation_df, dataset.get_feature_file, None, transform=transforms)
load_eval_desed = DataLoadDf(eval_desed_df, dataset.get_feature_file, None, transform=transforms)
self.list_dataset = [load_weak, load_synthetic, load_validation, load_eval_desed]
scaler = Scaler()
scaler.calculate_scaler(ConcatDataset(self.list_dataset))
transforms = get_transforms(cfg.max_frames, scaler)
for i in range(len(self.list_dataset)):
self.list_dataset[i].set_transform(transforms)
print(self.list_dataset)
fixed_segment=f_args.fixed_segment
)
if resume_training is None:
classes = dataset.classes
many_hot_encoder = ManyHotEncoder(classes)
else:
many_hot_encoder = ManyHotEncoder.load_state_dict(state["many_hot_encoder"])
classes = many_hot_encoder.labels
encode_function_label = many_hot_encoder.encode_weak
# Datasets
trans_fr = [ApplyLog(), ToTensor(), Unsqueeze(0)]
train_weak_df_fr = dfs["train"]
train_weak_dl_fr = DataLoadDf(train_weak_df_fr, encode_function_label, transform=Compose(trans_fr))
if type_positive != "label" or type_negative != "label":
unlabel_df_fr = dataset.get_df_feat_dir(cfg.unlabel, subpart_data=subpart_data, frames_in_sec=frames_in_sec)
unlabel_dl_fr = DataLoadDf(unlabel_df_fr, encode_function_label, transform=Compose(trans_fr))
datasets_mean = [train_weak_dl_fr, unlabel_dl_fr]
else:
datasets_mean = [train_weak_dl_fr]
# Normalize
if resume_training is None:
scaler = ScalerSum()
scaler.calculate_scaler(ConcatDataset(datasets_mean))
else:
scaler = ScalerSum.load_state_dict(state["scaler"])
LOG.debug(scaler.mean_)
dfs = get_dfs(dataset,
reduced_number_of_data,
separated_sources=use_separated_sources)
# 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_conv=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]))
logger.debug(f"scaler mean: {scaler.mean_}")
else:
scaler_args = ["global", "min-max"]
scaler = ScalerPerAudio(*scaler_args)
transforms = get_transforms(cfg.max_frames,
scaler,
def test_model(state, reduced_number_of_data, strore_predicitions_fname=None):
crnn_kwargs = state["model"]["kwargs"]
crnn = CRNN(**crnn_kwargs)
crnn.load(parameters=state["model"]["state_dict"])
LOG.info("Model loaded at epoch: {}".format(state["epoch"]))
pooling_time_ratio = state["pooling_time_ratio"]
crnn.load(parameters=state["model"]["state_dict"])
scaler = Scaler()
scaler.load_state_dict(state["scaler"])
classes = cfg.classes
many_hot_encoder = ManyHotEncoder.load_state_dict(
state["many_hot_encoder"])
# ##############
# Validation
# ##############
crnn = crnn.eval()
[crnn] = to_cuda_if_available([crnn])
transforms_valid = get_transforms(cfg.max_frames, scaler=scaler)
# # 2018
# LOG.info("Eval 2018")
# eval_2018_df = dataset.initialize_and_get_df(cfg.eval2018, reduced_number_of_data)
# # Strong
# eval_2018_strong = DataLoadDf(eval_2018_df, dataset.get_feature_file, many_hot_encoder.encode_strong_df,
# transform=transforms_valid)
# predictions = get_predictions(crnn, eval_2018_strong, many_hot_encoder.decode_strong)
# compute_strong_metrics(predictions, eval_2018_df, pooling_time_ratio)
# # Weak
# eval_2018_weak = DataLoadDf(eval_2018_df, dataset.get_feature_file, many_hot_encoder.encode_weak,
# transform=transforms_valid)
# weak_metric = get_f_measure_by_class(crnn, len(classes), DataLoader(eval_2018_weak, batch_size=cfg.batch_size))
# LOG.info("Weak F1-score per class: \n {}".format(pd.DataFrame(weak_metric * 100, many_hot_encoder.labels)))
# LOG.info("Weak F1-score macro averaged: {}".format(np.mean(weak_metric)))
# Validation 2019
# LOG.info("Validation 2019 (original code)")
# b_dataset = B_DatasetDcase2019Task4(cfg.workspace,
# base_feature_dir=os.path.join(cfg.workspace, 'dataset', 'features'),
# save_log_feature=False)
# b_validation_df = b_dataset.initialize_and_get_df(cfg.validation, reduced_number_of_data)
# b_validation_df.to_csv('old.csv')
# b_validation_strong = B_DataLoadDf(b_validation_df,
# b_dataset.get_feature_file, many_hot_encoder.encode_strong_df,
# transform=transforms_valid)
# predictions2 = get_predictions(crnn, b_validation_strong, many_hot_encoder.decode_strong,
# save_predictions=strore_predicitions_fname)
# compute_strong_metrics(predictions2, b_validation_df, pooling_time_ratio)
# b_validation_weak = B_DataLoadDf(b_validation_df, b_dataset.get_feature_file, many_hot_encoder.encode_weak,
# transform=transforms_valid)
# weak_metric = get_f_measure_by_class(crnn, len(classes), DataLoader(b_validation_weak, batch_size=cfg.batch_size))
# LOG.info("Weak F1-score per class: \n {}".format(pd.DataFrame(weak_metric * 100, many_hot_encoder.labels)))
# LOG.info("Weak F1-score macro averaged: {}".format(np.mean(weak_metric)))
# ============================================================================================
# ============================================================================================
# ============================================================================================
dataset = DatasetDcase2019Task4(feature_dir=cfg.feature_dir,
local_path=cfg.workspace,
exp_tag=cfg.exp_tag,
save_log_feature=False)
# Validation 2019
LOG.info("Validation 2019")
validation_df = dataset.initialize_and_get_df(cfg.validation,
reduced_number_of_data)
validation_strong = DataLoadDf(validation_df,
dataset.get_feature_file,
many_hot_encoder.encode_strong_df,
transform=transforms_valid)
predictions = get_predictions(crnn,
validation_strong,
many_hot_encoder.decode_strong,
save_predictions=strore_predicitions_fname)
vdf = validation_df.copy()
vdf.filename = vdf.filename.str.replace('.npy', '.wav')
pdf = predictions.copy()
pdf.filename = pdf.filename.str.replace('.npy', '.wav')
compute_strong_metrics(pdf, vdf, pooling_time_ratio)
validation_weak = DataLoadDf(validation_df,
dataset.get_feature_file,
many_hot_encoder.encode_weak,
transform=transforms_valid)
weak_metric = get_f_measure_by_class(
crnn, len(classes),
DataLoader(validation_weak, batch_size=cfg.batch_size))
LOG.info("Weak F1-score per class: \n {}".format(
pd.DataFrame(weak_metric * 100, many_hot_encoder.labels)))
LOG.info("Weak F1-score macro averaged: {}".format(np.mean(weak_metric)))
if torch.cuda.is_available():
emb_model = emb_model.cuda()
emb_model.eval()
many_hot_encoder = ManyHotEncoder.load_state_dict(state['many_hot_encoder'])
encode_function_label = many_hot_encoder.encode_weak
scaler = ScalerSum.load_state_dict(state['scaler'])
frames_in_sec = cfg.frames_in_sec
transf = Compose([ApplyLog(), PadOrTrunc(nb_frames=cfg.frames), ToTensor(), Unsqueeze(0),
Normalize(scaler), Unsqueeze(1)])
test_fr = dataset.get_df_feat_dir(cfg.test2018, frames_in_sec=frames_in_sec, subpart_data=subpart_data)
print(len(test_fr))
test_dataset = DataLoadDf(test_fr, many_hot_encoder.encode_weak, transform=transf)
embed_set = "embedding"
embed_dir = "stored_data/embeddings"
embed_dir = os.path.join(embed_dir, embed_name, "embeddings")
create_folder(embed_dir)
fig_dir = os.path.join(embed_dir, "figures")
create_folder(fig_dir)
df_emb, embeddings = calculate_embedding(test_dataset, emb_model,
savedir=os.path.join(embed_dir, embed_set), concatenate="append")
print(embeddings.mean())
print(embeddings.var())
embeddings = sklearn.preprocessing.StandardScaler().fit_transform(embeddings.reshape(embeddings.shape[0], -1))
print("normalized")
print(embeddings.mean())
[weak_df, validation_df, synthetic_df])
# Be careful, frames is max_frames // pooling_time_ratio because max_pooling is applied on time axis in the model
many_hot_encoder = ManyHotEncoder(classes,
n_frames=cfg.max_frames //
pooling_time_ratio)
transforms = get_transforms(cfg.max_frames)
# Divide weak in train and valid
train_weak_df = weak_df.sample(frac=0.8, random_state=26)
valid_weak_df = weak_df.drop(train_weak_df.index).reset_index(drop=True)
train_weak_df = train_weak_df.reset_index(drop=True)
LOG.debug(valid_weak_df.event_labels.value_counts())
train_weak_data = DataLoadDf(train_weak_df,
dataset.get_feature_file,
many_hot_encoder.encode_strong_df,
transform=transforms)
# Divide synthetic in train and valid
filenames_train = synthetic_df.filename.drop_duplicates().sample(
frac=0.8, random_state=26)
train_synth_df = synthetic_df[synthetic_df.filename.isin(filenames_train)]
valid_synth_df = synthetic_df.drop(
train_synth_df.index).reset_index(drop=True)
# Put train_synth in frames so many_hot_encoder can work.
# Not doing it for valid, because not using labels (when prediction) and event based metric expect sec.
train_synth_df_frames = train_synth_df.copy()
train_synth_df_frames.onset = train_synth_df_frames.onset * cfg.sample_rate // cfg.hop_length // pooling_time_ratio
train_synth_df_frames.offset = train_synth_df_frames.offset * cfg.sample_rate // cfg.hop_length // pooling_time_ratio
LOG.debug(valid_synth_df.event_label.value_counts())
train_weak_df = dfs["train"]
classes = dataset.classes
many_hot_encoder = ManyHotEncoder(classes)
# ##############
# Triplet dataset
# #############
batch_size = cfg.batch_size
num_workers = cfg.num_workers
list_trans_fr = [ApplyLog(), ToTensor(), Unsqueeze(0)]
if args.segment:
list_trans_fr.append(Unsqueeze(0))
train_set = DataLoadDf(train_weak_df,
many_hot_encoder.encode_weak,
Compose(list_trans_fr),
return_indexes=False)
LOG.debug("len train : {}".format(len(train_set)))
# train_load = DataLoader(train_set, batch_size=batch_size, num_workers=num_workers, shuffle=True,
# drop_last=True, collate_fn=default_collate)
# scaler = Scaler()
scaler = ScalerSum()
scaler.calculate_scaler(train_set)
LOG.debug(scaler.mean_)
list_trans_fr.append(Normalize(scaler))
train_set.set_transform(Compose(list_trans_fr))
# Validation data
valid_weak_df = dfs["valid"]
if valid_weak_df is not None:
LOG.info("train_classes repartition: \n {}".format(
train_weak_df.event_labels.value_counts()))
classes = dataset.classes
many_hot_encoder = ManyHotEncoder(classes)
# Model
batch_size = cfg.batch_size
num_workers = cfg.num_workers
list_trans_fr = [ApplyLog(), ToTensor(), Unsqueeze(0)]
if args.segment:
list_trans_fr.append(Unsqueeze(0))
train_set = DataLoadDf(train_weak_df,
many_hot_encoder.encode_weak,
Compose(list_trans_fr),
return_indexes=False)
if args.balance:
train_sampler = CategoriesSampler(train_set.df.event_labels, classes,
round(cfg.batch_size / len(classes)))
train_load = DataLoader(train_set,
num_workers=num_workers,
batch_sampler=train_sampler)
else:
train_load = DataLoader(train_set,
num_workers=num_workers,
batch_size=batch_size,
shuffle=True)
train_sampler = train_load.batch_sampler
LOG.debug("len train : {}".format(len(train_set)))
scaler = ScalerSum()
reduced_number_of_data,
training=True)
validation_df = dataset.initialize_and_get_df(cfg.validation,
reduced_number_of_data,
training=True)
test_df = dataset.initialize_and_get_df(cfg.test,
reduced_number_of_data,
training=True)
many_hot_encoder = ManyHotEncoder(cfg.classes,
n_frames=cfg.max_frames //
pooling_time_ratio)
transforms = get_transforms(cfg.max_frames)
train_data = DataLoadDf(train_df,
dataset.get_feature_file,
many_hot_encoder.encode_weak,
transform=transforms)
validation_data = DataLoadDf(validation_df,
dataset.get_feature_file,
many_hot_encoder.encode_weak,
transform=transforms)
test_data = DataLoadDf(test_df,
dataset.get_feature_file,
many_hot_encoder.encode_weak,
transform=transforms)
list_dataset = [train_data]
batch_sizes = [cfg.batch_size]
# batch_sizes = [cfg.batch_size // len(list_dataset)] * len(list_dataset)
weak_mask = slice(cfg.batch_size)
strong_mask = None
# Put train_synth in frames so many_hot_encoder can work.
# Not doing it for valid, because not using labels (when prediction) and event based metric expect sec.
train_synth_df.onset = train_synth_df.onset * cfg.sample_rate // cfg.hop_length // pooling_time_ratio
train_synth_df.offset = train_synth_df.offset * cfg.sample_rate // cfg.hop_length // pooling_time_ratio
# print(train_synth_df)
# exit()
LOG.debug(valid_synth_df.event_label.value_counts())
LOG.debug(valid_w2s_df.event_labels.value_counts())
# LOG.info(train_synth_df)
# exit()
train_weak_data = DataLoadDf(train_weak_df,
dataset.get_feature_file,
many_hot_encoder.encode_strong_df,
transform=transforms)
# unlabel_data = DataLoadDf(unlabel_df, dataset.get_feature_file, many_hot_encoder.encode_strong_df,
# transform=transforms)
train_synth_data = DataLoadDf(train_synth_df,
dataset.get_feature_file,
many_hot_encoder.encode_strong_df,
transform=transforms)
#############################################################################
train_w2s_data = DataLoadDf(train_w2s_df,
dataset.get_feature_file,
many_hot_encoder.encode_strong_df,
transform=transforms)
def datasets_classif(model,
train_weak_embed,
valid_weak_dl_fr,
test_dl_fr,
args,
many_hot_encoder,
classes,
save_name="",
eval_dl=None):
encode_function_label = many_hot_encoder.encode_weak
num_workers = cfg.num_workers
model.eval()
embed_dir = "stored_data/embeddings"
embed_dir = os.path.join(embed_dir, save_name)
create_folder(embed_dir)
fig_dir = os.path.join(embed_dir, "figures")
create_folder(fig_dir)
if args.agg_time is not None:
trans_embedding = [ToTensor(), View(-1)]
else:
trans_embedding = [ToTensor()]
model = to_cuda_if_available(model)
embed_set = "final"
train_embed_dir = os.path.join(embed_dir, embed_set)
df_weak, embed_weak = calculate_embedding(train_weak_embed,
model,
savedir=train_embed_dir,
concatenate="append")
weak_embed = DataLoadDf(df_weak,
encode_function_label,
transform=Compose(trans_embedding))
LOG.info(f"len weak embed: {len(weak_embed)}")
weak_embed.set_transform(Compose(trans_embedding))
batch_size_classif = cfg.batch_size_classif
df_valid, embed_valid = calculate_embedding(valid_weak_dl_fr,
model,
savedir=train_embed_dir,
concatenate="append")
valid_embed = DataLoadDf(df_valid,
encode_function_label,
transform=Compose(trans_embedding))
embed_set = "final_test"
test_embed_dir = os.path.join(embed_dir, embed_set)
df_test_embed, emb_test = calculate_embedding(test_dl_fr,
model,
savedir=test_embed_dir,
concatenate="append")
test_embed = DataLoadDf(df_test_embed,
encode_function_label,
transform=Compose(trans_embedding))
if args.balance:
n_per_class = max(round(batch_size_classif / len(classes)), 1)
weak_sampler = CategoriesSampler(weak_embed.df.event_labels, classes,
n_per_class)
weak_embed_loader = DataLoader(weak_embed,
batch_sampler=weak_sampler,
num_workers=num_workers)
valid_sampler = CategoriesSampler(valid_embed.df.event_labels, classes,
n_per_class)
valid_embed_loader = DataLoader(valid_embed,
batch_sampler=valid_sampler,
num_workers=num_workers)
test_sampler = CategoriesSampler(test_embed.df.event_labels, classes,
n_per_class)
test_embed_loader = DataLoader(test_embed,
batch_sampler=test_sampler,
num_workers=num_workers)
else:
weak_embed_loader = DataLoader(weak_embed,
batch_size=batch_size_classif,
num_workers=num_workers,
shuffle=True,
drop_last=True)
valid_embed_loader = DataLoader(valid_embed,
batch_size=batch_size_classif,
shuffle=False,
num_workers=num_workers,
drop_last=False)
test_embed_loader = DataLoader(test_embed,
batch_size=batch_size_classif,
shuffle=False,
num_workers=num_workers,
drop_last=False)
if eval_dl is not None:
model = to_cuda_if_available(model)
embed_set = "final_eval"
eval_embed_dir = os.path.join(embed_dir, embed_set)
df_eval_embed, embed_eval = calculate_embedding(eval_dl,
model,
savedir=eval_embed_dir,
concatenate="append")
eval_embed = DataLoadDf(df_eval_embed,
encode_function_label,
transform=Compose(trans_embedding))
if args.balance:
eval_sampler = CategoriesSampler(eval_embed.df.event_labels,
classes, n_per_class)
eval_embed_loader = DataLoader(eval_embed,
batch_sampler=eval_sampler,
num_workers=num_workers)
else:
eval_embed_loader = DataLoader(eval_embed,
batch_size=batch_size_classif,
shuffle=False,
num_workers=num_workers,
drop_last=False)
else:
eval_embed_loader = None
model = to_cpu(model)
return {
"train": weak_embed_loader,
"valid": valid_embed_loader,
"test": test_embed_loader,
"eval": eval_embed_loader
}