def train_fold(base_model_path, save_dir, train_folds, val_folds, folds_data,
noisy_data):
train_transfrom = get_transforms(train=True,
size=CROP_SIZE,
wrap_pad_prob=WRAP_PAD_PROB)
mixer = RandomMixer([
SigmoidConcatMixer(sigmoid_range=(3, 12)),
AddMixer(alpha_dist='uniform')
],
p=[0.6, 0.4])
mixer = UseMixerWithProb(mixer, prob=MIXER_PROB)
curated_dataset = FreesoundDataset(folds_data,
train_folds,
transform=train_transfrom,
mixer=mixer)
noisy_dataset = FreesoundNoisyDataset(noisy_data,
transform=train_transfrom,
mixer=mixer)
train_dataset = RandomDataset([noisy_dataset, curated_dataset],
p=[NOISY_PROB, 1 - NOISY_PROB],
size=DATASET_SIZE)
val_dataset = FreesoundDataset(folds_data, val_folds,
get_transforms(False, CROP_SIZE))
train_loader = DataLoader(train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
drop_last=True,
num_workers=NUM_WORKERS)
val_loader = DataLoader(val_dataset,
batch_size=BATCH_SIZE * 2,
shuffle=False,
num_workers=NUM_WORKERS)
model = load_model(base_model_path, device=DEVICE)
model.set_lr(BASE_LR)
callbacks = [
MonitorCheckpoint(save_dir, monitor='val_lwlrap', max_saves=3),
CosineAnnealing(T_0=10, T_mult=2, eta_min=0.00001),
LoggingToFile(save_dir / 'log.txt'),
]
model.fit(train_loader,
val_loader=val_loader,
max_epochs=150,
callbacks=callbacks,
metrics=['multi_accuracy', 'lwlrap'])
def get_sed_inference_loader(df: pd.DataFrame, datadir: Path, config: dict):
transforms = get_transforms(config, "train")
if config["data"].get("denoised_audio_dir") is not None:
denoised_audio_dir = Path(config["data"]["denoised_audio_dir"])
else:
denoised_audio_dir = None # type: ignore
if config.get("dataset") is None:
dataset = datasets.PANNsSedDataset(
df, datadir, transforms, denoised_audio_dir)
elif config["dataset"]["name"] == "ChannelsSedDataset":
melspectrogram_parameters = config["dataset"]["melspectrogram_parameters"]
pcen_parameters = config["dataset"]["pcen_parameters"]
period = config["dataset"]["period"]
dataset = datasets.ChannelsSedDataset(
df, datadir, transforms, denoised_audio_dir,
melspectrogram_parameters,
pcen_parameters, period)
elif config["dataset"]["name"] == "NormalizedChannelsSedDataset":
melspectrogram_parameters = config["dataset"]["melspectrogram_parameters"]
pcen_parameters = config["dataset"]["pcen_parameters"]
period = config["dataset"]["period"]
dataset = datasets.NormalizedChannelsSedDataset(
df, datadir, transforms, denoised_audio_dir,
melspectrogram_parameters,
pcen_parameters, period)
loader = data.DataLoader(
dataset, batch_size=1, shuffle=False, num_workers=8)
return loader
def experiment_pred(experiment_dir, audio_id2spec):
print(f"Start predict: {experiment_dir}")
transforms = get_transforms(False, CROP_SIZE)
pred_lst = []
for fold in FOLDS:
print("Predict fold", fold)
fold_dir = experiment_dir / f'fold_{fold}'
model_path = get_best_model_path(fold_dir)
print("Model path", model_path)
predictor = Predictor(model_path, transforms, BATCH_SIZE,
CROP_SIZE, CROP_SIZE, DEVICE)
transforms = get_transforms(False, CROP_SIZE)
pred = fold_pred(predictor, audio_id2spec)
pred_lst.append(pred)
audio_id2pred = dict()
for audio_id in audio_id2spec:
pred = [p[audio_id] for p in pred_lst]
audio_id2pred[audio_id] = np.mean(pred, axis=0)
return audio_id2pred
def experiment_pred(experiment_dir, test_data):
print(f"Start predict: {experiment_dir}")
transforms = get_transforms(False, CROP_SIZE)
pred_df_lst = []
for fold in config.folds:
print("Predict fold", fold)
fold_dir = experiment_dir / f'fold_{fold}'
model_path = get_best_model_path(fold_dir)
print("Model path", model_path)
predictor = Predictor(model_path,
transforms,
BATCH_SIZE, (config.audio.n_mels, CROP_SIZE),
(config.audio.n_mels, CROP_SIZE // 4),
device=DEVICE)
pred_df = pred_test(predictor, test_data)
pred_df_lst.append(pred_df)
pred_df = gmean_preds_blend(pred_df_lst)
return pred_df
def train_fold(save_dir, train_folds, val_folds, folds_data, noisy_data,
corrected_noisy_data):
train_transfrom = get_transforms(train=True,
size=CROP_SIZE,
wrap_pad_prob=WRAP_PAD_PROB,
resize_scale=(0.8, 1.0),
resize_ratio=(1.7, 2.3),
resize_prob=0.33,
spec_num_mask=2,
spec_freq_masking=0.15,
spec_time_masking=0.20,
spec_prob=0.5)
mixer = RandomMixer([
SigmoidConcatMixer(sigmoid_range=(3, 12)),
AddMixer(alpha_dist='uniform')
],
p=[0.6, 0.4])
mixer = UseMixerWithProb(mixer, prob=MIXER_PROB)
curated_dataset = FreesoundDataset(folds_data,
train_folds,
transform=train_transfrom,
mixer=mixer)
noisy_dataset = FreesoundNoisyDataset(noisy_data,
transform=train_transfrom,
mixer=mixer)
corr_noisy_dataset = FreesoundCorrectedNoisyDataset(
corrected_noisy_data, transform=train_transfrom, mixer=mixer)
dataset_probs = [
NOISY_PROB, CORR_NOISY_PROB, 1 - NOISY_PROB - CORR_NOISY_PROB
]
print("Dataset probs", dataset_probs)
print("Dataset lens", len(noisy_dataset), len(corr_noisy_dataset),
len(curated_dataset))
train_dataset = RandomDataset(
[noisy_dataset, corr_noisy_dataset, curated_dataset],
p=dataset_probs,
size=DATASET_SIZE)
val_dataset = FreesoundDataset(folds_data, val_folds,
get_transforms(False, CROP_SIZE))
train_loader = DataLoader(train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
drop_last=True,
num_workers=NUM_WORKERS)
val_loader = DataLoader(val_dataset,
batch_size=BATCH_SIZE * 2,
shuffle=False,
num_workers=NUM_WORKERS)
model = FreesoundModel(PARAMS)
callbacks = [
MonitorCheckpoint(save_dir, monitor='val_lwlrap', max_saves=1),
ReduceLROnPlateau(monitor='val_lwlrap',
patience=6,
factor=0.6,
min_lr=1e-8),
EarlyStopping(monitor='val_lwlrap', patience=18),
LoggingToFile(save_dir / 'log.txt'),
]
model.fit(train_loader,
val_loader=val_loader,
max_epochs=700,
callbacks=callbacks,
metrics=['multi_accuracy', 'lwlrap'])
from src.datasets import WhaleDataset, RandomWhaleDataset
from src.argus_models import ArcfaceModel
from src.metrics import CosMAPatK
from src import config
experiment_name = 'arcface_resnet50_016'
experiment_dir = join(config.EXPERIMENTS_DIR, experiment_name)
train_val_csv_path = config.TRAIN_VAL_CSV_PATH
image_size = (96, 304)
num_workers = 8
batch_size = 128
balance_coef = 0.0
train_epoch_size = 50000
if __name__ == "__main__":
train_transforms = get_transforms(True, image_size)
train_dataset = RandomWhaleDataset(train_val_csv_path,
True,
balance_coef=balance_coef,
size=train_epoch_size,
**train_transforms)
val_transforms = get_transforms(False, image_size)
val_dataset = WhaleDataset(train_val_csv_path, False, **val_transforms)
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
num_workers=num_workers,
shuffle=True)
val_loader = DataLoader(val_dataset,
batch_size=batch_size,
num_workers=num_workers,
def train_fold(save_dir, train_folds, val_folds, folds_data):
train_transfrom = get_transforms(train=True,
size=CROP_SIZE,
wrap_pad_prob=0.0,
resize_scale=(0.8, 1.0),
resize_ratio=(1.7, 2.3),
resize_prob=0.0,
spec_num_mask=2,
spec_freq_masking=0.15,
spec_time_masking=0.20,
spec_prob=0.0)
val_transform = get_transforms(train=False, size=CROP_SIZE)
if MIXER_PROB:
mixer = get_mixer(mixer_prob=MIXER_PROB,
sigmoid_range=(3, 12),
alpha_dist='uniform',
random_prob=(0.6, 0.4))
else:
mixer = None
train_dataset = BirdsongDataset(folds_data,
folds=train_folds,
transform=train_transfrom,
mixer=mixer)
val_dataset = BirdsongDataset(folds_data,
folds=val_folds,
transform=val_transform)
train_loader = DataLoader(train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
drop_last=True,
num_workers=NUM_WORKERS)
val_loader = DataLoader(val_dataset,
batch_size=BATCH_SIZE * 2 // ITER_SIZE,
shuffle=False,
num_workers=NUM_WORKERS)
model = BirdsongModel(PARAMS)
if 'pretrained' in model.params['nn_module'][1]:
model.params['nn_module'][1]['pretrained'] = False
if USE_AMP:
initialize_amp(model)
model.set_device(DEVICES)
num_iterations = (5 * len(train_dataset)) // BATCH_SIZE
callbacks = [
MonitorCheckpoint(save_dir, monitor='val_loss', max_saves=1),
CosineAnnealingLR(T_max=num_iterations,
eta_min=0,
step_on_iteration=True),
EarlyStopping(monitor='val_loss', patience=12),
LoggingToFile(save_dir / 'log.txt'),
LoggingToCSV(save_dir / 'log.csv')
]
model.fit(train_loader,
val_loader=val_loader,
num_epochs=EPOCHS,
callbacks=callbacks,
metrics=['f1_score'])
del model
model_path = get_best_model_path(save_dir)
model = load_model(model_path)
val_dataset = BirdsongDataset(folds_data,
folds=val_folds + [config.n_folds],
transform=val_transform)
val_loader = DataLoader(val_dataset,
batch_size=BATCH_SIZE * 2 // ITER_SIZE,
shuffle=False,
num_workers=NUM_WORKERS)
model.set_device(DEVICES[0])
model.validate(val_loader,
metrics=['f1_score'],
callbacks=[
LoggingToFile(save_dir / 'log.txt'),
LoggingToCSV(save_dir / 'log.csv')
])
model_path = config.experiments_dir / 'corr_noisy_007/fold_0/model-072-0.886906.pth'
if not model_path.exists():
model_path.parent.mkdir(parents=True, exist_ok=True)
gdown.download(
'https://drive.google.com/uc?id=1vf6LtRHlpxCC-CjyCguM4JCrso2v3Tip',
str(model_path),
quiet=False)
DEVICE = 'cpu'
CROP_SIZE = 256
BATCH_SIZE = 16
TILE_STEP = 2
PREDICTOR = Predictor(model_path,
get_transforms(False, CROP_SIZE),
BATCH_SIZE, (config.audio.n_mels, CROP_SIZE),
(config.audio.n_mels, CROP_SIZE // TILE_STEP),
device=DEVICE)
signal_block_size = config.audio.sampling_rate
SPEC_BLOCK_SIZE = 64
spec_num = 4
SPEC_LST = [
np.zeros((config.audio.n_mels, SPEC_BLOCK_SIZE), dtype=np.float32)
] * spec_num
PREV_SIGNAL = np.zeros(signal_block_size, dtype=np.float32)
def audio_callback(indata, frames, time, status):
pred = np.concatenate(pred_lst)
target = np.concatenate(target_lst)
names = np.concatenate(image_names)
np.savez(
val_prediction_dir / f'preds-{NAME}.npz',
altered_pred=pred,
altered_target=target,
name=names,
)
if __name__ == "__main__":
if check_dir_not_exist(PREDICTION_DIR):
fold_dir = EXPERIMENT_DIR / f'fold_0'
model_path = get_best_model_path(fold_dir)
print("Model path", model_path)
predictor = Predictor(model_path,
batch_size=BATCH_SIZE,
transform=get_transforms(train=False),
device=DEVICE,
logits=LOGITS,
tta=TTA)
print("Test predict")
predict_test(predictor)
print("Val predict")
predict_validation(predictor)
args = parser.parse_args()
BATCH_SIZE = 16
CROP_SIZE = 256
DATASET_SIZE = 128 * 256
NOISY_PROB = 0.01
CORR_NOISY_PROB = 0.42
MIXER_PROB = 0.8
WRAP_PAD_PROB = 0.5
CORRECTIONS = True
train_transfrom = get_transforms(train=True,
size=CROP_SIZE,
wrap_pad_prob=WRAP_PAD_PROB,
resize_scale=(0.8, 1.0),
resize_ratio=(1.7, 2.3),
resize_prob=0.33,
spec_num_mask=2,
spec_freq_masking=0.15,
spec_time_masking=0.20,
spec_prob=0.5)
mixer = RandomMixer([
SigmoidConcatMixer(sigmoid_range=(3, 12)),
AddMixer(alpha_dist='uniform')
],
p=[0.6, 0.4])
mixer = UseMixerWithProb(mixer, prob=MIXER_PROB)
train_csv = pd.read_csv('./train_fold' + str(1) + '.csv')
val_csv = pd.read_csv('./val_fold' + str(1) + '.csv')
def get_loader(df: pd.DataFrame,
datadir: Path,
config: dict,
phase: str,
event_level_labels=None,
calltype_labels=None):
dataset_config = config["dataset"]
if dataset_config["name"] == "PANNsDataset":
transforms = get_transforms(config, phase)
loader_config = config["loader"][phase]
dataset = datasets.PANNsDataset(
df, datadir=datadir, transforms=transforms)
elif dataset_config["name"] == "PANNsMultiLabelDataset":
transforms = get_transforms(config, phase)
loader_config = config["loader"][phase]
period = dataset_config["params"][phase]["period"]
dataset = datasets.PANNsMultiLabelDataset(
df, datadir=datadir, transforms=transforms, period=period)
elif dataset_config["name"] == "MultiChannelDataset":
waveform_transforms = get_waveform_transforms(config, phase)
spectrogram_transforms = get_spectrogram_transforms(config, phase)
melspectrogram_parameters = dataset_config["params"]["melspectrogram_parameters"]
pcen_parameters = dataset_config["params"]["pcen_parameters"]
period = dataset_config["params"]["period"][phase]
loader_config = config["loader"][phase]
dataset = datasets.MultiChannelDataset(
df,
datadir=datadir,
img_size=dataset_config["img_size"],
waveform_transforms=waveform_transforms,
spectrogram_transforms=spectrogram_transforms,
melspectrogram_parameters=melspectrogram_parameters,
pcen_parameters=pcen_parameters,
period=period)
elif dataset_config["name"] == "LabelCorrectionDataset":
waveform_transforms = get_waveform_transforms(config, phase)
spectrogram_transforms = get_spectrogram_transforms(config, phase)
melspectrogram_parameters = dataset_config["params"]["melspectrogram_parameters"]
pcen_parameters = dataset_config["params"]["pcen_parameters"]
period = dataset_config["params"]["period"][phase]
n_segments = dataset_config["params"]["n_segments"][phase]
soft_label_dir = Path(dataset_config["params"]["soft_label_dir"])
threshold = dataset_config["params"].get("threshold", 0.5)
loader_config = config["loader"][phase]
dataset = datasets.LabelCorrectionDataset(
df,
datadir=datadir,
soft_label_dir=soft_label_dir,
img_size=dataset_config["img_size"],
waveform_transforms=waveform_transforms,
spectrogram_transforms=spectrogram_transforms,
melspectrogram_parameters=melspectrogram_parameters,
pcen_parameters=pcen_parameters,
period=period,
n_segments=n_segments,
threshold=threshold)
else:
raise NotImplementedError
loader = data.DataLoader(dataset, **loader_config)
return loader
def train_fold(save_dir,
train_folds,
val_folds,
local_rank=0,
distributed=False,
pretrain_dir=''):
folds_data = get_folds_data()
model = AlaskaModel(PARAMS)
model.params['nn_module'][1]['pretrained'] = False
if pretrain_dir:
pretrain_path = get_best_model_path(pretrain_dir)
if pretrain_path is not None:
print(f'Pretrain model path {pretrain_path}')
load_pretrain_weigths(model, pretrain_path)
else:
print(f"Pretrain model not found in '{pretrain_dir}'")
if USE_AMP:
initialize_amp(model)
if distributed:
model.nn_module = SyncBatchNorm.convert_sync_batchnorm(model.nn_module)
model.nn_module = DistributedDataParallel(
model.nn_module.to(local_rank),
device_ids=[local_rank],
output_device=local_rank)
if local_rank:
model.logger.disabled = True
else:
model.set_device(DEVICES)
if USE_EMA:
initialize_ema(model, decay=0.9999)
checkpoint = EmaMonitorCheckpoint
else:
checkpoint = MonitorCheckpoint
for epochs, stage in zip(TRAIN_EPOCHS, STAGE):
test_transform = get_transforms(train=False)
if stage == 'train':
mixer = RandomMixer([BitMix(gamma=0.25), EmptyMix()], p=[0., 1.])
train_transform = get_transforms(train=True)
else:
mixer = EmptyMix()
train_transform = get_transforms(train=False)
train_dataset = AlaskaDataset(folds_data,
train_folds,
transform=train_transform,
mixer=mixer)
val_dataset = AlaskaDataset(folds_data,
val_folds,
transform=test_transform)
val_sampler = AlaskaSampler(val_dataset, train=False)
if distributed:
train_sampler = AlaskaDistributedSampler(train_dataset)
else:
train_sampler = AlaskaSampler(train_dataset, train=True)
train_loader = DataLoader(train_dataset,
sampler=train_sampler,
num_workers=NUM_WORKERS,
batch_size=BATCH_SIZE)
val_loader = DataLoader(val_dataset,
sampler=val_sampler,
num_workers=NUM_WORKERS,
batch_size=VAL_BATCH_SIZE)
callbacks = []
if local_rank == 0:
callbacks += [
checkpoint(save_dir,
monitor='val_weighted_auc',
max_saves=5,
file_format=stage +
'-model-{epoch:03d}-{monitor:.6f}.pth'),
LoggingToFile(save_dir / 'log.txt'),
LoggingToCSV(save_dir / 'log.csv', append=True)
]
if stage == 'train':
callbacks += [
CosineAnnealingLR(T_max=epochs,
eta_min=get_lr(9e-6, WORLD_BATCH_SIZE))
]
elif stage == 'warmup':
warmup_iterations = epochs * (len(train_sampler) / BATCH_SIZE)
callbacks += [
LambdaLR(lambda x: x / warmup_iterations,
step_on_iteration=True)
]
if stage == 'train':
@argus.callbacks.on_epoch_start
def schedule_mixer_prob(state):
bitmix_prob = state.epoch / epochs
mixer.p = [bitmix_prob, 1 - bitmix_prob]
state.logger.info(f"Mixer probabilities {mixer.p}")
callbacks += [schedule_mixer_prob]
if distributed:
@argus.callbacks.on_epoch_complete
def schedule_sampler(state):
train_sampler.set_epoch(state.epoch + 1)
callbacks += [schedule_sampler]
metrics = ['weighted_auc', Accuracy('stegano'), Accuracy('quality')]
model.fit(train_loader,
val_loader=val_loader,
num_epochs=epochs,
callbacks=callbacks,
metrics=metrics)
def train_experiment(folds_data, noisy_data, num):
experiment_dir = SAVE_DIR / f'{num:04}'
np.random.seed(num)
random.seed(num)
random_params = {
'p_dropout': float(np.random.uniform(0.1, 0.3)),
'batch_size': int(np.random.choice([128])),
'lr': float(np.random.choice([0.001, 0.0006, 0.0003])),
'add_prob': float(np.random.uniform(0.0, 1.0)),
'noisy_prob': float(np.random.uniform(0.0, 1.0)),
'lsoft_beta': float(np.random.uniform(0.2, 0.8)),
'noisy_weight': float(np.random.uniform(0.3, 0.7)),
'patience': int(np.random.randint(2, 10)),
'factor': float(np.random.uniform(0.5, 0.8))
}
pprint(random_params)
params = {
'nn_module': ('SimpleKaggle', {
'num_classes': len(config.classes),
'dropout': random_params['p_dropout'],
'base_size': 64
}),
'loss': ('OnlyNoisyLSoftLoss', {
'beta': random_params['lsoft_beta'],
'noisy_weight': random_params['noisy_weight'],
'curated_weight': 1 - random_params['noisy_weight']
}),
'optimizer': ('Adam', {'lr': random_params['lr']}),
'device': 'cuda',
'amp': {
'opt_level': 'O2',
'keep_batchnorm_fp32': True,
'loss_scale': "dynamic"
}
}
pprint(params)
try:
train_transfrom = get_transforms(True, CROP_SIZE)
curated_dataset = FreesoundDataset(folds_data, TRAIN_FOLDS,
transform=train_transfrom,
add_prob=random_params['add_prob'])
noisy_dataset = FreesoundNoisyDataset(noisy_data,
transform=train_transfrom)
train_dataset = CombinedDataset(noisy_dataset, curated_dataset,
noisy_prob=random_params['noisy_prob'],
size=DATASET_SIZE)
val_dataset = FreesoundDataset(folds_data, VAL_FOLDS,
get_transforms(False, CROP_SIZE))
train_loader = DataLoader(train_dataset, batch_size=random_params['batch_size'],
shuffle=True, drop_last=True,
num_workers=NUM_WORKERS)
val_loader = DataLoader(val_dataset, batch_size=random_params['batch_size'] * 2,
shuffle=False, num_workers=NUM_WORKERS)
model = FreesoundModel(params)
callbacks = [
MonitorCheckpoint(experiment_dir, monitor='val_lwlrap', max_saves=1),
ReduceLROnPlateau(monitor='val_lwlrap',
patience=random_params['patience'],
factor=random_params['factor'],
min_lr=1e-8),
EarlyStopping(monitor='val_lwlrap', patience=20),
LoggingToFile(experiment_dir / 'log.txt'),
]
with open(experiment_dir / 'random_params.json', 'w') as outfile:
json.dump(random_params, outfile)
model.fit(train_loader,
val_loader=val_loader,
max_epochs=100,
callbacks=callbacks,
metrics=['multi_accuracy', 'lwlrap'])
except KeyboardInterrupt as e:
raise e
except BaseException as e:
print(f"Exception '{e}' with random params '{random_params}'")