def main (_folder, _csv_path_train, _imgs_folder_train, _lr_init, _sched_factor, _sched_min_lr, _sched_patience,
_batch_size, _epochs, _early_stop, _weights, _model_name, _pretrained, _save_folder, _csv_path_test,
_best_metric, _neurons_reducer_block, _comb_method, _comb_config, _use_meta_data):
meta_data_columns = ['age_approx', 'female', 'male', 'anterior torso', 'head/neck', "lateral torso",
'lower extremity', 'oral/genital', 'palms/soles', 'posterior torso', 'upper extremity']
_metric_options = {
'save_all_path': os.path.join(_save_folder, "best_metrics"),
'pred_name_scores': 'predictions_best_test.csv',
'normalize_conf_matrix': True}
_checkpoint_best = os.path.join(_save_folder, 'best-checkpoint/best-checkpoint.pth')
# Loading the csv file
csv_all_folders = pd.read_csv(_csv_path_train)
print("-" * 50)
print("- Loading validation data...")
val_csv_folder = csv_all_folders[ (csv_all_folders['folder'] == _folder) ]
train_csv_folder = csv_all_folders[ csv_all_folders['folder'] != _folder ]
# Loading validation data
val_imgs_id = val_csv_folder['image'].values
val_imgs_path = ["{}/{}.jpg".format(_imgs_folder_train, img_id) for img_id in val_imgs_id]
val_labels = val_csv_folder['diagnostic_number'].values
if _use_meta_data:
val_meta_data = val_csv_folder[meta_data_columns].values
print("-- Using {} meta-data features".format(len(meta_data_columns)))
else:
print("-- No metadata")
val_meta_data = None
val_data_loader = get_data_loader (val_imgs_path, val_labels, val_meta_data, transform=ImgEvalTransform(),
batch_size=_batch_size, shuf=True, num_workers=16, pin_memory=True)
print("-- Validation partition loaded with {} images".format(len(val_data_loader)*_batch_size))
print("- Loading training data...")
train_imgs_id = train_csv_folder['image'].values
train_imgs_path = ["{}/{}.jpg".format(_imgs_folder_train, img_id) for img_id in train_imgs_id]
train_labels = train_csv_folder['diagnostic_number'].values
if _use_meta_data:
train_meta_data = train_csv_folder[meta_data_columns].values
print("-- Using {} meta-data features".format(len(meta_data_columns)))
else:
print("-- No metadata")
train_meta_data = None
train_data_loader = get_data_loader (train_imgs_path, train_labels, train_meta_data, transform=ImgTrainTransform(),
batch_size=_batch_size, shuf=True, num_workers=16, pin_memory=True)
print("-- Training partition loaded with {} images".format(len(train_data_loader)*_batch_size))
print("-"*50)
####################################################################################################################
ser_lab_freq = get_labels_frequency(train_csv_folder, "diagnostic", "image")
_labels_name = ser_lab_freq.index.values
_freq = ser_lab_freq.values
####################################################################################################################
print("- Loading", _model_name)
model = set_model(_model_name, len(_labels_name), neurons_reducer_block=_neurons_reducer_block,
comb_method=_comb_method, comb_config=_comb_config, pretrained=_pretrained)
####################################################################################################################
if _weights == 'frequency':
_weights = (_freq.sum() / _freq).round(3)
loss_fn = nn.CrossEntropyLoss(weight=torch.Tensor(_weights).cuda())
# optimizer = optim.Adam(model.parameters(), lr=_lr_init)
optimizer = optim.SGD(model.parameters(), lr=_lr_init, momentum=0.9, weight_decay=0.001)
scheduler_lr = optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=_sched_factor, min_lr=_sched_min_lr,
patience=_sched_patience)
####################################################################################################################
print("- Starting the training phase...")
print("-" * 50)
fit_model (model, train_data_loader, val_data_loader, optimizer=optimizer, loss_fn=loss_fn, epochs=_epochs,
epochs_early_stop=_early_stop, save_folder=_save_folder, initial_model=None,
device=None, schedule_lr=scheduler_lr, config_bot=None, model_name="CNN", resume_train=False,
history_plot=True, val_metrics=["auc"], best_metric=_best_metric)
####################################################################################################################
# Testing the validation partition
print("- Evaluating the validation partition...")
test_model (model, val_data_loader, checkpoint_path=_checkpoint_best, loss_fn=loss_fn, save_pred=True,
partition_name='eval', metrics_to_comp='all', class_names=_labels_name, metrics_options=_metric_options,
apply_softmax=True, verbose=False)
####################################################################################################################
####################################################################################################################
print("- Loading test data...")
csv_test = pd.read_csv(_csv_path_test)
test_imgs_id = csv_test['image'].values
test_imgs_path = ["{}/{}.jpg".format(_imgs_folder_train, img_id) for img_id in test_imgs_id]
test_labels = csv_test['diagnostic_number'].values
if _use_meta_data:
test_meta_data = csv_test[meta_data_columns].values
print("-- Using {} meta-data features".format(len(meta_data_columns)))
else:
test_meta_data = None
print("-- No metadata")
_metric_options = {
'save_all_path': os.path.join(_save_folder, "test_pred"),
'pred_name_scores': 'predictions.csv',
'normalize_conf_matrix': True}
test_data_loader = get_data_loader(test_imgs_path, test_labels, test_meta_data, transform=ImgEvalTransform(),
batch_size=_batch_size, shuf=False, num_workers=16, pin_memory=True)
print("-" * 50)
# Testing the test partition
print("\n- Evaluating the validation partition...")
test_model(model, test_data_loader, checkpoint_path=None, metrics_to_comp="all",
class_names=_labels_name, metrics_options=_metric_options, save_pred=True, verbose=False)
def main(_folder, _csv_path_train, _imgs_folder_train, _lr_init, _sched_factor,
_sched_min_lr, _sched_patience, _batch_size, _epochs, _early_stop,
_weights, _model_name, _pretrained, _save_folder, _best_metric):
_metric_options = {
'save_all_path': os.path.join(_save_folder, "best_metrics"),
'pred_name_scores': 'predictions_best_test.csv',
'normalize_conf_matrix': True
}
_checkpoint_best = os.path.join(_save_folder,
'best-checkpoint/best-checkpoint.pth')
# Loading the csv file
csv_all_folders = pd.read_csv(_csv_path_train)
print("-" * 50)
print("- Loading validation data...")
val_csv_folder = csv_all_folders[(csv_all_folders['folder'] == _folder)]
train_csv_folder = csv_all_folders[csv_all_folders['folder'] != _folder]
# Loading validation data
val_labels = val_csv_folder['target_number'].values
val_labels_str = val_csv_folder['target'].values
val_imgs_path_ = val_csv_folder['Image_id'].values
val_imgs_path = [
"{}/{}/{}".format(_imgs_folder_train, lab, img_id)
for img_id, lab in zip(val_imgs_path_, val_labels_str)
]
val_meta_data = None
val_data_loader = get_data_loader(val_imgs_path,
val_labels,
val_meta_data,
transform=ImgEvalTransform(),
batch_size=_batch_size,
shuf=True,
num_workers=16,
pin_memory=True)
print("-- Validation partition loaded with {} images".format(
len(val_data_loader) * _batch_size))
print("- Loading training data...")
train_labels = train_csv_folder['target_number'].values
train_labels_str = train_csv_folder['target'].values
train_imgs_path_ = train_csv_folder['Image_id'].values
train_imgs_path = [
"{}/{}/{}".format(_imgs_folder_train, lab, img_id)
for img_id, lab in zip(train_imgs_path_, train_labels_str)
]
train_meta_data = None
train_data_loader = get_data_loader(train_imgs_path,
train_labels,
train_meta_data,
transform=ImgTrainTransform(),
batch_size=_batch_size,
shuf=True,
num_workers=16,
pin_memory=True)
print("-- Training partition loaded with {} images".format(
len(train_data_loader) * _batch_size))
print("-" * 50)
####################################################################################################################
ser_lab_freq = get_labels_frequency(train_csv_folder, "target", "Image_id")
_labels_name = ser_lab_freq.index.values
_freq = ser_lab_freq.values
####################################################################################################################
print("- Loading", _model_name)
model = set_model(_model_name, len(_labels_name), pretrained=_pretrained)
####################################################################################################################
if _weights == 'frequency':
_weights = (_freq.sum() / _freq).round(3)
loss_fn = nn.CrossEntropyLoss(weight=torch.Tensor(_weights).cuda())
optimizer = optim.SGD(model.parameters(),
lr=_lr_init,
momentum=0.9,
weight_decay=0.001)
scheduler_lr = optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
factor=_sched_factor,
min_lr=_sched_min_lr,
patience=_sched_patience)
####################################################################################################################
print("- Starting the training phase...")
print("-" * 50)
fit_model(model,
train_data_loader,
val_data_loader,
optimizer=optimizer,
loss_fn=loss_fn,
epochs=_epochs,
epochs_early_stop=_early_stop,
save_folder=_save_folder,
initial_model=None,
device=None,
schedule_lr=scheduler_lr,
config_bot=None,
model_name="CNN",
resume_train=False,
history_plot=True,
val_metrics=["auc"],
best_metric=_best_metric)
####################################################################################################################
# Testing the validation partition
print("- Evaluating the validation partition...")
test_model(model,
val_data_loader,
checkpoint_path=_checkpoint_best,
loss_fn=loss_fn,
save_pred=True,
partition_name='eval',
metrics_to_comp='all',
class_names=_labels_name,
metrics_options=_metric_options,
apply_softmax=True,
verbose=False)
def main(_folder, _csv_path_train, _imgs_folder_train, _lr_init, _sched_factor,
_sched_min_lr, _sched_patience, _batch_size, _epochs, _early_stop,
_weights, _model_name, _pretrained, _save_folder, _csv_path_test,
_best_metric, _neurons_reducer_block, _comb_method, _comb_config,
_use_meta_data, _metric_early_stop):
meta_data_columns = [
"smoke_False", "smoke_True", "drink_False", "drink_True",
"background_father_POMERANIA", "background_father_GERMANY",
"background_father_BRAZIL", "background_father_NETHERLANDS",
"background_father_ITALY", "background_father_POLAND",
"background_father_UNK", "background_father_PORTUGAL",
"background_father_BRASIL", "background_father_CZECH",
"background_father_AUSTRIA", "background_father_SPAIN",
"background_father_ISRAEL", "background_mother_POMERANIA",
"background_mother_ITALY", "background_mother_GERMANY",
"background_mother_BRAZIL", "background_mother_UNK",
"background_mother_POLAND", "background_mother_NORWAY",
"background_mother_PORTUGAL", "background_mother_NETHERLANDS",
"background_mother_FRANCE", "background_mother_SPAIN", "age",
"pesticide_False", "pesticide_True", "gender_FEMALE", "gender_MALE",
"skin_cancer_history_True", "skin_cancer_history_False",
"cancer_history_True", "cancer_history_False", "has_piped_water_True",
"has_piped_water_False", "has_sewage_system_True",
"has_sewage_system_False", "fitspatrick_3.0", "fitspatrick_1.0",
"fitspatrick_2.0", "fitspatrick_4.0", "fitspatrick_5.0",
"fitspatrick_6.0", "region_ARM", "region_NECK", "region_FACE",
"region_HAND", "region_FOREARM", "region_CHEST", "region_NOSE",
"region_THIGH", "region_SCALP", "region_EAR", "region_BACK",
"region_FOOT", "region_ABDOMEN", "region_LIP", "diameter_1",
"diameter_2", "itch_False", "itch_True", "itch_UNK", "grew_False",
"grew_True", "grew_UNK", "hurt_False", "hurt_True", "hurt_UNK",
"changed_False", "changed_True", "changed_UNK", "bleed_False",
"bleed_True", "bleed_UNK", "elevation_False", "elevation_True",
"elevation_UNK"
]
_metric_options = {
'save_all_path': os.path.join(_save_folder, "best_metrics"),
'pred_name_scores': 'predictions_best_test.csv',
'normalize_conf_matrix': True
}
_checkpoint_best = os.path.join(_save_folder,
'best-checkpoint/best-checkpoint.pth')
# Loading the csv file
csv_all_folders = pd.read_csv(_csv_path_train)
print("-" * 50)
print("- Loading validation data...")
val_csv_folder = csv_all_folders[(csv_all_folders['folder'] == _folder)]
train_csv_folder = csv_all_folders[csv_all_folders['folder'] != _folder]
# Loading validation data
val_imgs_id = val_csv_folder['img_id'].values
val_imgs_path = [
"{}/{}".format(_imgs_folder_train, img_id) for img_id in val_imgs_id
]
val_labels = val_csv_folder['diagnostic_number'].values
if _use_meta_data:
val_meta_data = val_csv_folder[meta_data_columns].values
print("-- Using {} meta-data features".format(len(meta_data_columns)))
else:
print("-- No metadata")
val_meta_data = None
val_data_loader = get_data_loader(val_imgs_path,
val_labels,
val_meta_data,
transform=ImgEvalTransform(),
batch_size=_batch_size,
shuf=True,
num_workers=16,
pin_memory=True)
print("-- Validation partition loaded with {} images".format(
len(val_data_loader) * _batch_size))
print("- Loading training data...")
train_imgs_id = train_csv_folder['img_id'].values
train_imgs_path = [
"{}/{}".format(_imgs_folder_train, img_id) for img_id in train_imgs_id
]
train_labels = train_csv_folder['diagnostic_number'].values
if _use_meta_data:
train_meta_data = train_csv_folder[meta_data_columns].values
print("-- Using {} meta-data features".format(len(meta_data_columns)))
else:
print("-- No metadata")
train_meta_data = None
train_data_loader = get_data_loader(train_imgs_path,
train_labels,
train_meta_data,
transform=ImgTrainTransform(),
batch_size=_batch_size,
shuf=True,
num_workers=16,
pin_memory=True)
print("-- Training partition loaded with {} images".format(
len(train_data_loader) * _batch_size))
print("-" * 50)
####################################################################################################################
ser_lab_freq = get_labels_frequency(train_csv_folder, "diagnostic",
"img_id")
_labels_name = ser_lab_freq.index.values
_freq = ser_lab_freq.values
print(ser_lab_freq)
####################################################################################################################
print("- Loading", _model_name)
model = set_model(_model_name,
len(_labels_name),
neurons_reducer_block=_neurons_reducer_block,
comb_method=_comb_method,
comb_config=_comb_config,
pretrained=_pretrained)
####################################################################################################################
if _weights == 'frequency':
_weights = (_freq.sum() / _freq).round(3)
loss_fn = nn.CrossEntropyLoss(weight=torch.Tensor(_weights).cuda())
optimizer = optim.SGD(model.parameters(),
lr=_lr_init,
momentum=0.9,
weight_decay=0.001)
scheduler_lr = optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
factor=_sched_factor,
min_lr=_sched_min_lr,
patience=_sched_patience)
####################################################################################################################
print("- Starting the training phase...")
print("-" * 50)
fit_model(model,
train_data_loader,
val_data_loader,
optimizer=optimizer,
loss_fn=loss_fn,
epochs=_epochs,
epochs_early_stop=_early_stop,
save_folder=_save_folder,
initial_model=None,
metric_early_stop=_metric_early_stop,
device=None,
schedule_lr=scheduler_lr,
config_bot=None,
model_name="CNN",
resume_train=False,
history_plot=True,
val_metrics=["balanced_accuracy"],
best_metric=_best_metric)
####################################################################################################################
# Testing the validation partition
print("- Evaluating the validation partition...")
test_model(model,
val_data_loader,
checkpoint_path=_checkpoint_best,
loss_fn=loss_fn,
save_pred=True,
partition_name='eval',
metrics_to_comp='all',
class_names=_labels_name,
metrics_options=_metric_options,
apply_softmax=True,
verbose=False)
####################################################################################################################
####################################################################################################################
print("- Loading test data...")
csv_test = pd.read_csv(_csv_path_test)
test_imgs_id = csv_test['img_id'].values
test_imgs_path = [
"{}/{}".format(_imgs_folder_train, img_id) for img_id in test_imgs_id
]
test_labels = csv_test['diagnostic_number'].values
if _use_meta_data:
test_meta_data = csv_test[meta_data_columns].values
print("-- Using {} meta-data features".format(len(meta_data_columns)))
else:
test_meta_data = None
print("-- No metadata")
_metric_options = {
'save_all_path': os.path.join(_save_folder, "test_pred"),
'pred_name_scores': 'predictions.csv',
'normalize_conf_matrix': True
}
test_data_loader = get_data_loader(test_imgs_path,
test_labels,
test_meta_data,
transform=ImgEvalTransform(),
batch_size=_batch_size,
shuf=False,
num_workers=16,
pin_memory=True)
print("-" * 50)
# Testing the test partition
print("\n- Evaluating the validation partition...")
test_model(model,
test_data_loader,
checkpoint_path=None,
metrics_to_comp="all",
class_names=_labels_name,
metrics_options=_metric_options,
save_pred=True,
verbose=False)