def main():
train_ds = MyImageFolder(root_dir="train/",
transform=config.train_transforms)
val_ds = MyImageFolder(root_dir="val/", transform=config.val_transforms)
train_loader = DataLoader(train_ds,
batch_size=config.BATCH_SIZE,
num_workers=config.NUM_WORKERS,
pin_memory=config.PIN_MEMORY,
shuffle=True)
val_loader = DataLoader(val_ds,
batch_size=config.BATCH_SIZE,
num_workers=config.NUM_WORKERS,
pin_memory=config.PIN_MEMORY,
shuffle=True)
loss_fn = nn.CrossEntropyLoss()
model = Net(net_version="b0", num_classes=10).to(config.DEVICE)
optimizer = optim.Adam(model.parameters(), lr=config.LEARNING_RATE)
scaler = torch.cuda.amp.GradScaler()
if config.LOAD_MODEL:
load_checkpoint(torch.load("my_checkpoint.pth.tar"), model, optimizer)
make_prediction(model, config.val_transforms, 'test/', config.DEVICE)
check_accuracy(val_loader, model, config.DEVICE)
for epoch in range(config.NUM_EPOCHS):
train_fn(train_loader, model, optimizer, loss_fn, scaler,
config.DEVICE)
check_accuracy(val_loader, model, config.DEVICE)
checkpoint = {
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_checkpoint(checkpoint)
def apply_KMM(trainX, trainY, testX, testY, window, source_pos, target_pos):
# Decision Tree
print("\n Kernel Mean Matching (Huang et al., 2006) ")
classifier = ImportanceWeightedClassifier(iwe='kmm', loss="dtree")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_DT_KMM, acc_DT_KMM_INFO = check_accuracy(testY, pred_naive)
# Logistic Regression
classifier = ImportanceWeightedClassifier(iwe='kmm', loss="logistic")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_LR_KMM, acc_LR_KMM_INFO = check_accuracy(testY, pred_naive)
# Naive Bayes Bernoulli
classifier = ImportanceWeightedClassifier(iwe='kmm', loss="berno")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_NB_KMM, acc_NB_KMM_INFO = check_accuracy(testY, pred_naive)
#
return pd.DataFrame(
[{
'window': window,
'source_position': source_pos,
'target_position': target_pos,
'acc_LR_KMM': acc_LR_KMM,
'acc_LR_KMM_INFO': str(acc_LR_KMM_INFO),
'acc_DT_KMM': acc_DT_KMM,
'acc_DT_KMM_INFO': str(acc_DT_KMM_INFO),
'acc_NB_KMM': acc_NB_KMM,
'acc_NB_KMM_INFO': str(acc_NB_KMM_INFO),
}]
)
def apply_SA(trainX, trainY, testX, testY, window, source_pos, target_pos):
# Decision Tree
print("\n Subspace Alignment (Fernando et al., 2013) ")
classifier = SubspaceAlignedClassifier(loss="dtree")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_DT_SA, acc_DT_SA_INFO = check_accuracy(testY, pred_naive)
# Logistic Regression
print("\n Subspace Alignment (Fernando et al., 2013) ")
classifier = SubspaceAlignedClassifier(loss="logistic")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_LR_SA, acc_LR_SA_INFO = check_accuracy(testY, pred_naive)
# Naive Bayes Bernoulli
print("\n Subspace Alignment (Fernando et al., 2013) ")
classifier = SubspaceAlignedClassifier(loss="berno")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_NB_SA, acc_NB_SA_INFO = check_accuracy(testY, pred_naive)
#
return pd.DataFrame(
[{
'window': window,
'source_position': source_pos,
'target_position': target_pos,
'acc_LR_SA': acc_LR_SA,
'acc_LR_SA_INFO': str(acc_LR_SA_INFO),
'acc_DT_SA': acc_DT_SA,
'acc_DT_SA_INFO': str(acc_DT_SA_INFO),
'acc_NB_SA': acc_NB_SA,
'acc_NB_SA_INFO': str(acc_NB_SA_INFO),
}]
)
def main():
model = UNET(in_channels=1, out_channels=1).to(device=DEVICE)
loss_fn = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
train_loader, val_loader = get_loaders(TRAIN_DIR, VAL_DIR, BATCH_SIZE,
NUM_WORKER, PIN_MEMORY)
if LOAD_MODEL:
load_checkpoint(torch.load("mycheckpoint.pth.tar"), model)
scaler = torch.cuda.amp.GradScaler()
for epoch in range(NUM_EPOCHS):
train_fn(train_loader, model, optimizer, loss_fn, scaler)
# TODO : save model
checkpoint = {
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_checkpoint(checkpoint)
# TODO : check acuuracy
check_accuracy(val_loader, model, device=DEVICE)
# TODO : Print results to folder
save_predictions_as_imgs(val_loader, model, folder='saved_imgs/')
def apply_NN(trainX, trainY, testX, testY, window, source_pos, target_pos):
# Decision Tree
print("\n Nearest-neighbour-based weighting (Loog, 2015) ")
classifier = ImportanceWeightedClassifier(iwe='nn', loss="dtree")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_DT_NN, acc_DT_NN_INFO = check_accuracy(testY, pred_naive)
# Logistic Regression
print("\n Nearest-neighbour-based weighting (Loog, 2015) ")
classifier = ImportanceWeightedClassifier(iwe='nn', loss="logistic")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_LR_NN, acc_LR_NN_INFO = check_accuracy(testY, pred_naive)
# Naive Bayes Bernoulli
print("\n Nearest-neighbour-based weighting (Loog, 2015) ")
classifier = ImportanceWeightedClassifier(iwe='nn', loss="berno")
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_NB_NN, acc_NB_NN_INFO = check_accuracy(testY, pred_naive)
#
return pd.DataFrame(
[{
'window': window,
'source_position': source_pos,
'target_position': target_pos,
'acc_LR_NN': acc_LR_NN,
'acc_LR_NN_INFO': str(acc_LR_NN_INFO),
'acc_DT_NN': acc_DT_NN,
'acc_DT_NN_INFO': str(acc_DT_NN_INFO),
'acc_NB_NN': acc_NB_NN,
'acc_NB_NN_INFO': str(acc_NB_NN_INFO),
}]
)
def main():
train_transform = A.Compose([
A.Resize(height=IMAGE_HEIGHT, width=IMAGE_WIDTH),
A.Rotate(limit=35, p=1.0),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.1),
A.Normalize(
mean=[0.0, 0.0, 0.0],
std=[1.0, 1.0, 1.0],
max_pixel_value=255.0,
),
ToTensorV2(),
], )
val_transforms = A.Compose([
A.Resize(height=IMAGE_HEIGHT, width=IMAGE_WIDTH),
A.Normalize(
mean=[0.0, 0.0, 0.0],
std=[1.0, 1.0, 1.0],
max_pixel_value=255.0,
),
ToTensorV2(),
], )
model = UNET(3, 1).to(DEVICE)
loss_fn = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
train_loader, val_loader = get_loaders(
TRAIN_IMG_DIR,
TRAIN_MASK_DIR,
VAL_IMG_DIR,
VAL_MASK_DIR,
BATCH_SIZE,
train_transform,
val_transforms,
NUM_WORKERS,
PIN_MEMORY,
)
check_accuracy(val_loader, model, device=DEVICE)
scaler = torch.cuda.amp.GradScaler()
for epoch in range(NUM_EPOCHS):
train_fn(train_loader, model, optimizer, loss_fn, scaler)
checkpoint = {
"state_dict": model.state_dict(),
"optimizer": optimizer.state_dict(),
}
save_checkpoint(checkpoint)
# check accuracy
check_accuracy(val_loader, model, device=DEVICE)
# print some examples to a folder
save_predictions_as_imgs(val_loader,
model,
folder="saved_images/",
device=DEVICE)
def main():
train_transform = A.Compose([
A.Resize(height=config.IMAGE_HEIGHT, width=config.IMAGE_WIDTH),
A.Rotate(limit=35, p=1.0),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.1),
A.normalize(mean=[0.0, 0.0, 0.0],
std=[1.0, 1.0, 1.0],
max_pixel_value=255.0),
ToTensorV2,
])
val_transform = A.Compose([
A.Resize(height=config.IMAGE_HEIGHT, width=config.IMAGE_WIDTH),
A.normalize(mean=[0.0, 0.0, 0.0],
std=[1.0, 1.0, 1.0],
max_pixel_value=255.0),
ToTensorV2,
])
model = UNet(in_channels=3, out_channels=1).to(config.DEVICE)
loss_fn = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters(), lr=config.LEARNING_RATE)
train_loader, val_loader = get_loaders(
config.TRAIN_IMAGE_DIR,
config.TRAIN_MASK_DIR,
config.VAL_IMG_DIR,
config.VAL_MASK_DIR,
config.BATCH_SIZE,
train_transform,
val_transform,
)
if config.LOAD_MODEL:
load_checkpoint(torch.load('my_checkpoint.pth.tar'), model)
scaler = torch.cuda.amp.GradScaler()
for epoch in range(config.NUM_EPOCHS):
train_fn(train_loader, model, optimizer, loss_fn, scaler)
# save model
checkpoint = {
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}
save_checkpoint(checkpoint)
# check acc
check_accuracy(val_loader, model, device=config.DEVICE)
# print some examples to a folder
save_predictions_as_imgs(val_loader,
model,
folder='saved_images',
device=config.DEVICE)
def train(seq_len,
window_size,
model_type,
params,
batch_size=16,
num_epochs=20,
print_every=5):
metrics = []
max_val_f2_score = 0.
best_model = None
train_data, validation_data = load_pytorch_data(seq_len, window_size)
if model_type == 'LSTM':
model = SimpleLSTM(INPUT_SIZE, params['lstm_hidden_size'])
elif model_type == 'CNN':
model = SimpleCNN(int(HZ * seq_len), params['cnn_hidden_size'])
else:
raise Exception('invalid model type')
optimizer = torch.optim.Adam(model.parameters(), lr=params['lr'])
criterion = torch.nn.CrossEntropyLoss(
weight=torch.tensor(params['loss_weights']))
print('starting training!')
for epoch in range(num_epochs):
print('starting epoch {}...'.format(epoch))
for iter, (X_batch, y_batch, idx) in enumerate(train_data):
X_batch = X_batch.float()
y_batch = y_batch.long()
output = model(X_batch)
output = torch.squeeze(output, 0)
loss = criterion(output, y_batch)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if iter % print_every == 0:
# print('Iter {} loss: {}'.format(iter, loss.item()))
f1_val, f2_val, precision_val, recall_val, accuracy_val = check_accuracy(
model, validation_data, False)
f1_train, f2_train, precision_train, recall_train, accuracy_train = check_accuracy(
model, train_data, False)
train_loss = loss.item()
metrics.append(
'{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format(
train_loss, f1_val, f2_val, precision_train,
recall_val, accuracy_val, f1_train, f2_train,
precision_train, recall_train, accuracy_train))
if f2_val > max_val_f2_score:
max_val_f2_score = f2_val
best_model = copy.deepcopy(model)
print('finished training!')
return best_model, max_val_f2_score, metrics
def apply_notl(trainX, trainY, testX, testY, window, source_pos, target_pos):
#######################
### SEMI-SUPERVISED ###
########################
# Label Propagation
label_prop_model = LabelPropagation(kernel='knn')
label_prop_model.fit(trainX, trainY)
Y_Pred = label_prop_model.predict(testX);
acc_ss_propagation, acc_ss_propagation_INFO = check_accuracy(testY, Y_Pred)
# Label Spreading
label_prop_models_spr = LabelSpreading(kernel='knn')
label_prop_models_spr.fit(trainX, trainY)
Y_Pred = label_prop_models_spr.predict(testX);
acc_ss_spreading, acc_ss_spreading_INFO = check_accuracy(testY, Y_Pred)
########################
#### WITHOUT TL ########
########################
# LogisticRegression
modelLR = LogisticRegression()
modelLR.fit(trainX, trainY)
predLR = modelLR.predict(testX)
accLR, acc_LR_INFO = check_accuracy(testY, predLR)
# DecisionTreeClassifier
modelDT = tree.DecisionTreeClassifier()
modelDT.fit(trainX, trainY)
predDT = modelDT.predict(testX)
accDT, acc_DT_INFO = check_accuracy(testY, predDT)
# BernoulliNB
modelNB = BernoulliNB()
modelNB.fit(trainX, trainY)
predND = modelNB.predict(testX)
accNB, acc_NB_INFO = check_accuracy(testY, predND)
#
return pd.DataFrame(
[{
'window': window,
'source_position': source_pos,
'target_position': target_pos,
'acc_SS_propagation': acc_ss_propagation,
'acc_SS_propagation_INFO':acc_ss_propagation_INFO,
'acc_SS_spreading': acc_ss_spreading,
'acc_SS_spreading_INFO':acc_ss_spreading_INFO,
'acc_LR':accLR,
'acc_LR_INFO': str(acc_LR_INFO),
'acc_DT': accDT,
'acc_DT_INFO': str(acc_DT_INFO),
'acc_NB': accNB,
'acc_NB_INFO': str(acc_NB_INFO)
}]
)
def main():
train_transforms = A.Compose([
A.Resize(height=IMAGE_HEIGHT, width=IMAGE_WIDTH),
A.Rotate(limit=35, p=1.0),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.1),
A.Normalize(mean=[0.0, 0.0, 0.0],
std=[1.0, 1.0, 1.0],
max_pixel_value=255.0),
ToTensorV2()
])
val_transforms = A.Compose([
A.Resize(height=IMAGE_HEIGHT, width=IMAGE_WIDTH),
A.Normalize(mean=[0.0, 0.0, 0.0],
std=[1.0, 1.0, 1.0],
max_pixel_value=255.0),
ToTensorV2()
])
model = UNET(in_channels=3, out_channels=1).to(DEVICE)
loss_fn = nn.BCEWithLogitsLoss()
loss_fn = ComboLoss({'bce': 0.4, 'dice': 0.5, 'focal': 0.1})
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
train_loader, val_loader = get_loaders(TRAIN_IMG_DIR, TRAIN_MASK_DIR,
VAL_IMG_DIR, VAL_MASK_DIR,
BATCH_SIZE, train_transforms,
val_transforms)
if LOAD_MODEL:
load_checkpoint(torch.load('checkpoint.pth'), model)
scaler = torch.cuda.amp.GradScaler()
for epoch in range(NUM_EPOCHS):
train_fn(train_loader, model, optimizer, loss_fn, scaler)
#save model
checkpoint = {
"state_dict": model.state_dict(),
"optimizer": optimizer.state_dict()
}
# save_checkpoint(checkpoint)
#check accuracy
check_accuracy(val_loader, model, DEVICE)
#print some example to a folder
save_predictions_as_imgs(val_loader, model, device=DEVICE)
def train_test_classifier():
w = ut.train_model(x_train, y_train, x_control_train, loss_function,
apply_fairness_constraints,
apply_accuracy_constraint, sep_constraint,
sensitive_attrs, sensitive_attrs_to_cov_thresh,
gamma)
train_score, test_score, correct_answers_train, correct_answers_test = ut.check_accuracy(
w, x_train, y_train, x_test, y_test, None, None)
distances_boundary_test = (np.dot(x_test, w)).tolist()
all_class_labels_assigned_test = np.sign(distances_boundary_test)
correlation_dict_test = ut.get_correlations(
None, None, all_class_labels_assigned_test, x_control_test,
sensitive_attrs)
cov_dict_test = ut.print_covariance_sensitive_attrs(
None, x_test, distances_boundary_test, x_control_test,
sensitive_attrs)
p_rule = ut.print_classifier_fairness_stats([test_score],
[correlation_dict_test],
[cov_dict_test],
sensitive_attrs[0])
eq_op_acc, chance_bin_zero, chance_bin_one = ut.get_eq_op_acc(
w, x_train, y_train, x_control_train, None)
eq_odds_acc = ut.get_eq_odds_acc(w, x_train, y_train, x_control_train,
None)
pred_rate_par_acc = ut.get_pred_rate_par_acc(w, x_train, y_train,
x_control_train, None)
demo_par_acc_f_cons = ut.get_dem_par_acc(w, x_train, y_train,
x_control_train, None)
return w, p_rule, test_score, eq_op_acc, eq_odds_acc, pred_rate_par_acc, demo_par_acc_f_cons
def train(self, X, y, x_sensitive, fairness_constraint):
self.x_sensitive = {"s1": x_sensitive}
self.X = ut.add_intercept(X)
self.y = y
if fairness_constraint == -1.0:
self.w = ut.train_model(self.X, self.y, self.x_sensitive,
lf._logistic_loss, 0, 0, 0, ["s1"],
{"s1": 0}, None)
else:
self.w = ut.train_model(self.X, self.y, self.x_sensitive,
lf._logistic_loss, 1, 0, 0, ["s1"],
{"s1": fairness_constraint}, None)
train_score, test_score, correct_answers_train, correct_answers_test = ut.check_accuracy(
self.w, self.X, self.y, self.X, self.y, None, None)
distances_boundary_test = (np.dot(self.X, self.w)).tolist()
all_class_labels_assigned_test = np.sign(distances_boundary_test)
correlation_dict_test = ut.get_correlations(
None, None, all_class_labels_assigned_test, self.x_sensitive,
["s1"])
correlation_dict = ut.get_avg_correlation_dict([correlation_dict_test])
non_prot_pos = correlation_dict["s1"][1][1]
prot_pos = correlation_dict["s1"][0][1]
p_rule = (prot_pos / non_prot_pos) * 100.0
return self.w, p_rule, 100.0 * test_score
def train(self, X, y, x_sensitive, fairness_constraint):
self.x_sensitive = {"s1": x_sensitive}
self.X = ut.add_intercept(X)
self.y = y
if fairness_constraint==-1.0:
self.w = ut.train_model(self.X, self.y, self.x_sensitive, lf._logistic_loss, 0, 0, 0,
["s1"], {"s1":0}, None)
else:
self.w = ut.train_model(self.X, self.y, self.x_sensitive, lf._logistic_loss, 1, 0, 0,
["s1"], {"s1": fairness_constraint}, None)
train_score, test_score, correct_answers_train, correct_answers_test = ut.check_accuracy(self.w,
self.X, self.y, self.X, self.y, None, None)
distances_boundary_test = (np.dot(self.X, self.w)).tolist()
all_class_labels_assigned_test = np.sign(distances_boundary_test)
correlation_dict_test = ut.get_correlations(None, None,
all_class_labels_assigned_test, self.x_sensitive, ["s1"])
correlation_dict = ut.get_avg_correlation_dict([correlation_dict_test])
non_prot_pos = correlation_dict["s1"][1][1]
prot_pos = correlation_dict["s1"][0][1]
p_rule = (prot_pos / non_prot_pos) * 100.0
return self.w, p_rule, 100.0*test_score
def main():
args = ParseArgs()
model = args.model
dataset_name = args.dataset_name
lambda_ = args.lambda_
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
trainloader, testloader, num_classes = Dataset(dataset_name)
model = torch.load(args.ckpt).to(device)
weights = [w for name, w in model.named_parameters() if "weight" in name]
num_features = sum([w.numel() for w in weights])
criterion = torch.nn.CrossEntropyLoss()
F, f, norm_l1_x = compute_F(
trainloader, model, weights, criterion, lambda_)
density = sum([torch.sum(w != 0).item() for w in weights]) / num_features
accuracy = check_accuracy(model, testloader)
print('F:', F)
print('f:', f)
print('density:', density)
print('validation accuracy:', accuracy)
def get_clf_stats(w, x_train, y_train, x_control_train, x_test, y_test, x_control_test, sensitive_attrs):
# compute distances to boundaries
distances_boundary_train = get_distance_boundary(w, x_train, None)
distances_boundary_test = get_distance_boundary(w, x_test, None)
# compute the class labels
all_class_labels_assigned_train = np.sign(distances_boundary_train)
all_class_labels_assigned_test = np.sign(distances_boundary_test)
train_score, test_score, correct_answers_train, correct_answers_test = ut.check_accuracy(None, x_train, y_train, x_test, y_test, all_class_labels_assigned_train, all_class_labels_assigned_test)
cov_all_train = {}
cov_all_test = {}
print "\n"
print "Overall Accuracy: %0.3f" % (test_score)
for s_attr in sensitive_attrs:
s_attr_to_fp_fn_train = get_fpr_fnr_sensitive_features(y_train, all_class_labels_assigned_train, x_control_train, sensitive_attrs, False)
cov_all_train[s_attr] = get_sensitive_attr_constraint_fpr_fnr_cov(None, x_train, y_train, distances_boundary_train, x_control_train[s_attr])
print_stats = True # only print stats for the test fold
s_attr_to_fp_fn_test = get_fpr_fnr_sensitive_features(y_test, all_class_labels_assigned_test, x_control_test, sensitive_attrs, False)
cov_all_test[s_attr] = get_sensitive_attr_constraint_fpr_fnr_cov(None, x_test, y_test, distances_boundary_test, x_control_test[s_attr])
return train_score, test_score, cov_all_train, cov_all_test, s_attr_to_fp_fn_train, s_attr_to_fp_fn_test
def part_loss(output, target, loss_f, join_channels=slice(0, 3), normal_channels=slice(3, None), train_reduced=False):
output_normal = output[:, normal_channels]
target_normal = target[:, normal_channels]
max_output, _ = output[:, join_channels].max(1, keepdim=True)
output = torch.cat((output_normal, max_output), dim=1)
max_target, _ = target[:, join_channels].max(1, keepdim=True)
target = torch.cat((target_normal, max_target), dim=1)
acc = check_accuracy(output, target)
return (loss_f(output, target) if train_reduced else loss_f(output_normal, target_normal)), acc
def train_test_classifier(): w = ut.train_model(x_train, y_train, x_control_train, loss_function, apply_fairness_constraints, apply_accuracy_constraint, sep_constraint, sensitive_attrs, sensitive_attrs_to_cov_thresh, gamma) train_score, test_score, correct_answers_train, correct_answers_test = ut.check_accuracy(w, x_train, y_train, x_test, y_test, None, None) distances_boundary_test = (np.dot(x_test, w)).tolist() all_class_labels_assigned_test = np.sign(distances_boundary_test) correlation_dict_test = ut.get_correlations(None, None, all_class_labels_assigned_test, x_control_test, sensitive_attrs) cov_dict_test = ut.print_covariance_sensitive_attrs(None, x_test, distances_boundary_test, x_control_test, sensitive_attrs) p_rule = ut.print_classifier_fairness_stats([test_score], [correlation_dict_test], [cov_dict_test], sensitive_attrs[0]) return w, p_rule, test_score
def get_clf_stats(w, x_train, y_train, x_control_train, x_test, y_test,
x_control_test, sensitive_attrs):
assert (len(sensitive_attrs) == 1
) # ensure that we have just one sensitive attribute
sensitive_attrs = list(sensitive_attrs)
s_attr = sensitive_attrs[
0] # for now, lets compute the accuracy for just one sensitive attr
# compute distance from boundary
distances_boundary_train = get_distance_boundary(w, x_train,
x_control_train[s_attr])
distances_boundary_test = get_distance_boundary(w, x_test,
x_control_test[s_attr])
# compute the class labels
all_class_labels_assigned_train = np.sign(distances_boundary_train)
all_class_labels_assigned_test = np.sign(distances_boundary_test)
train_score, test_score, correct_answers_train, correct_answers_test = ut.check_accuracy(
None, x_train, y_train, x_test, y_test,
all_class_labels_assigned_train, all_class_labels_assigned_test)
cov_all_train = {}
cov_all_test = {}
for s_attr in sensitive_attrs:
print_stats = False # we arent printing the stats for the train set to avoid clutter
# uncomment these lines to print stats for the train fold
# print "*** Train ***"
# print "Accuracy: %0.3f" % (train_score)
# print_stats = True
s_attr_to_fp_fn_train = get_fpr_fnr_sensitive_features(
y_train, all_class_labels_assigned_train, x_control_train,
sensitive_attrs, print_stats)
cov_all_train[s_attr] = get_sensitive_attr_constraint_fpr_fnr_cov(
None, x_train, y_train, distances_boundary_train,
x_control_train[s_attr])
#print("\n")
#print("Accuracy: %0.3f" % (test_score))
print_stats = False # only print stats for the test fold
s_attr_to_fp_fn_test = get_fpr_fnr_sensitive_features(
y_test, all_class_labels_assigned_test, x_control_test,
sensitive_attrs, print_stats)
cov_all_test[s_attr] = get_sensitive_attr_constraint_fpr_fnr_cov(
None, x_test, y_test, distances_boundary_test,
x_control_test[s_attr])
#print("\n")
return train_score, test_score, cov_all_train, cov_all_test, s_attr_to_fp_fn_train, s_attr_to_fp_fn_test
def main():
model = UNET(in_channels=3, out_channels=1).to(config.DEVICE)
BCE = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters(), lr=config.LEARNING_RATE)
train_loader, val_loader = get_loaders(
train_dir=config.TRAIN_IMG_DIR,
train_mask_dir=config.TRAIN_MASK_DIR,
val_dir=config.VAL_IMG_DIR,
val_mask_dir=config.VAL_MASK_DIR,
batch_size=config.BATCH_SIZE,
train_transform=config.train_transform,
val_transform=config.val_transform,
num_workers=config.NUM_WORKERS,
pin_memory=config.PIN_MEMORY,
)
if config.LOAD_MODEL:
load_checkpoint(torch.load(config.CHECKPOINT_PTH), model)
check_accuracy(val_loader, model)
scaler = torch.cuda.amp.GradScaler()
for epoch in range(config.NUM_EPOCHS):
train_fn(train_loader, model, optimizer, BCE, scaler, val_loader)
# save model
if config.SAVE_MODEL:
checkpoint = {
"state_dict": model.state_dict(),
"optimizer": optimizer.state_dict(),
}
save_checkpoint(checkpoint)
# check accuracy
check_accuracy(val_loader, model)
# print some example
save_predictions_as_imgs(val_loader, model, folder=config.SAVE_IMAGES)
def train_test_classifier():
w = ut.train_model(x_train, y_train, x_control_train, loss_function, apply_fairness_constraints, apply_accuracy_constraint, sep_constraint, sensitive_attrs, sensitive_attrs_to_cov_thresh, gamma)
train_score, test_score, correct_answers_train, correct_answers_test = ut.check_accuracy(w, x_train, y_train, x_test, y_test, None, None)
distances_boundary_test = np.dot(x_test, w)
distances_boundary_train = np.dot(x_train, w)
prob_test = [sigmoid(x) for x in distances_boundary_test]
prob_train = [sigmoid(x) for x in distances_boundary_train]
all_class_labels_assigned_test = np.sign(distances_boundary_test)
correlation_dict_test = ut.get_correlations(None, None, all_class_labels_assigned_test, x_control_test, sensitive_attrs)
cov_dict_test = ut.print_covariance_sensitive_attrs(None, x_test, distances_boundary_test, x_control_test, sensitive_attrs)
p_rule = ut.print_classifier_fairness_stats([test_score], [correlation_dict_test], [cov_dict_test], sensitive_attrs[0])
# return w, p_rule, test_score
return prob_train, prob_test
def main():
model = EfficientNet.from_pretrained("efficientnet-b7")
model._fc = nn.Linear(2560, 1)
train_dataset = CatDog(root="data/train/", transform=config.basic_transform)
test_dataset = CatDog(root="data/test/", transform=config.basic_transform)
train_loader = DataLoader(
train_dataset,
shuffle=True,
batch_size=config.BATCH_SIZE,
num_workers=config.NUM_WORKERS,
pin_memory=True,
)
test_loader = DataLoader(
test_dataset,
shuffle=False,
batch_size=config.BATCH_SIZE,
num_workers=config.NUM_WORKERS,
)
model = model.to(config.DEVICE)
scaler = torch.cuda.amp.GradScaler()
loss_fn = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(
model.parameters(), lr=config.LEARNING_RATE, weight_decay=config.WEIGHT_DECAY
)
if config.LOAD_MODEL and config.CHECKPOINT_FILE in os.listdir():
load_checkpoint(torch.load(config.CHECKPOINT_FILE), model)
for epoch in range(config.NUM_EPOCHS):
train_one_epoch(train_loader, model, loss_fn, optimizer, scaler)
check_accuracy(train_loader, model, loss_fn)
if config.SAVE_MODEL:
checkpoint = {"state_dict": model.state_dict(), "optimizer": optimizer.state_dict()}
save_checkpoint(checkpoint, filename=config.CHECKPOINT_FILE)
save_feature_vectors(model, train_loader, output_size=(1, 1), file="train_b7")
save_feature_vectors(model, test_loader, output_size=(1, 1), file="test_b7")
def apply_TCA(trainX, trainY, testX, testY, window, source_pos, target_pos):
####################################################
### Transfer Component Analysis (Pan et al, 2009)
###
# Decision Tree
print("\n Transfer Component Analysis (Pan et al, 2009)")
classifier = TransferComponentClassifier(loss="dtree", num_components=1)
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_DT_TCA, acc_DT_TCA_INFO = check_accuracy(testY, pred_naive)
# Logistic Regression
classifier = TransferComponentClassifier(loss="logistic", num_components=100)
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_LR_TCA, acc_LR_TCA_INFO = check_accuracy(testY, pred_naive)
# Naive Bayes Bernoulli
classifier = TransferComponentClassifier(loss="berno", num_components=1, l2=100, bandwidth=10, order=100.0, mu=100.0)
classifier.fit(trainX, trainY, testX)
pred_naive = classifier.predict(testX)
acc_NB_TCA, acc_NB_TCA_INFO = check_accuracy(testY, pred_naive)
return pd.DataFrame(
[{
'window': window,
'source_position': source_pos,
'target_position': target_pos,
'acc_LR_TCA': acc_LR_TCA,
'acc_LR_TCA_INFO': str(acc_LR_TCA_INFO),
'acc_DT_TCA': acc_DT_TCA,
'acc_DT_TCA_INFO': str(acc_DT_TCA_INFO),
'acc_NB_TCA': acc_NB_TCA,
'acc_NB_TCA_INFO': str(acc_NB_TCA_INFO),
}]
)
def main():
train_transform = tr.Compose([
tr.Resize((160, 240)),
tr.ToTensor(),
tr.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
model = UNET(in_channels=3, out_channels=3).to(DEVICE)
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
train_loader, val_loader = get_loaders(
TRAIN_IMG_DIR,
TRAIN_BLUR_DIR,
BATCH_SIZE,
train_transform=train_transform,
num_workers=NUM_WORKERS,
pin_memory=PIN_MEMORY,
)
if LOAD_MODEL:
load_checkpoint(torch.load("my_checkpoint.pth.tar"), model)
check_accuracy(val_loader, model, device=DEVICE)
scaler = torch.cuda.amp.GradScaler()
for epoch in range(NUM_EPOCHS):
train_model(train_loader, model, optimizer, loss_fn, scaler)
checkpoint = {
"state_dict": model.state_dict(),
"optimizer": optimizer.state_dict(),
}
save_checkpoint(checkpoint)
check_accuracy(val_loader, model, device=DEVICE)
def apply_ENSEMBLE(trainX, trainY, testX, testY, window, source_pos, target_pos):
classifier_SA_DT = SubspaceAlignedClassifier(loss="dtree")
classifier_SA_LR = SubspaceAlignedClassifier(loss="logistic")
classifier_SA_NB = SubspaceAlignedClassifier(loss="berno")
classifier_TCA_DT = TransferComponentClassifier(loss="dtree")
classifier_TCA_LR = TransferComponentClassifier(loss="logistic")
classifier_TCA_NB = TransferComponentClassifier(loss="berno")
classifier_NN_DT = ImportanceWeightedClassifier(iwe='nn', loss="dtree")
classifier_NN_LR = ImportanceWeightedClassifier(iwe='nn', loss="logistic")
classifier_NN_NB = ImportanceWeightedClassifier(iwe='nn', loss="berno")
classifier_KMM_DT = ImportanceWeightedClassifier(iwe='kmm', loss="dtree")
classifier_KMM_LR = ImportanceWeightedClassifier(iwe='kmm', loss="logistic")
classifier_KMM_NB = ImportanceWeightedClassifier(iwe='kmm', loss="berno")
#
eclf = EnsembleClassifier(clfs=[
#classifier_SA_DT,
#classifier_SA_LR,
#classifier_SA_NB,
#classifier_TCA_DT,
#classifier_TCA_LR,
classifier_TCA_NB,
classifier_NN_DT,
#classifier_NN_LR,
#classifier_NN_NB,
classifier_KMM_DT,
classifier_KMM_LR,
#classifier_KMM_NB
])
eclf.fit(trainX, trainY, testX)
pred = eclf.predict(testX)
acc_ENSEMBLE, acc_ENSEMBLE_INFO = check_accuracy(testY, pred)
#
return pd.DataFrame(
[{
'window': window,
'source_position': source_pos,
'target_position': target_pos,
'acc_ENSEMBLE': acc_ENSEMBLE,
'acc_ENSEMBLE_INFO': acc_ENSEMBLE_INFO,
}]
)
def get_clf_stats(w, x_train, y_train, x_control_train, x_test, y_test, x_control_test, sensitive_attrs):
assert(len(sensitive_attrs) == 1) # ensure that we have just one sensitive attribute
s_attr = sensitive_attrs[0] # for now, lets compute the accuracy for just one sensitive attr
# compute distance from boundary
distances_boundary_train = get_distance_boundary(w, x_train, x_control_train[s_attr])
distances_boundary_test = get_distance_boundary(w, x_test, x_control_test[s_attr])
# compute the class labels
all_class_labels_assigned_train = np.sign(distances_boundary_train)
all_class_labels_assigned_test = np.sign(distances_boundary_test)
train_score, test_score, correct_answers_train, correct_answers_test = ut.check_accuracy(None, x_train, y_train, x_test, y_test, all_class_labels_assigned_train, all_class_labels_assigned_test)
cov_all_train = {}
cov_all_test = {}
for s_attr in sensitive_attrs:
print_stats = False # we arent printing the stats for the train set to avoid clutter
# uncomment these lines to print stats for the train fold
# print "*** Train ***"
# print "Accuracy: %0.3f" % (train_score)
# print_stats = True
s_attr_to_fp_fn_train = get_fpr_fnr_sensitive_features(y_train, all_class_labels_assigned_train, x_control_train, sensitive_attrs, print_stats)
cov_all_train[s_attr] = get_sensitive_attr_constraint_fpr_fnr_cov(None, x_train, y_train, distances_boundary_train, x_control_train[s_attr])
print "\n"
print "Accuracy: %0.3f" % (test_score)
print_stats = True # only print stats for the test fold
s_attr_to_fp_fn_test = get_fpr_fnr_sensitive_features(y_test, all_class_labels_assigned_test, x_control_test, sensitive_attrs, print_stats)
cov_all_test[s_attr] = get_sensitive_attr_constraint_fpr_fnr_cov(None, x_test, y_test, distances_boundary_test, x_control_test[s_attr])
print "\n"
return train_score, test_score, cov_all_train, cov_all_test, s_attr_to_fp_fn_train, s_attr_to_fp_fn_test
def main():
parser = ArgumentParser(description='Validate model on .png files')
parser.add_argument('-m',
type=str,
required=False,
default=None,
help='Model to validate')
parser.add_argument('input_dir', type=str, help='Input directory')
parser.add_argument('target_dir', type=str, help='Output directory')
args = parser.parse_args()
device = get_device()
model, _, _ = load_model(load_config().model if args.m is None else args.m,
device=device)
loss_f = BCELoss()
count, loss_sum, acc_sum = 0, 0, 0
with torch.no_grad():
for fn in listdir(args.input_dir):
input_path = join(args.input_dir, fn)
target_path = join(args.target_dir, fn)
if fn.endswith('.png') and isfile(input_path) and isfile(
target_path):
print('Validating %s with target %s' %
(input_path, target_path))
data = image_to_tensor(imread(input_path), device=device)
target = image_to_probs(imread(target_path), device=device)
data = model(data).squeeze(0)
loss = loss_f(data, target).item()
acc = check_accuracy(data, target)
print('Loss %f, acc %s' % (loss, acc_to_str(acc)))
count += 1
acc_sum += acc
loss_sum += loss
print('\nSUMMARY\nLoss %f, acc %s' %
(loss_sum / count, acc_to_str(acc_sum)))
print(acc_to_details(acc_sum))
def main(train_dir,
val_dir,
checkpoint_dir,
batch_size,
image_size=512,
num_epochs=10,
checkpoint_name=None,
num_workers=1,
pin_memory=True,
log_dir="logs",
model_name=None,
train_csv=None,
val_csv=None):
# declare datasets
train_ds = DataFolder(root_dir=train_dir,
transform=transform(image_size, is_training=True),
csv_path=train_csv)
val_ds = DataFolder(root_dir=val_dir,
transform=transform(image_size, is_training=False),
csv_path=val_csv)
train_loader = DataLoader(train_ds,
batch_size=batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
shuffle=True)
val_loader = DataLoader(val_ds,
batch_size=batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
shuffle=True)
#init model
model = MainModel(128, model_name)
# configure parameter
loss_fn = nn.CrossEntropyLoss()
model = model.to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-4)
scaler = torch.cuda.amp.GradScaler()
# checkpoint = {'state_dict': model.state_dict(), 'optimizer': optimizer.state_dict()}
# save_checkpoint(checkpoint, os.path.join(checkpoint_dir, f"checkpoint_initialilze.pth.tar"))
# return
if checkpoint_name:
ckp_path = os.path.join(checkpoint_dir, checkpoint_name)
load_checkpoint(torch.load(ckp_path), model, optimizer)
check_accuracy(val_loader, model, device)
#training
for epoch in range(num_epochs):
train_fn(train_loader,
model,
optimizer,
loss_fn,
scaler,
device,
epoch,
log_dir=log_dir)
check_accuracy(val_loader, model, device)
checkpoint = {
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_checkpoint(
checkpoint,
os.path.join(checkpoint_dir, f"checkpoint_{epoch}.pth.tar"))
x.resize_(x.size()[0], 1, x.size()[1])
x, y = x.float(), y.long()
x, y = x.to(device), y.to(device)
# loss and predictions
scores = model(x)
loss = loss_fn(scores, y)
writer.add_scalar('loss', loss.item())
# print and save loss per 'print_every' times
if (t + 1) % opt.print_every == 0:
print('t = %d, loss = %.4f' % (t + 1, loss.item()))
# parameters update
optimizer.zero_grad()
loss.backward()
optimizer.step()
# save epoch loss and acc to train or val history
train_acc, _ = check_accuracy(model, tr_loader, device)
val_acc, _ = check_accuracy(model, val_loader, device)
# writer acc and weight to tensorboard
writer.add_scalars('acc', {
'train_acc': train_acc,
'val_acc': val_acc
}, epoch)
for name, param in model.named_parameters():
writer.add_histogram(name, param.clone().cpu().data.numpy(), epoch)
# save the best model
if val_acc > best_acc:
best_acc = val_acc
best_model_wts = copy.deepcopy(model.state_dict())
t1 = time.time()
print(t1 - t0)
# print results
def main():
# TODO: Might be worth trying the normalization from assignment 2
train_transform = A.Compose([
A.Resize(height=IMAGE_HEIGHT, width=IMAGE_WIDTH),
A.Rotate(limit=35, p=1.0),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.1),
A.Normalize(
mean=[0.0, 0.0, 0.0],
std=[1.0, 1.0, 1.0],
max_pixel_value=255.0,
),
ToTensorV2(),
], )
val_transforms = A.Compose([
A.Resize(height=IMAGE_HEIGHT, width=IMAGE_WIDTH),
A.Normalize(
mean=[0.0, 0.0, 0.0],
std=[1.0, 1.0, 1.0],
max_pixel_value=255.0,
),
ToTensorV2(),
], )
model = UNET(in_channels=3, out_channels=1).to(DEVICE)
"""
We're using with logitsLoss because we're not using sigmoid on the,
final output layer.
If we wanted to have several output channels, we'd change the loss_fn
to a cross entropy loss instead.
"""
loss_fn = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
train_loader, val_loader = get_loaders(
TRAIN_IMG_DIR,
TRAIN_MASK_DIR,
VAL_IMG_DIR,
VAL_MASK_DIR,
BATCH_SIZE,
train_transform,
val_transforms,
NUM_WORKERS,
PIN_MEMORY,
)
if LOAD_MODEL:
load_checkpoint(torch.load("my_checkpoint.pth.tar"), model)
scaler = torch.cuda.amp.GradScaler(
) # Scales the gradients to avoid underflow. Requires a GPU
for epoch in range(NUM_EPOCHS):
train_fn(train_loader, model, optimizer, loss_fn, scaler)
# save model
checkpoint = {
"state_dict": model.state_dict(),
"optimizer": optimizer.state_dict(),
}
save_checkpoint(checkpoint)
# check accuracy
check_accuracy(val_loader, model, device=DEVICE)
# print some examples to a folder
save_predictions_as_imgs(val_loader,
model,
folder="saved_images/",
device=DEVICE)
trained_models = [] # testing only
for file_name in files_out:
f = deepcopy(file_name)
f += ".pkl"
model = model_fromPickle(f)
if model is not None:
trained_models.append(model) # testing only
f = deepcopy(file_name)
f += ".csv"
output_files(model, ds1Test, f)
########## testing only #####################################
for index, model in enumerate(trained_models):
accuracy = check_accuracy(model, set1_xTest, set1_yTest)
print("Model {}: {}".format(index, accuracy))
##############################################################
# create dict to store some meta-data on params and performance after tuning in case we want access
# later for summary statistics
best = {'naive bayes': {}, 'decision trees': {}, 'random forests': {}}
# Get classifier and meta data on Naive Bayes for ds1
bern, best['naive bayes'] = bernoulli(set1_xTrain,
set1_yTrain,
set1_xTest,
set1_yTest,
data_set1=True,
combined=combined_ds1)
output_files(
def main():
args = parseArgs()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
trainloader, testloader, num_classes = Dataset(
args.dataset_name, args.batch_size)
model = selectModel(args.model)
model = model.to(device)
weights = [w for name, w in model.named_parameters() if "weight" in name]
num_features = sum([w.numel() for w in weights])
num_samples = len(trainloader) * trainloader.batch_size
criterion = nn.CrossEntropyLoss()
if args.optimizer == 'obproxsg':
optimizer = OBProxSG(model.parameters(), lr=args.learning_rate,
lambda_=args.lambda_, epochSize=len(trainloader), Np=5, No=5)
elif args.optimizer == 'obproxsg_plus':
optimizer = OBProxSG(model.parameters(), lr=args.learning_rate,
lambda_=args.lambda_, epochSize=len(trainloader), Np=int(args.max_epoch/10))
elif args.optimizer == 'proxsg':
optimizer = ProxSG(model.parameters(),
lr=args.learning_rate, lambda_=args.lambda_)
elif args.optimizer == 'proxsvrg':
optimizer = ProxSVRG(model.parameters(), lr=args.learning_rate,
lambda_=args.lambda_, epochSize=len(trainloader))
elif args.optimizer == 'rda':
optimizer = RDA(model.parameters(), lambda_=args.lambda_, gamma=20, epochSize=len(trainloader))
if args.optimizer != 'rda':
scheduler = StepLR(optimizer, step_size=60, gamma=0.1)
os.makedirs('results', exist_ok=True)
setting = '%s_%s_%s_%E' % (
args.optimizer, args.model, args.dataset_name, args.lambda_)
csvname = os.path.join('results', setting + '.csv')
print('Results are saving to the CSV file: %s.' % csvname)
csvfile = open(csvname, 'w', newline='')
fieldnames = ['epoch', 'F_value', 'f_value', 'norm_l1_x',
'density', 'validation_acc', 'train_time']
writer = csv.DictWriter(csvfile, fieldnames=fieldnames, delimiter=",")
writer.writeheader()
alg_start_time = time.time()
epoch = 0
while True:
epoch_start_time = time.time()
if epoch >= args.max_epoch:
break
if args.optimizer == 'proxsvrg':
optimizer.zero_grad()
for index, (X, y) in enumerate(trainloader):
X = X.to(device)
y = y.to(device)
y_pred = model.forward(X)
f1 = criterion(y_pred, y)
f1.backward()
optimizer.init_epoch()
for index, (X, y) in enumerate(trainloader):
X = X.to(device)
y = y.to(device)
# calculate grad_f_i
optimizer.set_weights_as_xs()
y_pred = model.forward(X)
f = criterion(y_pred, y)
optimizer.zero_grad()
f.backward()
optimizer.save_grad_f()
# calculate grad_f_hat_i
optimizer.set_weights_as_hat_x()
y_pred = model.forward(X)
f = criterion(y_pred, y)
optimizer.zero_grad()
f.backward()
optimizer.save_grad_f_hat()
optimizer.update_xs()
optimizer.step()
else:
for _, (X, y) in enumerate(trainloader):
X = X.to(device)
y = y.to(device)
y_pred = model.forward(X)
f = criterion(y_pred, y)
optimizer.zero_grad()
f.backward()
optimizer.step()
if args.optimizer != 'rda':
scheduler.step()
epoch += 1
train_time = time.time() - epoch_start_time
F, f, norm_l1_x = compute_F(
trainloader, model, weights, criterion, args.lambda_)
density = sum([torch.sum(w != 0).item()
for w in weights]) / num_features
accuracy = check_accuracy(model, testloader)
writer.writerow({'epoch': epoch, 'F_value': F, 'f_value': f, 'norm_l1_x': norm_l1_x,
'density': density, 'validation_acc': accuracy, 'train_time': train_time})
csvfile.flush()
print("Epoch {}: {:2f}seconds ...".format(epoch, train_time))
alg_time = time.time() - alg_start_time
writer.writerow({'train_time': alg_time / epoch})
os.makedirs('checkpoints', exist_ok=True)
torch.save(model, os.path.join('checkpoints', setting + '.pt'))
csvfile.close()
type=int,
default=3,
help='No of epochs to train the model')
parser.add_argument('--gpu',
default='gpu',
help='Determine GPU vs CPU for the neural network')
args = parser.parse_args()
data_dir = args.data_dir
save_dir = args.save_dir
arch = args.arch
learning_rate = args.learning_rate
hidden_units = args.hidden_units
epochs = args.epochs
gpu = args.gpu
if gpu == 'gpu':
is_gpu = True
else:
is_gpu = False
image_datasets, dataloaders = load_data(data_dir)
model = build_model(arch, hidden_units, is_gpu)
model = train_model(model, dataloaders['training'],
dataloaders['validation'], epochs, learning_rate,
is_gpu)
accuracy = check_accuracy(model, dataloaders['test'], is_gpu)
print("Accuracy: {:.2f}%".format(accuracy * 100))
save_checkpoint(model, hidden_units, epochs, image_datasets['training'],
save_dir)
y = y.to(config.DEVICE).view(-1).float()
# forward
scores = model(x).view(-1)
loss = loss_fn(scores, y)
losses.append(loss.item())
# backward
optimizer.zero_grad()
loss.backward()
# gradient descent or adam step
optimizer.step()
print(f"Loss: {sum(losses)/len(losses)}")
if config.SAVE_MODEL:
checkpoint = {
"state_dict": model.state_dict(),
"optimizer": optimizer.state_dict()
}
save_checkpoint(checkpoint, filename="linear.pth.tar")
preds, labels = check_accuracy(loader_val, model)
print(cohen_kappa_score(labels, preds, weights="quadratic"))
preds, labels = check_accuracy(loader, model)
print(cohen_kappa_score(labels, preds, weights="quadratic"))
make_prediction(model, loader_test, "test_preds.csv")
