def contrast_bn(epochs, lr):
train_loader_no_bn = DataLoader(64, data_type='train', scale=True)
valid_loader_no_bn = DataLoader(64, data_type='valid', scale=True)
test_loader_no_bn = DataLoader(64, data_type='test', scale=True)
model = Model(4, [28 * 28, 512, 512, 10],
initializer='xavier',
optimizer='sgd')
his_no_bn = model.train(train_loader_no_bn,
valid_loader_no_bn,
epochs,
learning_rate=lr)
pred, label = model.predict(test_loader_no_bn)
acc = np.sum(pred == label) / len(pred)
print('acc', acc)
cm = confusion_matrix(label.reshape(-1), pred.reshape(-1), 10)
cm_plot(cm, 'no_bn_cm, acc {:.3f}'.format(acc))
train_loader_bn = DataLoader(64, data_type='train', scale=True)
valid_loader_bn = DataLoader(64, data_type='valid', scale=True)
test_loader_bn = DataLoader(64, data_type='test', scale=True)
model2 = Model(4, [28 * 28, 512, 512, 10],
initializer='xavier',
optimizer='sgd')
his_bn = model2.train_bn(train_loader_bn,
valid_loader_bn,
epochs,
learning_rate=lr)
pred, label = model2.predict(test_loader_bn, bn=True)
acc = np.sum(pred == label) / len(pred)
print('acc', acc)
cm = confusion_matrix(label.reshape(-1), pred.reshape(-1), 10)
cm_plot(cm, 'bn_cm, acc {:.3f}'.format(acc))
plot_history(his_no_bn, his_bn, 'no bn', 'bn',
'bn and no bn training loss', 'bn and no bn validation loss')
def contrast_scale(epochs, lr):
train_loader_no_scale = DataLoader(64, data_type='train', scale=False)
valid_loader_no_scale = DataLoader(64, data_type='valid', scale=False)
test_loader_no_scale = DataLoader(64, data_type='test', scale=False)
model = Model(4, [28 * 28, 512, 512, 10], initializer='xavier')
his_no_scale = model.train(train_loader_no_scale,
valid_loader_no_scale,
epochs,
learning_rate=lr)
pred, label = model.predict(test_loader_no_scale)
acc = np.sum(pred == label) / len(pred)
print('acc', acc)
cm = confusion_matrix(label.reshape(-1), pred.reshape(-1), 10)
cm_plot(cm, 'no_scale_cm, acc {:.3f}'.format(acc))
train_loader_scale = DataLoader(64, data_type='train', scale=True)
valid_loader_scale = DataLoader(64, data_type='valid', scale=True)
test_loader_scale = DataLoader(64, data_type='test', scale=True)
model2 = Model(4, [28 * 28, 512, 512, 10], initializer='xavier')
his_scale = model2.train(train_loader_scale,
valid_loader_scale,
epochs,
learning_rate=lr)
pred, label = model2.predict(test_loader_scale)
acc = np.sum(pred == label) / len(pred)
print('acc', acc)
cm = confusion_matrix(label.reshape(-1), pred.reshape(-1), 10)
cm_plot(cm, 'scale_cm, acc {:.3f}'.format(acc))
plot_history(his_no_scale, his_scale, 'no scale', 'scale',
'sacle and no scale training loss',
'scale an no scale validation loss')
def hierarchical(encoder,
tsne,
true_data,
true_labels,
save_name="hierarchical.png"):
"""
1. Predicts labels using hierarchical clustering
2. Prints confusion_matrix
3. Prints t-SNE plot of prediction
"""
enc_output = encoder.predict(true_data)
# Hierarchical Clustering
labels = HierarchicalClustering()
predictions = labels.draw_dendogram(
enc_output,
title='Hierarchical Clustering Dendrogram',
savetitle="hierarchical.png")
# Confusion matrix of hierarchical clustering
confusion_matrix(true_labels,
predictions,
save_name="confusion_matrix_hierarchical.png")
# Visualize test predictions from hierarchical
true_data = np.reshape(true_data, (len(true_data), 64, 64))
visualize_class_predictions(true_data, true_labels, predictions)
def contrast_dropout(epochs, lr):
train_loader_no_dropout = DataLoader(64, data_type='train', scale=True)
valid_loader_no_dropout = DataLoader(64, data_type='valid', scale=True)
test_loader_no_dropout = DataLoader(64, data_type='test', scale=True)
model = Model(4, [28 * 28, 512, 512, 10], initializer='xavier')
his_no_dropout = model.train(train_loader_no_dropout,
valid_loader_no_dropout,
epochs,
learning_rate=lr)
pred, label = model.predict(test_loader_no_dropout)
acc = np.sum(pred == label) / len(pred)
print('acc', acc)
cm = confusion_matrix(label.reshape(-1), pred.reshape(-1), 10)
cm_plot(cm, 'no_dropout_cm, acc {:.3f}'.format(acc))
train_loader_dropout = DataLoader(64, data_type='train', scale=True)
valid_loader_dropout = DataLoader(64, data_type='valid', scale=True)
test_loader_dropout = DataLoader(64, data_type='test', scale=True)
model2 = Model(4, [28 * 28, 512, 512, 10], initializer='xavier')
his_dropout = model2.train(train_loader_dropout,
valid_loader_dropout,
epochs,
learning_rate=lr,
dropout_prob=0.3)
pred, label = model2.predict(test_loader_dropout)
acc = np.sum(pred == label) / len(pred)
print('acc', acc)
cm = confusion_matrix(label.reshape(-1), pred.reshape(-1), 10)
cm_plot(cm, 'dropout_cm, acc {:.3f}'.format(acc))
plot_history(his_no_dropout, his_dropout, 'no dropout', 'dropout',
'dropout and no dropout training loss',
'dropout and no dropout validation loss')
def test_knn():
data, _ = utils.read_table("combined_data_normalized.csv", True)
class_index = 4
predictors = [2, 3, 5, 9]
results = knn.knn_classifier(data, class_index, predictors, 5, 5)
accuracy = utils.compute_accuracy(results)
print(accuracy)
utils.confusion_matrix(results, "Crime Rate?",
"KNN Classifier Prediction of Crime Rate")
def test_naive_bayes():
data, header = utils.read_table("combined_data_normalized.csv", True)
class_index = 4
predictors = [2, 3, 5, 7]
results = bayes.naive_bayes_classifier(data, header, 10, class_index,
predictors, [2, 3, 5, 9])
accuracy = utils.compute_accuracy(results)
print(accuracy)
utils.confusion_matrix(results, "Crime Rate?",
"Naive Bayes Classifier Prediction of Crime Rate")
def test_random_forest():
data, header = utils.read_table("combined_data_discretized.csv", True)
class_index = 4
predictors = [2, 3, 5, 9]
results = rforest.random_forest_classifier(data, header, class_index,
predictors, 100, 25, 3)
accuracy = utils.compute_accuracy(results)
print(accuracy)
utils.confusion_matrix(
results, "Crime Rate?",
"Random Forest Classifier Prediction of Crime Rate")
def test_decision_tree():
data, header = utils.read_table("combined_data_discretized.csv", True)
class_index = 4
predictors = [2, 3, 5, 9]
results = dtree.decision_tree_classifier(data, header, class_index,
predictors, 30)
accuracy = utils.compute_accuracy(results)
print(accuracy)
utils.confusion_matrix(
results, "Crime Rate?",
"Decision Tree Classifier Prediction of Crime Rate")
def logistic(Xtrain, Ytrain, Xdev, Ydev, verbose=False, scoring='f1'):
"""
Trains a Logist Regression Model on the provided data. Scores the model
and returns both the model and the score. It also prints the optimal
hyperparameters. 5-fold cross validation is performed to tune l1 loss ratio
and C (regularization weight).
Inputs:
Xtrain
Ytrain
Xdev
Ydev
Returns:
float: the F1 on the dev data for the best model specifications.
LogisticRegression: the best trained model.
"""
print("\n========================\nTraining Logistic Regression\n")
if scoring == 'f1':
scoring = metrics.make_scorer(metrics.f1_score, average='binary')
logit = LogisticRegressionCV(l1_ratios=[.1, .5, .7, .9, .95, .99, 1],
Cs=[0.1, 1, 10],
max_iter=1e4,
solver='saga',
scoring=scoring,
penalty='elasticnet')
logit.fit(Xtrain, Ytrain)
best_score = logit.score(Xdev, Ydev)
Ydev_pred = logit.predict(Xdev)
num_coeff = len(logit.coef_[logit.coef_ != 0])
results = {
"F1": best_score,
"l1_ratio": logit.l1_ratio_[0],
"C": logit.C_[0],
"n_nonzero_weights": num_coeff,
"accuracy": metrics.accuracy_score(Ydev, Ydev_pred),
"precision": metrics.precision_score(Ydev, Ydev_pred,
average='binary'),
"recall": metrics.recall_score(Ydev, Ydev_pred, average='binary')
}
try:
print(results)
except Exception as e:
print(f"Error occured printing results: {e}")
if verbose:
print(f"There are {num_coeff} non-zero weights in the logistic " +
"regression model.")
utils.confusion_matrix(Ydev, Ydev_pred)
utils.roc_auc(logit, Xdev, Ydev)
utils.precision_recall(logit, Xdev, Ydev)
return results, logit
def heat_map(prds_all, msks_all):
if 'grss' in opt.data_dir:
y_labels = [
'Road', 'Tree', 'Red roof', 'Grey roof', 'Concrete\nroof',
'Vegetation'
]
sr_heatmap = normalize_rows(
confusion_matrix(prds_all, msks_all, opt.num_classes))
fig = plt.figure(figsize=(6, 6))
ax = sns.heatmap(sr_heatmap,
linewidth=0.5,
cmap='Blues',
annot=True,
yticklabels=y_labels,
xticklabels=False)
fig.savefig('heat_maps/' + exp_name + '/' + 'sr.png')
elif 'coffee' in opt.data_dir:
y_labels = ['non-coffee', 'coffee']
sr_heatmap = normalize_rows(
confusion_matrix(prds_all, msks_all, opt.num_classes))
sns.set(font_scale=1.3)
fig = plt.figure(figsize=(3.5, 3.5))
ax = sns.heatmap(sr_heatmap,
linewidth=0.5,
cmap='Blues',
annot=True,
yticklabels=y_labels,
xticklabels=False)
fig.savefig('heat_maps/' + exp_name + '/' + 'sr.png')
elif 'vaihingen' in opt.data_dir or 'task_test' in opt.data_dir:
y_labels = [
'Impervious\nsurfaces', 'Building', 'Low\nvegetation', 'Tree',
'Car'
]
sr_heatmap = normalize_rows(
confusion_matrix(prds_all, msks_all, opt.num_classes))
sr_heatmap = np.delete(sr_heatmap, -1, axis=0)
sr_heatmap = np.delete(sr_heatmap, -1, axis=1)
fig = plt.figure(figsize=(5, 5))
ax = sns.heatmap(sr_heatmap,
linewidth=0.5,
cmap='Blues',
annot=True,
yticklabels=y_labels,
xticklabels=False)
fig.savefig('heat_maps/' + exp_name + '/' + 'sr.png')
def kmean(encoder, tsne, true_data, true_label):
"""
1. Predicts labels using k-means clustering
2. Prints confusion_matrix
3. Prints accuracy
4. Prints t-SNE plot of prediction
"""
enc_output = encoder.predict(true_data)
kmean = KMeansClustering()
kmean.fit(enc_output)
pred = kmean.predict(enc_output)
accuracy(true_label, pred)
confusion_matrix(true_label, pred, save_name="confusion_matrix_kmean.png")
tsne.tsne_plot(true_data, pred, save_data_dir="kmean", save_name="kmean")
def _pred(self, x_df, y_df=None, data='confusion_matrix', format='tensor'):
assert isinstance(x_df, pd.DataFrame)
if y_df is not None: assert isinstance(y_df, pd.DataFrame)
x_tensor = utils.get_tensor(x_df)
y_tensor = self.df_to_tensor(y_df)
y_pred_prob = self.model(Variable(x_tensor))
# y_pred = x_tensor.mm(w1).clamp(min=0).mm(w2)
if data == 'loss':
loss = self.loss_fn(y_pred_prob, Variable(y_tensor))
result = loss.data[0]
return result
elif data == 'pred':
data_tensor = y_pred_prob.data
else: # data == 'confusion_matrix':
y_pred, total_type_num = utils.max_ix(y_pred_prob.data)
# total_type_num = len(self.ix_to_label)
data_tensor = utils.confusion_matrix(y_pred, y_tensor,
total_type_num)
if format == 'df':
result = pd.DataFrame(data_tensor.numpy())
elif format == 'np':
result = data_tensor.numpy()
else: # format == 'tensor':
result = data_tensor
return result
def eval(args):
device = torch.device(f"cuda:{args.device_id}")
model = AlexNet(n_cls = 100)
model.to(device)
model.load_state_dict(torch.load(args.pretrained_path))
model.eval()
test_loader = getLoaders(split="eval", batch_size = args.batch_size, num_workers=args.num_workers )
pred_arr = []
label_arr = []
with torch.no_grad():
for idx, (img, label) in tqdm(enumerate(test_loader),total= len(test_loader)):
img = img.to(device)
pred = model.pred(img)
# mean of softmax prob from 10 different aug
pred = pred.view(-1, 10, 100)
pred = pred.mean(dim = 1)
pred_arr.append(pred.detach().cpu().numpy())
label_arr.append(label.detach().numpy())
pred_np = np.concatenate(pred_arr)
label_np = np.concatenate(label_arr)
top_1 = utils.top_k_acc(k = 1, pred = pred_np, label= label_np)
top_5 = utils.top_k_acc(k = 5, pred = pred_np, label= label_np)
confusion = utils.confusion_matrix(100, pred_np, label_np)
torch.save({
"top_1": top_1,
"top_5": top_5,
"confusion": confusion,
}, "result.pth")
print(f"top_1: {top_1*100:.2f}, top_5: {top_5*100:.2f}")
def measure(self, X, y, threshold=0.5):
y_hat = self.predict(X)
TP, FP, FN, TN = utils.confusion_matrix(threshold, y_hat, y)
precision = float(TP) / (TP + FP)
recall = float(TP) / (TP + FN)
F1 = 2 * precision * recall / (precision + recall)
return F1
def fix_dropout(self, new_weights, old_weights, mask_prev, mask_next=None):
"""
Update old weight with new dropout weights.
New weights don't have the same format as old ones, so updating them is pretty tricky.
:param new_weights: Weights of the Net after dropout
:param old_weights: Weights of the Net before dropout (Real weights)
:param mask_prev: Mask used on previous nodes
:param mask_next: Mask used on next nodes
:return: New weights for the real Net
"""
if mask_next is not None:
new_w = old_weights
if mask_next is not None:
conf_matrix = utils.confusion_matrix(mask_next, mask_prev)
# Compute confusion matrix containing indexes of weights to update
indexes = np.argwhere(conf_matrix)
c = 0
for i in range(len(new_weights)):
for j in range(len(new_weights[i])):
new_w[indexes[c][0], indexes[c][1]] = new_weights[i, j]
c += 1
else:
new_w = self.fix_dropout(new_weights, old_weights, mask_prev,
np.ones(old_weights.shape[1]))
return new_w
def _decode(self, x, x_, attention_mask, threshold=0.5):
mask = attention_mask == 1
y = x.masked_select(mask).cpu().long().numpy()
y_ = x_.masked_select(mask).cpu().numpy()
y_ = np.where(y_ > threshold, 1, 0)
# np.array vector float
return confusion_matrix(y, y_)
def valid(self, epoch, val_loader):
self.G.eval()
with torch.no_grad():
# (tn, fp, fn, tp)
cm = utils.ConfusionMatrix()
for i, (input_, target_, _) in enumerate(val_loader):
input_ = input_.to(self.torch_device)
output_ = self.G(input_)
target_ = target_.to(self.torch_device)
ground_truth = target_.int().squeeze(1)
prediction = torch.argmax(output_, dim=1).int()
cm.update(
utils.confusion_matrix(prediction,
ground_truth,
reduce=False))
metric = 1.5 * cm.f2 + cm.accuracy
if metric > self.best_metric:
self.best_metric = metric
self.save(epoch)
self.logger.write(
"[Val] epoch: %d accuracy: %f f05: %f f1: %f f2: %f" %
(epoch, cm.accuracy, cm.f05, cm.f1, cm.f2))
def validate(valloader, train_state):
model = train_state.ema_model
criterion = train_state.criterion
losses = AverageMeter()
accuracy_meter = AverageMeter()
# switch to evaluate mode
model.eval()
n_classes = len(train_state.class_names)
confusion = torch.zeros(n_classes, n_classes)
end = time.time()
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(valloader):
if constants['use_cuda']:
inputs, targets = inputs.cuda(), targets.cuda(
non_blocking=True)
# compute output
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
acc = accuracy(outputs, targets)
batch_confusion = confusion_matrix(outputs, targets)
confusion += batch_confusion
losses.update(loss.item(), inputs.size(0))
accuracy_meter.update(acc.item(), inputs.size(0))
return (losses.avg, accuracy_meter.avg, confusion)
def main():
graph = utils.load_graph()
position = utils.get_positions(graph)
utils.make_dir('images/louvain')
true_communities = utils.get_labels(graph, list(graph.nodes))
utils.plot_communities(graph, position, true_communities, labels=True, title='Butterfly Similarity Network - True Communities', path='images/louvain/communities_true.png')
communities = utils.group_communities(louvain_clustering(graph))
utils.plot_communities(graph, position, communities, labels=False, title='Butterfly Similarity Network - Louvain Communities', path='images/louvain/communities_louvain.png')
graph_nodes = sorted(list(graph.nodes))
predictions = utils.predict_majority_class(graph, communities)
preds = [predictions[n] for n in graph_nodes]
labels = [graph.nodes[n]['label'] for n in graph_nodes]
utils.accuracy(preds, labels)
utils.confusion_matrix(preds, labels, 'Confusion Matrix - Majority Label Predictions from Louvain Communities', 'images/louvain/cm_louvain.png')
def main(args):
# Load and standardize data.
embedding_file_path = 'node2vec/embeddings/' + args.input
X_train, X_test, y_train, y_test = load_splits(embedding_file_path)
X_train, X_test = standardize_data(X_train, X_test)
# Train classifier and make predictions.
optimal_svc = hyperparameter_search(SVC(), X_train, y_train)
print('Cross Validation Accuracy:', optimal_svc.best_score_)
print('Optimal parameters:', optimal_svc.best_params_)
predictions = optimal_svc.predict(X_test)
# Report results.
utils.make_dir('images/svc')
cm_path = 'images/svc/cm_' + args.input[:-4] + '.png'
utils.accuracy(predictions, y_test)
utils.confusion_matrix(predictions, y_test, 'Confusion Matrix - SVC',
cm_path)
def getDECNetworkResults(dec, enc):
#Load test dataset
test_data = LoadDataset("dataset/kaggle_original_train/", 0)
test_data, test_label, val, val_label = test_data.load_data()
big_data = LoadDataset("dataset/kaggle_augmented_train_new/", 0)
big_data, _, _, _ = big_data.load_data()
# make save directory
os.makedirs(os.path.join("dec"), exist_ok=True)
os.chdir("dec")
encoded = enc.predict(test_data)
q, _ = dec.predict(test_data, verbose=0)
y_pred = q.argmax(1)
print(y_pred)
confusion_matrix(test_label.astype(np.int64), y_pred)
#Take prediction time
for i in range(20):
iterate = 5000 * (i + 1)
data = big_data[0:iterate, :]
print(data.shape)
print("KMEAN")
start = time.time()
q, _ = dec.predict(data, verbose=0)
y_pred = q.argmax(1)
end = time.time()
print(end - start)
train_x = np.reshape(test_data, (3720, 64, 64))
TSNE = TSNEAlgo()
TSNE.tsne_fit(encoded, perplexity=35)
TSNE.tsne_plot(train_x,
y_pred.astype(int),
save_name="Pred",
save_data_dir="dec")
TSNE.tsne_plot(train_x,
test_label.astype(int),
save_name="True",
save_data_dir="dec")
def main():
graph = utils.load_graph()
position = utils.get_positions(graph)
utils.make_dir('images/spectral')
true_communities = utils.get_labels(graph, list(graph.nodes))
utils.plot_communities(graph, position, true_communities, labels=True, title='Butterfly Similarity Network - True Communities', path='images/spectral/communities_true.png')
node_assignments = spectral_clustering(graph)
nodes_to_communities = {k:v for (k,v) in zip(range(len(node_assignments)), node_assignments)}
communities = utils.group_communities(nodes_to_communities)
utils.plot_communities(graph, position, communities, labels=False, title='Butterfly Similarity Network - Spectral Communities', path='images/spectral/communities_spectral.png')
graph_nodes = sorted(list(graph.nodes))
predictions = utils.predict_majority_class(graph, communities)
preds = [predictions[n] for n in graph_nodes]
labels = [graph.nodes[n]['label'] for n in graph_nodes]
utils.accuracy(preds, labels)
utils.confusion_matrix(preds, labels, 'Confusion Matrix - Spectral Clustering', 'images/spectral/cm_spectral.png')
def train_bottomupdnn_v1(epochs=40, init_epochs=10, lr=0.01):
print("======Train base DNN using clean data======")
model, train_loader, test_loader, optimizer, criterion = train_init(
"BaseDNN", lr=lr)
clean_loader, clean_train_loader, noisy_loader = get_dataloader_bu1()
model.mode = "BaseModel"
# Train base DNN model
for epoch in range(init_epochs):
train_epoch(model, clean_train_loader, test_loader, optimizer,
criterion, epoch)
# Estimate confusion matrix for new bottom-up DNN v1 model
y_clean, y_pred = get_predict_label(model, clean_loader)
cmatrix_clean = confusion_matrix(y_clean, y_pred)
y_noisy, y_pred = get_predict_label(model, noisy_loader)
cmatrix_noisy = confusion_matrix(y_noisy, y_pred)
rmatrix = compute_rmatrix(cmatrix_clean, cmatrix_noisy)
cmatrix = compute_estimate_confusion(rmatrix, y_noisy)
# Initialize new bottom-up DNN v1 model
new_model = models.ButtomUpDNN1(model.params)
new_model.confusion.weight = nn.Parameter(cmatrix)
new_model.confusion.weight.requires_grad = False
optimizer = optim.SGD(new_model.parameters(), lr=lr, momentum=0.9)
# Concatenate all clean dataset and get data loader
clean_dataset = torch.utils.data.ConcatDataset(
[clean_loader.dataset, clean_train_loader.dataset])
all_clean_loader = torch.utils.data.DataLoader(dataset=clean_dataset,
batch_size=batch_size,
shuffle=True)
# Train bottom-up DNN v1 model
for epoch in range(init_epochs, epochs):
train_epoch_bu1(new_model, all_clean_loader, noisy_loader, test_loader,
optimizer, criterion, epoch)
# Save estimated Q matrix
plt.matshow(cmatrix)
plt.colorbar()
plt.savefig("./imgs/estimated_Q.png")
def evaluate(self, X, y, with_plot=False):
pred = self.__fit(X, y)
if self.classification:
return confusion_matrix(y, pred)
else:
if with_plot:
plt.figure(figsize=(12, 8))
plt.scatter(y, pred)
plt.xlabel('y')
plt.ylabel('y_pred')
return np.mean((y - pred)**2)
def __model_evaluation(self, arr, p_dna):
wt_data, mt_data = arr
df_dna = confusion_matrix(wt=wt_data, mt=mt_data, predicted=p_dna)
'Convert to amino acid'
wt_aa = map(dna_to_aa, wt_data)
mt_aa = map(dna_to_aa, mt_data)
p_aa = map(dna_to_aa, p_dna)
df_aa = confusion_matrix(wt=wt_aa, mt=mt_aa, predicted=p_aa)
df = pd.concat([df_dna, df_aa], axis=1)
outfile = 'summary_6nodes_nopostproc.csv'
df.to_csv(outfile)
'Show histogram'
df['Accuracy'].hist(bins=20)
df['TP'].hist(bins=20)
df['#Mutate Positions'].hist(bins=20)
plt.show()
def train_model(train_config):
images,labels = image_load.read_12channel_images(train_config['target_path'],train_config['image_resize'])
kflod_container = image_load.KFoldContainer(images,labels,train_config['Kfold'])
utils.write_config(train_config,train_config['save_path']+'config.csv')
confuse_matrix = []
accuracy = []
duration = []
for k in range(train_config['Kfold']):
train_x,train_y,test_x,test_y = kflod_container.get_fold_k(k)
train_x,train_y = image_load.augimage(train_x,train_y)
print('Load Training {n:d} images'.format(n=len(train_x)))
print('Load Testing {n:d} images'.format(n=len(test_x)))
train_ds = tf.data.Dataset.from_tensor_slices((train_x,train_y))
train_ds = train_ds.shuffle(buffer_size=train_config['shuffle_buffer_size']).repeat().batch(train_config['batch_size'])
test_ds = tf.data.Dataset.from_tensor_slices((test_x,test_y))
test_ds = test_ds.repeat(1).batch(train_config['batch_size'])
model = models.model_7_3_12channel(train_config['image_resize'],train_config['l2_factor'])
model.compile(optimizer=tf.keras.optimizers.Adam(lr=train_config['lr']),
loss=tf.keras.losses.SparseCategoricalCrossentropy(),
metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])
save_name = f'{train_config["save_path"]}{train_config["model_name"]}-{k}'
cbs = [tf.keras.callbacks.EarlyStopping(patience=train_config['early_stop_patience']),
tf.keras.callbacks.ModelCheckpoint(monitor='val_sparse_categorical_accuracy',filepath=save_name,save_best_only=True,save_weights_only=True,verbose=1),
utils.HSLRSchedular(train_config['lr'],
watch_value_name=train_config['schedular_watch_name'],
max_reduce_time=train_config['schedular_max_reduce_time'],
reduce_factor=train_config['schedular_reduce_factor'],
restart_factor=train_config['schedular_restart_factor'],
patience=train_config['schedular_patience'],
verbose=0),
utils.HSTensorboard(log_dir=f'./logs/{save_name}/',embeddings_metadata=test_x)]
ct = time.time()
model.fit(train_ds,epochs=train_config['epochs'],steps_per_epoch=train_config['steps_per_epoch'],validation_data=test_ds,callbacks=cbs)
duration.append(time.time()-ct)
model.load_weights(save_name)
logits = model.predict(test_x)
cm,acc = utils.confusion_matrix(test_y,tf.argmax(logits,axis=1).numpy())
confuse_matrix.append(cm)
accuracy.append(acc)
print(f'finish training. k={k}, accuracy={acc:.2f}')
sio.savemat(train_config['save_path']+'result.mat',{'cm':np.array(confuse_matrix),
'accuracy':np.array(accuracy),
'duration':np.array(duration)})
def train(epoch):
print('\nEpoch: %d' % epoch)
global Train_acc
net.train()
train_loss = 0
correct = 0
total = 0
conf_mat = np.zeros((NUM_CLASSES, NUM_CLASSES))
if epoch > learning_rate_decay_start and learning_rate_decay_start >= 0:
frac = (epoch - learning_rate_decay_start) // learning_rate_decay_every
decay_factor = learning_rate_decay_rate**frac
current_lr = opt.lr * decay_factor
utils.set_lr(optimizer, current_lr) # set the decayed rate
else:
current_lr = opt.lr
print('learning_rate: %s' % str(current_lr))
for batch_idx, (inputs, targets) in enumerate(trainloader):
if use_cuda:
inputs, targets = inputs.cuda(), targets.cuda()
optimizer.zero_grad()
inputs, targets = Variable(inputs), Variable(targets)
outputs = net(inputs)
loss = criterion(outputs, targets)
loss.backward()
utils.clip_gradient(optimizer, 0.1)
optimizer.step()
train_loss += loss.item()
conf_mat += utils.confusion_matrix(outputs, targets, NUM_CLASSES)
acc = sum([conf_mat[i, i]
for i in range(conf_mat.shape[0])]) / conf_mat.sum()
uacc_per_class = [
conf_mat[i, i] / conf_mat[i].sum()
for i in range(conf_mat.shape[0])
]
unweighted_acc = sum(uacc_per_class) / len(uacc_per_class)
prec_per_class = [
conf_mat[i, i] / conf_mat[:, i].sum()
for i in range(conf_mat.shape[0])
]
average_precision = sum(prec_per_class) / len(prec_per_class)
utils.progress_bar(
batch_idx, len(trainloader),
'Loss: %.3f | Acc: %.3f%% | unweighted_Acc: %.3f%%' %
(train_loss / (batch_idx + 1), 100. * acc, 100. * unweighted_acc))
Train_acc = 100. * acc
def process_dataset(dataset, colors):
y = np.load("./data/" + dataset + '_labels.npy')
pred = np.load("./data/" + dataset + '_clasification.npy')
segments = np.load("./results/" + dataset + '_segments.npy')
test_mask = np.load("./data/" + dataset + '_test_mask.npy').reshape(
y.shape)
sc_pred = classify_segments(pred, segments)
sc_score = utils.balanced_score(y[test_mask], sc_pred[test_mask])
sc_cm = utils.confusion_matrix(y[test_mask], sc_pred[test_mask])
utils.save_json({"sc": sc_score}, dataset + "_sc_score")
utils.save_csv(sc_cm, dataset + "_sc_cm")
color_map = color_true_map(sc_pred, labels_colors=colors)
save_image(color_map, dataset + "_sc_clasification")
def spectral(encoder, tsne, true_data, true_label):
"""
1. Predicts labels using spectral clustering
2. Prints confusion_matrix
3. Prints accuracy
4. Prints t-SNE plot of prediction
"""
enc_output = encoder.predict(true_data)
model = SpectralClustering(n_clusters=5,
affinity='nearest_neighbors',
assign_labels='kmeans')
pred = model.fit_predict(enc_output)
accuracy(true_label, pred)
confusion_matrix(true_label,
pred,
save_name="confusion_matrix_spectral.png")
tsne.tsne_plot(true_data,
pred,
save_data_dir="spectral",
save_name="spectral")
tsne.tsne_plot(true_data,
true_label,
save_data_dir="true_label",
save_name="true_label")
def eval(loader, model, is_test=False, confusion_matrix=False, filename=None):
"""
Evaluate model performance on data object (graph) in loader.
:param - loader: torch_geometric DataLoader for BIOSNAP dataset
:param - model: trained GNN model ready for making predictions
:param - is_test: boolean indicating whether to evaluate on test or eval split
"""
model.eval()
data = [data for data in loader][0]
mask = data.test_mask if is_test else data.val_mask
with torch.no_grad():
pred = model(data).max(dim=1)[1][mask]
label = data.y[mask]
if confusion_matrix:
utils.make_dir(gnn_utils.images_dir)
title = 'Confusion Matrix - GNN'
path = gnn_utils.images_dir + 'cm_' + filename + '.png'
utils.confusion_matrix(pred.to('cpu'), label.to('cpu'), title, path)
correct = pred.eq(label).sum().item()
total = mask.sum().item()
return correct / total