def test_cca(net, cca_loss, optimizer, train_iter, val_iter, train_iterations,
val_iterations):
'''
Return:
test_loss: number
acc: number
'''
train_losses = []
train_data = []
train_label = []
net.train()
for iteration in range(train_iterations):
optimizer.zero_grad()
imgs, texts, labels = Variable_(next(train_iter))
o1, o2 = net.forward_cca(imgs, texts)
loss = cca_loss.loss(o1, o2)
train_losses.append(loss.item())
loss.backward()
torch.nn.utils.clip_grad_norm_(net.parameters(), 0.25)
optimizer.step()
train_data.append(o1) # o1.clone().detach().cpu().numpy())
train_label.append(labels) # labels.clone().detach().cpu().numpy())
train_loss = np.mean(train_losses)
losses = []
val_data = []
val_label = []
with torch.no_grad():
net.eval()
for iteration in range(val_iterations):
imgs, texts, labels = Variable_(next(val_iter))
o1, o2 = net.forward_cca(imgs, texts)
val_data.append(o1) # o1.detach().cpu().numpy())
val_label.append(labels) # labels.detach().cpu().numpy())
loss = cca_loss.loss(o1, o2)
losses.append(loss.item())
train_data = torch.cat(train_data, 0)
val_data = torch.cat(val_data, 0)
train_label = torch.cat(train_label, 0)
val_label = torch.cat(val_label, 0)
train_data = train_data.cpu().numpy()
val_data = val_data.cpu().numpy()
train_label = train_label.cpu().numpy()
val_label = val_label.cpu().numpy()
test_loss = np.mean(losses)
acc = svm_classify(train_data, val_data, train_label, val_label)
return train_loss, test_loss, acc
def fit(self, data):
print('len(data.items()): ', len(data.items()))
X, Y = [], []
for idx, embs in data.items():
for emb in embs:
X.append(emb)
Y.append(idx)
pairs = list(zip(X, Y))
shuffle(pairs)
# print('pairs: ', pairs)
X, Y = zip(*pairs)
print('len(X), len(Y): ', len(X), len(Y))
if len(X) > len(data.items())*2:
self.classify_model = svm_classify(X, Y)
else:
from sklearn import svm
clf = svm.SVC(gamma='scale')
clf.fit(X, Y)
self.classify_model = clf
# if a linear CCA should get applied on the learned features extracted from the networks
# it does not affect the performance on noisy MNIST significantly
apply_linear_cca = True
# end of parameters section
############
# Each view is stored in a gzip file separately. They will get downloaded the first time the code gets executed.
# Datasets get stored under the datasets folder of user's Keras folder
# normally under [Home Folder]/.keras/datasets/
data1 = load_data('noisymnist_view1.gz', 'https://www2.cs.uic.edu/~vnoroozi/noisy-mnist/noisymnist_view1.gz')
data2 = load_data('noisymnist_view2.gz', 'https://www2.cs.uic.edu/~vnoroozi/noisy-mnist/noisymnist_view2.gz')
# Building, training, and producing the new features by DCCA
model = create_model(layer_sizes1, layer_sizes2, input_shape1, input_shape2,
learning_rate, reg_par, outdim_size, use_all_singular_values)
model.summary()
model = train_model(model, data1, data2, epoch_num, batch_size)
new_data = test_model(model, data1, data2, outdim_size, apply_linear_cca)
# Training and testing of SVM with linear kernel on the view 1 with new features
[test_acc, valid_acc] = svm_classify(new_data, C=0.01)
print("Accuracy on view 1 (validation data) is:", valid_acc * 100.0)
print("Accuracy on view 1 (test data) is:", test_acc*100.0)
# Saving new features in a gzip pickled file specified by save_to
print('saving new features ...')
f1 = gzip.open(save_to, 'wb')
thepickle.dump(new_data, f1)
f1.close()
# if a linear CCA should get applied on the learned features extracted from the networks
# it does not affect the performance on noisy MNIST significantly
apply_linear_cca = True
# end of parameters section
############
# Each view is stored in a gzip file separately. They will get downloaded the first time the code gets executed.
# Datasets get stored under the datasets folder of user's Keras folder
# normally under [Home Folder]/.keras/datasets/
data1 = load_data('noisymnist_view1.gz', 'https://www2.cs.uic.edu/~vnoroozi/noisy-mnist/noisymnist_view1.gz')
data2 = load_data('noisymnist_view2.gz', 'https://www2.cs.uic.edu/~vnoroozi/noisy-mnist/noisymnist_view2.gz')
# Building, training, and producing the new features by DCCA
model = create_model(layer_sizes1, layer_sizes2, input_shape1, input_shape2,
learning_rate, reg_par, outdim_size, use_all_singular_values)
model.summary()
model = train_model(model, data1, data2, epoch_num, batch_size)
new_data = test_model(model, data1, data2, outdim_size, apply_linear_cca)
# Training and testing of SVM with linear kernel on the view 1 with new features
[test_acc, valid_acc] = svm_classify(new_data, C=0.01)
print("Accuracy on view 1 (validation data) is:", valid_acc * 100.0)
print("Accuracy on view 1 (test data) is:", test_acc*100.0)
# Saving new features in a gzip pickled file specified by save_to
print('saving new features ...')
f1 = gzip.open(save_to, 'wb')
thepickle.dump(new_data, f1)
f1.close()
model, embed_generator = create_model(numNodes, factors)
model.summary()
model.compile(optimizer='rmsprop', loss={'left_right_dot': LINE_loss})
model.fit_generator(data_gen,
samples_per_epoch=epoch_train_size,
nb_epoch=epoch_num,
verbose=1)
new_X = []
new_label = []
keys = list(labels_dict.keys())
np.random.shuffle(keys)
for k in keys:
v = labels_dict[k]
x = embed_generator.predict_on_batch(
[np.asarray([k]), np.asarray([k])])
new_X.append(x[0][0] + x[1][0])
new_label.append(labels_dict[k])
new_X = np.asarray(new_X, dtype=np.float32)
new_label = np.asarray(new_label, dtype=np.int32)
[train_acc, valid_acc, test_acc] = svm_classify(new_X, new_label,
split_ratios, svm_C)
print("Train Acc:", train_acc, " Validation Acc:", valid_acc, " Test Acc:",
test_acc)
#data1 = load_data('noisymnist_view1.gz', 'https://www2.cs.uic.edu/~vnoroozi/noisy-mnist/noisymnist_view1.gz')
#data2 = load_data('noisymnist_view2.gz', 'https://www2.cs.uic.edu/~vnoroozi/noisy-mnist/noisymnist_view2.gz')
# Building, training, and producing the new features by DCCA
model = create_model(layer_sizes1, layer_sizes2, input_shape1,
input_shape2, learning_rate, reg_par, outdim_size,
use_all_singular_values, batch_size)
model.summary()
model = train_model(model, data1, data2, epoch_num, batch_size)
new_data = tt_model(model, data1, data2, outdim_size, apply_linear_cca)
# Training and testing of SVM with linear kernel on the view 1 with new features
[test_acc, valid_acc], [test_p,
valid_p], [test_label,
valid_label] = svm_classify(new_data,
C=0.01,
view=1)
print("Accuracy on view 1 (validation data) is:", valid_acc * 100.0)
print("Accuracy on view 1 (test data) is:", test_acc * 100.0)
report_performance(test_label, test_p, 0.5)
# Training and testing of SVM with linear kernel on the view 2 with new features
[test_acc, valid_acc], [test_p,
valid_p], [test_label,
valid_label] = svm_classify(new_data,
C=0.01,
view=2)
print("Accuracy on view 2 (validation data) is:", valid_acc * 100.0)
print("Accuracy on view 2 (test data) is:", test_acc * 100.0)
report_performance(test_label, test_p, 0.5)
