def train_epoch(dataset, sess, model, train_fetches, valid_fetches, train_writer, test_writer):
global last_f1
global lr
starttime = datetime.datetime.now()
batch_iter=dataset.batch_iter()
for i,(X1_batch,X2_batch,y_batch) in tqdm(enumerate(batch_iter)):
global_step = sess.run(model.global_step)
if 0 == (global_step + 1) % FLAGS.valid_step:
valid_cost, precision, recall, f1 = valid_epoch(dataset, sess, model)
print('Global_step=%d: valid cost=%g; p=%g, r=%g, f1=%g' % (
global_step, valid_cost, precision, recall, f1))
print ('cost time:%s',(datetime.datetime.now()-starttime))
if f1 > last_f1 :
last_f1 = f1
saving_path = model.saver.save(sess, model_path+str(f1)+'_', global_step+1)
print('saved new model to %s ' % saving_path)
# training
y_batch = to_categorical(y_batch)
_batch_size = len(y_batch)
feed_dict = {model.X1_inputs: X1_batch, model.X2_inputs: X2_batch, model.y_inputs: y_batch,
model.batch_size: _batch_size, model.tst: False, model.keep_prob: FLAGS.keep_prob}
summary, _cost, _, _ = sess.run(train_fetches, feed_dict) # the cost is the mean cost of one batch
# valid per 500 steps
if 0 == (global_step + 1) % 500:
train_writer.add_summary(summary, global_step)
batch_id = np.random.randint(0, n_va_batches) # 随机选一个验证batch
[X1_batch, X2_batch, y_batch] =dataset.get_vali_item(batch_id)
y_batch = to_categorical(y_batch)
_batch_size = len(y_batch)
feed_dict = {model.X1_inputs: X1_batch, model.X2_inputs: X2_batch, model.y_inputs: y_batch,
model.batch_size: _batch_size, model.tst: True, model.keep_prob: 1.0}
summary, _cost = sess.run(valid_fetches, feed_dict)
test_writer.add_summary(summary, global_step)
print ('global_step:%d,loss:%f'%(global_step,_cost))
def load_cifar(path):
"""Loads CIFAR10 dataset.
# Returns
Tuple of Numpy arrays: `(x_train, y_train), (x_test, y_test)`.
"""
num_train_samples = 50000
x_train = np.empty((num_train_samples, 3, 32, 32), dtype='uint8')
y_train = np.empty((num_train_samples, ), dtype='uint8')
for i in range(1, 6):
fpath = os.path.join(path, 'data_batch_' + str(i))
(x_train[(i - 1) * 10000:i * 10000, :, :, :],
y_train[(i - 1) * 10000:i * 10000]) = load_batch(fpath)
fpath = os.path.join(path, 'test_batch')
x_test, y_test = load_batch(fpath)
y_train = np.reshape(y_train, (len(y_train), 1))
y_test = np.reshape(y_test, (len(y_test), 1))
# 归一化
# x_train = x_train.astype(np.float) / 255. - 1.
# x_test = x_test.astype(np.float) / 255. - 1.
mean = np.array([123.680, 116.779, 103.939])
x_train = x_train.astype(np.float) - mean[:, np.newaxis, np.newaxis]
x_test = x_test.astype(np.float) - mean[:, np.newaxis, np.newaxis]
return (x_train, to_categorical(y_train)), (x_test, to_categorical(y_test))
def get_xy(
image_list,
label_names,
batch_centers,
size,
fc_shape,
nlabels,
preload,
datatype
):
n_images = len(image_list)
centers, idx = centers_and_idx(batch_centers, n_images)
print(''.join([' '] * 15) + 'Loading x')
x = filter(lambda z: z.any(), get_patches_list(image_list, centers, size, preload))
x = np.concatenate(x)
print(''.join([' '] * 15) + '- Concatenation')
x[idx] = x
print(''.join([' '] * 15) + 'Loading y')
y = [np.array([l[c] for c in lc]) for l, lc in izip(labels_generator(label_names), centers)]
print(''.join([' '] * 15) + '- Concatenation')
y = np.concatenate(y)
y[idx] = y
y_fc = [np.asarray(get_patches(l, lc, fc_shape))
for l, lc in izip(labels_generator(label_names), centers)]
y_fc = np.concatenate(y_fc)
y_fc[idx] = y_fc
if nlabels <= 2:
y = y.astype(dtype=np.bool)
y_fc = y_fc.astype(dtype=np.bool)
y = [
to_categorical(y, num_classes=nlabels),
to_categorical(y_fc, num_classes=nlabels).reshape((len(y_fc), -1, nlabels))
]
return x.astype(dtype=datatype), y
def execute(self, inputs, targets, config_model=None, dataset_name=''):
model = FactoryModel(
config_model['name'],
'{}_{}_holdout-{}'.format(dataset_name, config_model['name'],
self.test_size), config_model['size'],
config_model['params'], config_model['init']).get_model()
if self.state.epochs > 0:
model().load_weights(self.state.weights)
else:
#initialize weight and history path for the model
self.state.weights += model().name + '.h5'
self.state.history += '{}/history_{}.csv'.format(
config_model['name'],
model().name)
dict_scores = {}
dict_scores['scores'] = {}
dict_scores['scores']['model'] = [model().name]
print('\n------[executing Hold out {} for {} model]------------------'.
format(self.test_size * 100,
model().name))
train_x, test_x, train_y, test_y = train_test_split(
inputs,
targets,
test_size=self.test_size,
random_state=0,
shuffle=False)
history = self.trainner.train_model(
train_x,
to_categorical(train_y),
model(),
validation_data=(test_x, to_categorical(test_y)),
init_epoch=self.state.epochs)
print(
'Avaluating model-------------------------------------------------------------'
)
scores = model().evaluate(test_x, to_categorical(test_y))
self.add_score(model().metrics_names, dict_scores['scores'], scores)
(fpr, tpr, auc) = get_roc_curve(to_categorical(test_y),
model().predict(test_x))
dict_scores['roc'] = (fpr, tpr, auc)
dict_scores['history'] = [history]
dict_scores['cm'] = confusion_matrix(
test_y, np.argmax(model().predict(test_x), axis=1))
print("Result for the {} model".format(model().name))
print(dict_scores['scores'])
self.state.status = True
return dict_scores
def dnn_mnist():
# load datasets
path = 'mnist.pkl.gz'
train_set, val_set, test_set = load_mnist_datasets(path)
X_train, y_train = train_set
X_val, y_val = val_set
X_test, y_test = test_set
# 转为稀疏分类
y_train, y_val, y_test = utils.to_categorical(
y_train,
10), utils.to_categorical(y_val, 10), utils.to_categorical(y_test, 10)
# bookeeping for best model based on validation set
best_val_acc = -1
mnist = Mnist()
# Train
batch_size = 32
lr = 1e-1
for epoch in range(10):
num_train = X_train.shape[0]
num_batch = num_train // batch_size
for batch in range(num_batch):
# get batch data
batch_mask = np.random.choice(num_train, batch_size)
X_batch = X_train[batch_mask]
y_batch = y_train[batch_mask]
# 前向及反向
mnist.forward(X_batch)
loss = mnist.backward(X_batch, y_batch)
if batch % 200 == 0:
print("Epoch %2d Iter %3d Loss %.5f" % (epoch, batch, loss))
# 更新梯度
for w in ["W1", "b1", "W2", "b2", "W3", "b3"]:
mnist.weights[w] -= lr * mnist.gradients[w]
train_acc = mnist.get_accuracy(X_train, y_train)
val_acc = mnist.get_accuracy(X_val, y_val)
if (best_val_acc < val_acc):
best_val_acc = val_acc
# store best model based n acc_val
print('Epoch finish. ')
print('Train acc %.3f' % train_acc)
print('Val acc %.3f' % val_acc)
print('-' * 30)
print('')
print('Train finished. Best acc %.3f' % best_val_acc)
test_acc = mnist.get_accuracy(X_test, y_test)
print('Test acc %.3f' % test_acc)
def __getitem__(self, index):
"""Return a (transformed) vrd_input and target sample from an integer index"""
key, rel = self.data[index]
subject_box = rel['subject']['bbox'] # [ymin, ymax, xmin, xmax]
object_box = rel['object']['bbox']
minbbox = [
min(subject_box[0], object_box[0]),
max(subject_box[1], object_box[1]),
min(subject_box[2], object_box[2]),
max(subject_box[3], object_box[3])
]
image = imread('/scratch/datasets/sg_dataset/sg_' + self.type_dataset +
'_images/' + key)
bboxes = [subject_box, object_box, minbbox]
list_image = [
image[bbox[0]:bbox[1], bbox[2]:bbox[3]] for bbox in bboxes
]
subject_visual_input, object_visual_input, union_visual_input = tuple(
transform(x) for x in list_image)
list_binary_image = [np.zeros_like(image) for _ in range(len(bboxes))]
for (binary_image, bbox) in zip(list_binary_image, bboxes):
binary_image[bbox[0]:bbox[1], bbox[2]:bbox[3]] = 1
subject_spatial_input, object_spatial_input, union_spatial_input = \
tuple(spatial_transform(x)[0, :, :].view(1, 32, 32) for x in list_binary_image)
predicate_spatial_feature = torch.cat(
[subject_spatial_input, object_spatial_input], 0)
object_word_feature = torch.FloatTensor(
index_to_emb_dict[rel['object']['category']])
subject_word_feature = torch.FloatTensor(
index_to_emb_dict[rel['subject']['category']])
if use_model == 1:
input_sample = union_visual_input, predicate_spatial_feature
target_sample = rel['subject']['category'], rel['object'][
'category'], rel['predicate']
elif use_model == 2:
input_sample = torch.FloatTensor(to_categorical(rel['subject']['category'], object_size)), \
torch.FloatTensor(to_categorical(rel['object']['category'], object_size)), \
union_visual_input, predicate_spatial_feature
target_sample = rel['predicate']
elif use_model == 3:
input_sample = (subject_word_feature, object_word_feature,
union_visual_input, predicate_spatial_feature)
target_sample = rel['predicate']
return input_sample, target_sample
def cv(hyper_params):
print 'next:', hyper_params
fold = 2
perm = np.random.permutation(len(ids_list))
chunk_size = len(perm) / fold
chunks = np.split(
perm, [chunk_size * (i + 1) for i, a in enumerate(range(fold))])
acc_total = 0.
model.set_params(**hyper_params)
graph = tf.Graph()
with graph.as_default():
# define variables and ops in `graph`
model.build_graph(constants.n_entity, constants.n_attr,
constants.word_vec_dim, constants.n_label)
init_op = tf.global_variables_initializer()
# actual cross validation
for i in range(fold):
test_indices = chunks[i]
train_indices = np.array(list(set(perm) - set(test_indices)))
ids_train, ids_val = ids_list[train_indices, :], ids_list[
test_indices, :]
ent_train, ent_val = ents[train_indices], ents[test_indices]
attr_train, attr_val = attrs[train_indices], attrs[test_indices]
pol_train, pol_val = pols[train_indices], pols[test_indices]
lens_train, lens_val = sent_lens[train_indices], sent_lens[
test_indices]
with tf.Session(graph=graph).as_default() as sess:
sess.run(init_op)
# train
model.fit(sess, ids_train, ent_train, attr_train, lens_train,
utils.to_categorical(pol_train), embedding_w,
ent_embedding, attr_embedding)
# validate
acc = model.eval(sess, ids_val, ent_val, attr_val, lens_val,
utils.to_categorical(pol_val))
utils.log('validation {}/{} acc: {}'.format(i + 1, fold, acc),
True)
acc_total += acc
utils.log(
'cv for params below end. mean validation acc: {}'.format(
acc_total / fold), True)
utils.log(hyper_params)
return acc_total / fold
def main():
data = pd.read_csv("blood.csv")
x_train = data.drop(["a"], axis=1)
y_train = data["a"]
x_train = pd.get_dummies(x_train)
target_map = dict()
if y_train.dtype == "object":
target_map = {val: i for (i, val) in enumerate(np.unique(y_train))}
y_train = y_train.map(target_map)
y_train = to_categorical(y_train)
x_train = x_train.values
x_train, x_test, y_train, y_test = train_test_split(x_train,
y_train,
test_size=0.2)
mean_train = x_train.mean(axis=0)
std_train = np.std(x_train, axis=0)
x_train = (x_train - mean_train) / std_train
x_test = (x_test - mean_train) / std_train
mlp = MultiLayerPerceptron()
mlp.add(Input(x_train.shape[1]))
mlp.add(Dense(32, activation="relu"))
mlp.add(Dense(2, activation="sigmoid"))
mlp.build()
mlp.fit(x_train,
y_train,
epoch=40,
lr=0.05,
validation_data=(x_test, y_test))
mlp.draw()
def evaluate_audio_tagging(fold):
"""Evaluate the audio tagging predictions and write results.
Args:
fold (int): The fold (validation set) to evaluate.
"""
import evaluation
# Load grouth truth data and predictions
dataset = cfg.to_dataset('training', preprocessed=False)
df = io.read_metadata(dataset.metadata_path)
df = df[df.fold == fold]
y_true = utils.to_categorical(df.label)
fold_str = 'training' + str(fold)
path = cfg.predictions_path.format('predictions', fold_str)
y_pred = pd.read_csv(path, index_col=0).values
# Mask out those that are not manually verified
mask = df.manually_verified == 1
y_pred = y_pred[mask]
y_true = y_true[mask]
# Evaluate audio tagging performance
scores = evaluation.evaluate_audio_tagging(y_true,
y_pred,
threshold=cfg.threshold)
# Ensure output directory exist and write results
os.makedirs(os.path.dirname(cfg.results_path), exist_ok=True)
output_path = cfg.results_path.format(fold_str)
scores.to_csv(output_path)
# Print scores to 3 decimal places
pd.options.display.float_format = '{:,.3f}'.format
print('\n' + str(scores))
def test(model, sess):
reviews, ent2idx, attr2idx, polarity2idx = load_semeval_reviews(
constants.test_filename)
# list of (ids, ent, attr, pol)
tuples = []
for review in reviews:
if len(review.tokens) <= 1:
# ids = [0] + [word2idx[tok] for tok in review.tokens]
ids = [word2idx[tok] for tok in review.tokens]
else:
ids = [word2idx[tok] for tok in review.tokens]
tuples_ = [(ids, ent2idx[op.ent], attr2idx[op.attr],
polarity2idx[op.polarity]) for op in review.opinions]
tuples.extend(tuples_)
unzipped = zip(*tuples)
ids = utils.pad_sequences(unzipped[0], maxlen=constants.max_sent_len)
sent_lens = np.array(map(
lambda x: len(x)
if len(x) < constants.max_sent_len else constants.max_sent_len,
unzipped[0]),
dtype='int32')
ents, attrs, pols = (np.array(x, dtype='int32') for x in unzipped[1:])
acc = model.eval(sess, ids, ents, attrs, sent_lens,
utils.to_categorical(pols))
utils.log('test accuracy: {}'.format(acc), True)
def main():
data = datasets.load_digits()
X = normalize(data.data)
y = data.target
# One-hot encoding of nominal y-values
y = to_categorical(y)
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
seed=8)
# Perceptron
clf = Perceptron(n_iterations=5000,
learning_rate=0.001,
loss=CrossEntropy,
activation_function=Sigmoid)
clf.fit(X_train, y_train)
y_pred = np.argmax(clf.predict(X_test), axis=1)
y_test = np.argmax(y_test, axis=1)
accuracy = accuracy_score(y_test, y_pred)
print("Accuracy:", accuracy)
# Reduce dimension to two using PCA and plot the results
Plot().plot_in_2d(X_test,
y_pred,
title="Perceptron",
accuracy=accuracy,
legend_labels=np.unique(y))
def transform_data(self, maxlen):
self.maxlen_userUtter = maxlen
# replace unknown words with <unk> in user utterance, and encode it
# using word id.
print '-------------------utter-----------------------'
self.userUtter_encodePad, self.userUtter_txt = vectorizing_zeropad(
self.userUtter_txt, self.maxlen_userUtter, self.word2id, prefix='')
# replace unknown tags with <tag-unk> in user slot tags, and encode it
# as 1hot matrix
print '-------------------tag-----------------------'
userTag_encodePad, self.userTag_txt = vectorizing_zeropad(
self.userTag_txt,
self.maxlen_userUtter,
self.userTag2id,
prefix='tag-')
self.userTag_1hotPad = to_categorical(userTag_encodePad,
self.userTag_vocab_size)
# replace unknown intents with <intent-unk> in user intents, and encode
# it as binary vec
print '-------------------intent-----------------------'
self.userIntent_vecBin, self.userIntent_txt = vectorizing_binaryVec(
self.userIntent_txt,
self.userIntent_vocab_size,
self.userIntent2id,
prefix='intent-')
def _preprocess_input_data(self, data_splits):
train_data, validation_data, test_data = data_splits
train_images, train_classes = train_data
validation_images, validation_classes = validation_data
test_images, test_classes = test_data
# encode outputs into one hot vectors
class_arg_splits = (train_classes, validation_classes, test_classes)
categorical_data = [
to_categorical(class_arg_split)
for class_arg_split in class_arg_splits
]
train_classes, validation_classes, test_classes = categorical_data
# flatten, add salt/pepper noise and normalize images
image_splits = (train_images, validation_images, test_images)
image_splits = [flatten(image_split) for image_split in image_splits]
image_splits = [
spice_up_images(image_split) for image_split in image_splits
]
image_splits = [
normalize_images(image_split) for image_split in image_splits
]
train_images, validation_images, test_images = image_splits
train_data = (train_images, train_classes)
validation_data = (validation_images, validation_classes)
test_data = (test_images, test_classes)
preprocessed_data_split = (train_data, validation_data, test_data)
return preprocessed_data_split
def main():
data = pd.read_csv('blood.csv')
x_train = data.drop(['a'], axis=1)
y_train = data['a']
x_train = pd.get_dummies(x_train)
target_map = dict()
if y_train.dtype == 'object':
target_map = {val: i for (i, val) in enumerate(np.unique(y_train))}
# print(target_map)
y_train = y_train.map(target_map)
y_train = to_categorical(y_train)
# print(y_train[:5])
x_train = x_train.values
# x_train = (x_train - x_train.min(axis=0)) / (x_train.max(axis=0) - x_train.min(axis=0))
# x_train = (x_train - x_train.mean(axis=0)) / np.std(x_train, axis=0)
x_train, x_test, y_train, y_test = train_test_split(x_train, y_train, test_size=0.2)
mean_train = x_train.mean(axis=0)
std_train = np.std(x_train, axis=0)
x_train = (x_train - mean_train)/std_train
x_test = (x_test - mean_train)/std_train
mlp = MultiLayerPerceptron()
mlp.add(Input(x_train.shape[1]))
mlp.add(Dense(32, activation='relu'))
mlp.add(Dense(2, activation='sigmoid'))
mlp.build()
mlp.fit(x_train, y_train, epoch=40, lr=0.05, validation_data=(x_test, y_test))
mlp.draw()
def valid_train_epoch(dataset, sess, model, train_sample):
"""Test on the train data."""
batch_iter = train_sample.train_sample_iter()
_costs = 0.0
num = 0
marked_labels_list = list()
for i, (X1_batch, X2_batch, y_batch) in enumerate(batch_iter):
num = num + 1
marked_labels_list.extend(y_batch)
y_batch = to_categorical(y_batch)
_batch_size = len(y_batch)
fetches = [model.loss]
feed_dict = {
model.X1_inputs: X1_batch,
model.X2_inputs: X2_batch,
model.y_inputs: y_batch,
model.batch_size: _batch_size,
model.is_training: False,
model.keep_prob: 1.0
}
_cost = sess.run(fetches, feed_dict)
_costs += _cost[0]
mean_cost = _costs / num
return mean_cost
def take_percentage_of_type(trainX, trainY, relation_type, percentage):
"""take instances squentially
assume input shuffled
remove the rest of instances from the dataset"""
if percentage >= 1.0:
return trainX, trainY
nb_classes = trainY.shape[1]
trainY = np.argmax(trainY, axis=1)
total_num = 0
for i in xrange(len(trainY)):
if trainY[i] == relation_type:
total_num += 1
current_num = 0
remove_list, retain_list = [], []
for i in xrange(len(trainY)):
if trainY[i] == relation_type:
current_num += 1
if current_num > total_num * percentage:
remove_list += [i] # other
else:
retain_list += [i]
else:
retain_list += [i]
trainY = utils.to_categorical(trainY, nb_classes=nb_classes)
trainY = trainY[retain_list]
for key in trainX.keys():
trainX[key] = trainX[key][retain_list]
return trainX, trainY
def __getitem__(self, index):
sentence = self.collection.sentences[index]
batch = self.build_batch(sentence)
batch_input, batch_output = tuple(zip(*tuple(batch)))
sents, starts_1, starts_2 = tuple(zip(*tuple(batch_input)))
sents = torch.tensor(sents)
starts_1 = to_categorical(torch.tensor(starts_1), sents.size()[-1])
starts_1.unsqueeze_(-2)
starts_2 = to_categorical(torch.tensor(starts_2), sents.size()[-1])
starts_2.unsqueeze_(-2)
batch_output = torch.tensor(batch_output)
return (sents, starts_1, starts_2), batch_output
def learn_toy(topology_file,
features_file,
orientation,
num_hops,
target_dims=None):
if target_dims is None:
target_dims = [-1]
graph = labeled_graph.DenseLabeledGraph()
graph.load_topology_from_text(topology_file)
graph.load_features_from_text(features_file)
graph.summary(verbose=1)
x = utils.extract_inputs(graph.features, target_dims)
y = utils.extract_targets(graph.features, target_dims)
num_nodes = int(tf.shape(x)[0])
num_features = tf.shape(x)[1]
p = graph.transition_matrix
y, num_classes = utils.to_categorical(tf.squeeze(y, axis=1))
if orientation == 'reversed':
num_hops *= 2
x = dcnn.diffuse_features(orientation, p, x, num_hops)
toy_dataset = tf.data.Dataset.from_tensor_slices((x, y)).repeat().batch(16)
model = dcnn.DCNN(num_features, num_hops, num_classes, bias=True)
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
model.compile(optimizer='adam', loss=loss_fn, metrics=['accuracy'])
model.fit(toy_dataset, epochs=10, steps_per_epoch=768)
model.evaluate(toy_dataset, steps=num_nodes)
def gradient_Wj(dist, X, y, Wj):
start = timeit.default_timer()
# gradient = np.zeros_like(Wj) # (26, 128)
if y.ndim < 2:
y = to_categorical(y, 26)
x_index = np.arange(0, X.shape[0])
y_deltas = (y - dist[x_index]).reshape((len(x_index), -1)).T
x_deltas = X[x_index, 1:].reshape((len(x_index), -1))
gradient = np.dot(y_deltas, x_deltas)
#print(gradient.shape)
# for j in range(X_train.shape[0]):
# y_delta = (y_train[j] - dist[j]).reshape((-1, 1))
# x_delta = X_train[j, 1:].reshape((1, -1))
# gradient += np.dot(y_delta, x_delta)
gradient /= len(X)
# for l2 regularization
gradient += Wj
#flattened_gradient = gradient.flatten()
# result = open(r'grad.txt', 'w')
# for g in flattened_gradient:
# result.write(str(g) + "\n")
# result.close()
stop = timeit.default_timer()
#print('gradient_Wj (s): ' + str(stop - start))
return gradient
def valid_epoch(dataset, sess, model):
"""Test on the valid data."""
batch_iter = dataset.batch_iter(False)
_costs = 0.0
predict_labels_list = list() # 所有的预测结果
marked_labels_list = list()
for i, (X1_batch, X2_batch, y_batch) in tqdm(enumerate(batch_iter)):
marked_labels_list.extend(y_batch)
y_batch = to_categorical(y_batch)
_batch_size = len(y_batch)
fetches = [model.loss, model.y_pred]
feed_dict = {
model.X1_inputs: X1_batch,
model.X2_inputs: X2_batch,
model.y_inputs: y_batch,
model.batch_size: _batch_size,
model.tst: True,
model.keep_prob: 1.0
}
_cost, predict_labels = sess.run(fetches, feed_dict)
_costs += _cost
predict_labels = map(labellist2id, predict_labels)
predict_labels_list.extend(predict_labels)
predict_label_and_marked_label_list = zip(predict_labels_list,
marked_labels_list)
precision, recall, f1 = score_eval(predict_label_and_marked_label_list)
mean_cost = _costs / n_va_batches
return mean_cost, precision, recall, f1
def train(input_df, max_feature_length, num_classes, embedding_size, learning_rate, batch_size, num_epochs, save_dir=None, print_summary=False):
# Stage 1: Convert raw texts into char-ids format && convert labels into one-hot vectors
X_train, X_test, y_train, y_test = get_input_data(input_df)
y_train = to_categorical(y_train, num_classes)
# Stage 2: Build Model
num_filters = [64, 128, 256, 512]
model = build_model(num_filters=num_filters, num_classes=num_classes, embedding_size=embedding_size, learning_rate=learning_rate)
# Stage 3: Training
save_dir = save_dir if save_dir is not None else 'checkpoints'
filepath = os.path.join(save_dir, "weights-{epoch:02d}-{val_acc:.2f}.hdf5")
checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')
if print_summary:
print(model.summary())
model.fit(
x=X_train,
y=y_train,
batch_size=batch_size,
epochs=num_epochs,
validation_split=0.33,
callbacks=[checkpoint],
shuffle=True,
verbose=True
)
def main():
data = datasets.load_digits()
X = normalize(data.data)
y = data.target
# convert the nominal y values to binary
y = to_categorical(y)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.4)
#mlp
clf = MultilayerPerceptron(n_hidden=16,
n_iterations=1000,
learning_rate=0.01)
clf.fit(X_train, y_train)
y_pred = np.argmax(clf.predict(X_test), axis=1)
y_test = np.argmax(y_test, axis=1)
accuracy = accuracy_score(y_test, y_pred)
print("Accuracy:", accuracy)
# Reduce dimension to two using PCA and plot the results
Plot().plot_in_2d(X_test,
y_pred,
title="Multilayer Perceptron",
accuracy=accuracy,
legend_labels=np.unique(y))
def do_eval(test_data_path, shuffle=False):
if FLAGS.load_model is None:
raise ValueError("You need to specify the model location by --load_model=[location]")
# Load Testing Data
question_1, question_2, labels = get_input_from_csv(test_data_path)
if shuffle:
question_1, question_2, labels = shuffle_data(question_1, question_2, labels)
# Load Pre-trained Model
if FLAGS.best_glove:
import en_core_web_md
nlp = en_core_web_md.load() # load best-matching version for Glove
else:
nlp = spacy.load('en')
embedding_matrix = load_glove_embeddings(nlp.vocab, n_unknown=FLAGS.num_unknown) # shape=(1071074, 300)
tf.logging.info('Build model ...')
esim = ESIM(embedding_matrix, FLAGS.max_length, FLAGS.num_hidden, FLAGS.num_classes, FLAGS.keep_prob, FLAGS.learning_rate)
if FLAGS.load_model:
model = esim.build_model(FLAGS.load_model)
else:
raise ValueError("You need to specify the model location by --load_model=[location]")
# Convert the "raw data" to word-ids format && convert "labels" to one-hot vectors
q1_test, q2_test = convert_questions_to_word_ids(question_1, question_2, nlp, max_length=FLAGS.max_length, tree_truncate=FLAGS.tree_truncate)
labels = to_categorical(np.asarray(labels, dtype='int32'))
scores = model.evaluate([q1_test, q2_test], labels, batch_size=FLAGS.batch_size, verbose=1)
print("=================== RESULTS =====================")
print("[*] LOSS OF TEST DATA: %.4f" % scores[0])
print("[*] ACCURACY OF TEST DATA: %.4f" % scores[1])
def mel_collate(batch):
""" Zero-pads model inputs and targets based on number of frames per step """
len_out = int(hparams.freq * ceil(float(hparams.seq_len / hparams.freq)))
mels = []
labels = []
labels_onehot = []
for mel, speaker, sample_len, frame_len in batch:
if frame_len < len_out:
len_pad = len_out - frame_len
x = np.pad(mel, ((0, len_pad), (0, 0)), 'constant')
else:
start = np.random.randint(frame_len - len_out + 1)
x = mel[start:start + len_out]
mels.append(x)
label = to_categorical(speaker, hparams.speaker_num)
labels.append(speaker)
labels_onehot.append(label)
mels = torch.FloatTensor(mels)
labels = torch.LongTensor(labels)
labels_onehot = torch.FloatTensor(labels_onehot)
return mels, labels, labels_onehot
def train(train_data, val_data, batch_size, n_epochs, save_dir=None):
# Stage 1: Read training data (csv) && Preprocessing them
tf.logging.info('Loading training and validataion data ...')
train_question_1, train_question_2, train_labels = get_input_from_csv(train_data)
# val_question_1, val_question_2, val_labels = get_input_from_csv(val_data)
# Stage 2: Load Pre-trained embedding matrix (Using GLOVE here)
tf.logging.info('Loading pre-trained embedding matrix ...')
if FLAGS.best_glove:
import en_core_web_md
nlp = en_core_web_md.load() # load best-matching version for Glove
else:
nlp = spacy.load('en')
embedding_matrix = load_glove_embeddings(nlp.vocab, n_unknown=FLAGS.num_unknown) # shape=(1071074, 300)
# Stage 3: Build Model
tf.logging.info('Build model ...')
esim = ESIM(embedding_matrix, FLAGS.max_length, FLAGS.num_hidden, FLAGS.num_classes, FLAGS.keep_prob, FLAGS.learning_rate)
if FLAGS.load_model:
model = esim.build_model(FLAGS.load_model)
else:
model = esim.build_model()
# Stage 4: Convert the "raw data" to word-ids format && convert "labels" to one-hot vectors
tf.logging.info('Converting questions into ids ...')
q1_train, q2_train = convert_questions_to_word_ids(train_question_1, train_question_2, nlp, max_length=FLAGS.max_length, tree_truncate=FLAGS.tree_truncate)
train_labels = to_categorical(np.asarray(train_labels, dtype='int32'))
# q1_val, q2_val = convert_questions_to_word_ids(val_question_1, val_question_2, nlp, max_length=FLAGS.max_length, tree_truncate=FLAGS.tree_truncate)
# val_labels = to_categorical(np.asarray(val_labels, dtype='int32'))
# Stage 5: Training
tf.logging.info('Start training ...')
callbacks = []
save_dir = save_dir if save_dir is not None else 'checkpoints'
filepath = os.path.join(save_dir, "weights-{epoch:02d}-{val_acc:.2f}.hdf5")
checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')
callbacks.append(checkpoint)
if FLAGS.tensorboard:
graph_dir = os.path.join('.', 'GRAPHs')
if not os.path.exists(graph_dir):
os.makedirs(graph_dir)
tb = TensorBoard(log_dir=graph_dir, histogram_freq=0, write_graph=True, write_images=True)
callbacks.append(tb)
model.fit(
x=[q1_train, q2_train],
y=train_labels,
batch_size=batch_size,
epochs=n_epochs,
# validation_data=([q1_val, q2_val], val_labels),
validation_split=0.2,
callbacks=callbacks,
shuffle=True,
verbose=FLAGS.verbose
)
def preprocess(data, params):
# change list of dictionary to dictionary of lists
data = dict(zip(data[0],zip(*[d.values() for d in data])))
Y = data.pop('y')
X = data
X = pad_features(X, params)
Y = to_categorical(Y, nb_classes=params['num_classes'])
return X, Y
def evaluate(self, data_gen, training=False):
preds_history = {}
targets_history = {}
scores = {}
for X, Y in data_gen:
if training: self.fit_on_batch(X, Y)
self.train(training)
Y_pred = self(X)
for target in Y.keys():
if Y[target] is None: continue
if target not in self.targets: continue
target_idx = self.targets.index(target)
if target not in preds_history:
preds_history[target] = []
targets_history[target] = []
ypred = Y_pred[target_idx][0].cpu().data.numpy()
if target != 'regression':
ytarget = to_categorical([Y[target]], num_classes=self.classes[target_idx])[0]
else:
ytarget = Y[target]
preds_history[target].append(ypred)
targets_history[target].append(ytarget)
for target in preds_history.keys():
preds_history[target] = np.array(preds_history[target])
targets_history[target] = np.array(targets_history[target])
if target != 'regression':
mask = targets_history[target].sum(axis=0) > 2
if mask.sum() < 2: continue
print (target, targets_history[target].shape[0])
score, baseline = 0.0, 0.0
if target == 'regression':
#score = pearsonr (targets_history[target], preds_history[target][:, 0])
#baseline = pearsonr (targets_history[target], shuffle(preds_history[target][:, 0]))
#scores[target] = "{0:0.3f}(p={1:0.4f})/{2:0.3f}(p={3:0.4f})".format(score[0], score[1], baseline[0], baseline[1])
score = concordance_index(targets_history[target], preds_history[target][:, 0],
event_observed=targets_history[target]<1.0)
baseline = concordance_index(targets_history[target], shuffle(preds_history[target][:, 0]),
event_observed=targets_history[target]<1.0)
scores[target] = "{0:0.3f}/{1:0.3f}".format(score, baseline)
else:
score = roc_auc_score (targets_history[target][:, mask], preds_history[target][:, mask])
baseline = roc_auc_score (targets_history[target][:, mask], shuffle(preds_history[target][:, mask]))
scores[target] = "{0:0.3f}/{1:0.3f}".format(score, baseline)
return scores
def fit(self, X, y):
y = to_categorical(y)
y_pred = np.zeros(np.shape(y))
for i in self.bar(range(self.n_estimators)):
tree = self.trees[i]
y_and_pred = np.concatenate((y, y_pred), axis=1)
tree.fit(X, y_and_pred)
update_pred = tree.predict(X)
y_pred -= np.multiply(self.learning_rate, update_pred)
def do_evaluation(eval_data, max_feature_length):
if FLAGS.load_model is None:
raise ValueError(
"You need to specify the model location by --load_model=[location]"
)
# Load Testing Data
comments = []
sentiments = []
contents = []
with open(eval_data, 'r') as f:
for line in f.readlines():
comment, sentiment, content = line.split(',')
comments.append(comment)
sentiments.append(sentiment)
contents.append(content)
for i in xrange(len(comments)):
X, y_sentiment, y_comment = get_input_data_from_text(
comments[i], sentiments[i], contents[i], max_feature_length)
y_sentiment = to_categorical(y_sentiment, FLAGS.n_sentiment_classes)
y_comment = to_categorical(y_comment, FLAGS.n_content_classes)
sentiment_model = build_model(
num_filters=num_filters,
num_classes=FLAGS.n_sentiment_classes,
sequence_max_length=FLAGS.max_feature_length,
embedding_size=FLAGS.embedding_size,
learning_rate=FLAGS.learning_rate,
load_pretrained_model=True)
comment_model = build_model(
num_filters=num_filters,
num_classes=FLAGS.n_content_classes,
sequence_max_length=FLAGS.max_feature_length,
embedding_size=FLAGS.embedding_size,
learning_rate=FLAGS.learning_rate,
load_pretrained_model=True)
sentiment_loss_and_history = sentiment_model.evaluate(
X, y_sentiment, batch_size=FLAGS.batch_size, verbose=1)
comment_loss_and_history = comment_model.evaluate(
X, y_sentiment, batch_size=FLAGS.batch_size, verbose=1)
def train(input_dim, z_dim, num_epochs, num_classes, batch_size, learning_rate, shuffle=False, data_dir=None, summary_dir=None):
# Load data
X_train = np.load(os.path.join(data_dir, 'data.npy'))
y_train = np.load(os.path.join(data_dir, 'label.npy'))
y_train = to_categorical(y_train, num_classes + 1)
print('Number of training images: %d' % X_train.shape[0])
with tf.Graph().as_default():
sess = tf.Session()
with sess.as_default():
# Build model
aae = AAE(input_dim, z_dim, num_classes, batch_size, learning_rate)
aae.build(G_type=FLAGS.G_type)
loss_summary_writer = tf.summary.FileWriter(summary_dir, sess.graph)
num_batches_per_epoch = X_train.shape[0] // batch_size
tf.logging.info('Create new session')
sess.run(tf.global_variables_initializer())
for epoch in range(num_epochs):
total_vae_loss, total_gen_loss, total_disc_loss = 0.0, 0.0, 0.0
for i in range(num_batches_per_epoch):
start_index = i * batch_size
end_index = min((i + 1) * batch_size, X_train.shape[0])
X_batch = X_train[start_index:end_index]
y_batch = y_train[start_index:end_index]
vae_loss, gen_loss, disc_loss = train_step(X_batch, y_batch, sess, aae, loss_summary_writer)
total_vae_loss += vae_loss
total_gen_loss += gen_loss
total_disc_loss += disc_loss
print("Epoch %3d ==> vae_loss: %.4f\tgen_loss: %.4f\tdisc_loss: %.4f" % (epoch, total_vae_loss / num_batches_per_epoch, total_gen_loss / num_batches_per_epoch, total_disc_loss / num_batches_per_epoch))
if FLAGS.plot:
indices = np.random.choice(X_train.shape[0], size=batch_size)
X_sample = X_train[indices]
y_sample = y_train[indices]
plot_reconstructed_images(X_sample, y_sample, aae, sess)
plot_generated_images(aae, sess)
indices = np.random.choice(X_train.shape[0], size=5000)
X_sample = X_train[indices]
y_sample = y_train[indices]
X_latent_space = aae.get_latent_space(sess, X_sample)
X_latent_space = X_latent_space.astype('float64')
plot_tsne(X_latent_space, y_sample)
def flow(self, mode='train'):
while True:
if mode =='train':
shuffle(self.train_keys)
keys = self.train_keys
elif mode == 'val' or mode == 'demo':
shuffle(self.validation_keys)
keys = self.validation_keys
else:
raise Exception('invalid mode: %s' % mode)
inputs = []
targets = []
for key in keys:
image_path = self.path_prefix + key
image_array = imread(image_path)
image_array = imresize(image_array, self.image_size)
num_image_channels = len(image_array.shape)
if num_image_channels != 3:
continue
ground_truth = self.ground_truth_data[key]
if self.do_random_crop:
image_array = self._do_random_crop(image_array)
image_array = image_array.astype('float32')
if mode == 'train' or mode == 'demo':
if self.ground_truth_transformer != None:
image_array, ground_truth = self.transform(
image_array,
ground_truth)
ground_truth = (
self.ground_truth_transformer.assign_boxes(
ground_truth))
else:
image_array = self.transform(image_array)[0]
inputs.append(image_array)
targets.append(ground_truth)
if len(targets) == self.batch_size:
inputs = np.asarray(inputs)
targets = np.asarray(targets)
# this will not work for boxes
targets = to_categorical(targets)
if mode == 'train' or mode == 'val':
inputs = self.preprocess_images(inputs)
yield self._wrap_in_dictionary(inputs, targets)
if mode == 'demo':
yield self._wrap_in_dictionary(inputs, targets)
inputs = []
targets = []
def fit(self, X, y):
y = to_categorical(y)
super(GBDTClassifier, self).fit(X, y)
