def accuracy_fn(target, output):
"""
function: 获取准确值
:param target: label
:param output: prediction
:return: accuracy
"""
accuracy = Accuracy()
output = round(output)
accuracy.update_state(y_true=target, y_pred=output)
accuracy_value = accuracy.result().numpy()
return accuracy_value
def report(model, ds_test_images, ds_test_labels, threshold=0.5):
""""""
true_iter = ds_test_labels.as_numpy_iterator()
y_true = np.hstack(list(true_iter))
y_pred = model.predict(ds_test_images).flatten()
class_labels = np.unique(y_true)
depth = class_labels.shape[0]
y_true_oh = tf.one_hot(y_true, depth=depth)
y_pred_oh = tf.one_hot(np.where(y_pred < threshold, 0, 1), depth=depth)
results = {'Accuracy': [], 'Precision': [], 'Recall': []}
m = Accuracy()
_ = m.update_state(y_true, np.around(y_pred).astype(int))
results['Accuracy'].append(m.result().numpy())
results['Precision'].append(" ")
results['Recall'].append(" ")
prec = [Precision(class_id=n) for n in class_labels]
rec = [Recall(class_id=n) for n in class_labels]
for p, r in zip(prec, rec):
p.update_state(y_true_oh, y_pred_oh)
r.update_state(y_true_oh, y_pred_oh)
results['Accuracy'].append(" ")
results['Precision'].append(p.result().numpy())
results['Recall'].append(r.result().numpy())
row_labels = [
'All' if i == 0 else f'Class {i-1}' for i in range(depth + 1)
]
return pd.DataFrame(data=results, index=row_labels)
def plot_confusion_matrix(y_true, y_pred):
cm = confusion_matrix(y_true, y_pred, normalize='true')
plt.imshow(cm, cmap='Blues')
plt.xlabel('predictions')
plt.ylabel('ground truth')
# predict
y_pred = np.argmax(dnn.predict(x=test_x, batch_size=256), axis=1)
y_true = np.argmax(test_y, axis=1)
accuracy = Accuracy()
# frame-by-frame at the state level
accuracy.update_state(y_true, y_pred)
print('Frame-by-frame accuracy at the state level: {:.2f}%'.format(
accuracy.result().numpy() * 100))
plt.figure()
plot_confusion_matrix(y_true, y_pred)
plt.title('Frame-by-frame confusion matrix at the state level')
# frame-by-frame at the phoneme level
y_pred_phones = states2phones(y_pred, phones, stateList)
y_true_phones = states2phones(y_true, phones, stateList)
accuracy.reset_states()
accuracy.update_state(y_true_phones, y_pred_phones)
print('Frame-by-frame accuracy at the phoneme level: {:.2f}%'.format(
accuracy.result().numpy() * 100))
plt.figure()
plot_confusion_matrix(y_true_phones, y_pred_phones)
plt.title('Frame-by-frame confusion matrix at the phoneme level')
class ParallelLSTMTextClassifier(Model):
EPOCHS = 1
logger = logging.getLogger('tensorflow')
logger.setLevel(logging.INFO)
def __init__(self,
lstm_units=100,
char_vocab=26,
char_embed_size=100,
cnn_filters=300,
cnn_kernel_size=5,
dropout_rate=0.2,
max_char_len=10,
lr=1e-4):
super().__init__()
self.char_embedding = Embedding(char_vocab + 1, char_embed_size)
self.word_embedding = tf.Variable(
np.load('../data/embedding.npy'),
dtype=tf.float32,
name='pretrained_embedding',
trainable=False,
)
self.char_cnn = Conv1D(filters=cnn_filters,
kernel_size=cnn_kernel_size,
activation='elu',
padding='same')
self.embed_drop = Dropout(dropout_rate)
self.embed_fc = Dense(cnn_filters, 'elu', name='embed_fc')
self.word_cnn = Conv1D(filters=cnn_filters,
kernel_size=cnn_kernel_size,
activation='elu',
padding='same')
self.word_drop = Dropout(dropout_rate)
self.max_char_len = max_char_len
self.char_embed_size = char_embed_size
self.cnn_filters = cnn_filters
self.attentive_pooling = KernelAttentivePooling(dropout_rate)
self.dropout_l1 = Dropout(dropout_rate)
self.dropout_l2 = Dropout(dropout_rate)
self.blstm_l1 = Bidirectional(LSTM(lstm_units, return_sequences=True))
self.blstm_l2 = Bidirectional(LSTM(lstm_units, return_sequences=True))
self.dropout_op = Dropout(dropout_rate)
self.units = 2 * lstm_units
self.fc = Dense(units=self.units, activation='elu')
self.out_linear = Dense(2)
self.optimizer = Adam(lr)
self.accuracy = Accuracy()
self.decay_lr = tf.optimizers.schedules.ExponentialDecay(
lr, 1000, 0.95)
self.logger = logging.getLogger('tensorflow')
self.logger.setLevel(logging.INFO)
def call(self, inputs, training=False):
words, chars = inputs
if words.dtype != tf.int32:
words = tf.cast(words, tf.int32)
masks = tf.sign(words)
batch_size = tf.shape(words)[0]
word_len = tf.shape(words)[1]
chars = self.char_embedding(chars)
chars = tf.reshape(
chars,
(batch_size * word_len, self.max_char_len, self.char_embed_size))
chars = self.char_cnn(chars)
chars = tf.reduce_max(chars, 1)
chars = tf.reshape(chars, (batch_size, word_len, self.cnn_filters))
words = tf.nn.embedding_lookup(self.word_embedding, words)
x = tf.concat((words, chars), axis=-1)
x = tf.reshape(x, (batch_size * 10 * 10, 10, self.embedding.shape[-1]))
x = self.dropout_l1(x, training=training)
x = self.blstm_l1(x)
x = tf.reduce_max(x, 1)
x = tf.reshape(x, (batch_size * 10, 10, self.units))
x = self.dropout_l2(x, training=training)
x = self.blstm_l2(x)
x = tf.reduce_max(x, 1)
x = tf.reshape(x, (batch_size, 10, self.units))
masks = tf.reshape(tf.sign(tf.reduce_sum(masks, 1)),
(self.batch_size, 10))
x = self.attentive_pooling((x, masks), training=training)
x = self.dropout_op(x, training=training)
x = self.fc(x)
x = self.out_linear(x)
return x
def fit(self, data, epochs=EPOCHS):
t0 = time.time()
step = 0
epoch = 1
while epoch <= epochs:
for texts, labels in data:
with tf.GradientTape() as tape:
logits = self.call(texts, training=True)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=labels,
logits=logits,
label_smoothing=.2,
))
self.optimizer.lr.assign(self.decay_lr(step))
grads = tape.gradient(loss, self.trainable_variables)
grads, _ = tf.clip_by_global_norm(grads, 5.)
self.optimizer.apply_gradients(
zip(grads, self.trainable_variables))
if step % 100 == 0:
self.logger.info(
"Step {} | Loss: {:.4f} | Spent: {:.1f} secs | LR: {:.6f}"
.format(step,
loss.numpy().item(),
time.time() - t0,
self.optimizer.lr.numpy().item()))
t0 = time.time()
step += 1
epoch += 1
return True
def evaluate(self, data):
self.accuracy.reset_states()
for texts, labels in data:
logits = self.call(texts, training=False)
y_pred = tf.argmax(logits, axis=-1)
self.accuracy.update_state(y_true=labels, y_pred=y_pred)
accuracy = self.accuracy.result().numpy()
self.logger.info("Evaluation Accuracy: {:.3f}".format(accuracy))
self.logger.info("Accuracy: {:.3f}".format(accuracy))
def cnn_1d_pipeline(train_data, train_labels, val_data, val_labels, test_data,
test_labels, classes, weights):
"""This function is the main pipeline for the 1D CNN.
Parameters :
------------
xxx_data : np.array
Data of the train, validation and test datasets
xxx_labels : np.array
Labels of the train, validation and test datasets
classes : np.array
Classes of the dataset
weights : np.array
Weights applied for each class during training
Returns :
-----------
Accuracy : float
Precision : float
Recall : float
F1-Score : float
Metrics evaluated on the test set
"""
# Reshaping the data and labels
train_data = train_data.reshape(
train_data.shape[0] * train_data.shape[1] * train_data.shape[2],
train_data.shape[3])
train_labels = train_labels.reshape(
train_labels.shape[0] * train_labels.shape[1] * train_labels.shape[2])
val_data = val_data.reshape(
val_data.shape[0] * val_data.shape[1] * val_data.shape[2],
val_data.shape[3])
val_labels = val_labels.reshape(val_labels.shape[0] * val_labels.shape[1] *
val_labels.shape[2])
test_shape = test_labels.shape
test_data = test_data.reshape(
test_data.shape[0] * test_data.shape[1] * test_data.shape[2],
test_data.shape[3])
test_labels = test_labels.reshape(
test_labels.shape[0] * test_labels.shape[1] * test_labels.shape[2])
#Calculating the weights of each pixel
train_weights = np.ones((train_labels.shape[0]))
for i in range(len(classes)):
train_weights[train_labels == classes[i]] = weights[i]
val_weights = np.ones((val_labels.shape[0]))
for i in range(len(classes)):
val_weights[val_labels == classes[i]] = weights[i]
test_weights = np.ones((test_labels.shape[0]))
for i in range(len(classes)):
test_weights[test_labels == classes[i]] = weights[i]
# One-Hot Encoding the labels
train_labels = tf.one_hot(train_labels, depth=17, dtype=tf.int8).numpy()
val_labels = tf.one_hot(val_labels, depth=17, dtype=tf.int8).numpy()
# Creating a data generator class
class DataGenerator(tf.keras.utils.Sequence):
def __init__(self,
data,
labels,
weights,
batch_size=32,
n_classes=10,
shuffle=True):
'Initialization'
self.data = data
self.labels = labels
self.weights = weights
self.batch_size = batch_size
self.shuffle = shuffle
self.on_epoch_end()
def __len__(self):
'Denotes the number of batches per epoch'
return int(np.floor(len(self.data) / self.batch_size))
def __getitem__(self, index):
'Generate one batch of data'
# Generate indexes of the batch
indexes = self.indexes[index * self.batch_size:(index + 1) *
self.batch_size]
# Get data and labels
data_yield = self.data[indexes]
labels_yield = self.labels[indexes]
weights_yield = self.weights[indexes]
return data_yield, labels_yield, weights_yield
def on_epoch_end(self):
'Updates indexes after each epoch'
self.indexes = np.arange(len(self.data))
if self.shuffle == True:
np.random.shuffle(self.indexes)
## -- Training the model --
batch_size = 3000
steps_per_epoch = 20
train_generator = DataGenerator(train_data,
train_labels,
train_weights,
batch_size,
shuffle=True)
val_generator = DataGenerator(val_data,
val_labels,
val_weights,
batch_size,
shuffle=True)
checkpoint_filepath = "tmp/checkpoint_salinas_convnet"
model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(
filepath=checkpoint_filepath,
save_weights_only=True,
monitor='val_loss',
mode='min',
save_best_only=True)
model = cnn_1D(shape=(train_data.shape[1], 1),
kernel_size=7,
nb_filters_0=64,
nb_dense_neurons=500,
output_channels=17,
kernel_reg=0.0009875,
dense_reg=0.0012559,
dropout=0)
model.compile(loss="categorical_crossentropy",
optimizer=Adam(learning_rate=0.00074582),
metrics=["accuracy"])
print(model.summary())
# Training the model
history = model.fit(train_generator,
steps_per_epoch=steps_per_epoch,
epochs=500,
validation_data=val_generator,
callbacks=[
model_checkpoint_callback,
tf.keras.callbacks.EarlyStopping(
monitor='val_accuracy', patience=30)
])
# Making a prediction on the test data
y_pred = model.predict(test_data, batch_size=batch_size)
y_pred = np.argmax(y_pred[:, 1:], axis=1) + 1
# Calculating accuracy
accuracy = Accuracy()
accuracy.update_state(test_labels, y_pred)
pred_accuracy = accuracy.result()
print(pred_accuracy)
# Computing the confusion matrix
conf = confusion_matrix(test_labels, y_pred)
print(conf)
y_pred = y_pred.reshape(test_shape)
y_pred[test_labels.reshape(test_shape) == 0] = 0
y_pred = y_pred.reshape(test_shape[0] * test_shape[1] * test_shape[2])
# Calculating metrics
accuracy = Accuracy()
accuracy.update_state(test_labels, y_pred, sample_weight=test_weights)
recall = tf.keras.metrics.Recall()
recall.update_state(
tf.one_hot(test_labels, depth=17).numpy().flatten(),
tf.one_hot(y_pred, depth=17).numpy().flatten())
precision = tf.keras.metrics.Precision()
precision.update_state(
tf.one_hot(test_labels, depth=17).numpy().flatten(),
tf.one_hot(y_pred, depth=17).numpy().flatten())
return (accuracy.result().numpy(), precision.result().numpy(),
recall.result().numpy(),
2 * precision.result().numpy() * recall.result().numpy() /
(precision.result().numpy() + recall.result().numpy()))
def price_direction_accuracy(targets_directs, predicts_directs):
accur = Accuracy()
accur.update_state(targets_directs, predicts_directs)
return accur.result().numpy()
def cnn_1d_pipeline(train_data, train_labels, val_data, val_labels, test_data, test_labels, classes, weights):
"""This function is the main pipeline for the 1D CNN.
Parameters :
------------
xxx_data : np.array
Data of the train, validation and test datasets
xxx_labels : np.array
Labels of the train, validation and test datasets
classes : np.array
Classes of the dataset
weights : np.array
Weights applied for each class during training
Returns :
-----------
Accuracy : float
Precision : float
Recall : float
F1-Score : float
Metrics evaluated on the test set
"""
# Reshaping the data
train_data = train_data.reshape(train_data.shape[0]*train_data.shape[1]*train_data.shape[2], train_data.shape[3])
train_labels = train_labels.reshape(train_labels.shape[0]*train_labels.shape[1]*train_labels.shape[2])
val_data = val_data.reshape(val_data.shape[0]*val_data.shape[1]*val_data.shape[2], val_data.shape[3])
val_labels = val_labels.reshape(val_labels.shape[0]*val_labels.shape[1]*val_labels.shape[2])
test_shape = test_labels.shape
test_data = test_data.reshape(test_data.shape[0]*test_data.shape[1]*test_data.shape[2], test_data.shape[3])
test_labels = test_labels.reshape(test_labels.shape[0]*test_labels.shape[1]*test_labels.shape[2])
# Computing the weights of each pixel
train_weights = np.ones((train_labels.shape[0]))
for i in range(len(classes)):
train_weights[train_labels == classes[i]] = weights[i]
val_weights = np.ones((val_labels.shape[0]))
for i in range(len(classes)):
val_weights[val_labels == classes[i]] = weights[i]
test_weights = np.ones((test_labels.shape[0]))
for i in range(len(classes)):
test_weights[test_labels == classes[i]] = weights[i]
# One-Hot Encoding the labels
train_labels = tf.one_hot(train_labels, depth=6, dtype=tf.int8).numpy()
val_labels = tf.one_hot(val_labels, depth=6, dtype=tf.int8).numpy()
# Creating a data generator class
class DataGenerator(tf.keras.utils.Sequence):
def __init__(self, data, labels, weights, batch_size=32, n_classes=10, shuffle=True):
'Initialization'
self.data = data
self.labels = labels
self.weights = weights
self.batch_size = batch_size
self.shuffle = shuffle
self.on_epoch_end()
def __len__(self):
'Denotes the number of batches per epoch'
return int(np.floor(len(self.data) / self.batch_size))
def __getitem__(self, index):
'Generate one batch of data'
# Generate indexes of the batch
indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size]
# Get data and labels
data_yield = self.data[indexes]
labels_yield = self.labels[indexes]
weights_yield = self.weights[indexes]
return data_yield, labels_yield, weights_yield
def on_epoch_end(self):
'Updates indexes after each epoch'
self.indexes = np.arange(len(self.data))
if self.shuffle == True:
np.random.shuffle(self.indexes)
## -- Training the model --
batch_size = 30000
steps_per_epoch = 100
train_generator = DataGenerator(train_data, train_labels, train_weights, batch_size, shuffle=True)
val_generator = DataGenerator(val_data, val_labels, val_weights, batch_size, shuffle=True)
checkpoint_filepath = "tmp/checkpoint_SPARCS_cnn_1D"
model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(
filepath=checkpoint_filepath,
save_weights_only=True,
monitor='val_loss',
mode='min',
save_best_only=True)
model = cnn_1D(shape=(10, 1), kernel_size=3, nb_filters_0=128, nb_dense_neurons=180, kernel_reg=0.0083, dense_reg=0.054, output_channels=6, dropout=0.0, learning_rate=0.000322)
print(model.summary())
# Training the model
history = model.fit(train_generator,
steps_per_epoch=steps_per_epoch,
epochs=500,
validation_data = val_generator,
validation_steps = 5,
callbacks=[EarlyStopping(monitor='val_accuracy', patience=40)]
)
# Making a prediction
y_pred = model.predict(test_data, batch_size=30000)
y_pred = np.argmax(y_pred, axis=1)
# Computing accuracy
accuracy = Accuracy()
accuracy.update_state(test_labels, y_pred, sample_weight=test_weights)
pred_accuracy = accuracy.result()
print(pred_accuracy)
# Computing the confusion matrix
print("Computing confusion matrix")
conf = confusion_matrix(test_labels, y_pred)
print(conf)
np.save("logs/metrics/SPARCS/cnn_1D/confusion_matrix", conf)
y_pred = y_pred.reshape(test_shape)
print(y_pred.shape)
# Plotting result images on the test dataset
c = 0
cmap = plt.get_cmap('viridis', 6)
for image in y_pred[:30]:
plt.imshow(image+1e-5, vmin=0, vmax=6, cmap=cmap)
plt.colorbar()
plt.savefig(f"images/SPARCS/cnn_1D/pred{c}_raw")
plt.clf()
plt.imshow(test_labels.reshape(test_shape)[c]+1e-5, vmin=0, vmax=6, cmap=cmap)
plt.colorbar()
plt.savefig(f"images/SPARCS/cnn_1D/GT{c}")
plt.clf()
c+=1
# Applying a median filter on the results
for i in range(len(y_pred)):
y_pred[i] = medfilt(y_pred[i], kernel_size = 3)
# Plotting the median filtered result images
c = 0
for image in y_pred[:30]:
plt.imshow(image+1e-5, vmin=0, vmax=6, cmap=cmap)
plt.colorbar()
plt.savefig(f"images/SPARCS/cnn_1D/pred{c}_filtered")
plt.clf()
c+=1
# Computing metrics on the test dataset
y_pred = y_pred.reshape(test_shape[0]*test_shape[1]*test_shape[2])
accuracy = Accuracy()
accuracy.update_state(test_labels, y_pred, sample_weight=test_weights)
test_labels = tf.one_hot(test_labels, depth=6).numpy().flatten()
y_pred = tf.one_hot(y_pred, depth=6).numpy().flatten()
recall = tf.keras.metrics.Recall()
recall.update_state(test_labels, y_pred)
precision = tf.keras.metrics.Precision()
precision.update_state(test_labels, y_pred)
# Printing metrics
test_accuracy = accuracy.result().numpy()
test_recall = recall.result().numpy()
test_precision = precision.result().numpy()
test_f1 = 2/(1/test_recall + 1/test_precision)
print("Test accuracy = ", test_accuracy)
print("Test recall =", test_recall)
print("Test precision=", test_precision)
print("Test f1 =", test_f1)
class DualStudent(Model):
""""
Dual Student for Automatic Speech Recognition (ASR).
How to train: 1) set the optimizer by means of compile(), 2) use train()
How to test: use test()
Remarks:
- Do not use fit() by Keras, use train()
- Do not use evaluate() by Keras, use test()
- Compiled metrics and loss (i.e. set by means of compile()) are not used
Original proposal for image classification: https://arxiv.org/abs/1909.01804
"""
def __init__(self,
n_classes,
n_hidden_layers=3,
n_units=96,
consistency_loss='mse',
consistency_scale=10,
stabilization_scale=100,
xi=0.6,
padding_value=0.,
sigma=0.01,
schedule='rampup',
schedule_length=5,
version='mono_directional'):
"""
Constructs a Dual Student model.
:param n_classes: number of classes (i.e. number of units in the last layer of each student)
:param n_hidden_layers: number of hidden layers in each student (i.e. LSTM layers)
:param n_units: number of units for each hidden layer
:param consistency_loss: one of 'mse', 'kl'
:param consistency_scale: maximum value of weight for consistency constraint
:param stabilization_scale: maximum value of weight for stabilization constraint
:param xi: threshold for stable sample
:param padding_value: value used to pad input sequences (used as mask_value for Masking layer)
:param sigma: standard deviation for noisy augmentation
:param schedule: type of schedule for lambdas, one of 'rampup', 'triangular_cycling', 'sinusoidal_cycling'
:param schedule_length:
:param version: one of:
- 'mono_directional': both students have mono-directional LSTM layers
- 'bidirectional: both students have bidirectional LSTM layers
- 'imbalanced': one student has mono-directional LSTM layers, the other one bidirectional
"""
super(DualStudent, self).__init__()
# store parameters
self.n_classes = n_classes
self.padding_value = padding_value
self.n_units = n_units
self.n_hidden_layers = n_hidden_layers
self.xi = xi
self.consistency_scale = consistency_scale
self.stabilization_scale = stabilization_scale
self.sigma = sigma
self.version = version
self.schedule = schedule
self.schedule_length = schedule_length
self._lambda1 = None
self._lambda2 = None
# schedule for lambdas
if schedule == 'rampup':
self.schedule_fn = sigmoid_rampup
elif schedule == 'triangular_cycling':
self.schedule_fn = triangular_cycling
elif schedule == 'sinusoidal_cycling':
self.schedule_fn = sinusoidal_cycling
else:
raise ValueError('Invalid schedule')
# loss
self._loss_cls = SparseCategoricalCrossentropy() # classification loss
self._loss_sta = MeanSquaredError() # stabilization loss
if consistency_loss == 'mse':
self._loss_con = MeanSquaredError() # consistency loss
elif consistency_loss == 'kl':
self._loss_con = KLDivergence()
else:
raise ValueError('Invalid consistency metric')
# metrics for training
self._loss1 = Mean(
name='loss1') # we want to average the loss for each batch
self._loss2 = Mean(name='loss2')
self._loss1_cls = Mean(name='loss1_cls')
self._loss2_cls = Mean(name='loss2_cls')
self._loss1_con = Mean(name='loss1_con')
self._loss2_con = Mean(name='loss2_con')
self._loss1_sta = Mean(name='loss1_sta')
self._loss2_sta = Mean(name='loss2_sta')
self._acc1 = SparseCategoricalAccuracy(name='acc1')
self._acc2 = SparseCategoricalAccuracy(name='acc2')
# metrics for testing
self._test_loss1 = Mean(name='test_loss1')
self._test_loss2 = Mean(name='test_loss2')
self._test_acc1_train_phones = SparseCategoricalAccuracy(
name='test_acc1_train_phones')
self._test_acc2_train_phones = SparseCategoricalAccuracy(
name='test_acc2_train_phones')
self._test_acc1 = Accuracy(name='test_acc1')
self._test_acc2 = Accuracy(name='test_acc2')
self._test_per1 = PhoneErrorRate(name='test_per1')
self._test_per2 = PhoneErrorRate(name='test_per2')
# compose students
if version == 'mono_directional':
lstm_types = ['mono_directional', 'mono_directional']
elif version == 'bidirectional':
lstm_types = ['bidirectional', 'bidirectional']
elif version == 'imbalanced':
lstm_types = ['mono_directional', 'bidirectional']
else:
raise ValueError('Invalid student version')
self.student1 = self._get_student('student1', lstm_types[0])
self.student2 = self._get_student('student2', lstm_types[1])
# masking layer (just to use compute_mask and remove padding)
self.mask = Masking(mask_value=self.padding_value)
def _get_student(self, name, lstm_type):
student = Sequential(name=name)
student.add(Masking(mask_value=self.padding_value))
if lstm_type == 'mono_directional':
for i in range(self.n_hidden_layers):
student.add(LSTM(units=self.n_units, return_sequences=True))
elif lstm_type == 'bidirectional':
for i in range(self.n_hidden_layers):
student.add(
Bidirectional(
LSTM(units=self.n_units, return_sequences=True)))
else:
raise ValueError('Invalid LSTM version')
student.add(Dense(units=self.n_classes, activation="softmax"))
return student
def _noisy_augment(self, x):
return x + tf.random.normal(shape=x.shape, stddev=self.sigma)
def call(self, inputs, training=False, student='student1', **kwargs):
"""
Feed-forwards inputs to one of the students.
This function is called internally by __call__(). Do not use it directly, use the model as callable. You may
prefer to use pad_and_predict() instead of this, because it pads the sequences and splits in batches. For a big
dataset, it is strongly suggested that you use pad_and_predict().
:param inputs: tensor of shape (batch_size, n_frames, n_features)
:param training: boolean, whether the call is in inference mode or training mode
:param student: one of 'student1', 'student2'
:return: tensor of shape (batch_size, n_frames, n_classes), softmax activations (probabilities)
"""
if student == 'student1':
return self.student1(inputs, training=training)
elif student != 'student1':
return self.student2(inputs, training=training)
else:
raise ValueError('Invalid student')
def build(self, input_shape):
super(DualStudent, self).build(input_shape)
self.student1.build(input_shape)
self.student2.build(input_shape)
def train(self,
x_labeled,
x_unlabeled,
y_labeled,
x_val=None,
y_val=None,
n_epochs=10,
batch_size=32,
shuffle=True,
evaluation_mapping=None,
logs_path=None,
checkpoints_path=None,
initial_epoch=0,
seed=None):
"""
Trains the students with both labeled and unlabeled data (semi-supervised learning).
:param x_labeled: numpy array of numpy arrays (n_frames, n_features), features corresponding to y_labeled.
'n_frames' can vary, padding is added to make x_labeled a tensor.
:param x_unlabeled: numpy array of numpy arrays of shape (n_frames, n_features), features without labels.
'n_frames' can vary, padding is added to make x_unlabeled a tensor.
:param y_labeled: numpy array of numpy arrays of shape (n_frames,), labels corresponding to x_labeled.
'n_frames' can vary, padding is added to make y_labeled a tensor.
:param x_val: like x_labeled, but for validation set
:param y_val: like y_labeled, but for validation set
:param n_epochs: integer, number of training epochs
:param batch_size: integer, batch size
:param shuffle: boolean, whether to shuffle at each epoch or not
:param evaluation_mapping: dictionary {training label -> test label}, the test phones should be a subset of the
training phones
:param logs_path: path where to save logs for TensorBoard
:param checkpoints_path: path to a directory. If the directory contains checkpoints, the latest checkpoint is
restored.
:param initial_epoch: int, initial epoch from which to start the training. It can be used together with
checkpoints_path to resume the training from a previous run.
:param seed: seed for the random number generator
"""
# set seed
if seed is not None:
np.random.seed(seed)
tf.random.set_seed(seed)
# show summary
self.build(input_shape=(None, ) + x_labeled[0].shape)
self.student1.summary()
self.student2.summary()
# setup for logs
train_summary_writer = None
if logs_path is not None:
train_summary_writer = tf.summary.create_file_writer(logs_path)
# setup for checkpoints
checkpoint = None
if checkpoints_path is not None:
checkpoint = tf.train.Checkpoint(optimizer=self.optimizer,
model=self)
checkpoint_path = tf.train.latest_checkpoint(checkpoints_path)
if checkpoint_path is not None:
checkpoint.restore(checkpoint_path)
checkpoint_path = Path(checkpoints_path) / 'ckpt'
checkpoint_path = str(checkpoint_path)
# compute batch sizes
labeled_batch_size = ceil(
len(x_labeled) / (len(x_unlabeled) + len(x_labeled)) * batch_size)
unlabeled_batch_size = batch_size - labeled_batch_size
n_batches = min(ceil(len(x_unlabeled) / unlabeled_batch_size),
ceil(len(x_labeled) / labeled_batch_size))
# training loop
for epoch in trange(initial_epoch, n_epochs, desc='epochs'):
# ramp up lambda1 and lambda2
self._lambda1 = self.consistency_scale * self.schedule_fn(
epoch, self.schedule_length)
self._lambda2 = self.stabilization_scale * self.schedule_fn(
epoch, self.schedule_length)
# shuffle training set
if shuffle:
indices = np.arange(
len(x_labeled)
) # get indices to shuffle coherently features and labels
np.random.shuffle(indices)
x_labeled = x_labeled[indices]
y_labeled = y_labeled[indices]
np.random.shuffle(x_unlabeled)
for i in trange(n_batches, desc='batches'):
# select batch
x_labeled_batch = select_batch(x_labeled, i,
labeled_batch_size)
x_unlabeled_batch = select_batch(x_unlabeled, i,
unlabeled_batch_size)
y_labeled_batch = select_batch(y_labeled, i,
labeled_batch_size)
# pad batch
x_labeled_batch = pad_sequences(x_labeled_batch,
padding='post',
value=self.padding_value,
dtype='float32')
x_unlabeled_batch = pad_sequences(x_unlabeled_batch,
padding='post',
value=self.padding_value,
dtype='float32')
y_labeled_batch = pad_sequences(y_labeled_batch,
padding='post',
value=-1)
# convert to tensors
x_labeled_batch = tf.convert_to_tensor(x_labeled_batch)
x_unlabeled_batch = tf.convert_to_tensor(x_unlabeled_batch)
y_labeled_batch = tf.convert_to_tensor(y_labeled_batch)
# train step
self._train_step(x_labeled_batch, x_unlabeled_batch,
y_labeled_batch)
# put metrics in dictionary (easy management)
train_metrics = {
self._loss1.name: self._loss1.result(),
self._loss2.name: self._loss2.result(),
self._loss1_cls.name: self._loss1_cls.result(),
self._loss2_cls.name: self._loss2_cls.result(),
self._loss1_con.name: self._loss1_con.result(),
self._loss2_con.name: self._loss2_con.result(),
self._loss1_sta.name: self._loss1_sta.result(),
self._loss2_sta.name: self._loss2_sta.result(),
self._acc1.name: self._acc1.result(),
self._acc2.name: self._acc2.result(),
}
metrics = {'train': train_metrics}
# test on validation set
if x_val is not None and y_val is not None:
val_metrics = self.test(x_val,
y_val,
evaluation_mapping=evaluation_mapping)
metrics['val'] = val_metrics
# print metrics
for dataset, metrics_ in metrics.items():
print(f'Epoch {epoch + 1} - ', dataset, ' - ', sep='', end='')
for k, v in metrics_.items():
print(f'{k}: {v}, ', end='')
print()
# save logs
if train_summary_writer is not None:
with train_summary_writer.as_default():
for dataset, metrics_ in metrics.items():
for k, v in metrics_.items():
tf.summary.scalar(k, v, step=epoch)
# save checkpoint
if checkpoint is not None:
checkpoint.save(file_prefix=checkpoint_path)
# reset metrics
self._loss1.reset_states()
self._loss2.reset_states()
self._loss1_cls.reset_states()
self._loss2_cls.reset_states()
self._loss1_con.reset_states()
self._loss2_con.reset_states()
self._loss1_sta.reset_states()
self._loss2_sta.reset_states()
self._acc1.reset_states()
self._acc2.reset_states()
"""
If you want to use graph execution, pad the whole dataset externally and uncomment the decorator below.
If you uncomment the decorator without padding the dataset, the graph will be compiled for each batch,
because train() pads at batch level and so the batches have different shapes. This would result in worse
performance compared to eager execution.
"""
# @tf.function
def _train_step(self, x_labeled, x_unlabeled, y_labeled):
# noisy augmented batches (TODO: improvement with data augmentation instead of noise)
B1_labeled = self._noisy_augment(x_labeled)
B2_labeled = self._noisy_augment(x_labeled)
B1_unlabeled = self._noisy_augment(x_unlabeled)
B2_unlabeled = self._noisy_augment(x_unlabeled)
# compute masks (to remove padding)
mask_labeled = self.mask.compute_mask(x_labeled)
mask_unlabeled = self.mask.compute_mask(x_unlabeled)
y_labeled = y_labeled[mask_labeled] # remove padding from labels
# forward pass
with tf.GradientTape(persistent=True) as tape:
# predict augmented labeled samples (for classification and consistency constraint)
prob1_labeled_B1 = self.student1(B1_labeled, training=True)
prob1_labeled_B2 = self.student1(B2_labeled, training=True)
prob2_labeled_B1 = self.student2(B1_labeled, training=True)
prob2_labeled_B2 = self.student2(B2_labeled, training=True)
# predict augmented unlabeled samples (for consistency and stabilization constraints)
prob1_unlabeled_B1 = self.student1(B1_unlabeled, training=True)
prob1_unlabeled_B2 = self.student1(B2_unlabeled, training=True)
prob2_unlabeled_B1 = self.student2(B1_unlabeled, training=True)
prob2_unlabeled_B2 = self.student2(B2_unlabeled, training=True)
# remove padding
prob1_labeled_B1 = prob1_labeled_B1[mask_labeled]
prob1_labeled_B2 = prob1_labeled_B2[mask_labeled]
prob2_labeled_B1 = prob2_labeled_B1[mask_labeled]
prob2_labeled_B2 = prob2_labeled_B2[mask_labeled]
prob1_unlabeled_B1 = prob1_unlabeled_B1[mask_unlabeled]
prob1_unlabeled_B2 = prob1_unlabeled_B2[mask_unlabeled]
prob2_unlabeled_B1 = prob2_unlabeled_B1[mask_unlabeled]
prob2_unlabeled_B2 = prob2_unlabeled_B2[mask_unlabeled]
# compute classification losses
L1_cls = self._loss_cls(y_labeled, prob1_labeled_B1)
L2_cls = self._loss_cls(y_labeled, prob2_labeled_B2)
# concatenate labeled and unlabeled probability predictions (for consistency loss)
prob1_labeled_unlabeled_B1 = tf.concat(
[prob1_labeled_B1, prob1_unlabeled_B1], axis=0)
prob1_labeled_unlabeled_B2 = tf.concat(
[prob1_labeled_B2, prob1_unlabeled_B2], axis=0)
prob2_labeled_unlabeled_B1 = tf.concat(
[prob2_labeled_B1, prob2_unlabeled_B1], axis=0)
prob2_labeled_unlabeled_B2 = tf.concat(
[prob2_labeled_B2, prob2_unlabeled_B2], axis=0)
# compute consistency losses
L1_con = self._loss_con(prob1_labeled_unlabeled_B1,
prob1_labeled_unlabeled_B2)
L2_con = self._loss_con(prob2_labeled_unlabeled_B1,
prob2_labeled_unlabeled_B2)
# prediction
P1_unlabeled_B1 = tf.argmax(prob1_unlabeled_B1, axis=-1)
P1_unlabeled_B2 = tf.argmax(prob1_unlabeled_B2, axis=-1)
P2_unlabeled_B1 = tf.argmax(prob2_unlabeled_B1, axis=-1)
P2_unlabeled_B2 = tf.argmax(prob2_unlabeled_B2, axis=-1)
# confidence (probability of predicted class)
M1_unlabeled_B1 = tf.reduce_max(prob1_unlabeled_B1, axis=-1)
M1_unlabeled_B2 = tf.reduce_max(prob1_unlabeled_B2, axis=-1)
M2_unlabeled_B1 = tf.reduce_max(prob2_unlabeled_B1, axis=-1)
M2_unlabeled_B2 = tf.reduce_max(prob2_unlabeled_B2, axis=-1)
# stable samples (masks to index probabilities)
R1 = tf.logical_and(
P1_unlabeled_B1 == P1_unlabeled_B2,
tf.logical_or(M1_unlabeled_B1 > self.xi,
M1_unlabeled_B2 > self.xi))
R2 = tf.logical_and(
P2_unlabeled_B1 == P2_unlabeled_B2,
tf.logical_or(M2_unlabeled_B1 > self.xi,
M2_unlabeled_B2 > self.xi))
R12 = tf.logical_and(R1, R2)
# stabilities
epsilon1 = MSE(prob1_unlabeled_B1[R12], prob1_unlabeled_B2[R12])
epsilon2 = MSE(prob2_unlabeled_B1[R12], prob2_unlabeled_B2[R12])
# compute stabilization losses
L1_sta = self._loss_sta(
prob1_unlabeled_B1[R12][epsilon1 > epsilon2],
prob2_unlabeled_B1[R12][epsilon1 > epsilon2])
L2_sta = self._loss_sta(
prob1_unlabeled_B2[R12][epsilon1 < epsilon2],
prob2_unlabeled_B2[R12][epsilon1 < epsilon2])
L1_sta += self._loss_sta(
prob1_unlabeled_B1[tf.logical_and(tf.logical_not(R1), R2)],
prob2_unlabeled_B1[tf.logical_and(tf.logical_not(R1), R2)])
L2_sta += self._loss_sta(
prob1_unlabeled_B2[tf.logical_and(R1, tf.logical_not(R2))],
prob2_unlabeled_B2[tf.logical_and(R1, tf.logical_not(R2))])
# compute complete losses
L1 = L1_cls + self._lambda1 * L1_con + self._lambda2 * L1_sta
L2 = L2_cls + self._lambda1 * L2_con + self._lambda2 * L2_sta
# backward pass
gradients1 = tape.gradient(L1, self.student1.trainable_variables)
gradients2 = tape.gradient(L2, self.student2.trainable_variables)
self.optimizer.apply_gradients(
zip(gradients1, self.student1.trainable_variables))
self.optimizer.apply_gradients(
zip(gradients2, self.student2.trainable_variables))
del tape # to release memory (persistent tape)
# update metrics
self._loss1.update_state(L1)
self._loss2.update_state(L2)
self._loss1_cls.update_state(L1_cls)
self._loss2_cls.update_state(L2_cls)
self._loss1_con.update_state(L1_con)
self._loss2_con.update_state(L2_con)
self._loss1_sta.update_state(L1_sta)
self._loss2_sta.update_state(L2_sta)
self._acc1.update_state(y_labeled, prob1_labeled_B1)
self._acc2.update_state(y_labeled, prob2_labeled_B2)
def test(self, x, y, batch_size=32, evaluation_mapping=None):
"""
Tests the model (both students).
:param x: numpy array of numpy arrays (n_frames, n_features), features corresponding to y_labeled.
'n_frames' can vary, padding is added to make x a tensor.
:param y: numpy array of numpy arrays of shape (n_frames,), labels corresponding to x_labeled.
'n_frames' can vary, padding is added to make y a tensor.
:param batch_size: integer, batch size
:param evaluation_mapping: dictionary {training label -> test label}, the test phones should be a subset of the
training phones
:return: dictionary {metric_name -> value}
"""
# test batch by batch
n_batches = ceil(len(x) / batch_size)
for i in trange(n_batches, desc='test batches'):
# select batch
x_batch = select_batch(x, i, batch_size)
y_batch = select_batch(y, i, batch_size)
# pad batch
x_batch = pad_sequences(x_batch,
padding='post',
value=self.padding_value,
dtype='float32')
y_batch = pad_sequences(y_batch, padding='post', value=-1)
# convert to tensors
x_batch = tf.convert_to_tensor(x_batch)
y_batch = tf.convert_to_tensor(y_batch)
# test step
self._test_step(x_batch, y_batch, evaluation_mapping)
# put metrics in dictionary (easy management)
test_metrics = {
self._test_loss1.name:
self._test_loss1.result(),
self._test_loss2.name:
self._test_loss2.result(),
self._test_acc1_train_phones.name:
self._test_acc1_train_phones.result(),
self._test_acc2_train_phones.name:
self._test_acc2_train_phones.result(),
self._test_acc1.name:
self._test_acc1.result(),
self._test_acc2.name:
self._test_acc2.result(),
self._test_per1.name:
self._test_per1.result(),
self._test_per2.name:
self._test_per2.result(),
}
# reset metrics
self._test_loss1.reset_states()
self._test_loss2.reset_states()
self._test_acc1_train_phones.reset_states()
self._test_acc2_train_phones.reset_states()
self._test_acc1.reset_states()
self._test_acc2.reset_states()
self._test_per1.reset_states()
self._test_per2.reset_states()
return test_metrics
# @tf.function # see note in _train_step()
def _test_step(self, x, y, evaluation_mapping):
# compute mask (to remove padding)
mask = self.mask.compute_mask(x)
# forward pass
y_prob1_train_phones = self.student1(x, training=False)
y_prob2_train_phones = self.student2(x, training=False)
y_pred1_train_phones = tf.argmax(y_prob1_train_phones, axis=-1)
y_pred2_train_phones = tf.argmax(y_prob2_train_phones, axis=-1)
y_train_phones = tf.identity(y)
# map labels to set of test phones
if evaluation_mapping is not None:
y = tf.numpy_function(map_labels,
[y_train_phones, evaluation_mapping],
[tf.float32])
y_pred1 = tf.numpy_function(
map_labels, [y_pred1_train_phones, evaluation_mapping],
[tf.float32])
y_pred2 = tf.numpy_function(
map_labels, [y_pred2_train_phones, evaluation_mapping],
[tf.float32])
else:
y = y_train_phones
y_pred1 = y_pred1_train_phones
y_pred2 = y_pred2_train_phones
# update phone error rate
self._test_per1.update_state(y, y_pred1, mask)
self._test_per2.update_state(y, y_pred2, mask)
# remove padding
y_pred1 = y_pred1[mask]
y_pred2 = y_pred2[mask]
y_prob1_train_phones = y_prob1_train_phones[mask]
y_prob2_train_phones = y_prob2_train_phones[mask]
y_train_phones = y_train_phones[mask]
y = y[mask]
# compute loss
loss1 = self._loss_cls(y_train_phones, y_prob1_train_phones)
loss2 = self._loss_cls(y_train_phones, y_prob2_train_phones)
# update loss
self._test_loss1.update_state(loss1)
self._test_loss2.update_state(loss2)
# update accuracy using training phones
self._test_acc1_train_phones.update_state(y_train_phones,
y_prob1_train_phones)
self._test_acc2_train_phones.update_state(y_train_phones,
y_prob2_train_phones)
# update accuracy using test phones
self._test_acc1.update_state(y, y_pred1)
self._test_acc2.update_state(y, y_pred2)
class ParallelLSTMTextClassifier(Model):
EPOCHS = 1
logger = logging.getLogger('tensorflow')
logger.setLevel(logging.INFO)
def __init__(self, lstm_units=100, lr=1e-4, dropout_rate=0.2):
super().__init__()
self.embedding = tf.Variable(np.load('../data/embedding.npy'),
dtype=tf.float32,
name='pretrained_embedding',
trainable=False)
self.dropout_l1 = Dropout(dropout_rate)
self.dropout_l2 = Dropout(dropout_rate)
self.dropout_l3 = Dropout(dropout_rate)
self.blstm_l1 = Bidirectional(LSTM(lstm_units, return_sequences=True))
self.blstm_l2 = Bidirectional(LSTM(lstm_units, return_sequences=True))
self.blstm_l3 = Bidirectional(LSTM(lstm_units, return_sequences=True))
self.dropout_op = Dropout(dropout_rate)
self.units = 2 * lstm_units
self.fc = Dense(units=self.units, activation='elu')
self.out_linear = Dense(2)
self.optimizer = Adam(lr)
self.decay_lr = ExponentialDecay(lr, 1000, 0.90)
self.accuracy = Accuracy()
self.logger = logging.getLogger('tensorflow')
self.logger.setLevel(logging.INFO)
def call(self, inputs, training=False):
if inputs.dtype != tf.int32:
inputs = tf.cast(inputs, tf.int32)
batch_size = tf.shape(inputs)[0]
x = tf.nn.embedding_lookup(self.embedding, inputs)
x = tf.reshape(x, (batch_size * 10 * 10, 10, self.embedding.shape[-1]))
x = self.dropout_l1(x, training=training)
x = self.blstm_l1(x)
x = tf.reduce_max(x, 1)
x = tf.reshape(x, (batch_size * 10, 10, self.units))
x = self.dropout_l2(x, training=training)
x = self.blstm_l2(x)
x = tf.reduce_max(x, 1)
x = tf.reshape(x, (batch_size, 10, self.units))
x = self.dropout_l3(x, training=training)
x = self.blstm_l3(x)
x = tf.reduce_max(x, 1)
x = self.dropout_op(x, training=training)
x = self.fc(x)
x = self.out_linear(x)
return x
def fit(self, data, epochs=EPOCHS):
t0 = time.time()
step = 0
epoch = 1
while epoch <= epochs:
for texts, labels in data:
with tf.GradientTape() as tape:
logits = self.call(texts, training=True)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=labels,
logits=logits,
label_smoothing=.2,
))
self.optimizer.lr.assign(self.decay_lr(step))
grads = tape.gradient(loss, self.trainable_variables)
grads, _ = tf.clip_by_global_norm(grads, 5.)
self.optimizer.apply_gradients(
zip(grads, self.trainable_variables))
if step % 100 == 0:
self.logger.info(
"Step {} | Loss: {:.4f} | Spent: {:.1f} secs | LR: {:.6f}"
.format(step,
loss.numpy().item(),
time.time() - t0,
self.optimizer.lr.numpy().item()))
t0 = time.time()
step += 1
epoch += 1
return True
def evaluate(self, data):
self.accuracy.reset_states()
for texts, labels in data:
logits = self.call(texts, training=False)
y_pred = tf.argmax(logits, axis=-1)
self.accuracy.update_state(y_true=labels, y_pred=y_pred)
accuracy = self.accuracy.result().numpy()
self.logger.info("Evaluation Accuracy: {:.3f}".format(accuracy))
self.logger.info("Accuracy: {:.3f}".format(accuracy))