def Highway1(network, num_out, drop_prob=1.0):
dense1 = tflearn.fully_connected(network, 64, activation='elu', regularizer='L2', weight_decay=0.001)
highway = dense1
for i in range(10):
highway = tflearn.highway(highway, 64, activation='elu',regularizer='L2', weight_decay=0.001, transform_dropout=0.7)
network = tflearn.fully_connected(highway, num_out, activation='softmax')
return network
def Highway1(network, scale=False):
if scale is True:
network = scale(network)
dense1 = tflearn.fully_connected(network, 64, activation='elu', regularizer='L2', weight_decay=0.001)
highway = dense1
for i in range(10):
highway = tflearn.highway(highway, 64, activation='elu',regularizer='L2', weight_decay=0.001, transform_dropout=0.7)
network = tflearn.fully_connected(highway, 3, activation='sigmoid')
return network
def do_dnn(): #optimizer, loss, activation):
# print(y1.shape, testy1.shape)
#data_prep = tflearn.data_preprocessing.DataPreprocessing()
#data_prep.add_samplewise_zero_center()
#data_prep.add_samplewise_stdnorm()
# Building deep neural network
xshape = X.shape
activation = "relu6"
input_layer = tflearn.input_data(shape=[None, xshape[1]])# , data_preprocessing=data_prep)
dense1 = tflearn.fully_connected(input_layer, 40, activation=activation,
regularizer='L2', weight_decay=0.001)
# install a deep network of highway layers
highway = dense1
for i in range(1):
highway = tflearn.highway(highway, 40, activation=activation,
regularizer='L2', weight_decay=0.001, transform_dropout=0.726)
softmax = tflearn.fully_connected(highway, 4, activation="softmax")
#final = tflearn.layers.merge_outputs(highway)
# Regression using SGD with learning rate decay and accuracy
# sgd = tflearn.SGD(learning_rate=.001, lr_decay=0.96, decay_step=1000)
# top_k = tflearn.metrics.Top_k(1)
# lossfunc = tflearn.objectives.softmax_categorical_crossentropy(softmax, )
#adam = tflearn.optimizers.Adam(learning_rate=.001, beta1=.9, beta2=.99, epsilon=math.pow(10, -8))
# .001 is default, radically raised? Might be wrong. .1 and .00001 didn't work with 10k batches
adagrad = tflearn.optimizers.AdaGrad(learning_rate=.036)
# must use batchsize = 1
top3 = tflearn.metrics.Top_k(3)
net = tflearn.regression(softmax, optimizer=adagrad, metric="accuracy", loss="mean_square")# , #n_classes=11,
# to_one_hot=True, batch_size=100000) # metric=normal_R2 #categorical_crossentropy
# Training
model = tflearn.DNN(net, tensorboard_verbose=0, checkpoint_path='/home/thorbinator/PycharmProjects/production/saves/save',
best_checkpoint_path='/home/thorbinator/PycharmProjects/production/saves/best',
max_checkpoints=4, best_val_accuracy=.31)
# model.load('./saves/save-30000')
#print (final.get_shape().as_list() , y1.get_shape().as_list(), X.get_shape().as_list())
model.fit(X, Y, validation_set=(testX, testY), n_epoch=99999, show_metric=True, run_id="th_run",
snapshot_epoch=True, batch_size=1, snapshot_step=75000)
train_y = y_frames[split:]
print x_frames.shape
print x_frames.shape[1], x_frames.shape[2]
net = tflearn.input_data([None, x_frames.shape[1], x_frames.shape[2]])
print net.get_shape().as_list()
net = bidirectional_rnn(net, BasicLSTMCell(number_hidden),
BasicLSTMCell(number_hidden))
net = dropout(net, 0.8)
fc = tflearn.fully_connected(net, highway_size, activation='elu',
regularizer='L2', weight_decay=0.001)
net = fc
for i in xrange(highway_layer_amount):
net = tflearn.highway(net, highway_size, activation='elu',
regularizer='L2', weight_decay=0.001,
transform_dropout=0.8)
net = tflearn.fully_connected(net, y_frames.shape[1], activation='elu')
net = tflearn.regression(net, optimizer='adam', learning_rate=learning_rate,
loss='mean_square')
class MonitorCallback(tflearn.callbacks.Callback):
def __init__(self, model, tf_id):
self.lowest_loss = np.inf
self.number_saves = 0
self.model = model
self.tf_id = tf_id
X, Y, testX, testY = mnist.load_data(one_hot=True)
# Building deep neural network
input_layer = tflearn.input_data(shape=[None, 784])
dense1 = tflearn.fully_connected(input_layer,
64,
activation='elu',
regularizer='L2',
weight_decay=0.001)
#install a deep network of highway layers
highway = dense1
for i in range(10):
highway = tflearn.highway(highway,
64,
activation='elu',
regularizer='L2',
weight_decay=0.001,
transform_dropout=0.8)
softmax = tflearn.fully_connected(highway, 10, activation='softmax')
# Regression using SGD with learning rate decay and Top-3 accuracy
sgd = tflearn.SGD(learning_rate=0.1, lr_decay=0.96, decay_step=1000)
top_k = tflearn.metrics.Top_k(3)
net = tflearn.regression(softmax,
optimizer=sgd,
metric=top_k,
loss='categorical_crossentropy')
# Training
model = tflearn.DNN(net, tensorboard_verbose=0)
X, Y, testX, testY = mnist.load_data(one_hot=True)
# Building deep neural network
input_layer = tflearn.input_data(shape=[None, 784])
dense1 = tflearn.fully_connected(input_layer,
64,
activation='elu',
regularizer='L2',
weight_decay=0.001)
#install a deep network of highway layers
highway = dense1
for i in range(10):
highway = tflearn.highway(highway,
64,
activation='elu',
regularizer='L2',
weight_decay=0.001)
dropout2 = tflearn.dropout(highway, 0.5)
softmax = tflearn.fully_connected(dropout2, 10, activation='softmax')
# Regression using SGD with learning rate decay and Top-3 accuracy
sgd = tflearn.SGD(learning_rate=0.1, lr_decay=0.96, decay_step=1000)
top_k = tflearn.metrics.Top_k(3)
net = tflearn.regression(softmax,
optimizer=sgd,
metric=top_k,
loss='categorical_crossentropy')
# Training
def modelevaluation():
input_layer = tflearn.input_data(shape=[None, 392]) # , data_preprocessing=data_prep)
dense1 = tflearn.fully_connected(input_layer, 100, activation=activation,
regularizer='L2', weight_decay=0.001)
# install a deep network of highway layers
highway = dense1
for i in range(100):
highway = tflearn.highway(highway, 100, activation=activation,
regularizer='L2', weight_decay=0.001, transform_dropout=0.8)
softmax = tflearn.fully_connected(highway, 10, activation="softmax")
# final = tflearn.layers.merge_outputs(highway)
# Regression using SGD with learning rate decay and accuracy
# sgd = tflearn.SGD(learning_rate=.001, lr_decay=0.96, decay_step=1000)
# top_k = tflearn.metrics.Top_k(1)
# lossfunc = tflearn.objectives.softmax_categorical_crossentropy(softmax, )
adam = tflearn.optimizers.Adam(learning_rate=.00001, beta1=.9, beta2=.99, epsilon=math.pow(10, -8))
# .001 is default, radically raised? Might be wrong. .1 and .00001 didn't work with 10k batches
top3 = tflearn.metrics.Top_k(3)
net = tflearn.regression(softmax, optimizer=adam, metric='accuracy', loss="categorical_crossentropy") # , #n_classes=11,
# to_one_hot=True, batch_size=100000) # metric=normal_R2
# Training
model = tflearn.DNN(net, tensorboard_verbose=0,
# checkpoint_path='/home/thorbinator/PycharmProjects/production/saves/save',
best_checkpoint_path='/home/thorbinator/PycharmProjects/production/saves/best',
max_checkpoints=4, best_val_accuracy=.31)
model.load('./saves/save-300000')
minitestY, minitestX = [], []
for index, value in enumerate(testY):
if index < 100: minitestY.append(value)
for index, value in enumerate(testX):
if index < 100: minitestX.append(value)
minitestY, minitestX = np.array(minitestY), np.array(minitestX)
results = model.evaluate(minitestX, minitestY,
batch_size=1)
for key in sorted(results):
print(key)
updownscore = 0
score4bin = 0
bad4bin = 0
offbyone = 0
predicted = []
offbytwo = 0
zerobound = .0001 # 0.000550902932254
for index, result in enumerate(model.predict(minitestX)):
if label < 0:
if label < -zerobound:
label_mod = 0 # "bigdown"
else:
label_mod = 1 # "down"
else:
if label > zerobound:
label_mod = 3 # "bigup"
else:
label_mod = 2 # "up"
if result < 0:
if result < -zerobound:
pred4bin = 0 # "bigdown"
else:
pred4bin = 1 # "down"
else:
if result > zerobound:
pred4bin = 3 # "bigup"
else:
pred4bin = 2 # "up"
if pred4bin == label_mod:
score4bin += 1
if abs(pred4bin - label_mod) == 3:
bad4bin += 1
if abs(pred4bin - label_mod) == 2:
offbytwo += 1
if abs(pred4bin - label_mod) == 1:
offbyone += 1
if (label > 0) == (result > 0):
updownscore += 1
if index % 100 == 0: print(label, ":", result, ":", label_mod, ":", pred4bin)
predicted.extend(result)
print("updown correctness:", updownscore / len(testY))
print("4 bins correctness:", score4bin / len(testY))
print("off by one category:", offbyone / len(testY))
print("off by two categories:", offbytwo / len(testY))
print("really bad predictions:", bad4bin / len(testY))
sns.set(style="darkgrid")
sns.jointplot(np.array(predicted), testY, kind="reg", color="r", xlim=(-.01, .01), ylim=(-.01, .01), size=40,
space=0).plot_marginals(sns.distplot, bins=40)
plt.show()
import tflearn
# Data loading and preprocessing
import tflearn.datasets.mnist as mnist
X, Y, testX, testY = mnist.load_data(one_hot=True)
# Building deep neural network
input_layer = tflearn.input_data(shape=[None, 784])
dense1 = tflearn.fully_connected(input_layer, 64, activation='elu',
regularizer='L2', weight_decay=0.001)
#install a deep network of highway layers
highway = dense1
for i in range(10):
highway = tflearn.highway(highway, 64, activation='elu',
regularizer='L2', weight_decay=0.001)
dropout2 = tflearn.dropout(highway, 0.5)
softmax = tflearn.fully_connected(dropout2, 10, activation='softmax')
# Regression using SGD with learning rate decay and Top-3 accuracy
sgd = tflearn.SGD(learning_rate=0.1, lr_decay=0.96, decay_step=1000)
top_k = tflearn.metrics.Top_k(3)
net = tflearn.regression(softmax, optimizer=sgd, metric=top_k,
loss='categorical_crossentropy')
# Training
model = tflearn.DNN(net, tensorboard_verbose=0)
model.fit(X, Y, n_epoch=20, validation_set=(testX, testY),
show_metric=True, run_id="highway_dense_model")
elif evSample > 1000:
labels[i] = 2
elif evSample > 500:
labels[i] = 1
else:
labels[i] = 0
TrainingSetFeatures = preprocessor(TrainingSetFeatures)
categorizeLabels(TrainingSetLabels)
TrainingSetLabels = to_categorical(TrainingSetLabels, 9)
#create a test set from the number of samples and traning set
net = tflearn.input_data(shape=[None, 12])
net = tflearn.fully_connected(net, 32)
net = tflearn.highway(net, 32, activation="LeakyReLu", name="ReLuLayer")
net = tflearn.fully_connected(net, 32)
net = tflearn.fully_connected(net, 9, activation="softmax")
net = tflearn.regression(net, learning_rate=0.005)
#adam = tflearn.Optimizer()
#net = tflearn.regression(net, learning_rate=0.001, optimizer=adam)
# Define model
model = tflearn.DNN(net,
clip_gradients=1.0,
tensorboard_verbose=3,
tensorboard_dir='./tmp/weather1.log')
# Start training (apply gradient descent algorithm)
model.fit(TrainingSetFeatures,
else:
labels[i] = 0
TrainingSetFeatures = preprocessor(TrainingSetFeatures)
categorizeLabels(TrainingSetLabels)
TrainingSetLabels = to_categorical(TrainingSetLabels, 5)
#create a test set from the number of samples and traning set
net = tflearn.input_data(shape=[None, 12])
net = tflearn.fully_connected(net,
32,
weights_init='xavier',
activation="softsign",
name='First_Fully_Connected')
net = tflearn.highway(net, 32, activation="softsign", name="highwayLayer")
net = tflearn.fully_connected(net,
32,
weights_init='xavier',
activation="softsign",
name='Third_Fully_Connected')
net = tflearn.fully_connected(net,
5,
activation="softmax",
name='Final_Fully_Connected')
# todo: confusion matrix
adam = tflearn.Adam()
net = tflearn.regression(net, learning_rate=0.001, optimizer=adam)
# Define model
