validation_accuracy,
finished=True)
if show_confusion_matrix:
predicted = np.argmax(output, axis=1)
actual = np.argmax(val_labels, axis=1)
output = [[] for _ in range(num_classes)]
count_matrix = [[0] * num_classes for _ in range(num_classes)]
for i in range(len(actual)):
output[actual[i]].append(predicted[i])
count_matrix[actual[i]][predicted[i]] += 1
print
print
max_length = max(
2, len(str(max(map(len, validation_sequence_groups)))))
print ' ' * (max_length - 2) + 'Predicted ', ''.join(
map(lambda x: str(x) + ' ' * (max_length - len(str(x)) + 1),
range(num_classes)))
print 'Actual\t', '-' * ((num_classes + 1) * (max_length + 1) + 1)
for i in range(num_classes):
print ' '*(5-len(str(i))), i, '\t|' + ' ' * (max_length - 1), \
''.join(map(lambda x: (str(x) if x else '-') + ' ' * (max_length - len(str(x)) + 1), \
count_matrix[i])) + '| ', np.mean(np.array(output[i]) == i)
reprint_header = (
epoch +
1) % 10 == 0 and epoch < num_epochs - 1 or show_confusion_matrix
table.finalize(divider=not reprint_header)
if reprint_header:
table.print_header()
def train_net(filename, overwrite, should_graph):
# If overwrite is false and a saved model already exists, then simply
# return
if not overwrite and model_exists():
print("Model already exists (did you mean to include --overwrite?)")
return
# Otherwise, we train and save the model
# Load the data
data, labels = load_data(filename)
# Create batches
assert validation_set_size <= len(
data), 'validation_set_size must be smaller than len(data)'
training_data = data[:len(data) - validation_set_size]
training_labels = labels[:len(labels) - validation_set_size]
validation_data = data[len(data) - validation_set_size:]
validation_labels = labels[len(labels) - validation_set_size:]
#print('Training data: ' + str(len(training_data)))
#print('Validation data: ' + str(len(validation_data)))
assert float(
len(training_data)
) / batch_size % 1 == 0, 'batch_size must evenly divide len(training_data)'
#assert float(validation_set_size) / batch_size % 1 == 0, 'batch_size must evenly divide validation_set_size'
num_training_batches = len(training_data) / batch_size
#num_validation_batches = validation_set_size / batch_size
num_validation_batches = 1
training_data_batches = []
training_label_batches = []
validation_data_batches = []
validation_label_batches = []
for i in range(num_training_batches):
training_data_batches.append(training_data[i * batch_size:(i + 1) *
batch_size])
training_label_batches.append(training_labels[i * batch_size:(i + 1) *
batch_size])
for i in range(num_validation_batches):
validation_data_batches.append(validation_data[i * batch_size:(i + 1) *
batch_size])
validation_label_batches.append(
validation_labels[i * batch_size:(i + 1) * batch_size])
# Build model and get variable handles
train_op, x, y, out, loss, accuracy, weights, biases = model(
learning_rate, dropout)
# Initialize environment
initialize = tf.global_variables_initializer()
# Session config
config = tf.ConfigProto(device_count={'GPU': 1 if use_GPU == True else 0})
# Run model
done = False
epoch = 0
iteration = 0
sustained_loss = 0.0
loss_values = []
validation_accuracy_values = []
max_accuracy_values = []
max_accuracy = 0.0
max_accuracy_weights = None
max_accuracy_biases = None
with tf.Session(config=config) as session:
session.run(initialize)
print '=========='
print 'GPU ' + ('enabled' if use_GPU else 'disabled')
print
# Show weight initialization
if show_weights:
weights_val = session.run(weights)
display_weights(weights_val)
layout = [
dict(name='Ep.', width=4, align='center'),
dict(name='Batch',
width=2 * len(str(num_training_batches)) + 1,
suffix='/' + str(num_training_batches)),
dict(name='Loss', width=8),
dict(name='Val Acc', width=6, suffix='%'),
dict(name='Max Acc', width=6, suffix='%'),
dict(name='Time', width=progress_bar_size + 2, align='center'),
]
table = DynamicConsoleTable(layout)
table.print_header()
while not done:
epoch += 1
# Trains on the data, in batches
for i in range(num_training_batches):
iteration += 1
data_batch = training_data_batches[i]
label_batch = training_label_batches[i]
_, loss_val = session.run([train_op, loss],
feed_dict={
x: data_batch,
y: label_batch
})
sustained_loss = decay_rate * sustained_loss + (
1.0 - decay_rate) * loss_val
if len(loss_values) == loss_saved_iterations:
loss_values.pop(0)
if iteration >= loss_starting_iteration:
loss_values.append(loss_val)
data_batch = validation_data_batches[iteration %
num_validation_batches]
label_batch = validation_label_batches[iteration %
num_validation_batches]
validation_accuracy = 0.0
for j in range(num_validation_batches):
data_batch = validation_data_batches[j]
label_batch = validation_label_batches[j]
accuracy_val, out_val = session.run([accuracy, out],
feed_dict={
x: data_batch,
y: label_batch
})
if epoch >= max_epochs:
for j in range(len(label_batch)):
print label_batch[j], np.argmax(out_val[j])
validation_accuracy += accuracy_val
validation_accuracy /= num_validation_batches
if len(validation_accuracy_values
) == accuracy_saved_iterations:
validation_accuracy_values.pop(0)
if iteration >= accuracy_starting_iteration:
validation_accuracy_values.append(validation_accuracy)
if validation_accuracy > max_accuracy:
weights_val, biases_val = session.run([weights, biases])
max_accuracy = validation_accuracy
max_accuracy_weights = weights_val
max_accuracy_biases = biases_val
# Save weights
tf.add_to_collection("vars", out)
tf.add_to_collection("vars", x)
saver = tf.train.Saver()
saver.save(session, 'model')
if len(max_accuracy_values) == accuracy_saved_iterations:
max_accuracy_values.pop(0)
if iteration >= accuracy_starting_iteration:
max_accuracy_values.append(max_accuracy)
progress = int(
math.ceil(progress_bar_size * float(
(iteration - 1) % num_training_batches) /
max(1, num_training_batches - 1)))
progress_string = '[' + '#' * progress + ' ' * (
progress_bar_size - progress) + ']'
if iteration % num_training_batches == 0:
progress_string = time.strftime("%I:%M:%S %p",
time.localtime())
table.update(epoch, (iteration - 1) % num_training_batches + 1,
sustained_loss, validation_accuracy * 100,
max_accuracy * 100, progress_string)
# Termination condition
if sustained_loss < loss_threshold:
done = True
break
update_output(iteration, weights_val, loss_values,
validation_accuracy_values, max_accuracy_values)
table.finalize()
# Termination condition
if epoch >= max_epochs or sustained_loss < loss_threshold:
done = True
update_output(iteration,
weights_val,
loss_values,
validation_accuracy_values,
max_accuracy_values,
override=True)
n = 100
plt.figure('FC weights')
#plt.plot(weights_val['out'][:,1])
plt.plot(
np.sum([
max_accuracy_weights['out'][k * n:(k + 1) * n, 1]
for k in range(num_kernels)
],
axis=0))
#plt.plot(np.sum([weights_val['out'][0:1*n,1], weights_val['out'][1*n:2*n,1], weights_val['out'][2*n:3*n,1], weights_val['out'][3*n:4*n,1], weights_val['out'][4*n:5*n,1], weights_val['out'][5*n:6*n,1]], axis=0))
plt.show()
plt.pause(0)
def train(self,
training_data,
training_labels,
validation_data=None,
validation_labels=None,
skip_evaluation=False,
loss_fn=None,
optimizer_fn=None,
accuracy_fn=None,
max_epochs=float('inf'),
batch_size=1,
validation_set_size=None,
validation_interval=1,
loss_threshold=0.0,
sustained_loss_decay_rate=0.9,
row_output_interval=None):
assert loss_fn, 'Must specify a loss_fn (a function that takes (out, y) as input)'
assert optimizer_fn, 'Must specify a optimizer_fn (a function that takes loss as input)'
if validation_data is not None:
validation_set_size = (validation_set_size and \
min(validation_set_size, len(validation_data))) or \
len(validation_data)
assert len(training_data) == len(training_labels), \
'Number of training data and training labels do not match'
if not skip_evaluation and (validation_data is not None
or validation_labels is not None):
assert validation_data is not None and validation_labels is not None and \
len(validation_data) == len(validation_labels), \
'Number of validation data and validation labels do not match'
else:
skip_evaluation = True
if not skip_evaluation:
assert accuracy_fn, \
'Must specify an accuracy_fn (a function that takes (out, y) as input),' + \
' in order to evaluate the validation set'
if len(training_data) % batch_size != 0:
print 'WARNING: batch_size does not evenly divide len(training_data).' + \
'Some training data will be unused'
validation_data = np.array(validation_data)
validation_labels = np.array(validation_labels)
num_training_batches = len(training_data) / batch_size
training_data_indices = np.random.permutation(len(training_data))
training_data_permuted = np.array(training_data)[training_data_indices]
training_labels_permuted = np.array(
training_labels)[training_data_indices]
training_data_batches = []
training_label_batches = []
for i in range(num_training_batches):
training_data_batches.append(
training_data_permuted[i * batch_size:(i + 1) * batch_size])
training_label_batches.append(
training_labels_permuted[i * batch_size:(i + 1) * batch_size])
row_output_interval = row_output_interval or num_training_batches
y = tf.placeholder(tf.int64, [None])
loss = loss_fn(self.out, y)
optimizer = optimizer_fn(loss)
accuracy = accuracy_fn(self.out, y) if accuracy_fn else None
self.skip_evaluation = skip_evaluation
layout = [
dict(name='Ep.', width=3, align='center'),
dict(name='Batch', width=2*len(str(num_training_batches))+1,
suffix='/'+str(num_training_batches), align='center'),
dict(name='Loss', width=8, align='center')] + \
([dict(name='Val Acc', width=7, suffix='%', align='center'),
dict(name='Max Acc', width=7, suffix='%', align='center')] \
if not self.skip_evaluation else []) + \
[dict(name='Progress/Timestamp', width=self.progress_bar_size+2, align='center'),
dict(name='Elapsed (s)', width=7, align='center')]
# Initialize environment
initialize = tf.global_variables_initializer()
# Session config
config = tf.ConfigProto(
device_count={'GPU': 1 if self.use_gpu == True else 0})
# Run model
done = False
epoch = 0
iteration = 0
sustained_loss = 0.0
loss_values = []
sustained_loss_values = []
last_validation_accuracy = 0.0
validation_accuracy_values = []
max_accuracy_values = []
max_accuracy = 0.0
start_time = time.time()
with tf.Session(config=config) as session:
session.run(initialize)
print '=========='
print 'GPU ' + ('enabled' if self.use_gpu else 'disabled')
print
table = DynamicConsoleTable(layout)
table.print_header()
multiple_rows_per_epoch = row_output_interval < num_training_batches
while not done:
epoch += 1
if self.use_sound:
self.sounds.alert()
# Trains on the data, in batches
for i in range(num_training_batches):
iteration += 1
data_batch = training_data_batches[i]
labels_batch = training_label_batches[i]
_, loss_val = session.run([optimizer, loss],
feed_dict={
self.x: data_batch,
y: labels_batch
})
sustained_loss = sustained_loss_decay_rate * sustained_loss + \
(1.0 - sustained_loss_decay_rate) * loss_val
if len(loss_values) == self.loss_display_saved_iterations:
loss_values.pop(0)
sustained_loss_values.pop(0)
if iteration == self.loss_display_starting_iteration:
sustained_loss = loss_val
if iteration >= self.loss_display_starting_iteration:
loss_values.append(loss_val)
sustained_loss_values.append(sustained_loss)
validation_accuracy = last_validation_accuracy
if not skip_evaluation and iteration % validation_interval == 0:
validation_set_indices = np.random.choice(
np.arange(len(validation_data)),
size=validation_set_size,
replace=False)
validation_data_batch = validation_data[
validation_set_indices]
validation_labels_batch = validation_labels[
validation_set_indices]
validation_accuracy = session.run(
accuracy,
feed_dict={
self.x: validation_data_batch,
y: validation_labels_batch
})
last_validation_accuracy = validation_accuracy
if len(validation_accuracy_values
) == self.accuracy_display_saved_iterations:
validation_accuracy_values.pop(0)
if iteration >= self.accuracy_display_starting_iteration:
validation_accuracy_values.append(
validation_accuracy)
if validation_accuracy > max_accuracy:
max_accuracy = validation_accuracy
if self.use_sound:
self.sounds.success()
if len(max_accuracy_values
) == self.accuracy_display_saved_iterations:
max_accuracy_values.pop(0)
if iteration >= self.accuracy_display_starting_iteration:
max_accuracy_values.append(max_accuracy)
progress = int(math.ceil(self.progress_bar_size * \
float((iteration - 1) % num_training_batches) /\
(num_training_batches - 1)))
elapsed = time.time() - start_time
progress_string = '[' + '#' * progress + ' ' * \
(self.progress_bar_size - progress) + ']'
if iteration % num_training_batches == 0 or \
iteration % row_output_interval == 0:
progress_string = time.strftime(
"%I:%M:%S %p", time.localtime())
if not self.skip_evaluation:
table.update(
epoch, (iteration - 1) % num_training_batches + 1,
sustained_loss, validation_accuracy * 100,
max_accuracy * 100, progress_string, elapsed)
else:
table.update(
epoch, (iteration - 1) % num_training_batches + 1,
sustained_loss, progress_string, elapsed)
if iteration % num_training_batches == 0 or \
iteration % row_output_interval == 0:
heavy = False
if multiple_rows_per_epoch and iteration % num_training_batches == 0:
heavy = True
table.finalize(heavy=heavy)
# Termination condition
if sustained_loss < loss_threshold:
done = True
break
self.update_output(iteration, loss_values,
sustained_loss_values,
validation_accuracy_values,
max_accuracy_values)
# Termination condition
if epoch >= max_epochs or sustained_loss < loss_threshold:
done = True
self.update_output(iteration,
loss_values,
sustained_loss_values,
validation_accuracy_values,
max_accuracy_values,
override=True)
plt.pause(0)
last_date = None
for i in range(len(total_balances)):
(date, balance) = total_balances[i]
current_month = date.year * 12 + date.month
current_year = date.year
new_month = current_month != last_month
new_year = current_year != last_year
last_month = current_month
last_year = current_year
# if i > 0:
# table.finalize(heavy=new_month and output_table_dividers, divider=last_date != date)
# if new_year and output_table_dividers:
# table.print_divider(heavy=True)
if i > 0:
table.finalize(divider=last_date != date
and not (new_month and output_table_dividers))
if new_month and output_table_dividers:
table.print_message(months_full[date.month - 1] + ' ' + str(date.year),
heavy=True)
table.print_header()
previous_balance = total_balances[i - 1][1] if i > 0 else 0
checking_balance = checking_balances[i][1]
savings_balance = savings_balances[i][1]
credit_balance = credit_balances[i][1]
increased = total_balance_diffs[i] >= 0
args = [
date.strftime("%m/%d/%Y") if last_date != date else '',
# months[date.month-1] + ' ' + str(date.day) + ' ' + str(date.year) if last_date != date else '',
meta[i-1][1][0],
('+' if increased else '') + format_balance(total_balance_diffs[i]),