def fit(n_windows, win_width, rand_state, data_set, data_labels, filename='weights.pkl'):
# Permuting data
rng = np.random.RandomState(8000)
indices = rng.permutation(len(data_set))
data_set = np.array(data_set)
data_labels = np.array(data_labels)
data_set, data_labels = data_set[indices], data_labels[indices]
print str(len(data_set)) + " all samples"
train_len = int(len(data_set) * 9.0 / 10.0)
valid_len = len(data_set) - train_len
print "Train: " + str(train_len)
print "Validate: " + str(valid_len)
# Splitting fs
train_dir = fs.File("training.hdf5", "a")
train_data = train_dir.create_dataset('train_data', shape=((train_len + 1) * n_windows, 41, 41), dtype='i')
train_labels = train_dir.create_dataset('train_labels', shape=((train_len + 1) * n_windows,), dtype='i')
valid_dir = fs.File("validating.hdf5", "a")
valid_data = valid_dir.create_dataset('valid_data', shape=((valid_len + 1) * n_windows, 41, 41), dtype='i')
valid_labels = valid_dir.create_dataset('valid_labels', shape=((valid_len + 1) * n_windows,), dtype='i')
counter = 0
next_counter = 0
for iter, data_sample in enumerate(data_set):
if iter % 10000 == 0:
print iter
windows = WinExt.get_windows(data_sample, n_windows, win_width, rand_state)
for window in windows:
# First windows part for training
# Second part for validation
if iter < train_len:
train_data[counter] = window
train_labels[counter] = data_labels[iter]
counter += 1
else:
valid_data[next_counter] = window
valid_labels[next_counter] = data_labels[iter]
next_counter += 1
# Setting real length
train_len = counter
valid_len = next_counter
print "Size of train is " + str(train_len)
print "Size of valid is " + str(valid_len)
print "Extracting has finished its work..."
batch_size = 500
if train_len % batch_size != 0: # if the last batch is not full, just don't use the remainder
whole = (train_len / batch_size) * batch_size
train_len = whole
if valid_len % batch_size != 0:
whole = (valid_len / batch_size) * batch_size
valid_len = whole
n_train_batches = train_len / batch_size
n_valid_batches = valid_len / batch_size
print "Building net..."
nkerns = [5, 10, 20]
x = T.dtensor3('x') # dtensor3 for 3d array
y = T.ivector('y') # the labels are presented as 1D vector of [int] labels
rng = np.random.RandomState(8000)
classifier = CNN(
rng=rng,
input=x,
nkerns=nkerns,
batch_size=batch_size
)
# 4 layer is logistic
cost = classifier.layer4.negative_log_likelihood(y)
# create a list of gradients for all model parameters
grads = T.grad(cost, classifier.params)
learning_rate = 0.3 / float(n_train_batches)
updates = [
(param_i, param_i - learning_rate * grad_i)
for param_i, grad_i in zip(classifier.params, grads)
]
data_tr = theano.shared(
np.asarray(np.zeros((batch_size, 41, 41), dtype=np.int), dtype=theano.config.floatX),
borrow=True)
labels_tr = theano.shared(
np.asarray(np.zeros(batch_size, dtype=np.int), dtype='int32'),
borrow=True)
data_val = theano.shared(
np.asarray(np.zeros((batch_size, 41, 41), dtype=np.int), dtype=theano.config.floatX),
borrow=True)
labels_val = theano.shared(
np.asarray(np.zeros(batch_size, dtype=np.int), dtype='int32'),
borrow=True)
# When we will call them, updates will be recalculated, and errors(cost) will be returned
# for valid set we don't have to do random shuffle
validate_model = theano.function(
inputs=[],
outputs=classifier.errors(y),
givens={
x: data_val,
y: labels_val
}
)
# indices - for random shuffle
train_model = theano.function(
inputs=[],
outputs=cost,
updates=updates,
givens={
x: data_tr,
y: labels_tr
}
)
# ONLY if interrupted
file = open(filename, 'rb')
classifier.__setstate__(cPickle.load(file))
file.close()
print "Training..."
# GDM with batches
epoch = 14
n_epochs = 30
min_error = 100.0
errors = []
indices = rng.permutation(train_len)
while epoch < n_epochs:
print "================= " + str(epoch + 1) + " epoch =============== "
for minibatch_index in range(n_train_batches):
if minibatch_index % 50 == 0:
print str(minibatch_index) + " batch"
data_tr.set_value(np.array(
[train_data[indices[minibatch_index * batch_size + i]] for i in range(batch_size)]), borrow=True)
labels_tr.set_value(np.array(
[train_labels[indices[minibatch_index * batch_size + i]] for i in range(batch_size)]), borrow=True)
train_model()
# compute zero-one loss on validation set
validation_losses = []
for i in range(n_valid_batches):
data_val.set_value(np.array(
valid_data[i * batch_size: (i + 1) * batch_size]), borrow=True)
labels_val.set_value(np.array(
valid_labels[i * batch_size: (i + 1) * batch_size]), borrow=True)
validation_losses.append(validate_model())
this_validation_loss = np.mean(validation_losses) * 100
errors.append(this_validation_loss)
print str(this_validation_loss) + "% error"
if this_validation_loss < min_error:
min_error = this_validation_loss
save_parameters(classifier, filename)
epoch += 1
print "Shuffling..."
indices = rng.permutation(train_len)
show_errors(errors, 'You heading goes here...')
# Cleaning data
train_dir.clear()
valid_dir.clear()
train_dir.close()
valid_dir.close()
def fit(n_windows, win_width, rand_state, data_set, data_labels, filename="LR_weights.pkl"):
# Permuting data
rng = np.random.RandomState(8000)
indices = rng.permutation(len(data_set))
data_set = np.array(data_set)
data_labels = np.array(data_labels)
data_set, data_labels = data_set[indices], data_labels[indices]
print str(len(data_set)) + " all samples"
train_len = int(len(data_set) * 9.0 / 10.0)
valid_len = len(data_set) - train_len
print "Train: " + str(train_len)
print "Validate: " + str(valid_len)
# Splitting fs
train_dir = fs.File("LR_training.hdf5", "a")
train_data = train_dir.create_dataset("LR_train_data", shape=((train_len + 1) * n_windows, 41, 41), dtype="i")
train_labels = train_dir.create_dataset("LR_train_labels", shape=((train_len + 1) * n_windows,), dtype="i")
valid_dir = fs.File("LR_validating.hdf5", "a")
valid_data = valid_dir.create_dataset("LR_valid_data", shape=((valid_len + 1) * n_windows, 41, 41), dtype="i")
valid_labels = valid_dir.create_dataset("LR_valid_labels", shape=((valid_len + 1) * n_windows,), dtype="i")
counter = 0
next_counter = 0
for iter, data_sample in enumerate(data_set):
if iter % 10000 == 0:
print iter
windows = WinExt.get_windows(data_sample, n_windows, win_width, rand_state)
for window in windows:
# First windows part for training
# Second part for validation
if iter < train_len:
train_data[counter] = window
train_labels[counter] = data_labels[iter]
counter += 1
else:
valid_data[next_counter] = window
valid_labels[next_counter] = data_labels[iter]
next_counter += 1
# Setting real length
train_len = counter
valid_len = next_counter
print "Size of train is " + str(train_len)
print "Size of valid is " + str(valid_len)
print "Extracting has finished its work..."
batch_size = 500
if train_len % batch_size != 0: # if the last batch is not full, just don't use the remainder
whole = (train_len / batch_size) * batch_size
train_len = whole
if valid_len % batch_size != 0:
whole = (valid_len / batch_size) * batch_size
valid_len = whole
n_train_batches = train_len / batch_size
n_valid_batches = valid_len / batch_size
data_tr = theano.shared(
np.asarray(np.zeros((batch_size, 41, 41), dtype=np.int), dtype=theano.config.floatX), borrow=True
)
labels_tr = theano.shared(np.asarray(np.zeros(batch_size, dtype=np.int), dtype="int32"), borrow=True)
data_val = theano.shared(
np.asarray(np.zeros((batch_size, 41, 41), dtype=np.int), dtype=theano.config.floatX), borrow=True
)
labels_val = theano.shared(np.asarray(np.zeros(batch_size, dtype=np.int), dtype="int32"), borrow=True)
print "Building logistic regression classifier..."
x = T.dtensor3("x") # dtensor3 for 3d array
y = T.ivector("y") # the labels are presented as 1D vector of [int] labels
rng = np.random.RandomState(8000)
classifier = LogisticRegression(input=x.flatten(2), n_in=41 * 41, n_out=2)
cost = classifier.negative_log_likelihood(y)
learning_rate = 0.03 # 0.3 / float(n_train_batches)
g_W = T.grad(cost=cost, wrt=classifier.W)
g_b = T.grad(cost=cost, wrt=classifier.b)
# start-snippet-3
# specify how to update the parameters of the model as a list of
# (variable, update expression) pairs.
updates = [(classifier.W, classifier.W - learning_rate * g_W), (classifier.b, classifier.b - learning_rate * g_b)]
validate_model = theano.function(inputs=[], outputs=classifier.errors(y), givens={x: data_val, y: labels_val})
# indices - for random shuffle
train_model = theano.function(
inputs=[], outputs=classifier.errors(y), updates=updates, givens={x: data_tr, y: labels_tr}
)
print "Training..."
# GDM with batches
epoch = 0
n_epochs = 30
min_error = 100.0
errors = []
indices = rng.permutation(train_len)
while epoch < n_epochs:
print "================= " + str(epoch + 1) + " epoch =============== "
for minibatch_index in range(n_train_batches):
if minibatch_index % 50 == 0:
print str(minibatch_index) + " batch"
data_tr.set_value(
np.array([train_data[indices[minibatch_index * batch_size + i]] for i in range(batch_size)]),
borrow=True,
)
labels_tr.set_value(
np.array([train_labels[indices[minibatch_index * batch_size + i]] for i in range(batch_size)]),
borrow=True,
)
train_model()
# compute zero-one loss on validation set
validation_losses = []
for i in range(n_valid_batches):
data_val.set_value(np.array(valid_data[i * batch_size : (i + 1) * batch_size]), borrow=True)
labels_val.set_value(np.array(valid_labels[i * batch_size : (i + 1) * batch_size]), borrow=True)
validation_losses.append(validate_model())
this_validation_loss = np.mean(validation_losses) * 100
errors.append(this_validation_loss)
if this_validation_loss < min_error:
print str(this_validation_loss) + "% error"
min_error = this_validation_loss
save_parameters(classifier, filename)
epoch += 1
print "Shuffling..."
indices = rng.permutation(train_len)
show_errors(errors, "LogReg: 4 windows, h=41")
# Cleaning data
train_dir.clear()
valid_dir.clear()
train_dir.close()
valid_dir.close()
def predict(n_windows, win_width, rand_state, test_dataset, labels, filename='weights.pkl'):
false_positive = 0.0
true_positive = 0.0
false_negative = 0.0
true_negative = 0.0
print "Test size is " + str(len(test_dataset))
rng = np.random.RandomState(8000)
x = T.matrix('x')
y = T.ivector('y') # the labels are presented as 1D vector of [int] labels
index = T.lscalar()
classifier = CNN(
rng=rng,
input=x,
nkerns=[5, 10, 20],
batch_size=1 # Only for one image will predict
)
file = open(filename, 'rb')
classifier.__setstate__(cPickle.load(file))
file.close()
result = []
for iter, image in enumerate(test_dataset):
if iter % 1000 == 0:
print iter
windows = WinExt.get_windows(image, n_windows, win_width, rand_state)
test_data = theano.shared(np.asarray(windows, dtype=theano.config.floatX), borrow=True)
test_labels = theano.shared(np.asarray(np.zeros(len(windows)), dtype='int32'), borrow=True)
predict_model = theano.function(
inputs=[index],
outputs=classifier.layer4.pred_probs(),
givens={
x: test_data[index],
y: test_labels
},
on_unused_input='ignore' # because we dont need test lables,for now
)
predicts = [predict_model(i) for i in range(len(windows))]
windows_predictions = np.array(predicts).reshape((len(windows), 2))
avg_0 = 0.0
avg_1 = 0.0
for win_prediction in windows_predictions:
avg_0 += win_prediction[0]
avg_1 += win_prediction[1]
avg_0 /= len(windows_predictions)
avg_1 /= len(windows_predictions)
result.append(np.argmax([avg_0, avg_1])) # Predicting on average answers
if result[iter] != labels[iter]:
if labels[iter] == 0:
false_positive += 1
else:
false_negative += 1
image_save(n_windows, image, iter)
print str(iter) + " image proba:"
print windows_predictions
else:
if labels[iter] == 0:
true_negative += 1
else:
true_positive += 1
print "for negative class"
print "recall: "
print float(true_negative / (true_negative + false_positive))
print "precision: "
print float(true_negative / (true_negative + false_negative))
print "for positive class"
print "recall: "
print float(true_positive / (true_positive + false_negative))
print "precision: "
print float(true_positive / (true_positive + false_positive))
print "Accuracy " + str(T.mean(T.neq(result, labels)).eval())