def test_plot_conv_activity(self, net_fitted, X_train):
from nolearn.lasagne.visualize import plot_conv_activity
plot_conv_activity(net_fitted.layers_['conv1'], X_train[:1])
plot_conv_activity(net_fitted.layers_['conv2'], X_train[10:11],
figsize=(3, 4))
plt.clf()
plt.cla()
def test_plot_conv_activity(self, net_fitted, X_train):
from nolearn.lasagne.visualize import plot_conv_activity
plot_conv_activity(net_fitted.layers_['conv1'], X_train[:1])
plot_conv_activity(net_fitted.layers_['conv2'], X_train[10:11],
figsize=(3, 4))
plt.clf()
plt.cla()
def predict(path, debug=False):
global net, image_size
# img = load_image(path)
X = np.asarray([path])
img = np.asarray([load_image(path)], dtype="float32")
cats = load_classes()
probs = net.predict_proba(X)
if debug == True:
for layer in net.layers:
name = layer[0].__name__
filename = os.path.splitext(path)[0]
try:
os.mkdir("debug/" + filename)
except:
pass
if "2D" in name:
real_name = layer[1]["name"]
try:
plt = plot_conv_activity(net.layers_[real_name], img)
plt.savefig("debug/" + filename + "/" + real_name +
"_activity.png")
plt.close()
#plt = plot_conv_weights(net.layers_[real_name]) DISABLED: Eats too much memory, pretty useless anyhow
#plt.savefig("debug/"+filename+"/"+real_name+"_weights.png")
except Exception as e:
print(e)
for i, x in enumerate(probs[0]):
print(cats[i].name, x)
result = [(cats[np.argmax(x)].id, cats[np.argmax(x)].name, x[np.argmax(x)])
for x in probs]
return result[0]
name = exp_name + '_' + str(date.today())
with open('models/conv_net_'+name+'.pkl', 'wb') as f:
cPickle.dump(conv_net, f, -1)
conv_net.save_params_to('models/params_'+name)
# ----- Train set ----
train_predictions = conv_net.predict_proba(X)
make_submission_file(train_predictions[:sample_size], images_id[:sample_size],
output_filepath='models/training_'+name+'.csv')
# ----- Test set ----
X_test, _, images_id_test = load_numpy_arrays(args['test_file'])
print "Test:"
print "X_test.shape:", X_test.shape
predictions = conv_net.predict_proba(X_test)
make_submission_file(predictions, images_id_test,
output_filepath='submissions/submission_'+name+'.csv')
# ----- Make plots ----
plot_loss(conv_net, "models/loss_"+name+".png", show=False)
plot_conv_weights(conv_net.layers_[1], figsize=(4, 4))
plt.savefig('models/weights_'+name+'.png')
plot_conv_activity(conv_net.layers_[1], X[0:1])
plt.savefig('models/activity_'+name+'.png')
plot_occlusion(conv_net, X[:5], y[:5])
plt.savefig('models/occlusion_'+name+'.png')
def plot_conv_activity(self, net, X, **kwargs):
from nolearn.lasagne.visualize import plot_conv_activity
plot_conv_activity(net.layers_['conv1'], X, **kwargs)
plt.clf()
plt.cla()
from nolearn.lasagne.visualize import plot_occlusion, plot_saliency
draw_to_notebook(net0)
plot_loss(net0)
#plot helps determine if we are overfitting:
#If the train loss is much lower than the validation loss,
#we should probably do something to regularize the net.
# visualize layer weights
plot_conv_weights(net0.layers_[1], figsize = (4,4))
#If the weights just look like noise, we might have to do something
#(e.g. use more filters so that each can specialize better).
# visualize layers' activities
x = X_train[0:1] # an image in the bc01 format (so use X[0:1] instead of just X[0]).
plot_conv_activity(net0.layers_[1], x)
plot_occlusion(net0, X_train[:5], y_train[:5])
plot_saliency(net0, X_train[:5])
from nolearn.lasagne import PrintLayerInfo
layers1 = [
(InputLayer, {'shape': (None, 1, 28, 28)}),
(Conv2DLayer, {'num_filters': 32, 'filter_size': (3, 3)}),
(MaxPool2DLayer, {'pool_size': (2, 2)}),
(Conv2DLayer, {'num_filters': 64, 'filter_size': (3, 3)}),
(Conv2DLayer, {'num_filters': 64, 'filter_size': (3, 3)}),
def test_visualize_functions_with_cnn(mnist):
# this test simply tests that no exception is raised when using
# the plotting functions
from nolearn.lasagne import NeuralNet
from nolearn.lasagne.visualize import plot_conv_activity
from nolearn.lasagne.visualize import plot_conv_weights
from nolearn.lasagne.visualize import plot_loss
from nolearn.lasagne.visualize import plot_occlusion
X, y = mnist
X_train, y_train = X[:100].reshape(-1, 1, 28, 28), y[:100]
X_train = X_train.reshape(-1, 1, 28, 28)
num_epochs = 3
nn = NeuralNet(
layers=[
('input', InputLayer),
('conv1', Conv2DLayer),
('conv2', Conv2DLayer),
('pool2', MaxPool2DLayer),
('conv3', Conv2DLayer),
('conv4', Conv2DLayer),
('pool4', MaxPool2DLayer),
('hidden1', DenseLayer),
('output', DenseLayer),
],
input_shape=(None, 1, 28, 28),
output_num_units=10,
output_nonlinearity=softmax,
more_params=dict(
conv1_filter_size=(5, 5),
conv1_num_filters=16,
conv2_filter_size=(3, 3),
conv2_num_filters=16,
pool2_ds=(3, 3),
conv3_filter_size=(3, 3),
conv3_num_filters=16,
conv4_filter_size=(3, 3),
conv4_num_filters=16,
pool4_ds=(2, 2),
hidden1_num_units=16,
),
update=nesterov_momentum,
update_learning_rate=0.01,
update_momentum=0.9,
max_epochs=num_epochs,
)
nn.fit(X_train, y_train)
plot_loss(nn)
plot_conv_weights(nn.layers_['conv1'])
plot_conv_weights(nn.layers_['conv2'], figsize=(1, 2))
plot_conv_activity(nn.layers_['conv3'], X_train[:1])
plot_conv_activity(nn.layers_['conv4'], X_train[10:11], figsize=(3, 4))
plot_occlusion(nn, X_train[:1], y_train[:1])
plot_occlusion(nn,
X_train[2:4],
y_train[2:4],
square_length=3,
figsize=(5, 5))
# clear figures from memory
plt.clf()
plt.cla()
def plot_conv_activity(self, net, X, **kwargs):
from nolearn.lasagne.visualize import plot_conv_activity
plot_conv_activity(net.layers_["conv1"], X, **kwargs)
plt.clf()
plt.cla()
def test_visualize_functions_with_cnn(mnist):
# this test simply tests that no exception is raised when using
# the plotting functions
from nolearn.lasagne import NeuralNet
from nolearn.lasagne.visualize import plot_conv_activity
from nolearn.lasagne.visualize import plot_conv_weights
from nolearn.lasagne.visualize import plot_loss
from nolearn.lasagne.visualize import plot_occlusion
X, y = mnist
X_train, y_train = X[:100].reshape(-1, 1, 28, 28), y[:100]
X_train = X_train.reshape(-1, 1, 28, 28)
num_epochs = 3
nn = NeuralNet(
layers=[
('input', InputLayer),
('conv1', Conv2DLayer),
('conv2', Conv2DLayer),
('pool2', MaxPool2DLayer),
('conv3', Conv2DLayer),
('conv4', Conv2DLayer),
('pool4', MaxPool2DLayer),
('hidden1', DenseLayer),
('output', DenseLayer),
],
input_shape=(None, 1, 28, 28),
output_num_units=10,
output_nonlinearity=softmax,
more_params=dict(
conv1_filter_size=(5, 5), conv1_num_filters=16,
conv2_filter_size=(3, 3), conv2_num_filters=16,
pool2_ds=(3, 3),
conv3_filter_size=(3, 3), conv3_num_filters=16,
conv4_filter_size=(3, 3), conv4_num_filters=16,
pool4_ds=(2, 2),
hidden1_num_units=16,
),
update=nesterov_momentum,
update_learning_rate=0.01,
update_momentum=0.9,
max_epochs=num_epochs,
)
nn.fit(X_train, y_train)
plot_loss(nn)
plot_conv_weights(nn.layers_['conv1'])
plot_conv_weights(nn.layers_['conv2'], figsize=(1, 2))
plot_conv_activity(nn.layers_['conv3'], X_train[:1])
plot_conv_activity(nn.layers_['conv4'], X_train[10:11], figsize=(3, 4))
plot_occlusion(nn, X_train[:1], y_train[:1])
plot_occlusion(nn, X_train[2:4], y_train[2:4], square_length=3,
figsize=(5, 5))
# clear figures from memory
plt.clf()
plt.cla()