def fit_mnist_mlp(checkpoint=None):
"""Train a three-layer MLP using MNIST data.
128-128-10, prelu_shelf: 2% error
400-400-10, prelu_shelf: 1.55% error
"""
layers = [["InputLayer", {"name": "input"}], # May specify explicitly or leave this off.
["FCLayer", {"name": "fc1", "n_units": 400, "activation": "prelu_shelf", "l2": 0.001}],
["DropoutLayer", {"name": "DO-fc1", "dropout_p": 0.5}],
["FCLayer", {"name": "fc2", "n_units": 400, "activation": "prelu_shelf", "l2": 0.001}],
["DropoutLayer", {"name": "DO-fc2", "dropout_p": 0.5}],
["ClassificationOutputLayer", {"name": "output", "n_classes": 10}]]
classifier = nets.NNClassifier(layers, name="MNIST MLP", random_state=42)
# Specify how the learning rate changes during training.
lr_rule = {"rule": "stalled", "initial_value": 0.1, "multiply_by": 0.25, "interval": 5}
# Specify how the momentum changes during training.
momentum_rule = {"rule": "stalled", "initial_value": 0.7, "decrease_by": -0.1,
"final_value": 0.95, "interval": 5}
mnist_data = files.read_pickle(sample_data.mnist)
classifier.fit(mnist_data[0], n_epochs=50, valid=mnist_data[1], test=mnist_data[2],
augmentation=None, checkpoint=checkpoint, sgd_type="nag",
lr_rule=lr_rule, momentum_rule=momentum_rule, batch_size=128,
train_loss="nll", valid_loss="nll", test_loss=["nll", "error"])
return classifier
def fit_mnist_conv(checkpoint=None):
"""Train a convolutional neural network using MNIST data.
(3x3)x32 - (maxpool 2) - (2x2)x32 - (2x2)x32 - (maxpool2) - 400 - 400 - 10 :
Trains to 0.71% error after 148 epochs.
"""
batch_size = 128 # Let the Readers know how many images to supply with each epoch.
# Set up the data inputs. Demonstrate explicit use of the "Data" object, and the
# ability to alter data as it's read out. In this case, we want to reshape the
# training images to be square so that they're easier to shift as part of data augmentation.
def make_square(img_batch):
return img_batch.reshape((-1, 28, 28))
mnist_data = files.read_pickle(sample_data.mnist)
train = readers.Data(*mnist_data[0], batch_size=batch_size, alter_features=make_square)
valid = readers.Data(*mnist_data[1], batch_size=batch_size, alter_features=make_square,
allow_partial_batch=True)
test = readers.Data(*mnist_data[2], batch_size=batch_size, alter_features=make_square,
allow_partial_batch=True)
# Define the network architecture.
layers = [ # Now an input layer is needed to translate the flat array to a square image.
["InputImageLayer", {"name": "input", "n_images": 1, "n_pixels": (28, 28)}],
["BC01ToC01BLayer"], # Needed for optional running with CUDA-convnet.
["ConvLayer", {"name": "conv1", "n_output_maps": 32, "filter_shape": (3, 3),
"activation": "prelu"}],
["MaxPool2DLayer", {"name": "pool1", "pool_shape": (2, 2)}],
["ConvLayer", {"name": "conv2", "n_output_maps": 32, "filter_shape": (2, 2),
"activation": "prelu"}],
["ConvLayer", {"name": "conv3", "n_output_maps": 32, "filter_shape": (2, 2),
"activation": "prelu"}],
["MaxPool2DLayer", {"name": "pool3", "pool_shape": (2, 2)}],
["C01BToBC01Layer"], # Switch back to Theano ordering
["FCLayer", {"name": "fc1", "n_units": 400, "activation": "prelu_shelf", "l2": 0.01}],
["DropoutLayer", {"name": "DO-fc1", "dropout_p": 0.5}],
["FCLayer", {"name": "fc2", "n_units": 400, "activation": "prelu_shelf", "l2": 0.01}],
["DropoutLayer", {"name": "DO-fc2", "dropout_p": 0.5}],
["ClassificationOutputLayer", {"name": "output", "n_classes": 10}]
]
# Instantiate the network and start training.
classifier = nets.NNClassifier(layers, n_in=train.features.shape, name="MNIST ConvNet",
batch_size=batch_size, random_state=42)
lr_rule = {"rule": "stalled", "initial_value": 0.05, "multiply_by": 0.5,
"interval": 8, "final_value": 1e-3}
momentum_rule = {"rule": "stalled", "initial_value": 0.5, "decrease_by": -0.1,
"final_value": 0.95, "interval": 8}
classifier.fit(train, n_epochs=200, valid=valid, test=test,
augmentation=augment.create_shift_augment(3, 3),
checkpoint=checkpoint, sgd_type="nag",
lr_rule=lr_rule, momentum_rule=momentum_rule, batch_size=batch_size,
train_loss="nll", valid_loss="nll", test_loss=["nll", "error"])
return classifier
def fit_mnist_small_mlp(n_epochs, checkpoint=None):
"""Train a one-layer MLP using MNIST data. Use for testing.
"""
lr_rule = {"rule": "stalled", "initial_value": 0.1, "multiply_by": 0.25, "interval": 5}
momentum_rule = {"rule": "stalled", "initial_value": 0.7, "decrease_by": -0.1,
"final_value": 0.95, "interval": 5}
# Demonstrate automatic partitioning of data into training, validation, and test sets.
mnist_data = files.read_pickle(sample_data.mnist)
mnist_features = np.concatenate([mnist_data[0][0], mnist_data[1][0], mnist_data[2][0]])
mnist_labels = np.concatenate([mnist_data[0][1], mnist_data[1][1], mnist_data[2][1]])
layers = [["FCLayer", {"name": "fc2", "n_units": 100, "activation": "prelu_shelf", "l2": 0.001}],
["DropoutLayer", {"name": "DO-fc2", "dropout_p": 0.5}],
["ClassificationOutputLayer", {"name": "output", "n_classes": 10}]]
classifier = nets.NNClassifier(layers, name="MNIST MLP", random_state=42)
classifier.fit(mnist_features, mnist_labels, valid=1/7, test=1/7, n_epochs=n_epochs,
augmentation=None, checkpoint=checkpoint, sgd_type="nag",
lr_rule=lr_rule, momentum_rule=momentum_rule, batch_size=100,
train_loss="nll", valid_loss="nll", test_loss=["nll", "error"])
return classifier
def load_and_train(fname, checkpoint=None):
"""Test reloading a model in progress and continuing."""
mnist_data = files.read_pickle(sample_data.mnist)
c = files.checkpoint_read(fname)
c.fit(mnist_data[0], valid=mnist_data[1], test=mnist_data[2], checkpoint=checkpoint)
