convnets.py

#!/usr/bin/env python

from __future__ import print_function

import sys
import gc
import os
import time
import yaml
import traceback

import numpy as np
import theano
import theano.tensor as T

import lasagne
from data_utils import MRIDataIterator

def build_cnn(input_var=None, batch_size=5, meta_data_input_var=None):
    """number_of_buckets:   Is the number of histogram buckets we have created.
                            We treat these like layers for the convolution,
                            filling in the missing layers with 0s. We also throw
                            out slices that are probably from the same location
    """

    # Input layer, as usual:
    # (number of frames in cardiac cycle x number_of_buckets x image_width x image_height)
    # (30 x 10 x 64 x 64)
    network = lasagne.layers.InputLayer(shape=(None, 1, 64, 64),
                                        input_var=input_var)
    # metadata = gender, age (in years), lower_bound, mean, upper_bound for age group
    meta_data_network = lasagne.layers.InputLayer(shape=(None, 8), input_var=meta_data_input_var)
    # This time we do not apply input dropout, as it tends to work less well
    # for convolutional layers.

    # Convolutional layer with 32 kernels of size 5x5. Strided and padded
    # convolutions are supported as well; see the docstring.
    network = lasagne.layers.Conv2DLayer(
            network, num_filters=64, filter_size=(3, 3),
            nonlinearity=lasagne.nonlinearities.rectify,
            W=lasagne.init.GlorotUniform())

    network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2))
    network = lasagne.layers.DropoutLayer(network, p=.25)

    # Another convolution with 32 5x5 kernels, and another 2x2 pooling:
    network = lasagne.layers.Conv2DLayer(
            network, num_filters=96, filter_size=(3, 3),
            nonlinearity=lasagne.nonlinearities.rectify)
    network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2))
    network = lasagne.layers.DropoutLayer(network, p=.25)

    # Another convolution with 32 5x5 kernels, and another 2x2 pooling:
    network = lasagne.layers.Conv2DLayer(
            network, num_filters=128, filter_size=(2, 2),
            nonlinearity=lasagne.nonlinearities.rectify)
    network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2))
    network = lasagne.layers.DropoutLayer(network, p=.25)

    #should now be (30 x (64*64*10))
    network = lasagne.layers.Conv2DLayer(
            network, num_filters=256, filter_size=(2, 2),
            nonlinearity=lasagne.nonlinearities.rectify)
    network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2))
    network = lasagne.layers.DropoutLayer(network, p=.25)
    network = lasagne.layers.FlattenLayer(network, 2)

    print("After flatter, dims: {}".format(network.output_shape))

    # need to get it to (1 x 30 x (64*64*10))
    # With batching need to somehow reshape from (30*batch_size, 1, 64, 64) to
    # (batch_size, 30, 1024)
    network = lasagne.layers.ReshapeLayer(network, (-1, 30, [1]))

    #print("After reshape, dims: {}".format(network.output_shape))

    network = lasagne.layers.LSTMLayer(
        network, 1024, grad_clipping=100,
        nonlinearity=lasagne.nonlinearities.tanh)

    #print("After lstm, dims: {}".format(network.output_shape))

    # The l_forward layer creates an output of dimension (batch_size, SEQ_LENGTH, N_HIDDEN)
    # Since we are only interested in the final prediction, we isolate that quantity and feed it to the next layer.
    # The output of the sliced layer will then be of size (batch_size, N_HIDDEN)
    # network = lasagne.layers.SliceLayer(network, -1, 1)
    # print("After slice, dims: {}".format(network.output_shape))

    # A fully-connected layer of 1024 units with 50% dropout on its inputs:
    network = lasagne.layers.DenseLayer(
            lasagne.layers.dropout(network, p=.5),
            num_units=712,
            nonlinearity=lasagne.nonlinearities.tanh)

    meta_data_network = lasagne.layers.DenseLayer(
            lasagne.layers.dropout(meta_data_network, p=.5),
            num_units=256,
            nonlinearity=lasagne.nonlinearities.tanh)

    meta_data_network = lasagne.layers.DenseLayer(
            lasagne.layers.dropout(meta_data_network, p=.5),
            num_units=512,
            nonlinearity=lasagne.nonlinearities.tanh)

    meta_data_network = lasagne.layers.DenseLayer(
            lasagne.layers.dropout(meta_data_network, p=.5),
            num_units=768,
            nonlinearity=lasagne.nonlinearities.tanh)

    # And, finally, the 600-unit output layer with 50% dropout on its inputs:
    network = lasagne.layers.DenseLayer(
            lasagne.layers.dropout(network, p=.5),
            num_units=768,
            nonlinearity=lasagne.nonlinearities.tanh)

    network = lasagne.layers.ElemwiseSumLayer([network, meta_data_network])

    network = lasagne.layers.DenseLayer(
            lasagne.layers.dropout(network, p=.5),
            num_units=1024,
            nonlinearity=lasagne.nonlinearities.softmax)

    network = lasagne.layers.DenseLayer(
            lasagne.layers.dropout(network, p=.25),
            num_units=600,
            nonlinearity=lasagne.nonlinearities.softmax)

    return network


def compose_functions(scope="default", batch_size=5):
    # Prepare Theano variables for inputs and targets
    input_var = T.tensor4(scope + 'inputs')
    meta_data_input_var = T.matrix(scope + 'metadata_inputs')
    target_var = T.ivector(scope + 'targets')
    #x_printed = theano.printing.Print('this is a very important value')(target_var)
    network = build_cnn(input_var, batch_size, meta_data_input_var)

    # Create a loss expression for training, i.e., a scalar objective we want
    # to minimize (for our multi-class problem, it is the cross-entropy loss):
    prediction = lasagne.layers.get_output(network)
    #prediction_printed = theano.printing.Print('this is a very important prediction')(prediction)
    loss = lasagne.objectives.categorical_crossentropy(prediction, target_var).mean()
    # loss = loss.mean()
    # We could add some weight decay as well here, see lasagne.regularization.

    # Create update expressions for training, i.e., how to modify the
    # parameters at each training step. Here, we'll use Stochastic Gradient
    # Descent (SGD) with Nesterov momentum, but Lasagne offers plenty more.
    params = lasagne.layers.get_all_params(network, trainable=True)
    updates = lasagne.updates.nesterov_momentum(
            loss, params, learning_rate=0.01, momentum=0.9)

    # Create a loss expression for validation/testing. The crucial difference
    # here is that we do a deterministic forward pass through the network,
    # disabling dropout layers.

    #TODO: this should actually be CPRS

    test_prediction = lasagne.layers.get_output(network, deterministic=True)
    test_loss = lasagne.objectives.categorical_crossentropy(test_prediction,
                                                            target_var)
    test_loss = test_loss.mean()
    # As a bonus, also create an expression for the classification accuracy:
    # test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var),
                    #   dtype=theano.config.floatX)
    # print(len(test_prediction))
    # print(test_prediction.shape)
    # v = np.array(range(test_prediction.shape[0]))
    # h = (v >= V_m) * -1.
    # sq_dists = (test_prediction - h)**2
    # test_acc = T.sqr(T.sum(p, h))

    # Compile a function performing a training step on a mini-batch (by giving
    # the updates dictionary) and returning the corresponding training loss:
    train_fn = theano.function([input_var, target_var, meta_data_input_var], loss, updates=updates)

    # Compile a second function computing the validation loss and accuracy:
    val_fn = theano.function([input_var, target_var, meta_data_input_var], [test_loss, test_prediction])
    return network, train_fn, val_fn



# ############################## Main program ################################
# Everything else will be handled in our main program now. We could pull out
# more functions to better separate the code, but it wouldn't make it any
# easier to read.

def main(num_epochs=30):
    # Load the dataset
    print("Creating data iterator...")
    with open("config.yml", 'r') as ymlfile:
        cfg = yaml.load(ymlfile)
    train_dir = cfg['dataset_paths']['train_data']
    train_labels = cfg['dataset_paths']['train_labels']
    batch_size = 5

    mriIter = MRIDataIterator(train_dir, train_labels)

    network, train_fn, val_fn = compose_functions("systole", batch_size) #systole
    network_dia, train_fn_dia, val_fn_dia = compose_functions("diastole", batch_size)

    if os.path.exists('model-sys.npz'):
        with np.load('model-sys.npz') as f:
            param_values = [f['arr_%d' % i] for i in range(len(f.files))]
            lasagne.layers.set_all_param_values(network, param_values)
    if os.path.exists('model-dia.npz'):
        with np.load('model-dia.npz') as f:
            param_values = [f['arr_%d' % i] for i in range(len(f.files))]
            lasagne.layers.set_all_param_values(network_dia, param_values)

    # Finally, launch the training loop.
    print("Starting training...")
    # We iterate over epochs:
    for epoch in range(num_epochs):
        # In each epoch, we do a full pass over the training data:
        train_err_sys = 0
        train_err_dia = 0
        train_batches = 0
        training_index = 1
        validation_index = mriIter.last_training_index + 1

        start_time = time.time()
        while mriIter.has_more_training_data(training_index + batch_size):
            gc.collect()
            print("Training index %s" % training_index)
            inputs, systole, diastole, metadata = mriIter.get_median_bucket_data(training_index, batch_size, return_gender_age=True)
            train_err_sys += train_fn(inputs, systole, metadata)
            train_err_dia += train_fn_dia(inputs, diastole, metadata)
            train_batches += batch_size
            training_index += batch_size

        augmented_training_index = training_index
        while (augmented_training_index < 500):
            gc.collect()
            print("Augmented training index: %s" % augmented_training_index)
            inputs, systole, diastole, metadata= mriIter.get_augmented_data(augmented_training_index, training_index - batch_size, return_gender_age=True)
            train_err_sys += train_fn(inputs, systole, metadata)
            train_err_dia += train_fn_dia(inputs, diastole, metadata)
            augmented_training_index += batch_size

        # And a full pass over the validation data:
        val_err_sys = 0
        val_acc_sys = 0
        val_err_dia = 0
        val_acc_dia = 0
        val_batches = 0
        while mriIter.has_more_data(validation_index):
            gc.collect()
            print("Validation index %s" % validation_index)
            inputs, systole, diastole, metadata= mriIter.get_median_bucket_data(validation_index, batch_size, return_gender_age=True)
            # systole, diastole = targets
            err, prediction = val_fn(inputs, systole, metadata)
            y = 0
            for prob_set in prediction:
                prob_dist = np.cumsum(prob_set)
                v = np.array(range(prediction.shape[1]))
                heavy = (v >= systole[y])
                sq_dists = (prob_dist - heavy)**2
                # print(prediction.shape)
                val_err_sys += err
                val_acc_sys += (sum(sq_dists) / 600.)
                y += 1

            err, prediction = val_fn_dia(inputs, systole, metadata)
            y = 0
            for prob_set in prediction:
                prob_dist = np.cumsum(prob_set)
                v = np.array(range(prediction.shape[1]))
                heavy = (v >= diastole[y])
                sq_dists = (prob_dist - heavy)**2
                val_err_dia += err
                val_acc_dia += (sum(sq_dists) / 600.)
                y += 1
            val_batches += batch_size
            validation_index += batch_size

        # Then we print the results for this epoch:
        print("Epoch {} of {} took {:.3f}s".format(
            epoch + 1, num_epochs, time.time() - start_time))
        # print("  training loss:\t\t{:.6f}".format(train_err / train_batches))
        # print("  validation loss:\t\t{:.6f}".format(val_err / val_batches))
        print("Validation Sum Sqrts Systolic: {}".format(val_acc_sys))
        print("Validation Sum Sqrts Diastolic: {}".format(val_acc_dia))
        print("Train Err Systole: {}".format(train_err_sys))
        print("Train Err Diastole: {}".format(train_err_dia))
        print("CRPS:\t\t{:.6f} %".format(
            (val_acc_sys + val_acc_dia) / (val_batches) * .5))


        # Optionally, you could now dump the network weights to a file like this:
        np.savez('model-sys.npz', *lasagne.layers.get_all_param_values(network))
        np.savez('model-dia.npz', *lasagne.layers.get_all_param_values(network_dia))
    #
    # And load them again later on like this:
    # with np.load('model.npz') as f:
    #     param_values = [f['arr_%d' % i] for i in range(len(f.files))]
    # lasagne.layers.set_all_param_values(network, param_values)


if __name__ == '__main__':
    if ('--help' in sys.argv) or ('-h' in sys.argv):
        print("Trains a neural network on MNIST using Lasagne.")
        print("Usage: %s [MODEL [EPOCHS]]" % sys.argv[0])
        print()
        print("EPOCHS: number of training epochs to perform (default: 1)")
    else:
        kwargs = {}
        if len(sys.argv) > 1:
            kwargs['num_epochs'] = int(sys.argv[1])
        main(**kwargs)