def test_one_hot_from_labels_int(rng):
nc = 19
labels = rng.randint(nc, size=1000)
yref = np.zeros((len(labels), nc))
yref[np.arange(len(labels)), labels] = 1
y0 = one_hot_from_labels(labels)
y1 = one_hot_from_labels(labels, classes=nc+5)
assert np.array_equal(y0, yref)
assert np.array_equal(y0, y1[:, :nc])
assert (y1[:, nc:] == 0).all()
def test_one_hot_from_labels_int(rng):
nc = 19
labels = rng.randint(nc, size=1000)
yref = np.zeros((len(labels), nc))
yref[np.arange(len(labels)), labels] = 1
y0 = one_hot_from_labels(labels)
y1 = one_hot_from_labels(labels, classes=nc + 5)
assert np.array_equal(y0, yref)
assert np.array_equal(y0, y1[:, :nc])
assert (y1[:, nc:] == 0).all()
def test_one_hot_from_labels_skip(rng):
labels = 2*rng.randint(4, size=1000)
yref = np.zeros((len(labels), labels.max()+1))
yref[np.arange(len(labels)), labels] = 1
y = one_hot_from_labels(labels)
assert np.array_equal(y, yref)
def test_one_hot_from_labels_skip(rng):
labels = 2 * rng.randint(4, size=1000)
yref = np.zeros((len(labels), labels.max() + 1))
yref[np.arange(len(labels)), labels] = 1
y = one_hot_from_labels(labels)
assert np.array_equal(y, yref)
def test_one_hot_from_labels_float(rng):
classes = rng.uniform(0, 9, size=11)
inds = rng.randint(len(classes), size=1000)
labels = classes[inds]
yref = np.zeros((len(labels), len(classes)))
yref[np.arange(len(labels)), inds] = 1
y = one_hot_from_labels(labels, classes=classes)
assert np.array_equal(y, yref)
def test_one_hot_from_labels_float(rng):
classes = rng.uniform(0, 9, size=11)
inds = rng.randint(len(classes), size=1000)
labels = classes[inds]
yref = np.zeros((len(labels), len(classes)))
yref[np.arange(len(labels)), inds] = 1
y = one_hot_from_labels(labels, classes=classes)
assert np.array_equal(y, yref)
tau_rc=tau_rc,
amp=amp,
learners=learners), fh)
print("Saved %r" % filename)
else:
with open(filename, 'rb') as fh:
filedata = dill.load(fh)
globals().update(filedata)
# --- sanity test (to check that everything saved fine)
# for learner in learners:
# errors = learner.test(test_set)
# print('%s: %0.3f' % (learner.name, errors.mean()))
# --- plot weighted selectivity (rows=learners, cols=layers, traces=neurons)
testT = one_hot_from_labels(testY, classes=10).astype(float)
testW = testT / testT.sum(0)
test_outs = [learner.network.forward(testX)[1] for learner in learners]
# fig_sel = plt.figure(figsize=(6.4, 3.5))
# rows = len(learners)
# cols = learners[0].network.n_layers
# for i, outs in enumerate(test_outs):
# # outs = outs[1:-1]
# outs = outs[1:]
# outs[-1][:] = softmax(outs[-1], axis=1)
# for j, h in enumerate(outs):
# h = h / (h.sum(0) + 1e-16)
# R = np.dot(h.T, testT) # responses: neurons x classes
# n_hid = 6000
# n_hid = 12000
# n_hid = 15000
# n_hid = 21000
# --- load the data
print("Loading data (augment=%r)" % (augment, ))
train, test = load_cifar10('~/data/cifar-10-python.tar.gz')
train_images, train_labels = train
test_images, test_labels = test
train_images = (1. / 128) * train_images - 1
test_images = (1. / 128) * test_images - 1
del train, test
train_targets = one_hot_from_labels(train_labels, classes=10)
test_targets = one_hot_from_labels(test_labels, classes=10)
assert train_images.ndim == test_images.ndim == 2
n_vis = train_images.shape[1]
n_out = train_targets.shape[1]
shape = (3, 32, 32)
c = shape[0]
assert np.prod(shape) == n_vis
per_batch = 10000
if not augment:
def batches():
return itertools.izip(train_images.reshape(-1, per_batch, n_vis),
train_targets.reshape(-1, per_batch, n_out))
rng = np.random.RandomState(8)
# rng = np.random
# --- data
mnist = load_mnist('~/data/mnist.pkl.gz')
(Xtrain, Ytrain), (Xtest, Ytest) = mnist
labels = np.unique(Ytrain)
n_labels = len(labels)
def preprocess(images):
images[:] *= 2
images[:] -= 1
preprocess(Xtrain), preprocess(Xtest)
Ttrain = one_hot_from_labels(Ytrain, classes=n_labels)
Ttest = one_hot_from_labels(Ytest, classes=n_labels)
# --- params
# dhids = [600, 300]
dhids = [500, 500]
if 1:
# epochs = 0.1
# epochs = 0.5
# epochs = 1.5
# epochs = 3
epochs = 6
# epochs = 10
# epochs = 15
# --- load data
s_in = (28, 28)
n_in = np.prod(s_in)
n_out = 10
train, test = load_mnist('~/data/mnist.pkl.gz')
# train = (train[0][:1000], train[1][:1000])
# train = (train[0][:10000], train[1][:10000])
trainX, trainY = train
testX, testY = test
for images in [trainX, testX]:
images[:] = 2 * images - 1 # normalize to -1 to 1
trainT = one_hot_from_labels(trainY, classes=10)
testT = one_hot_from_labels(testY, classes=10)
assert trainX.shape[1] == n_in
assert trainT.shape[1] == n_out
# --- NEF network
# filename = None
# filename = 'combined_offline_mnist_nef.dil'
# filename = 'combined_offline_mnist_test.dil'
# filename = 'combined_offline_mnist_sgd25.dil'
# filename = 'combined_offline_mnist_sgd25vanilla.dil'
# filename = 'combined_offline_mnist_sgd50vanilla.dil'
# filename = 'combined_offline_mnist_sgd50vanillareset.dil'
filename = 'combined_offline_mnist_sgd50nesterov.dil'