def train_model(lrmodel, opt, cost, X, Y, devX, devY, devscores):
"""
Train model, using pearsonr on dev for early stopping
"""
done = False
best = -1.0
r = np.arange(1, 6)
train_set = ArrayIterator(X=X, y=Y, make_onehot=False)
valid_set = ArrayIterator(X=devX, y=devY, make_onehot=False)
eval_epoch = 10
while not done:
callbacks = Callbacks(lrmodel, eval_set=valid_set)
lrmodel.fit(train_set, optimizer=opt, num_epochs=eval_epoch,
cost=cost, callbacks=callbacks)
# Every 10 epochs, check Pearson on development set
yhat = np.dot(lrmodel.get_outputs(valid_set), r)
score = pearsonr(yhat, devscores)[0]
if score > best:
neon_logger.display('Dev Pearson: {}'.format(score))
best = score
bestlrmodel = copy.copy(lrmodel)
else:
done = True
eval_epoch += 10
yhat = np.dot(bestlrmodel.get_outputs(valid_set), r)
score = pearsonr(yhat, devscores)[0]
neon_logger.display('Dev Pearson: {}'.format(score))
return bestlrmodel
def test_iterator():
print('Testing iterator based data loader')
parser = NeonArgparser(__doc__)
args = parser.parse_args()
(X_train, y_train), (X_test, y_test), nclass = load_cifar10_imgs(path=args.data_dir)
train = ArrayIterator(X_train, y_train, nclass=nclass, lshape=(3, 32, 32))
test = ArrayIterator(X_test, y_test, nclass=nclass, lshape=(3, 32, 32))
return run(args, train, test)
def test_dataset(backend_default, data):
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=data)
train_set = ArrayIterator(X_train, y_train, nclass=nclass)
train_set.be = NervanaObject.be
for i in range(2):
for X_batch, y_batch in train_set:
neon_logger.display("Xshape: {}, yshape: {}".format(X_batch.shape, y_batch.shape))
train_set.index = 0
def test_dataset(backend_default, data):
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=data)
train_set = ArrayIterator(X_train, y_train, nclass=nclass)
train_set.be = NervanaObject.be
for i in range(2):
for X_batch, y_batch in train_set:
print X_batch.shape, y_batch.shape
train_set.index = 0
def test_dataset(backend_default, data):
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=data)
train_set = ArrayIterator(X_train, y_train, nclass=nclass)
train_set.be = NervanaObject.be
for i in range(2):
for X_batch, y_batch in train_set:
print X_batch.shape, y_batch.shape
train_set.index = 0
def test_iterator():
print('Testing data iterator')
NervanaObject.be.gen_rng(args.rng_seed)
image_dir = os.path.join(args.data_dir, 'cifar-extracted')
train_manifest, val_manifest = ingest_cifar10(out_dir=image_dir)
(X_train, y_train), (X_test, y_test), nclass = load_cifar10_imgs(
train_manifest, val_manifest)
train = ArrayIterator(X_train, y_train, nclass=nclass, lshape=(3, 32, 32))
test = ArrayIterator(X_test, y_test, nclass=nclass, lshape=(3, 32, 32))
return run(args, train, test)
def load_data_to_array_iterator(self):
"""
Load data into ArrayIterators
Returns:
tuple(:obj:`neon.data.ArrayIterators`): ArrayIterator train_set, ArrayIterator test_set
"""
X_train, y_train, X_test, y_test = self.load_data_from_file()
train_set = ArrayIterator(X=X_train, y=y_train, nclass=2)
if X_test.size == 0 and y_test.size == 0:
test_set = None
else:
test_set = ArrayIterator(X=X_test, y=y_test, nclass=2)
return train_set, test_set
def wsd_classify(x_test, y_test=None):
"""
classifiy target word. output all word senses ranked according to the most probable sense
Args:
x_test(numpy.ndarray): input x data for inference
y_test: input y data for inference
Returns:
str: predicted values by the model
"""
# test set
x_test = np.array(x_test)
if y_test is not None:
y_test = np.array(y_test)
test_set = ArrayIterator(X=x_test, y=y_test, make_onehot=False)
mlp_clf = MostCommonWordSense(args.rounding, args.callback_args,
args.epochs)
# load existing model
mlp_clf.load(args.model_prm)
results = mlp_clf.get_outputs(test_set)
return results
def predict_batch(batch):
inp = ArrayIterator(X=batch,
y=None,
nclass=2,
lshape=params.frame_shape,
make_onehot=False)
return model.get_outputs(inp)
def load_data(self):
# load up the mnist data set
# split into train and tests sets
(X_train,
y_train), (X_test,
y_test), nclass = load_mnist(path=self.args.data_dir)
# setup a training set iterator
self.train_set = ArrayIterator(X_train,
y_train,
nclass=nclass,
lshape=(1, 28, 28))
# setup a validation data set iterator
self.valid_set = ArrayIterator(X_test,
y_test,
nclass=nclass,
lshape=(1, 28, 28))
def get_offset_predicted_naive(predictions):
(replay_memory.clean_values_toone(predictions)[0] * 30) + 30
input_iter = ArrayIterator(X=batch, y=None, make_onehot=False)
outputs = mlp.get_outputs(input_iter)
outputs_ = 30 * outputs[0, 0]
print "outputs", outputs_
return int(round(outputs_))
def test_model_get_outputs(backend_default, data):
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=data)
train_set = ArrayIterator(X_train[:backend_default.bsz * 3])
init_norm = Gaussian(loc=0.0, scale=0.1)
layers = [Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin()),
Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))]
mlp = Model(layers=layers)
out_list = []
mlp.initialize(train_set)
for x, t in train_set:
x = mlp.fprop(x)
out_list.append(x.get().T.copy())
ref_output = np.vstack(out_list)
train_set.reset()
output = mlp.get_outputs(train_set)
assert np.allclose(output, ref_output)
def get_offset_predicted(predictions):
batch = np.zeros((bot_params.batch_size, 1))
batch[0, 0] = replay_memory.clean_values_toone(predictions)[0]
input_iter = ArrayIterator(X=batch, y=None, make_onehot=False)
outputs = mlp.get_outputs(input_iter)
outputs_ = 30 * outputs[0, 0]
print "outputs", outputs_
return int(round(outputs_))
def most_common_word_train(x_train, y_train, x_valid, y_valid):
"""
Train an MLP model, save it and evaluate it
Args:
x_train: x data for training
y_train: y data for training
x_valid: x data for validation
y_valid: x data for validation
Returns:
str: reslts, predicted values by the model
"""
# train set
x_train = np.array(x_train)
y_train1 = np.array(y_train)
train_set = ArrayIterator(X=x_train, y=y_train1, make_onehot=False)
# validation set
x_valid = np.array(x_valid)
y_valid1 = np.array(y_valid)
valid_set = ArrayIterator(X=x_valid, y=y_valid1, make_onehot=False)
mlp_model = MostCommonWordSense(args.rounding, args.callback_args,
args.epochs)
# build model
mlp_model.build()
# train
mlp_model.fit(valid_set, train_set)
# save model
mlp_model.save(args.model_prm)
# evaluation
error_rate = mlp_model.eval(valid_set)
logger.info('Mis-classification error on validation set= %0.1f',
error_rate * 100)
reslts = mlp_model.get_outputs(valid_set)
return reslts
def visualize(model, data, max_fm, filename):
data_shape = data.shape
data = data.reshape((data.shape[0], -1))
dataset = ArrayIterator(data, lshape=data_shape[1:])
deconv_file = h5py.File("no_file", driver='core', backing_store=False)
deconv = DeconvCallback(dataset, dataset, max_fm=max_fm, dataset_pct=100)
deconv.on_train_end(deconv_file, model)
deconv_data = h5_deconv_data(deconv_file)
deconv_summary_page(filename, deconv_data, max_fm)
def predict(input_img):
# model.set_batch_size(1)
x_new = np.zeros((params.batch_size, input_img.size), dtype=np.float16)
x_new[0] = mem.prepare_image(input_img)
inp = ArrayIterator(X=x_new,
y=None,
nclass=2,
lshape=params.frame_shape,
make_onehot=False)
qvalues = model.get_outputs(inp)
return qvalues[0][0]
def test_model_get_outputs(backend_default, data):
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=data)
train_set = ArrayIterator(X_train[:backend_default.bsz * 3])
init_norm = Gaussian(loc=0.0, scale=0.1)
layers = [
Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin()),
Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))
]
mlp = Model(layers=layers)
out_list = []
mlp.initialize(train_set)
for x, t in train_set:
x = mlp.fprop(x)
out_list.append(x.get().T.copy())
ref_output = np.vstack(out_list)
train_set.reset()
output = mlp.get_outputs(train_set)
assert np.allclose(output, ref_output)
def DeepCascadeLearning(modelLayers,X_train,Y_train,callbacks,init_uni=Uniform(low=-0.1, high=0.1),
testIterator=None,epochs=2,
cost=GeneralizedCost(costfunc=CrossEntropyMulti()),
opt_gdm=GradientDescentMomentum(learning_rate=0.01,momentum_coef=0.9)):
importantLayersIndexes = list()
i = 0
outputLayer = Affine(nout=10, init=init_uni, activation=Softmax())
modelToPredict = None
for currentLayer in modelLayers:
if(np.shape(currentLayer)):
currentLayer = currentLayer[0]
if((currentLayer.classnm == 'Convolution') or (currentLayer.classnm == 'Affine')):
importantLayersIndexes.append(i)
i += 1
for i in importantLayersIndexes:
modelToTrain = list()
for currentLayer in modelLayers[i:importantLayersIndexes[i+1]]:
modelToTrain.append(currentLayer)
modelToTrain.append(outputLayer)
modelToTrain = Model(modelToTrain)
if(modelToPredict == None):
trainIterator = ArrayIterator(X_train, Y_train, nclass=10, lshape=(3,32,32))
x = trainIterator.__iter__()
callbacks = Callbacks(modelToTrain)
modelToTrain.fit(trainIterator, optimizer=opt_gdm, num_epochs=epochs, cost=GeneralizedCost(costfunc=CrossEntropyMulti()), callbacks=callbacks)
else:
tmpIterator = ArrayIterator(X_train,lshape=(3,32,32))
tmpTrain = modelToPredict.get_outputs(tmpIterator)
tmpIterator = ArrayIterator(tmpTrain[0:20],Y_train[0:20],nclass=10,lshape=(32,30,30))
modelToTrain.fit(tmpIterator, optimizer=opt_gdm, num_epochs=epochs, cost=cost)
if modelToPredict == None:
modelToPredict = list()
else:
modelToPredict = modelToPredict.layers.layers
for currentLayer in modelToTrain.layers.layers[0:-2]:
modelToPredict.append(currentLayer)
modelToPredict = Model(modelToPredict)
return modelToPredict
def it_was_correct(last_in, last_out):
# print "naive", replay_memory.clean_values_toone(last_in)[0], last_out
offset_memory.add_episode(last_in, last_out)
print "osize", offset_memory.outputs.size
if offset_memory.outputs is not None and offset_memory.outputs.size % bot_params.batch_size == 0:
X, y = offset_memory.get_dataset()
train = ArrayIterator(X=X, y=y, make_onehot=False)
mlp.fit(train,
optimizer=optimizer,
num_epochs=1,
cost=cost,
callbacks=Callbacks(mlp))
mlp.save_params(bot_params.aim_weights_path)
def train_regressor(orig_wordvecs, w2v_W, w2v_vocab):
"""
Return regressor to map word2vec to RNN word space
Function modified from:
https://github.com/ryankiros/skip-thoughts/blob/master/training/tools.py
"""
# Gather all words from word2vec that appear in wordvecs
d = defaultdict(lambda: 0)
for w in w2v_vocab.keys():
d[w] = 1
shared = OrderedDict()
count = 0
for w in list(orig_wordvecs.keys())[:-2]:
if d[w] > 0:
shared[w] = count
count += 1
# Get the vectors for all words in 'shared'
w2v = np.zeros((len(shared), 300), dtype='float32')
sg = np.zeros((len(shared), 620), dtype='float32')
for w in shared.keys():
w2v[shared[w]] = w2v_W[w2v_vocab[w]]
sg[shared[w]] = orig_wordvecs[w]
train_set = ArrayIterator(X=w2v, y=sg, make_onehot=False)
layers = [
Linear(nout=620, init=Gaussian(loc=0.0, scale=0.1)),
Bias(init=Constant(0.0))
]
clf = Model(layers=layers)
# regression model is trained using default global batch size
cost = GeneralizedCost(costfunc=SumSquared())
opt = GradientDescentMomentum(0.1, 0.9, gradient_clip_value=5.0)
callbacks = Callbacks(clf)
clf.fit(train_set,
num_epochs=20,
optimizer=opt,
cost=cost,
callbacks=callbacks)
return clf
from neon.layers import GeneralizedCost, Affine, Sequential, MergeMultistream
from neon.models import Model
from neon.optimizers import GradientDescentMomentum
from neon.transforms import Rectlin, Logistic, CrossEntropyBinary
from neon.callbacks.callbacks import Callbacks
from neon.util.argparser import NeonArgparser
# parse the command line arguments
parser = NeonArgparser(__doc__)
args = parser.parse_args()
# hyperparameters
num_epochs = args.epochs
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=args.data_dir)
train_set = ArrayIterator([X_train, X_train], y_train, nclass=nclass, lshape=(1, 28, 28))
valid_set = ArrayIterator([X_test, X_test], y_test, nclass=nclass, lshape=(1, 28, 28))
# weight initialization
init_norm = Gaussian(loc=0.0, scale=0.01)
# initialize model
path1 = Sequential(layers=[Affine(nout=100, init=init_norm, activation=Rectlin()),
Affine(nout=100, init=init_norm, activation=Rectlin())])
path2 = Sequential(layers=[Affine(nout=100, init=init_norm, activation=Rectlin()),
Affine(nout=100, init=init_norm, activation=Rectlin())])
layers = [MergeMultistream(layers=[path1, path2], merge="stack"),
Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))]
Conv(fshape=(4, 4, 16), init=init_uni, activation=Rectlin()),
Pooling(fshape=2, strides=2),
Conv(fshape=(4, 4, 32), init=init_uni, activation=Rectlin()),
Pooling(fshape=2, strides=2),
Conv(fshape=(4, 4, 32), init=init_uni, activation=Rectlin()),
Pooling(fshape=2, strides=2),
Affine(nout=500, init=init_uni, activation=Rectlin()),
Affine(nout=11, init=init_uni, activation=Softmax())
]
model = Model(layers)
model.load_params('model.pkl')
data = readfile('PreImage', 'label.csv')
X_test = data.test_data
test_set = ArrayIterator(X_test, None, nclass=11, lshape=(1, 200, 200))
true = data.test_label
out = model.get_outputs(test_set)
row = len(X_test)
pred = np.zeros((row, 1))
i = 0
while i < row:
pred[i] = out[i].argmax()
i = i + 1
pred = pred + 1
loss = abs(true - pred)
print(loss)
count = 0
for i in range(len(loss)):
if loss[i] != 0:
count = count + 1
from neon.initializers import Gaussian
from neon.layers import Conv, Pooling, GeneralizedCost, Affine
from neon.optimizers import GradientDescentMomentum, MultiOptimizer, Schedule
from neon.transforms import Rectlin, Softmax, CrossEntropyMulti
from neon.models import Model
from neon.data import ArrayIterator
import numpy as np
parser = NeonArgparser(__doc__)
args = parser.parse_args()
NervanaObject.be.enable_winograd = 4
# setup data provider
X_train = np.random.uniform(-1, 1, (128, 3*224*224))
y_train = np.random.uniform(-1, 1, (128, 1000))
train = ArrayIterator(X_train, y_train, nclass=1000, lshape=(3, 224, 224))
layers = [Conv((11, 11, 64), init=Gaussian(scale=0.01),
activation=Rectlin(), padding=3, strides=4),
Pooling(3, strides=2),
Conv((5, 5, 192), init=Gaussian(scale=0.01), activation=Rectlin(), padding=2),
Pooling(3, strides=2),
Conv((3, 3, 384), init=Gaussian(scale=0.03), activation=Rectlin(), padding=1),
Conv((3, 3, 256), init=Gaussian(scale=0.03), activation=Rectlin(), padding=1),
Conv((3, 3, 256), init=Gaussian(scale=0.03), activation=Rectlin(), padding=1),
Pooling(3, strides=2),
Affine(nout=4096, init=Gaussian(scale=0.01), activation=Rectlin()),
Affine(nout=4096, init=Gaussian(scale=0.01), activation=Rectlin()),
Affine(nout=1000, init=Gaussian(scale=0.01), activation=Softmax())]
model = Model(layers=layers)
from neon.data import ArrayIterator, load_cifar10
from neon.initializers import Uniform
from neon.layers import GeneralizedCost, Affine
from neon.models import Model
from neon.optimizers import GradientDescentMomentum
from neon.transforms import Misclassification, CrossEntropyBinary, Logistic, Rectlin
from neon.callbacks.callbacks import Callbacks
from neon.util.argparser import NeonArgparser
# parse the command line arguments
parser = NeonArgparser(__doc__)
args = parser.parse_args()
(X_train, y_train), (X_test, y_test), nclass = load_cifar10(path=args.data_dir)
train = ArrayIterator(X_train, y_train, nclass=nclass, lshape=(3, 32, 32))
test = ArrayIterator(X_test, y_test, nclass=nclass, lshape=(3, 32, 32))
init_uni = Uniform(low=-0.1, high=0.1)
opt_gdm = GradientDescentMomentum(learning_rate=0.01, momentum_coef=0.9)
# set up the model layers
layers = [
Affine(nout=200, init=init_uni, activation=Rectlin()),
Affine(nout=10, init=init_uni, activation=Logistic(shortcut=True))
]
cost = GeneralizedCost(costfunc=CrossEntropyBinary())
mlp = Model(layers=layers)
drop_layers = {ky: -1 for ky in ["drop6", "drop7"]}
for l in drop_layers:
drop_layer = md.getlayer(l)
drop_layers[l] = drop_layer['config']['keep']
drop_layer['config']['keep'] = 1.0
md = dict(md)
# generate a fake input
IM_SIZE = (be.bsz, 3, 224, 224)
np.random.seed(1)
im = np.random.randint(-150, 150, IM_SIZE)
fake_labels = np.zeros((IM_SIZE[0], 10))
# need this iterator to initialize the model
train = ArrayIterator(im.reshape(IM_SIZE[0], -1).copy(),
fake_labels,
nclass=10,
lshape=IM_SIZE[1:])
# deserialize the neon model
model = Model(md, train)
# make sure the deserialization is setting the compat mode correctly
assert be.compat_mode == 'caffe'
# generate a fake input
im_neon = be.array(im.reshape(IM_SIZE[0], -1).astype(np.float32).T.copy())
out_neon = model.fprop(im_neon)
# get the neon layer output order using the layer names
l_map = []
loss_lay = model.layers.layers[-1].name
polar_predict, filenames = load_data(os.path.join(NACPolarDir,
'Unlabelled Regions'),
is_augment=False)
files = []
locations = []
for filename in filenames:
files.append(filename.split('/')[-1])
locations.append(filename.split(files[-1])[0])
log('Data loaded from ' + os.path.join(NACPolarDir, 'Unlabelled Regions'))
classes = ["Crater", "Not Crater"]
inference_set = ArrayIterator(polar_predict,
None,
nclass=2,
lshape=(1, 32, 32))
out = model.get_outputs(inference_set)
labels = np.argmax(out, axis=1)
count = []
threshold = []
threshold = np.subtract(
1,
np.divide(
np.multiply(-400, np.power((np.arange(100).astype(float)), 2)),
10 * np.min(
np.multiply(-400, np.power(
(np.arange(100).astype(float)), 2)))))
from neon.backends import gen_backend
# parser = NeonArgparser(__doc__)
# args = parser.parse_args(gen_be=False)
be = gen_backend(batch_size=100, backend='gpu')
traindir = 'train'
imwidth = 256
data = np.load("data.npy")
coordinate = np.load("coordinate.npy")
# X_train, y_train = data[0:1], coordinate[0:256]
X_train, y_train = np.asarray(data[0:3000]), np.asarray(coordinate[0:3000])
train_set = ArrayIterator(X_train[0:2000],
y_train[0:2000],
make_onehot=False,
lshape=(3, 256, 256))
eval_set = ArrayIterator(X_train[1000:2000],
y_train[1000:2000],
make_onehot=False,
lshape=(3, 256, 256))
test_set = ArrayIterator(X_train[2000:2500],
y_train[2000:2500],
make_onehot=False,
lshape=(3, 256, 256))
# weight initialization
init_norm = Gaussian(loc=0.0, scale=0.01)
# setup model layers
def evaluate(model, vocab, data_path, seed=1234, evaltest=False, vocab_size_layer=20002):
"""
Run experiment
"""
neon_logger.display('Preparing SICK evaluation data...')
# Check if SICK data exists in specified directory, otherwise download
sick_data = SICK(path=data_path)
train, dev, test, scores = sick_data.load_eval_data()
np.random.seed(seed)
shuf_idxs = np.random.permutation(range(len(train[0])))
train_A_shuf = train[0][shuf_idxs]
train_B_shuf = train[1][shuf_idxs]
scores_shuf = scores[0][shuf_idxs]
train_A_tok = tokenize_input(train_A_shuf, vocab=vocab)
train_B_tok = tokenize_input(train_B_shuf, vocab=vocab)
dev_A_tok = tokenize_input(dev[0], vocab=vocab)
dev_B_tok = tokenize_input(dev[1], vocab=vocab)
# Get iterator from tokenized data
train_set_A = SentenceEncode(train_A_tok, [], len(train_A_tok), vocab_size_layer,
max_len=30, index_from=2)
train_set_B = SentenceEncode(train_B_tok, [], len(train_B_tok), vocab_size_layer,
max_len=30, index_from=2)
# Compute embeddings using iterator
trainA = model.get_outputs(train_set_A)
trainB = model.get_outputs(train_set_B)
dev_set_A = SentenceEncode(dev_A_tok, [], len(dev_A_tok), vocab_size_layer,
max_len=30, index_from=2)
dev_set_B = SentenceEncode(dev_B_tok, [], len(dev_B_tok), vocab_size_layer,
max_len=30, index_from=2)
devA = model.get_outputs(dev_set_A)
devB = model.get_outputs(dev_set_B)
trainF = np.c_[np.abs(trainA - trainB), trainA * trainB]
devF = np.c_[np.abs(devA - devB), devA * devB]
trainY = encode_labels(scores_shuf, ndata=len(trainF))
devY = encode_labels(scores[1], ndata=len(devF))
lrmodel, opt, cost = prepare_model(ninputs=trainF.shape[1])
neon_logger.display('Training the regression model...')
bestlrmodel = train_model(lrmodel, opt, cost, trainF,
trainY, devF, devY, scores[1][:len(devF)])
if evaltest:
test_A_tok = tokenize_input(test[0], vocab=vocab)
test_B_tok = tokenize_input(test[1], vocab=vocab)
test_set_A = SentenceEncode(test_A_tok, [], len(test_A_tok), vocab_size_layer,
max_len=30, index_from=2)
test_set_B = SentenceEncode(test_B_tok, [], len(test_B_tok), vocab_size_layer,
max_len=30, index_from=2)
testA = model.get_outputs(test_set_A)
testB = model.get_outputs(test_set_B)
testF = np.c_[np.abs(testA - testB), testA * testB]
neon_logger.display('Evaluating using vectors and linear regression model')
r = np.arange(1, 6)
yhat = np.dot(bestlrmodel.get_outputs(ArrayIterator(testF)), r)
pr = pearsonr(yhat, scores[2][:len(yhat)])[0]
sr = spearmanr(yhat, scores[2][:len(yhat)])[0]
se = np.mean((yhat - scores[2][:len(yhat)]) ** 2)
neon_logger.display('Test Pearson: ' + str(pr))
neon_logger.display('Test Spearman: ' + str(sr))
neon_logger.display('Test MSE: ' + str(se))
return yhat
def main():
# setup the model and run for num_epochs saving the last state only
# this is at the top so that the be is generated
model = gen_model(args.backend)
# setup data iterators
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=args.data_dir)
NN = batch_size*5 # avoid partial mini batches
if args.backend == 'nervanacpu' or args.backend == 'cpu':
# limit data since cpu backend runs slower
train = ArrayIterator(X_train[:NN], y_train[:NN],
nclass=nclass, lshape=(1, 28, 28))
valid = ArrayIterator(X_test[:NN], y_test[:NN],
nclass=nclass, lshape=(1, 28, 28))
else:
train = ArrayIterator(X_train, y_train, nclass=nclass, lshape=(1, 28, 28))
valid = ArrayIterator(X_test, y_test, nclass=nclass, lshape=(1, 28, 28))
# serialization related
cost = GeneralizedCost(costfunc=CrossEntropyBinary())
opt_gdm = GradientDescentMomentum(learning_rate=0.1, momentum_coef=0.9)
checkpoint_model_path = os.path.join('./', 'test_oneshot.pkl')
checkpoint_schedule = 1 # save at every step
callbacks = Callbacks(model)
callbacks.add_callback(SerializeModelCallback(checkpoint_model_path,
checkpoint_schedule,
history=2))
# run the fit all the way through saving a checkpoint e
model.fit(train,
optimizer=opt_gdm,
num_epochs=num_epochs,
cost=cost,
callbacks=callbacks)
# setup model with same random seed run epoch by epoch
# serializing and deserializing at each step
model = gen_model(args.backend)
cost = GeneralizedCost(costfunc=CrossEntropyBinary())
opt_gdm = GradientDescentMomentum(learning_rate=0.1, momentum_coef=0.9)
# reset data iterators
train.reset()
valid.reset()
checkpoint_model_path = os.path.join('./', 'test_manyshot.pkl')
checkpoint_schedule = 1 # save at evey step
for epoch in range(num_epochs):
# _0 points to state at end of epoch 0
callbacks = Callbacks(model)
callbacks.add_callback(SerializeModelCallback(checkpoint_model_path,
checkpoint_schedule,
history=num_epochs))
model.fit(train,
optimizer=opt_gdm,
num_epochs=epoch+1,
cost=cost,
callbacks=callbacks)
# load saved file
prts = os.path.splitext(checkpoint_model_path)
fn = prts[0] + '_%d' % epoch + prts[1]
model.load_params(fn) # load the saved weights
# compare test_oneshot_<num_epochs>.pkl to test_manyshot_<num_epochs>.pkl
if not compare_model_pickles('test_oneshot_%d.pkl' % (num_epochs-1),
'test_manyshot_%d.pkl' % (num_epochs-1)):
print 'No Match'
sys.exit(1)
else:
print 'Match'
config = Config()
# setup backend
be = gen_backend(backend=config.backend,
batch_size=config.batch_size,
rng_seed=config.rng_seed,
datatype=np.float32,
stochastic_round=False)
# generate data
X = np.random.rand(config.batch_size, config.image_width**2 * 3)
y = np.random.randint(config.ydim, size=config.batch_size)
# setup a training set iterator
data = ArrayIterator(X,
y,
nclass=config.ydim,
lshape=(3, config.image_width, config.image_width))
# setup weight initialization function
init_norm = Gaussian(loc=0.0, scale=0.01)
init1 = Gaussian(scale=0.01)
init2 = Gaussian(scale=0.03)
init_uni = GlorotUniform()
# setup model layers
# no local response normalization
relu = Rectlin()
layers = []
layers.append(
Conv((11, 11, 96),
strides=4,
init_emb = Uniform(-0.1 / embedding_dim, 0.1 / embedding_dim)
h5f = h5py.File(fname_h5, 'r')
reviews, h5train, h5valid = h5f['reviews'], h5f['train'], h5f['valid']
ntrain, nvalid, nclass = reviews.attrs[
'ntrain'], reviews.attrs['nvalid'], reviews.attrs['nclass']
# make train dataset
Xy = h5train[:ntrain]
X = [xy[1:] for xy in Xy]
y = [xy[0] for xy in Xy]
X_train, y_train = get_paddedXY(
X, y, vocab_size=vocab_size, sentence_length=sentence_length)
train_set = ArrayIterator(X_train, y_train, nclass=nclass)
# make valid dataset
Xy = h5valid[:nvalid]
X = [xy[1:] for xy in Xy]
y = [xy[0] for xy in Xy]
X_valid, y_valid = get_paddedXY(
X, y, vocab_size=vocab_size, sentence_length=sentence_length)
valid_set = ArrayIterator(X_valid, y_valid, nclass=nclass)
# initialization
init_glorot = GlorotUniform()
# define layers
from neon.benchmark import Benchmark
from neon.data import ArrayIterator
from neon.initializers import Gaussian
from neon.layers import Conv, Affine, Pooling
from neon.models import Model
gen_backend(batch_size=16)
test_layers = [
Conv((2, 2, 4), init=Gaussian(scale=0.01), padding=3, strides=4),
Pooling(3, strides=2),
Affine(nout=10, init=Gaussian(scale=0.01))
]
NervanaObject.be.enable_winograd = 4
x = np.random.uniform(-1, 1, (16, 3 * 2 * 2))
y = np.random.randint(0, 9, (16, 10))
test_dataset = ArrayIterator(x, y, nclass=10, lshape=(3, 2, 2))
def test_empty_dataset():
model = Model(test_layers)
b = Benchmark(model=model)
with pytest.raises(ValueError):
b.time([], niterations=5, inference=True)
def test_fw_bw_no_cost_or_optimizer():
model = Model(test_layers)
model.initialize(test_dataset)
b = Benchmark(model=model)
with pytest.raises(RuntimeError):
b.time(test_dataset, niterations=1)
from neon.util.argparser import NeonArgparser parser = NeonArgparser(__doc__) args = parser.parse_args() from neon.data import MNIST from neon.data import ArrayIterator mnist = MNIST() (X_train, y_train), (X_test, y_test), nclass = mnist.load_data() #print "X_test: %s" % X_test[1] # setup training set iterator train_set = ArrayIterator(X_train, y_train, nclass=nclass) # setup test set iterator test_set = ArrayIterator(X_test, y_test, nclass=nclass) #Initialize weights to small random numbers with Gaussian from neon.initializers import Gaussian init_norm = Gaussian(loc=0.0, scale=0.01) #Affine is a FC network with 100 hidden units from neon.layers import Affine #We will use ReLu for hidden units and SoftMax for output units. Softmax is used to ensure that #all outputs sum up to 1 and are within the range of [0, 1] from neon.transforms import Rectlin, Softmax layers = []
from neon.layers import Conv, Pooling, GeneralizedCost, Deconv
from neon.models import Model
from neon.optimizers import GradientDescentMomentum
from neon.transforms import Rectlin, SumSquared
from neon.callbacks.callbacks import Callbacks
from neon.util.argparser import NeonArgparser
# parse the command line arguments
parser = NeonArgparser(__doc__)
args = parser.parse_args()
# Load dataset
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=args.data_dir)
# Set input and target to X_train
train = ArrayIterator(X_train, lshape=(1, 28, 28))
# Initialize the weights and the learning rule
init_uni = Uniform(low=-0.1, high=0.1)
opt_gdm = GradientDescentMomentum(learning_rate=0.001, momentum_coef=0.9)
# Strided conv autoencoder
bn = False
layers = [
Conv((4, 4, 8), init=init_uni, activation=Rectlin(), batch_norm=bn),
Pooling(2),
Conv((4, 4, 32), init=init_uni, activation=Rectlin(), batch_norm=bn),
Pooling(2),
Deconv(fshape=(4, 4, 8),
init=init_uni,
activation=Rectlin(),
def test_model_serialize(backend_default, data):
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=data)
train_set = ArrayIterator([X_train, X_train],
y_train,
nclass=nclass,
lshape=(1, 28, 28))
init_norm = Gaussian(loc=0.0, scale=0.01)
# initialize model
path1 = Sequential([
Conv((5, 5, 16),
init=init_norm,
bias=Constant(0),
activation=Rectlin()),
Pooling(2),
Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())
])
path2 = Sequential([
Affine(nout=100,
init=init_norm,
bias=Constant(0),
activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())
])
layers = [
MergeMultistream(layers=[path1, path2], merge="stack"),
Affine(nout=20, init=init_norm, batch_norm=True, activation=Rectlin()),
Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))
]
tmp_save = 'test_model_serialize_tmp_save.pickle'
mlp = Model(layers=layers)
mlp.optimizer = GradientDescentMomentum(learning_rate=0.1,
momentum_coef=0.9)
mlp.cost = GeneralizedCost(costfunc=CrossEntropyBinary())
mlp.initialize(train_set, cost=mlp.cost)
n_test = 3
num_epochs = 3
# Train model for num_epochs and n_test batches
for epoch in range(num_epochs):
for i, (x, t) in enumerate(train_set):
x = mlp.fprop(x)
delta = mlp.cost.get_errors(x, t)
mlp.bprop(delta)
mlp.optimizer.optimize(mlp.layers_to_optimize, epoch=epoch)
if i > n_test:
break
# Get expected outputs of n_test batches and states of all layers
outputs_exp = []
pdicts_exp = [l.get_params_serialize() for l in mlp.layers_to_optimize]
for i, (x, t) in enumerate(train_set):
outputs_exp.append(mlp.fprop(x, inference=True))
if i > n_test:
break
# Serialize model
mlp.save_params(tmp_save, keep_states=True)
# Load model
mlp = Model(tmp_save)
mlp.initialize(train_set)
outputs = []
pdicts = [l.get_params_serialize() for l in mlp.layers_to_optimize]
for i, (x, t) in enumerate(train_set):
outputs.append(mlp.fprop(x, inference=True))
if i > n_test:
break
# Check outputs, states, and params are the same
for output, output_exp in zip(outputs, outputs_exp):
assert np.allclose(output.get(), output_exp.get())
for pd, pd_exp in zip(pdicts, pdicts_exp):
for s, s_e in zip(pd['states'], pd_exp['states']):
if isinstance(s, list): # this is the batch norm case
for _s, _s_e in zip(s, s_e):
assert np.allclose(_s, _s_e)
else:
assert np.allclose(s, s_e)
for p, p_e in zip(pd['params'], pd_exp['params']):
assert type(p) == type(p_e)
if isinstance(p, list): # this is the batch norm case
for _p, _p_e in zip(p, p_e):
assert np.allclose(_p, _p_e)
elif isinstance(p, np.ndarray):
assert np.allclose(p, p_e)
else:
assert p == p_e
os.remove(tmp_save)