def test_model_get_outputs(backend_default, data):
dataset = MNIST(path=data)
train_set = dataset.train_iter
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 allclose_with_out(output, ref_output[:output.shape[0], :])
# test model benchmark inference
mlp.benchmark(train_set, inference=True, niterations=5)
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)
# test model benchmark inference
mlp.benchmark(train_set, inference=True, niterations=5)
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)
# test model benchmark inference
mlp.benchmark(train_set, inference=True, niterations=5)
def test_model_get_outputs(backend_default, data):
dataset = MNIST(path=data)
train_set = dataset.train_iter
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[:output.shape[0], :])
# test model benchmark inference
mlp.benchmark(train_set, inference=True, niterations=5)
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)
weight_sched = Schedule([22, 44, 65], (1/250.)**(1/3.))
opt_gdm = GradientDescentMomentum(0.01, 0.0, wdecay=0.0005, schedule=weight_sched)
opt = MultiOptimizer({'default': opt_gdm})
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
model.benchmark(train, cost=cost, optimizer=opt, niterations=10, nskip=5)
valmetric = TopKMisclassification(k=5)
# dummy optimizer for benchmarking
# training implementation coming soon
opt_gdm = GradientDescentMomentum(0.0, 0.0)
opt_biases = GradientDescentMomentum(0.0, 0.0)
opt = MultiOptimizer({'default': opt_gdm, 'Bias': opt_biases})
# setup cost function as CrossEntropy
cost = Multicost(costs=[GeneralizedCost(costfunc=CrossEntropyMulti()),
GeneralizedCost(costfunc=CrossEntropyMulti()),
GeneralizedCost(costfunc=CrossEntropyMulti())],
weights=[1, 0., 0.]) # We only want to consider the CE of the main path
assert os.path.exists(args.model_file), 'script requires the trained weights file'
model.load_params(args.model_file)
model.initialize(test, cost)
print 'running speed benchmark...'
model.benchmark(test, cost, opt)
print '\nCalculating performance on validation set...'
test.reset()
mets = model.eval(test, metric=valmetric)
print 'Validation set metrics:'
print 'LogLoss: %.2f, Accuracy: %.1f %% (Top-1), %.1f %% (Top-5)' % (mets[0],
(1.0-mets[1])*100,
(1.0-mets[2])*100)
# setup data provider
X_train = np.random.uniform(-1, 1, (64, 3*224*224))
y_train = np.random.uniform(-1, 1, (64, 1000))
train = ArrayIterator(X_train, y_train, nclass=1000, lshape=(3, 224, 224))
layers = [Conv((3, 3, 64), init=Gaussian(scale=0.01), activation=Rectlin(), padding=1),
Pooling(2, strides=2),
Conv((3, 3, 128), init=Gaussian(scale=0.01), activation=Rectlin(), padding=1),
Pooling(2, strides=2),
Conv((3, 3, 256), init=Gaussian(scale=0.01), activation=Rectlin(), padding=1),
Conv((3, 3, 256), init=Gaussian(scale=0.01), activation=Rectlin(), padding=1),
Pooling(2, strides=2),
Conv((3, 3, 512), init=Gaussian(scale=0.01), activation=Rectlin(), padding=1),
Conv((3, 3, 512), init=Gaussian(scale=0.01), activation=Rectlin(), padding=1),
Pooling(2, strides=2),
Conv((3, 3, 512), init=Gaussian(scale=0.01), activation=Rectlin(), padding=1),
Conv((3, 3, 512), init=Gaussian(scale=0.01), activation=Rectlin(), padding=1),
Pooling(2, 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)
weight_sched = Schedule([22, 44, 65], (1/250.)**(1/3.))
opt_gdm = GradientDescentMomentum(0.01, 0.0, wdecay=0.0005, schedule=weight_sched)
opt = MultiOptimizer({'default': opt_gdm})
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
model.benchmark(train, cost=cost, optimizer=opt, niterations=10, nskip=1)
import os from neon.util.argparser import NeonArgparser from neon.util.persist import load_obj from neon.transforms import Misclassification, CrossEntropyMulti from neon.optimizers import GradientDescentMomentum from neon.layers import GeneralizedCost from neon.models import Model from neon.data import DataLoader, ImageParams # parse the command line arguments (generates the backend) parser = NeonArgparser(__doc__) args = parser.parse_args() # setup data provider test_dir = os.path.join(args.data_dir, 'val') shape = dict(channel_count=3, height=32, width=32) test_params = ImageParams(center=True, flip=False, **shape) common = dict(target_size=1, nclasses=10) test_set = DataLoader(set_name='val', repo_dir=test_dir, media_params=test_params, **common) model = Model(load_obj(args.model_file)) cost = GeneralizedCost(costfunc=CrossEntropyMulti()) opt = GradientDescentMomentum(0.1, 0.9, wdecay=0.0001) model.initialize(test_set, cost=cost) acc = 1.0 - model.eval(test_set, metric=Misclassification())[0] print 'Accuracy: %.1f %% (Top-1)' % (acc*100.0) model.benchmark(test_set, cost=cost, optimizer=opt)
# dummy optimizer for benchmarking
# training implementation coming soon
opt_gdm = GradientDescentMomentum(0.0, 0.0)
opt_biases = GradientDescentMomentum(0.0, 0.0)
opt = MultiOptimizer({'default': opt_gdm, 'Bias': opt_biases})
# setup cost function as CrossEntropy
cost = Multicost(
costs=[
GeneralizedCost(costfunc=CrossEntropyMulti()),
GeneralizedCost(costfunc=CrossEntropyMulti()),
GeneralizedCost(costfunc=CrossEntropyMulti())
],
weights=[1, 0., 0.]) # We only want to consider the CE of the main path
assert os.path.exists(
args.model_file), 'script requires the trained weights file'
model.load_params(args.model_file)
model.initialize(test, cost)
print 'running speed benchmark...'
model.benchmark(test, cost, opt)
print '\nCalculating performance on validation set...'
test.reset()
mets = model.eval(test, metric=valmetric)
print 'Validation set metrics:'
print 'LogLoss: %.2f, Accuracy: %.1f %% (Top-1), %.1f %% (Top-5)' % (
mets[0], (1.0 - mets[1]) * 100, (1.0 - mets[2]) * 100)
def main():
# larger batch sizes may not fit on GPU
parser = NeonArgparser(__doc__, default_overrides={'batch_size': 4})
parser.add_argument("--bench", action="store_true", help="run benchmark instead of training")
parser.add_argument("--num_classes", type=int, default=12, help="number of classes in the annotation")
parser.add_argument("--height", type=int, default=256, help="image height")
parser.add_argument("--width", type=int, default=512, help="image width")
args = parser.parse_args(gen_be=False)
# check that image dimensions are powers of 2
if((args.height & (args.height - 1)) != 0):
raise TypeError("Height must be a power of 2.")
if((args.width & (args.width - 1)) != 0):
raise TypeError("Width must be a power of 2.")
(c, h, w) = (args.num_classes, args.height, args.width)
# need to use the backend with the new upsampling layer implementation
be = NervanaGPU_Upsample(rng_seed=args.rng_seed,
device_id=args.device_id)
# set batch size
be.bsz = args.batch_size
# couple backend to global neon object
NervanaObject.be = be
shape = dict(channel_count=3, height=h, width=w, subtract_mean=False)
train_params = ImageParams(center=True, flip=False,
scale_min=min(h, w), scale_max=min(h, w),
aspect_ratio=0, **shape)
test_params = ImageParams(center=True, flip=False,
scale_min=min(h, w), scale_max=min(h, w),
aspect_ratio=0, **shape)
common = dict(target_size=h*w, target_conversion='read_contents',
onehot=False, target_dtype=np.uint8, nclasses=args.num_classes)
train_set = PixelWiseImageLoader(set_name='train', repo_dir=args.data_dir,
media_params=train_params,
shuffle=False, subset_percent=100,
index_file=os.path.join(args.data_dir, 'train_images.csv'),
**common)
val_set = PixelWiseImageLoader(set_name='val', repo_dir=args.data_dir,media_params=test_params,
index_file=os.path.join(args.data_dir, 'val_images.csv'), **common)
# initialize model object
layers = gen_model(c, h, w)
segnet_model = Model(layers=layers)
# configure callbacks
callbacks = Callbacks(segnet_model, eval_set=val_set, **args.callback_args)
opt_gdm = GradientDescentMomentum(1.0e-6, 0.9, wdecay=0.0005, schedule=Schedule())
opt_biases = GradientDescentMomentum(2.0e-6, 0.9, schedule=Schedule())
opt_bn = GradientDescentMomentum(1.0e-6, 0.9, schedule=Schedule())
opt = MultiOptimizer({'default': opt_gdm, 'Bias': opt_biases, 'BatchNorm': opt_bn})
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
if args.bench:
segnet_model.initialize(train_set, cost=cost)
segnet_model.benchmark(train_set, cost=cost, optimizer=opt)
sys.exit(0)
else:
segnet_model.fit(train_set, optimizer=opt, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
# get the trained segnet model outputs for valisation set
outs_val = segnet_model.get_outputs(val_set)
with open('outputs.pkl', 'w') as fid:
pickle.dump(outs_val, fid, -1)
