def main():
# parse the command line arguments
parser = NeonArgparser(__doc__)
args = parser.parse_args()
logger = logging.getLogger()
logger.setLevel(args.log_thresh)
#Set up batch iterator for training images
train = ImgMaster(repo_dir='dataTmp_optFlow_BW',
set_name='train',
inner_size=240,
subset_pct=100)
val = ImgMaster(repo_dir='dataTmp_optFlow_BW',
set_name='train',
inner_size=240,
subset_pct=100,
do_transforms=False)
test = ImgMaster(repo_dir='dataTestTmp_optFlow_BW',
set_name='train',
inner_size=240,
subset_pct=100,
do_transforms=False)
train.init_batch_provider()
val.init_batch_provider()
test.init_batch_provider()
print "Constructing network..."
#Create AlexNet architecture
model = constuct_network()
# Optimzer for model
opt = Adadelta()
# configure callbacks
valmetric = TopKMisclassification(k=5)
callbacks = Callbacks(model,
train,
eval_set=test,
metric=valmetric,
**args.callback_args)
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
#flag = input("Press Enter if you want to begin training process.")
print "Training network..."
model.fit(train,
optimizer=opt,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
mets = model.eval(test, metric=valmetric)
print 'Validation set metrics:'
print 'LogLoss: %.2f, Accuracy: %.1f %%0 (Top-1), %.1f %% (Top-5)' % (
mets[0], (1.0 - mets[1]) * 100, (1.0 - mets[2]) * 100)
return
parser.add_argument('--model_tree', help='Whether or not to train tree of classifiers',
default=False, type=bool)
parser.add_argument('--freeze', type=int, help='Layers to freeze starting from end', default=0)
parser.add_argument('--dataset_dir', help='Directory containing images folder and label text files')
args = parser.parse_args()
# setup data provider
train_set_options = dict(repo_dir=args.data_dir,
inner_size=224,
dtype=args.datatype,
subset_pct=100)
test_set_options = dict(repo_dir=args.data_dir,
inner_size=224,
dtype=args.datatype,
subset_pct=20)
train = ImageLoader(set_name='train', **train_set_options)
test = ImageLoader(set_name='train', do_transforms=False, **test_set_options)
model, cost, opt = create_model(args.model_type, args.model_tree, args.freeze, args.dataset_dir,
args.model_file, train)
# configure callbacks
valmetric = TopKMisclassification(k=5)
valmetric.name = 'root_misclass'
# If freezing layers, load model in create_model
if args.freeze > 0:
args.callback_args['model_file'] = None
callbacks = Callbacks(model, train, eval_set=test, metric=valmetric, **args.callback_args)
model.fit(train, optimizer=opt, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
model, cost = create_network()
rseed = 0 if args.rng_seed is None else args.rng_seed
# setup data provider
assert 'train' in args.manifest, "Missing train manifest"
assert 'val' in args.manifest, "Missing validation manifest"
train = make_alexnet_train_loader(args.manifest['train'], args.manifest_root,
model.be, args.subset_pct, rseed)
valid = make_validation_loader(args.manifest['val'], args.manifest_root,
model.be, args.subset_pct)
sched_weight = Schedule([10], change=0.1)
opt = GradientDescentMomentum(0.01, 0.9, wdecay=0.0005, schedule=sched_weight)
# configure callbacks
valmetric = TopKMisclassification(k=5)
callbacks = Callbacks(model,
eval_set=valid,
metric=valmetric,
**args.callback_args)
if args.deconv:
callbacks.add_deconv_callback(train, valid)
model.fit(train,
optimizer=opt,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
def on_epoch_end(self, epoch):
self.valid_set.reset()
allmetrics = TopKMisclassification(k=5)
stats = mlp.eval(self.valid_set, metric=allmetrics)
print ", ".join(allmetrics.metric_names) + ": " + ", ".join(
map(str, stats.flatten()))
def main():
# parse the command line arguments
parser = NeonArgparser(__doc__)
args = parser.parse_args()
logger = logging.getLogger()
logger.setLevel(args.log_thresh)
#Set up batch iterator for training images
print "Setting up data batch loaders..."
train = ImgMaster(repo_dir='dataTmp',
set_name='train',
inner_size=120,
subset_pct=100)
val = ImgMaster(repo_dir='dataTmp',
set_name='train',
inner_size=120,
subset_pct=100,
do_transforms=False)
test = ImgMaster(repo_dir='dataTestTmp',
set_name='train',
inner_size=120,
subset_pct=100,
do_transforms=False)
train.init_batch_provider()
val.init_batch_provider()
test.init_batch_provider()
print "Constructing network..."
#Create AlexNet architecture
model = constuct_network()
#model.load_weights(args.model_file)
# drop weights LR by 1/250**(1/3) at epochs (23, 45, 66), drop bias LR by 1/10 at epoch 45
weight_sched = Schedule([22, 44, 65, 90, 97], (1 / 250.)**(1 / 3.))
opt_gdm = GradientDescentMomentum(0.01,
0.9,
wdecay=0.005,
schedule=weight_sched)
opt_biases = GradientDescentMomentum(0.04,
1.0,
schedule=Schedule([130], .1))
opt = MultiOptimizer({'default': opt_gdm, 'Bias': opt_biases})
# configure callbacks
valmetric = TopKMisclassification(k=5)
callbacks = Callbacks(model,
train,
eval_set=val,
metric=valmetric,
**args.callback_args)
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
#flag = input("Press Enter if you want to begin training process.")
print "Training network..."
model.fit(train,
optimizer=opt,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
mets = model.eval(test, metric=valmetric)
print 'Validation set metrics:'
print 'LogLoss: %.2f, Accuracy: %.1f %%0 (Top-1), %.1f %% (Top-5)' % (
mets[0], (1.0 - mets[1]) * 100, (1.0 - mets[2]) * 100)
test.exit_batch_provider()
val.exit_batch_provider()
train.exit_batch_provider()
# set up 3x3 conv stacks with different feature map sizes
for nofm in [64, 128, 256, 512, 512]:
layers.append(Conv((3, 3, nofm), **conv_params))
layers.append(Conv((3, 3, nofm), **conv_params))
if nofm > 128:
layers.append(Conv((3, 3, nofm), **conv_params))
if args.vgg_version == 'E':
layers.append(Conv((3, 3, nofm), **conv_params))
layers.append(Pooling(2, strides=2))
layers.append(Affine(nout=4096, init=initfc, bias=Constant(0), activation=relu))
layers.append(Dropout(keep=0.5))
layers.append(Affine(nout=4096, init=initfc, bias=Constant(0), activation=relu))
layers.append(Dropout(keep=0.5))
layers.append(Affine(nout=1000, init=initfc, bias=Constant(0), activation=Softmax()))
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
model = Model(layers=layers)
# configure callbacks
top5 = TopKMisclassification(k=5)
callbacks = Callbacks(model, eval_set=test, metric=top5, **args.callback_args)
model.load_params(args.model_file)
mets=model.eval(test, metric=TopKMisclassification(k=5))
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)
layers.append(Conv((3, 3, nofm), **conv_params))
layers.append(Conv((3, 3, nofm), **conv_params))
layers.append(Pooling(3, strides=2))
for nofm in [256, 512, 512]:
layers.append(Conv((3, 3, nofm), **conv_params))
layers.append(Conv((3, 3, nofm), **conv_params))
layers.append(Conv((3, 3, nofm), **conv_params))
layers.append(Pooling(3, strides=2))
else:
raise ValueError("Invalid specification for VGG model")
layers.append(Affine(nout=4096, init=init1, batch_norm=True, activation=relu))
layers.append(Dropout(keep=0.5))
layers.append(Affine(nout=4096, init=init1, batch_norm=True, activation=relu))
layers.append(Dropout(keep=0.5))
layers.append(Affine(nout=1000, init=init1, bias=Constant(0), activation=Softmax()))
cost = GeneralizedCost(costfunc=CrossEntropyMulti(scale=cost_scale))
opt = MultiOptimizer({'default': opt_gdm, 'Bias': opt_biases})
mlp = Model(layers=layers)
# configure callbacks
callbacks = Callbacks(mlp, train, args, eval_set=test, metric=TopKMisclassification(k=5))
mlp.fit(train, optimizer=opt, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
test.exit_batch_provider()
train.exit_batch_provider()