def train(args, hyper_params, model, opt, data_set):
# setup cost function as CrossEntropy
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
callbacks = Callbacks(model, **args.callback_args)
callbacks.add_callback(EpochEndCallback())
data_set.set_mode('train')
model.fit(data_set, optimizer=opt,
num_epochs=hyper_params.num_epochs, cost=cost, callbacks=callbacks)
return
random_seed = args.rng_seed if args.rng_seed else 0
# load up the mnist data set, padding images to size 32
dataset = MNIST(path=args.data_dir, sym_range=True, size=32, shuffle=True)
train = dataset.train_iter
# create a GAN
model, cost = create_model(dis_model=args.dmodel, gen_model=args.gmodel,
cost_type='wasserstein', noise_type='normal',
im_size=32, n_chan=1, n_noise=128,
n_gen_ftr=args.n_gen_ftr, n_dis_ftr=args.n_dis_ftr,
depth=4, n_extra_layers=4,
batch_norm=True, dis_iters=5,
wgan_param_clamp=0.01, wgan_train_sched=True)
# setup optimizer
optimizer = RMSProp(learning_rate=2e-4, decay_rate=0.99, epsilon=1e-8)
# configure callbacks
callbacks = Callbacks(model, **args.callback_args)
fdir = ensure_dirs_exist(os.path.join(os.path.dirname(os.path.realpath(__file__)), 'results/'))
fname = os.path.splitext(os.path.basename(__file__))[0] +\
'_[' + datetime.now().strftime('%Y-%m-%d-%H-%M-%S') + ']'
im_args = dict(filename=os.path.join(fdir, fname), hw=32,
num_samples=args.batch_size, nchan=1, sym_range=True)
callbacks.add_callback(GANPlotCallback(**im_args))
callbacks.add_callback(GANCostCallback())
# model fit
model.fit(train, optimizer=optimizer, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
model = Model(layers=SSD(ssd_config=train_config['ssd_config'], dataset=train_set))
cost = MBoxLoss(num_classes=train_set.num_classes)
if args.model_file is None:
load_vgg_weights(model, cache_dir)
else:
model.load_params(args.model_file)
if args.lr_step is None:
args.lr_step = [40, 80, 120]
base_lr = 0.0001 * be.bsz * args.lr_scale
schedule = Schedule(args.lr_step, 0.1)
opt_w = GradientDescentMomentum(base_lr, momentum_coef=0.9, wdecay=0.0005, schedule=schedule)
opt_b = GradientDescentMomentum(base_lr, momentum_coef=0.9, schedule=schedule)
opt = MultiOptimizer({'default': opt_w, 'Bias': opt_b})
# hijack the eval callback arg here
eval_freq = args.callback_args.pop('eval_freq')
callbacks = Callbacks(model, **args.callback_args)
callbacks.add_callback(MAP_Callback(eval_set=val_set, epoch_freq=eval_freq))
if args.image_sample_dir is not None:
callbacks.add_callback(ssd_image_callback(eval_set=val_set, image_dir=args.image_sample_dir,
epoch_freq=eval_freq, num_images=args.num_images,
classes=val_config['class_names']))
model.fit(train_set, optimizer=opt, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
#final model
mlp = Model(layers=layers)
logger.info("model construction complete...")
"""
model training and classification accurate rate
"""
#model training and results
callbacks = Callbacks(mlp,train, args, eval_set=valid,metric=Misclassification())
#add lost and metric call backs facilitate more diagnostic
callbacks.add_callback(MetricCallback(mlp,eval_set=train,metric=Misclassification(),epoch_freq=args.evaluation_freq))
callbacks.add_callback(MetricCallback(mlp,eval_set=valid,metric=Misclassification(),epoch_freq=args.evaluation_freq))
#run the model
mlp.fit(train, optimizer=opt_gdm, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
#final classification accuracy
t_mis_rate=mlp.eval(train, metric=Misclassification())*100
v_mis_rate=mlp.eval(valid, metric=Misclassification())*100
#test_mis_rate=mlp.eval(test, metric=Misclassification())*100
print ('Train Misclassification error = %.1f%%' %t_mis_rate)
print ('Valid Miscladdifcaiton error = %.1f%%' %v_mis_rate)
#print ('Test Miscladdifcaiton error = %.1f%%' %test_mis_rate)
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'
model, cost = create_network(args.depth)
# setup data provider
train = make_train_loader(args.manifest['train'], args.manifest_root, model.be,
args.subset_pct, random_seed)
test = make_validation_loader(args.manifest['val'], args.manifest_root,
model.be, args.subset_pct)
# tune batch norm parameters on subset of train set with no augmentations
tune_set = make_tuning_loader(args.manifest['train'], args.manifest_root,
model.be)
# configure callbacks
valmetric = TopKMisclassification(k=5)
callbacks = Callbacks(model,
eval_set=test,
metric=valmetric,
**args.callback_args)
callbacks.add_callback(BatchNormTuneCallback(tune_set), insert_pos=0)
# begin training
opt = GradientDescentMomentum(0.1,
0.9,
wdecay=0.0001,
schedule=Schedule([82, 124], 0.1))
model.fit(train,
optimizer=opt,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
opt_biases = GradientDescentMomentum(args.rate_init[1], args.momentum[1],
schedule=weight_sched, stochastic_round=args.rounding)
opt_fixed = GradientDescentMomentum(0.0, 1.0, wdecay=0.0)
opt = MultiOptimizer({'default': opt_gdm, 'Bias': opt_biases, 'DOG': opt_fixed})
# configure cost and test metrics
cost = GeneralizedCost(costfunc=(CrossEntropyBinary() \
if train.parser.independent_labels else CrossEntropyMulti()))
metric = EMMetric(oshape=test.parser.oshape, use_softmax=not train.parser.independent_labels) if test else None
# configure callbacks
if not args.neon_progress:
args.callback_args['progress_bar'] = False
callbacks = Callbacks(model, eval_set=test, metric=metric, **args.callback_args)
if not args.neon_progress:
callbacks.add_callback(EMEpochCallback(args.callback_args['eval_freq'],train.nmacrobatches),insert_pos=None)
# xxx - thought of making this an option but not clear that it slows anything down?
#callbacks.add_hist_callback() # xxx - not clear what information this conveys
if args.save_best_path:
callbacks.add_save_best_state_callback(args.save_best_path)
model.fit(train, optimizer=opt, num_epochs=num_epochs, cost=cost, callbacks=callbacks)
print('Model training complete for %d epochs!' % (args.epochs,))
#test.stop(); train.stop()
elif args.write_output:
# write_output mode, must have model loaded
if args.data_config:
test = EMDataIterator(args.data_config, write_output=args.write_output,
chunk_skip_list=args.chunk_skip_list, dim_ordering=args.dim_ordering,
weights=[1, 1, 1])
# setup optimizer
schedule_w = StepSchedule(step_config=[5], change=[0.001 / 10])
schedule_b = StepSchedule(step_config=[5], change=[0.002 / 10])
opt_w = GradientDescentMomentum(0.001, 0.9, wdecay=0.0005, schedule=schedule_w)
opt_b = GradientDescentMomentum(0.002, 0.9, wdecay=0.0005, schedule=schedule_b)
opt_skip = GradientDescentMomentum(0.0, 0.0)
optimizer = MultiOptimizer({'default': opt_w, 'Bias': opt_b,
'skip': opt_skip, 'skip_bias': opt_skip})
# if training a new model, seed the image model conv layers with pre-trained weights
# otherwise, just load the model file
if args.model_file is None:
util.load_vgg_all_weights(model, args.data_dir)
callbacks = Callbacks(model, eval_set=train_set, **args.callback_args)
callbacks.add_callback(TrainMulticostCallback())
# model.benchmark(train_set, optimizer=optimizer, cost=cost)
model.fit(train_set, optimizer=optimizer, cost=cost, num_epochs=args.epochs, callbacks=callbacks)
# Scale the bbox regression branch linear layer weights before saving the model
model = util.scale_bbreg_weights(model, [0.0, 0.0, 0.0, 0.0],
[0.1, 0.1, 0.2, 0.2], train_set.num_classes)
if args.save_path is not None:
save_obj(model.serialize(keep_states=True), args.save_path)
def benchmark(self):
for d in self.devices:
b = d if (self.backends is None) or (
"mkl" not in self.backends) else "mkl"
print("Use {} as backend.".format(b))
# Set up backend
# backend: 'cpu' for single cpu, 'mkl' for cpu using mkl library, and 'gpu' for gpu
be = gen_backend(backend=b,
batch_size=self.batch_size,
rng_seed=542,
datatype=np.float32)
# Make iterators
neon_train_set = ArrayIterator(X=np.asarray(
[t.flatten().astype('float32') / 255 for t in self.x_train]),
y=np.asarray(self.y_train),
make_onehot=True,
nclass=self.class_num,
lshape=(3, self.resize_size[0],
self.resize_size[1]))
neon_valid_set = ArrayIterator(X=np.asarray(
[t.flatten().astype('float32') / 255 for t in self.x_valid]),
y=np.asarray(self.y_valid),
make_onehot=True,
nclass=self.class_num,
lshape=(3, self.resize_size[0],
self.resize_size[1]))
neon_test_set = ArrayIterator(X=np.asarray([
t.flatten().astype('float32') / 255 for t in self.testImages
]),
y=np.asarray(self.testLabels),
make_onehot=True,
nclass=self.class_num,
lshape=(3, self.resize_size[0],
self.resize_size[1]))
# Initialize model object
self.neon_model = SelfModel(layers=self.constructCNN())
# Costs
neon_cost = GeneralizedCost(costfunc=CrossEntropyMulti())
# Model summary
self.neon_model.initialize(neon_train_set, neon_cost)
print(self.neon_model)
# Learning rules
neon_optimizer = SGD(0.01,
momentum_coef=0.9,
schedule=ExpSchedule(0.2))
# neon_optimizer = RMSProp(learning_rate=0.0001, decay_rate=0.95)
# # Benchmark for 20 minibatches
# d[b] = self.neon_model.benchmark(neon_train_set, cost=neon_cost, optimizer=neon_optimizer)
# Reset model
# self.neon_model = None
# self.neon_model = Model(layers=layers)
# self.neon_model.initialize(neon_train_set, neon_cost)
# Callbacks: validate on validation set
callbacks = Callbacks(
self.neon_model,
eval_set=neon_valid_set,
metric=Misclassification(3),
output_file=
"{}saved_data/{}/{}/callback_data_neon_{}_{}_{}by{}_{}.h5".
format(self.root, self.network_type, d, b, self.dataset,
self.resize_size[0], self.resize_size[1],
self.preprocessing))
callbacks.add_callback(
SelfCallback(eval_set=neon_valid_set,
test_set=neon_test_set,
epoch_freq=1))
# Fit
start = time.time()
self.neon_model.fit(neon_train_set,
optimizer=neon_optimizer,
num_epochs=self.epoch_num,
cost=neon_cost,
callbacks=callbacks)
print("Neon training finishes in {:.2f} seconds.".format(
time.time() - start))
# Result
# results = self.neon_model.get_outputs(neon_valid_set)
# Print error on validation set
start = time.time()
neon_error_mis = self.neon_model.eval(
neon_valid_set, metric=Misclassification()) * 100
print(
'Misclassification error = {:.1f}%. Finished in {:.2f} seconds.'
.format(neon_error_mis[0],
time.time() - start))
# start = time.time()
# neon_error_top3 = self.neon_model.eval(neon_valid_set, metric=TopKMisclassification(3))*100
# print('Top 3 Misclassification error = {:.1f}%. Finished in {:.2f} seconds.'.format(neon_error_top3[2], time.time() - start))
# start = time.time()
# neon_error_top5 = self.neon_model.eval(neon_valid_set, metric=TopKMisclassification(5))*100
# print('Top 5 Misclassification error = {:.1f}%. Finished in {:.2f} seconds.'.format(neon_error_top5[2], time.time() - start))
self.neon_model.save_params(
"{}saved_models/{}/{}/neon_weights_{}_{}_{}by{}_{}.prm".format(
self.root, self.network_type, d, b, self.dataset,
self.resize_size[0], self.resize_size[1],
self.preprocessing))
# Print error on test set
start = time.time()
neon_error_mis_t = self.neon_model.eval(
neon_test_set, metric=Misclassification()) * 100
print(
'Misclassification error = {:.1f}% on test set. Finished in {:.2f} seconds.'
.format(neon_error_mis_t[0],
time.time() - start))
# start = time.time()
# neon_error_top3_t = self.neon_model.eval(neon_test_set, metric=TopKMisclassification(3))*100
# print('Top 3 Misclassification error = {:.1f}% on test set. Finished in {:.2f} seconds.'.format(neon_error_top3_t[2], time.time() - start))
# start = time.time()
# neon_error_top5_t = self.neon_model.eval(neon_test_set, metric=TopKMisclassification(5))*100
# print('Top 5 Misclassification error = {:.1f}% on test set. Finished in {:.2f} seconds.'.format(neon_error_top5_t[2], time.time() - start))
cleanup_backend()
self.neon_model = None
# setup cost function as CrossEntropy
cost = GeneralizedGANCost(costfunc=GANCost(func="modified"))
# setup optimizer
optimizer = Adam(learning_rate=0.0005, beta_1=0.5)
# initialize model
noise_dim = (2, 7, 7)
gan = GAN(layers=layers, noise_dim=noise_dim, k=args.kbatch)
# configure callbacks
callbacks = Callbacks(gan, eval_set=valid_set, **args.callback_args)
fdir = ensure_dirs_exist(
os.path.join(os.path.dirname(os.path.realpath(__file__)), 'results/'))
fname = os.path.splitext(os.path.basename(__file__))[0] +\
'_[' + datetime.now().strftime('%Y-%m-%d-%H-%M-%S') + ']'
im_args = dict(filename=os.path.join(fdir, fname),
hw=27,
num_samples=args.batch_size,
nchan=1,
sym_range=True)
callbacks.add_callback(GANPlotCallback(**im_args))
callbacks.add_callback(GANCostCallback())
# run fit
gan.fit(train_set,
optimizer=optimizer,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
init=glorot,
activation=Rectlinclip(),
batch_norm=True,
reset_cells=True,
depth=depth),
Affine(hidden_size, init=glorot, activation=Rectlinclip()),
Affine(nout=nout, init=glorot, activation=Identity())
]
model = Model(layers=layers)
opt = GradientDescentMomentum(learning_rate,
momentum,
gradient_clip_norm=gradient_clip_norm,
stochastic_round=False,
nesterov=True)
callbacks = Callbacks(model, eval_set=dev, **args.callback_args)
# Print validation set word error rate at the end of every epoch
pcb = WordErrorRateCallback(dev, argmax_decoder, max_tscrpt_len, epoch_freq=1)
callbacks.add_callback(pcb)
cost = GeneralizedCost(costfunc=CTC(max_tscrpt_len, nout=nout))
# Fit the model
model.fit(train,
optimizer=opt,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
Conv((1, 1, 16), **conv),
Pooling(8, op="avg"),
Activation(Softmax())]
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
mlp = Model(layers=layers)
if args.model_file:
import os
assert os.path.exists(args.model_file), '%s not found' % args.model_file
mlp.load_weights(args.model_file)
# configure callbacks
callbacks = Callbacks(mlp, train_set, eval_set=valid_set, **args.callback_args)
if args.deconv:
callbacks.add_deconv_callback(train_set, valid_set)
callbacks.add_callback(
MetricCallback(
valid_set,
Misclassification()
)
)
mlp.fit(train_set, optimizer=opt_gdm, num_epochs=num_epochs, cost=cost, callbacks=callbacks)
import logging
logger = logging.getLogger(__name__)
logger.critical('Misclassification error = %.1f%%' % (mlp.eval(valid_set, metric=Misclassification())*100))
layers.append(Conv((2, 2, nchan), strides=2, **common))
for idx in range(16):
layers.append(Conv((3, 3, nchan), **common))
if nchan > 16:
nchan /= 2
for idx in range(15):
layers.append(Deconv((3, 3, nchan), **common))
layers.append(Deconv((4, 4, nchan), strides=2, **common))
layers.append(Deconv((3, 3, 1), init=init, activation=Logistic(shortcut=True)))
cost = GeneralizedCost(costfunc=SumSquared())
mlp = Model(layers=layers)
callbacks = Callbacks(mlp, train, **args.callback_args)
evaluator = Evaluator(callbacks.callback_data, mlp, test, imwidth, args.epochs,
args.data_dir, point_num)
callbacks.add_callback(evaluator)
mlp.fit(train,
optimizer=opt,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
train.exit_batch_provider()
preds = evaluator.get_outputs()
paths = np.genfromtxt(os.path.join(args.test_data_dir, 'val_file.csv'),
dtype=str)[1:]
basenames = [os.path.basename(path) for path in paths]
filenames = [path.split(',')[0] for path in basenames]
filenames.sort()
content = []
for i, filename in enumerate(filenames):
cost = GeneralizedCost(costfunc=(CrossEntropyBinary() \
if train.parser.independent_labels else CrossEntropyMulti()))
metric = EMMetric(
oshape=test.parser.oshape,
use_softmax=not train.parser.independent_labels) if test else None
# configure callbacks
if not args.neon_progress:
args.callback_args['progress_bar'] = False
callbacks = Callbacks(model,
eval_set=test,
metric=metric,
**args.callback_args)
if not args.neon_progress:
callbacks.add_callback(EMEpochCallback(
args.callback_args['eval_freq'], train.nmacrobatches),
insert_pos=None)
# xxx - thought of making this an option but not clear that it slows anything down?
#callbacks.add_hist_callback() # xxx - not clear what information this conveys
if args.save_best_path:
callbacks.add_save_best_state_callback(args.save_best_path)
model.fit(train,
optimizer=opt,
num_epochs=num_epochs,
cost=cost,
callbacks=callbacks)
print('Model training complete for %d epochs!' % (args.epochs, ))
#test.stop(); train.stop()
elif args.write_output:
Pooling(3, strides=2)]
# Structure of the deep residual part of the network:
# args.depth modules of 2 convolutional layers each at feature map depths
# of 64, 128, 256, 512
nfms = list(itt.chain.from_iterable(
[itt.repeat(2**(x + 6), r) for x, r in enumerate(stages)]))
strides = [-1] + [1 if cur == prev else 2 for cur,
prev in zip(nfms[1:], nfms[:-1])]
for nfm, stride in zip(nfms, strides):
layers.append(module_factory(nfm, stride))
layers.append(Pooling('all', op='avg'))
layers.append(Conv(**conv_params(1, train.nclass, relu=False)))
layers.append(Activation(Softmax()))
model = Model(layers=layers)
weight_sched = Schedule([30, 60], 0.1)
opt = GradientDescentMomentum(0.1, 0.9, wdecay=0.0001, schedule=weight_sched)
# configure callbacks
valmetric = TopKMisclassification(k=5)
callbacks = Callbacks(model, eval_set=test, metric=valmetric, **args.callback_args)
callbacks.add_callback(BatchNormTuneCallback(tune), insert_pos=0)
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
model.fit(train, optimizer=opt, num_epochs=args.epochs,
cost=cost, callbacks=callbacks)
]
model = Model(layers=layers)
# define optimizer
opt_w = GradientDescentMomentum(learning_rate=0.01,
momentum_coef=0.9,
wdecay=0.0005)
opt_b = GradientDescentMomentum(learning_rate=0.01, momentum_coef=0.9)
opt = MultiOptimizer({'default': opt_w, 'Bias': opt_b}, name='multiopt')
# configure callbacks
callbacks = Callbacks(model,
eval_set=valid_set,
metric=Misclassification(),
**args.callback_args)
callbacks.add_callback(
TrainByStageCallback(model,
valid_set,
Misclassification(),
max_patience=10))
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
logger.info('Training ...')
model.fit(train_set,
optimizer=opt,
num_epochs=250,
cost=cost,
callbacks=callbacks)
print('Accuracy = %.2f%%' %
(100. - model.eval(valid_set, metric=Misclassification()) * 100))
if args.model_file:
import os
assert os.path.exists(args.model_file), '%s not found' % args.model_file
mlp.load_weights(args.model_file)
# configure callbacks
callbacks = Callbacks(mlp, train, output_file=args.output_file)
if args.validation_freq:
class TopKMetrics(Callback):
def __init__(self, valid_set, epoch_freq=args.validation_freq):
super(TopKMetrics, self).__init__(epoch_freq=epoch_freq)
self.valid_set = valid_set
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()))
callbacks.add_callback(TopKMetrics(test))
if args.save_path:
checkpoint_schedule = range(1, args.epochs)
callbacks.add_serialize_callback(checkpoint_schedule, args.save_path, history=2)
mlp.fit(train, optimizer=opt, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
test.exit_batch_provider()
train.exit_batch_provider()
def train(self,
content,
targets,
test_content=None,
test_targets=None,
has_features=False,
serialize=0,
save_path=None,
learning_rate=0.001,
epochs=5):
"""
:param content: numerical content returned from gen_training_set
:param targets: numerical targets returned from gen_training_set
:param test_content: separate test set to be used for evaluation
:param test_targets: ""
:param test_features: ""
:param features: features arte a list of float lists that can also be in string form, but will be converted
to arrays of floats that must be of the same length as the features specified when creating the classifier.
:param receiver_address: deprecated in favor of features, but left in for testing
:param serialize:
:param save_path:
:param model_file:
:param holdout_pct:
:param learning_rate:
:param epochs:
:return:
"""
multicost = self.exclusive_classes is not None and self.overlapping_classes is not None
if multicost:
metric = MultiMetric(Misclassification(), 0)
elif self.overlapping_classes is None:
metric = Misclassification()
else:
metric = AverageLogLoss()
print('Training neural networks on {} samples for {} epochs.'.format(
len(targets[0]), epochs))
if not test_content is None and not test_targets is None:
valid = BatchIterator(test_content,
targets=test_targets,
steps=[1] if has_features else [1, 1])
else:
valid = None
train = BatchIterator(content,
targets=targets,
steps=[1] if has_features else [1, 1])
callbacks = Callbacks(self.neuralnet,
train_set=train,
multicost=multicost,
metric=metric,
eval_freq=None if valid is None else 1,
eval_set=valid)
#save_path=save_path, serialize=serialize)
if serialize:
callbacks.add_save_best_state_callback(save_path)
if not valid is None:
if self.exclusive_classes is not None:
print('Starting misclassification error = {:.03}%'.format(
self.neuralnet.eval(valid, metric)[0] * 100))
callbacks.add_callback(MisclassificationTest(valid, metric))
else:
print('Starting average logloss = {:.04}'.format(
self.neuralnet.eval(valid, metric)[0]))
callbacks.add_callback(LogLossTest(valid, metric))
if hasattr(self.optimizer, 'learning_rate'):
self.optimizer.learning_rate = learning_rate
print('learning rate = {}'.format(learning_rate))
self.fit(train, self.optimizer, epochs, callbacks)
layers.append(Conv((2, 2, nchan), strides=2, **common))
for idx in range(16):
layers.append(Conv((3, 3, nchan), **common))
if nchan > 16:
nchan /= 2
for idx in range(15):
layers.append(Deconv((3, 3, nchan), **common))
layers.append(Deconv((4, 4, nchan), strides=2, **common))
layers.append(Deconv((3, 3, 1), init=init, activation=Logistic(shortcut=True)))
cost = GeneralizedCost(costfunc=SumSquared())
mlp = Model(layers=layers)
callbacks = Callbacks(mlp, train, **args.callback_args)
evaluator = Evaluator(callbacks.callback_data, mlp, test, imwidth, args.epochs,
args.data_dir, point_num)
callbacks.add_callback(evaluator)
mlp.fit(train, optimizer=opt, num_epochs=args.epochs, cost=cost,
callbacks=callbacks)
train.exit_batch_provider()
preds = evaluator.get_outputs()
paths = np.genfromtxt(os.path.join(args.test_data_dir, 'val_file.csv'),
dtype=str)[1:]
basenames = [os.path.basename(path) for path in paths]
filenames = [path.split(',')[0] for path in basenames]
filenames.sort()
content = []
for i, filename in enumerate(filenames):
item = {
"annotations":
[
opt_b = GradientDescentMomentum(0.002, 0.9, wdecay=0.0005, schedule=schedule_b)
opt_skip = GradientDescentMomentum(0.0, 0.0)
optimizer = MultiOptimizer({
'default': opt_w,
'Bias': opt_b,
'skip': opt_skip,
'skip_bias': opt_skip
})
# if training a new model, seed the image model conv layers with pre-trained weights
# otherwise, just load the model file
if args.model_file is None:
util.load_vgg_all_weights(model, cache_dir)
callbacks = Callbacks(model, eval_set=train_set, **args.callback_args)
callbacks.add_callback(TrainMulticostCallback())
model.fit(train_set,
optimizer=optimizer,
cost=cost,
num_epochs=args.epochs,
callbacks=callbacks)
# Scale the bbox regression branch linear layer weights before saving the model
model = util.scale_bbreg_weights(model, [0.0, 0.0, 0.0, 0.0],
[0.1, 0.1, 0.2, 0.2], train_set.num_classes)
if args.save_path is not None:
save_obj(model.serialize(keep_states=True), args.save_path)
