def test_smoothL1_random_derivative(backend_default, fargs):
s1, s2, m = fargs
shape = (s1, s2)
magnitude = m
outputs = (np.random.random(shape) - 0.5) * magnitude
targets = np.random.random(shape)
x = outputs - targets
expected_result = np.zeros(shape)
I1, J1 = np.where(abs(x) < 1)
I2, J2 = np.where(abs(x) >= 1)
expected_result[I1, J1] = x[I1, J1]
expected_result[I2, J2] = np.sign(x[I2, J2])
compare_tensors(SmoothL1Loss(), outputs,
targets, expected_result,
deriv=True, tol=1e-5)
def test_smoothL1_ones(backend_default, fargs):
s1, s2, m = fargs
shape = (s1, s2)
outputs = np.ones(shape) * m
targets = np.ones(shape) * m
x = outputs - targets
expected_result = np.zeros(shape)
I1, J1 = np.where(abs(x) < 1)
I2, J2 = np.where(abs(x) >= 1)
expected_result[I1, J1] = 0.5 * x[I1, J1]**2
expected_result[I2, J2] = abs(x[I2, J2]) - 0.5
expected_result = np.sum(expected_result, axis=0, keepdims=True)
compare_tensors(SmoothL1Loss(), outputs,
targets, expected_result,
deriv=False, tol=1e-5)
def test_smoothL1_zeros(backend_default, fargs):
s1, s2, m, sigma = fargs
sigma2 = sigma**2
shape = (s1, s2)
outputs = np.zeros(shape)
targets = np.zeros(shape)
x = outputs - targets
expected_result = np.zeros(shape)
I1, J1 = np.where(abs(x) < 1.0 / sigma2)
I2, J2 = np.where(abs(x) >= 1.0 / sigma2)
expected_result[I1, J1] = 0.5 * x[I1, J1]**2 * sigma2
expected_result[I2, J2] = abs(x[I2, J2]) - 0.5 / sigma2
expected_result = np.sum(expected_result, axis=0, keepdims=True)
compare_tensors(SmoothL1Loss(sigma=sigma),
outputs,
targets,
expected_result,
deriv=False,
tol=1e-5)
model = create_frcn_model(frcn_fine_tune)
# setup optimizer
opt_w = GradientDescentMomentum(
0.001 * learning_rate_scale, 0.9, wdecay=0.0005)
opt_b = GradientDescentMomentum(0.002 * learning_rate_scale, 0.9)
optimizer = MultiOptimizer({'default': opt_w, 'Bias': opt_b})
# 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:
load_vgg_weights(model, args.data_dir)
cost = Multicost(costs=[GeneralizedCostMask(costfunc=CrossEntropyMulti()),
GeneralizedCostMask(costfunc=SmoothL1Loss())],
weights=[1, 1])
callbacks = Callbacks(model, eval_set=test_set, **args.callback_args)
model.fit(train_set, optimizer=optimizer,
num_epochs=num_epochs, cost=cost, callbacks=callbacks)
# Fast R-CNN model requires scale the bbox regression branch linear layer weights
# before saving the model
model = scale_bbreg_weights(
model, train_set.bbtarget_means, train_set.bbtarget_stds)
save_obj(model.serialize(keep_states=True), args.save_path)
neon_logger.display('running eval...')
rois_per_img=rpn_rois_per_img,
inference=False)
config['subset_fraction'] = float(args.subset_pct / 100.0)
train_set = faster_rcnn.build_dataloader(config, frcn_rois_per_img)
# build the Faster-RCNN model
model = faster_rcnn.build_model(train_set, frcn_rois_per_img, inference=False)
# set up cost different branches, respectively
weights = 1.0 / (rpn_rois_per_img)
roi_w = 1.0 / (frcn_rois_per_img)
frcn_tree_cost = Multicost(costs=[
GeneralizedCostMask(costfunc=CrossEntropyMulti(), weights=roi_w),
GeneralizedCostMask(costfunc=SmoothL1Loss(), weights=roi_w)
],
weights=[1, 1])
cost = Multicost(costs=[
GeneralizedCostMask(costfunc=CrossEntropyMulti(), weights=weights),
GeneralizedCostMask(costfunc=SmoothL1Loss(sigma=3.0), weights=weights),
frcn_tree_cost,
],
weights=[1, 1, 1])
# setup optimizer
schedule_w = StepSchedule(step_config=[10], change=[0.001 / 10])
schedule_b = StepSchedule(step_config=[10], change=[0.002 / 10])
opt_w = GradientDescentMomentum(0.001, 0.9, wdecay=0.0005, schedule=schedule_w)