def train_model(self, x_train, y_train, x_val, y_val, force_device=None):
print('training {} model (3 epochs, default setting)'.format(
self.__class__.__name__))
x_train, y_train, x_val, y_val = helpers.force_device(
self, (x_train, y_train, x_val, y_val), force_device)
self.model.train(x_train,
y_train,
Xval=x_val,
Yval=y_val,
batchsize=5,
lrate=0.005,
status=500) # train the model
self.model.train(
x_train,
y_train,
Xval=x_val,
Yval=y_val,
batchsize=5,
lrate=0.001,
status=500
) # slower training once the model has converged somewhat
self.model.train(x_train,
y_train,
Xval=x_val,
Yval=y_val,
batchsize=5,
lrate=0.0005,
status=500) # one last epoch
def train_model(self, x_train, y_train, x_val, y_val, force_device=None):
print('training {} model (3 epochs, decreasing batch size per epoch)'.
format(self.__class__.__name__))
x_train, y_train, x_val, y_val = helpers.force_device(
self, (x_train, y_train, x_val, y_val), force_device)
self.model.train(x_train,
y_train,
Xval=x_val,
Yval=y_val,
batchsize=128,
lrate=0.005,
status=1000,
iters=50000,
convergence=10) # train the model
self.model.train(
x_train,
y_train,
Xval=x_val,
Yval=y_val,
batchsize=64,
lrate=0.001,
status=1000,
iters=50000,
convergence=10
) # slower training once the model has converged somewhat
self.model.train(x_train,
y_train,
Xval=x_val,
Yval=y_val,
batchsize=32,
lrate=0.0005,
status=1000,
iters=50000,
convergence=10) # one last epoch
self.model.train(x_train,
y_train,
Xval=x_val,
Yval=y_val,
batchsize=8,
lrate=0.0005,
status=1000,
iters=50000,
convergence=10) # one last last epoch
def train_model(self, x_train, y_train, x_val, y_val, force_device=None):
x_train, y_train, x_val, y_val = helpers.force_device(
self, (x_train, y_train, x_val, y_val), force_device)
print('training {} model (quick test)'.format(self.__class__.__name__))
self.model.train(x_train, y_train, iters=10)
def evaluate_model(self,
x_test,
y_test,
force_device=None,
lower_upper=None):
"""
test model and computes relevance maps
Parameters:
-----------
x_test: array - shaped such that it is ready for consumption by the model
y_test: array - expected test labels
target_shape: list or tuple - the target output shape of the test data and relevance maps.
force_device: str - (optional) force execution of the evaluation either on cpu or gpu.
accepted values: "cpu", "gpu" respectively. None does nothing.
lower_upper: (array of float, array of float) - (optional): lower and upper bounds of the inputs, for LRP_zB.
automagically inferred from x_test.
arrays should match the feature dimensionality of the inputs, including broadcastable axes.
e.g. if x_test is shaped (N, featuredims), then the bounds should be shaped (1, featuredims)
Returns:
--------
results, packed in dictionary, as numpy arrays
"""
assert isinstance(
self.model, Sequential
), "self.model should be modules.sequential.Sequentialm but is {}. ensure correct type by converting model after training.".format(
type(self.model))
# remove the softmax output of the model.
# this does not change the ranking of the outputs but is required for most LRP methods
# self.model is required to be a modules.Sequential
results = {} #prepare results dictionary
#force model to specific device, if so desired.
x_test, y_test = helpers.force_device(self, (x_test, y_test),
force_device)
print('...forward pass for {} test samples for model performance eval'.
format(x_test.shape[0]))
y_pred = self.model.forward(x_test)
#evaluate accuracy and loss on cpu-copyies of prediction vectors
y_pred_c, y_test_c = helpers.arrays_to_numpy(y_pred, y_test)
results['acc'] = helpers.accuracy(y_test_c, y_pred_c)
results['loss_l1'] = helpers.l1loss(y_test_c, y_pred_c)
results['y_pred'] = y_pred_c
#NOTE: drop softmax layer AFTER forward for performance measures to obtain competetive loss values
self.model.drop_softmax_output_layer()
#NOTE: second forward pass without softmax for relevance computation
print('...forward pass for {} test samples (without softmax) for LRP'.
format(x_test.shape[0]))
y_pred = self.model.forward(
x_test) # this is also a requirement for LRP
# prepare initial relevance vectors for actual class and dominantly predicted class, on model-device (gpu or cpu)
R_init_act = y_pred * y_test #assumes y_test to be binary matrix
y_dom = (y_pred == y_pred.max(axis=1, keepdims=True))
R_init_dom = y_pred * y_dom #assumes prediction maxima are unique per sample
# compute epsilon-lrp for all model layers
for m in self.model.modules:
m.set_lrp_parameters(lrp_var='epsilon', param=1e-5)
print('...lrp (eps) for actual classes')
results['R_pred_act_epsilon'] = self.model.lrp(R_init_act)
print('...lrp (eps) for dominant classes')
results['R_pred_dom_epsilon'] = self.model.lrp(R_init_dom)
# eps + zB (lowest convolution/flatten layer) for all models here.
# infer lower and upper bounds from data, if not given
if not lower_upper:
print(
' ...inferring per-channel lower and upper bounds for zB from test data. THIS IS PROBABLY NOT OPTIMAL'
)
lower_upper = helpers.get_channel_wise_bounds(x_test)
else:
print(' ...using input lower and upper bounds for zB')
if self.use_gpu:
lower_upper = helpers.arrays_to_cupy(*lower_upper)
else:
lower_upper = helpers.arrays_to_numpy(*lower_upper)
# configure the lowest weighted layer to be decomposed with zB. This should be the one nearest to the input.
# We are not just taking the first layer, since the MLP models are starting with a Flatten layer for reshaping the data.
for m in self.model.modules:
if isinstance(m, (Linear, Convolution)):
m.set_lrp_parameters(lrp_var='zB', param=lower_upper)
break
print('...lrp (eps + zB) for actual classes')
results['R_pred_act_epsilon_zb'] = self.model.lrp(R_init_act)
print('...lrp (eps + zB) for dominant classes')
results['R_pred_dom_epsilon_zb'] = self.model.lrp(R_init_dom)
# compute CNN composite rules, if model has convolution layes
has_convolutions = False
for m in self.model.modules:
has_convolutions = has_convolutions or isinstance(m, Convolution)
if has_convolutions:
# convolution layers found.
# epsilon-lrp with flat decomposition in the lowest convolution layers
# process lowest convolution layer with FLAT lrp
# for "normal" cnns, this should overwrite the previously set zB rule
for m in self.model.modules:
if isinstance(m, Convolution):
m.set_lrp_parameters(lrp_var='flat')
break
print('...lrp (eps+flat) for actual classes')
results['R_pred_act_epsilon_flat'] = self.model.lrp(R_init_act)
print('...lrp (eps+flat) for dominant classes')
results['R_pred_dom_epsilon_flat'] = self.model.lrp(R_init_dom)
# preparing alpha2beta-1 for those layers
for m in self.model.modules:
if isinstance(m, Convolution):
m.set_lrp_parameters(lrp_var='alpha', param=2.0)
print('...lrp (composite:alpha=2) for actual classes')
results['R_pred_act_composite_alpha2'] = self.model.lrp(R_init_act)
print('...lrp (composite:alpha=2) for dominant classes')
results['R_pred_dom_composite_alpha2'] = self.model.lrp(R_init_dom)
# process lowest convolution layer with FLAT lrp
for m in self.model.modules:
if isinstance(m, Convolution):
m.set_lrp_parameters(lrp_var='flat')
break
print('...lrp (composite:alpha=2+flat) for actual classes')
results['R_pred_act_composite_alpha2_flat'] = self.model.lrp(
R_init_act)
print('...lrp (composite:alpha=2+flat) for dominant classes')
results['R_pred_dom_composite_alpha2_flat'] = self.model.lrp(
R_init_dom)
#process lowest convolution layer with zB lrp
for m in self.model.modules:
if isinstance(m, Convolution):
m.set_lrp_parameters(lrp_var='zB', param=lower_upper)
break
print('...lrp (composite:alpha=2+zB) for actual classes')
results['R_pred_act_composite_alpha2_zB'] = self.model.lrp(
R_init_act)
print('...lrp (composite:alpha=2+zB) for dominant classes')
results['R_pred_dom_composite_alpha2_zB'] = self.model.lrp(
R_init_dom)
# switching alpha1beta0 for those layers
for m in self.model.modules:
if isinstance(m, Convolution):
m.set_lrp_parameters(lrp_var='alpha', param=1.0)
print('...lrp (composite:alpha=1) for actual classes')
results['R_pred_act_composite_alpha1'] = self.model.lrp(R_init_act)
print('...lrp (composite:alpha=1) for dominant classes')
results['R_pred_dom_composite_alpha1'] = self.model.lrp(R_init_dom)
# process lowest convolution layer with FLAT lrp
for m in self.model.modules:
if isinstance(m, Convolution):
m.set_lrp_parameters(lrp_var='flat')
break
print('...lrp (composite:alpha=1+flat) for actual classes')
results['R_pred_act_composite_alpha1_flat'] = self.model.lrp(
R_init_act)
print('...lrp (composite:alpha=1+flat) for dominant classes')
results['R_pred_dom_composite_alpha1_flat'] = self.model.lrp(
R_init_dom)
#process lowest convolution layer with zB lrp
for m in self.model.modules:
if isinstance(m, Convolution):
m.set_lrp_parameters(lrp_var='zB', param=lower_upper)
break
print('...lrp (composite:alpha=1+zB) for actual classes')
results['R_pred_act_composite_alpha1_zB'] = self.model.lrp(
R_init_act)
print('...lrp (composite:alpha=1+zB) for dominant classes')
results['R_pred_dom_composite_alpha1_zB'] = self.model.lrp(
R_init_dom)
print('...copying collected results to CPU and reshaping if necessary')
for key in results.keys():
tmp = helpers.arrays_to_numpy(results[key])[0]
if key.startswith('R'):
tmp = self.postprocess_relevance(tmp)[0]
results[key] = tmp
return results