def make_model(input_dim, layer_dims, nclasses, args):
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = input_dim + int(not args.nobias)
elif args.h_type == 'fcc':
#args.nconcepts += int(not args.nobias)
conceptizer = image_fcc_conceptizer(
input_dim, args.nconcepts,
args.concept_dim) #, sparsity = sparsity_l)
else:
raise ValueError('Unrecognized h_type')
#model_path, log_path, results_path = generate_dir_names('uci', args)
parametrizer = dfc_parametrizer(input_dim, *layer_dims, args.nconcepts,
args.theta_dim)
aggregator = additive_scalar_aggregator(args.concept_dim, nclasses)
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
if args.theta_reg_type == 'unreg':
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ=1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ=2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ=3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecoginzed theta_reg_type')
return trainer
def main():
np.random.seed(2018)
args = parse_args()
args.nclasses = 1
args.theta_dim = args.nclasses
args.print_freq = 100
args.epochs = 10
train_loader, valid_loader, test_loader, train, valid, test, data = load_cancer_data(
)
layer_sizes = (10, 10, 5)
input_dim = 30
# model
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = input_dim + int(not args.nobias)
elif args.h_type == 'fcc':
args.nconcepts += int(not args.nobias)
conceptizer = image_fcc_conceptizer(
11, args.nconcepts, args.concept_dim) #, sparsity = sparsity_l)
else:
raise ValueError('Unrecognized h_type')
model_path, log_path, results_path = generate_dir_names('cancer', args)
parametrizer = dfc_parametrizer(input_dim, *layer_sizes, args.nconcepts,
args.theta_dim)
aggregator = additive_scalar_aggregator(args.concept_dim, args.nclasses)
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
if args.theta_reg_type == 'unreg':
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ=1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ=2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ=3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecoginzed theta_reg_type')
trainer.train(train_loader,
valid_loader,
epochs=args.epochs,
save_path=model_path)
trainer.plot_losses(save_path=results_path)
# Load Best One
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'))
model = checkpoint['model']
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
results = {}
train_acc = trainer.validate(train_loader, fold='train')
valid_acc = trainer.validate(valid_loader, fold='valid')
test_acc = trainer.validate(test_loader, fold='test')
results['train_accuracy'] = train_acc
results['valid_accuracy'] = valid_acc
results['test_accuracy'] = test_acc
print('Train accuracy: {:8.2f}'.format(train_acc))
print('Valid accuracy: {:8.2f}'.format(valid_acc))
print('Test accuracy: {:8.2f}'.format(test_acc))
#noise_stability_plots(model, test_tds, cuda = args.cuda, save_path = results_path)
lips, argmaxes = sample_local_lipschitz(model,
test,
mode=2,
top_k=10,
max_distance=1)
results['test_discrete_glip'] = lips
results['test_discrete_glip_argmaxes'] = argmaxes
print('Local discrete g-Lipschitz estimate: {:8.2f}'.format(lips.mean()))
pointwise_test_loader = dataloader.DataLoader(test,
shuffle=True,
batch_size=1,
num_workers=4)
# Need to run in mode one, h is identity here
Lips_mean, Lips = estimate_dataset_lipschitz(model,
pointwise_test_loader,
continuous=True,
mode=1,
eps=0.2,
tol=1e-2,
maxpoints=100,
maxit=1e3,
log_interval=10,
patience=5,
cuda=args.cuda,
verbose=True)
results['test_cont_glip'] = Lips
print('Local dcontinuous g-Lipschitz estimate: {:8.2f}'.format(
Lips.mean()))
pickle.dump(results, open(results_path + '/model_metrics.pkl', "wb"))
def main():
args = parse_args()
np.random.seed(args.seed)
torch.manual_seed(args.seed)
args.nclasses = 10
args.theta_dim = args.nclasses
if (args.theta_arch == 'simple') or ('vgg' in args.theta_arch):
H, W = 32, 32
else:
# Need to resize to have access to torchvision's models
H, W = 224, 224
args.input_dim = H * W
model_path, log_path, results_path = generate_dir_names('cifar', args)
train_loader, valid_loader, test_loader, train_tds, test_tds = load_cifar_data(
batch_size=args.batch_size,
num_workers=args.num_workers,
resize=(H, W))
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = args.input_dim + int(not args.nobias)
elif args.h_type == 'cnn':
# biase. They treat it like any other concept.
#args.nconcepts += int(not args.nobias)
conceptizer = image_cnn_conceptizer(
args.input_dim, args.nconcepts, args.concept_dim,
nchannel=3) #, sparsity = sparsity_l)
else:
#args.nconcepts += int(not args.nobias)
conceptizer = image_fcc_conceptizer(
args.input_dim, args.nconcepts, args.concept_dim,
nchannel=3) #, sparsity = sparsity_l)
if args.theta_arch == 'simple':
parametrizer = image_parametrizer(args.input_dim,
args.nconcepts,
args.theta_dim,
nchannel=3,
only_positive=args.positive_theta)
elif 'vgg' in args.theta_arch:
parametrizer = vgg_parametrizer(
args.input_dim,
args.nconcepts,
args.theta_dim,
arch=args.theta_arch,
nchannel=3,
only_positive=args.positive_theta
) #torchvision.models.alexnet(num_classes = args.nconcepts*args.theta_dim)
else:
parametrizer = torchvision_parametrizer(
args.input_dim,
args.nconcepts,
args.theta_dim,
arch=args.theta_arch,
nchannel=3,
only_positive=args.positive_theta
) #torchvision.models.alexnet(num_classes = args.nconcepts*args.theta_dim)
aggregator = additive_scalar_aggregator(args.concept_dim, args.nclasses)
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
# if not args.train and args.load_model:
# checkpoint = torch.load(os.path.join(model_path,'model_best.pth.tar'), map_location=lambda storage, loc: storage)
# checkpoint.keys()
# model = checkpoint['model']
#
#
if args.theta_reg_type in ['unreg', 'none', None]:
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ=1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ=2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ=3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecoginzed theta_reg_type')
if args.train or not args.load_model or (not os.path.isfile(
os.path.join(model_path, 'model_best.pth.tar'))):
trainer.train(train_loader,
valid_loader,
epochs=args.epochs,
save_path=model_path)
trainer.plot_losses(save_path=results_path)
else:
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'),
map_location=lambda storage, loc: storage)
checkpoint.keys()
model = checkpoint['model']
trainer = VanillaClassTrainer(
model, args) # arbtrary trained, only need to compuyte val acc
#trainer.validate(test_loader, fold = 'test')
model.eval()
All_Results = {}
### 0. Concept Grid for Visualization
#concept_grid(model, test_loader, top_k = 10, cuda = args.cuda, save_path = results_path + '/concept_grid.pdf')
### 1. Single point lipshiz estimate via black box optim (for fair comparison)
# with other methods in which we have to use BB optimization.
features = None
classes = [str(i) for i in range(10)]
expl = gsenn_wrapper(model,
mode='classification',
input_type='image',
multiclass=True,
feature_names=features,
class_names=classes,
train_data=train_loader,
skip_bias=True,
verbose=False)
### Debug single input
# x = next(iter(train_tds))[0]
# attr = expl(x, show_plot = False)
# pdb.set_trace()
# #### Debug multi input
# x = next(iter(test_loader))[0] # Transformed
# x_raw = test_loader.dataset.test_data[:args.batch_size,:,:]
# attr = expl(x, x_raw = x_raw, show_plot = True)
# #pdb.set_trace()
# #### Debug argmaz plot_theta_stability
if args.h_type == 'input':
x = next(iter(test_tds))[0].numpy()
y = next(iter(test_tds))[0].numpy()
x_raw = (test_tds.test_data[0].float() / 255).numpy()
y_raw = revert_to_raw(x)
att_x = expl(x, show_plot=False)
att_y = expl(y, show_plot=False)
lip = 1
lipschitz_argmax_plot(x_raw, y_raw, att_x, att_y,
lip) # save_path=fpath)
#pdb.set_trace()
### 2. Single example lipschitz estimate with Black Box
do_bb_stability_example = True
if do_bb_stability_example:
print('**** Performing lipschitz estimation for a single point ****')
idx = 0
print('Example index: {}'.format(idx))
#x = train_tds[idx][0].view(1,28,28).numpy()
x = next(iter(test_tds))[0].numpy()
#x_raw = (test_tds.test_data[0].float()/255).numpy()
x_raw = (test_tds.test_data[0] / 255)
#x_raw = next(iter(train_tds))[0]
# args.optim = 'gp'
# args.lip_eps = 0.1
# args.lip_calls = 10
Results = {}
lip, argmax = expl.local_lipschitz_estimate(x,
bound_type='box_std',
optim=args.optim,
eps=args.lip_eps,
n_calls=4 * args.lip_calls,
njobs=1,
verbose=2)
#pdb.set_trace()
Results['lip_argmax'] = (x, argmax, lip)
# .reshape(inputs.shape[0], inputs.shape[1], -1)
att = expl(x, None, show_plot=False) #.squeeze()
# .reshape(inputs.shape[0], inputs.shape[1], -1)
att_argmax = expl(argmax, None, show_plot=False) #.squeeze()
#pdb.set_trace()
Argmax_dict = {'lip': lip, 'argmax': argmax, 'x': x}
fpath = os.path.join(results_path, 'argmax_lip_gp_senn.pdf')
if args.h_type == 'input':
lipschitz_argmax_plot(x_raw,
revert_to_raw(argmax),
att,
att_argmax,
lip,
save_path=fpath)
pickle.dump(Argmax_dict,
open(results_path + '/argmax_lip_gp_senn.pkl', "wb"))
#noise_stability_plots(model, test_tds, cuda = args.cuda, save_path = results_path)
### 3. Local lipschitz estimate over multiple samples with Black BOx Optim
do_bb_stability = True
if do_bb_stability:
print(
'**** Performing black-box lipschitz estimation over subset of dataset ****'
)
maxpoints = 20
#valid_loader 0 it's shuffled, so it's like doing random choice
mini_test = next(iter(valid_loader))[0][:maxpoints].numpy()
lips = expl.estimate_dataset_lipschitz(mini_test,
n_jobs=-1,
bound_type='box_std',
eps=args.lip_eps,
optim=args.optim,
n_calls=args.lip_calls,
verbose=2)
Stability_dict = {'lips': lips}
pickle.dump(Stability_dict,
open(results_path + '_stability_blackbox.pkl', "wb"))
All_Results['stability_blackbox'] = lips
pickle.dump(
All_Results,
open(results_path + '_combined_metrics.pkl'.format(dataname), "wb"))
def main():
args = parse_args()
args.nclasses = 10
args.theta_dim = args.nclasses
model_path, log_path, results_path = generate_dir_names('mnist', args)
train_loader, valid_loader, test_loader, train_tds, test_tds = load_mnist_data(
batch_size=args.batch_size, num_workers=args.num_workers
)
# Set h_type
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = 28*28 + int(not args.nobias)
elif args.h_type == 'cnn':
conceptizer = image_cnn_conceptizer(28*28, args.nconcepts, args.concept_dim) #, sparsity = sparsity_l)
else:
conceptizer = image_fcc_conceptizer(28*28, args.nconcepts, args.concept_dim) #, sparsity = sparsity_l)
# Initialize model
parametrizer = image_parametrizer(28*28, args.nconcepts, args.theta_dim, only_positive = args.positive_theta)
aggregator = additive_scalar_aggregator(args.concept_dim, args.nclasses)
model = GSENN(conceptizer, parametrizer, aggregator)
# If load_model == True, load existing model
if args.load_model:
checkpoint = torch.load(os.path.join(model_path,'model_best.pth.tar'), map_location=lambda storage, loc: storage)
checkpoint.keys()
model = checkpoint['model']
# Specify theta regression type
if args.theta_reg_type in ['unreg','none', None]:
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ = 1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ = 2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ = 3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecoginzed theta_reg_type')
# Train model
if not args.load_model and args.train:
trainer.train(train_loader, valid_loader, epochs = args.epochs, save_path = model_path)
trainer.plot_losses(save_path=results_path)
# If nothing is specified, load model and use VanillaClassTrainer
else:
checkpoint = torch.load(os.path.join(model_path,'model_best.pth.tar'), map_location=lambda storage, loc: storage)
checkpoint.keys()
model = checkpoint['model']
trainer = VanillaClassTrainer(model, args)
print("Done training/ loading model")
# Evaluation
### 1. Single point lipshiz estimate via black box optim
# All methods tested with BB optim for fair comparison)
features = None
classes = [str(i) for i in range(10)]
model.eval()
expl = gsenn_wrapper(model,
mode = 'classification',
input_type = 'image',
multiclass=True,
feature_names = features,
class_names = classes,
train_data = train_loader,
skip_bias = True,
verbose = False)
# Make noise stability plots, (Figure 4 paper)
print("Results_path", results_path)
# noise_stability_plots(model, test_tds, cuda = args.cuda, save_path = results_path)
# One sample
# plt.imshow(test_tds[1][0].reshape(28,28))
# plt.show()
print(test_tds)
distances = defaultdict(list)
print(distances)
scale = 0.5
print(distances)
for i in tqdm(range(1000)):
x = Variable(test_tds[i][0].view(1,1,28,28), volatile = True)
true_class = test_tds[i][1][0].item()
pred = model(x)
theta = model.thetas.data.cpu().numpy().squeeze()
klass = pred.data.max(1)[1]
deps = theta[:,klass].squeeze()
# print("prediction", klass)
# print("dependencies", deps)
# Add noise to sample and repeat
noise = Variable(scale*torch.randn(x.size()), volatile = True)
pred = model(noise)
theta = model.thetas.data.cpu().numpy().squeeze()
klass_noise = pred.data.max(1)[1]
deps_noise = theta[:,klass].squeeze()
dist = np.linalg.norm(deps - deps_noise)
distances[true_class].append(dist)
print(distances)
def main():
args = parse_args()
args.nclasses = 1
args.theta_dim = args.nclasses
args.print_freq = 100
args.epochs = 10
train_loader, valid_loader, test_loader, train, valid, test, data, x_train, x_valid, x_test = load_cancer_data(
)
layer_sizes = (10, 10, 5)
input_dim = 30
# model
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = input_dim + int(not args.nobias)
elif args.h_type == 'fcc':
args.nconcepts += int(not args.nobias)
conceptizer = image_fcc_conceptizer(
11, args.nconcepts, args.concept_dim) #, sparsity = sparsity_l)
else:
raise ValueError('Unrecognized h_type')
args.theta_arch = "dummy"
model_path, log_path, results_path = generate_dir_names('cancer', args)
parametrizer = dfc_parametrizer(input_dim, *layer_sizes, args.nconcepts,
args.theta_dim)
aggregator = additive_scalar_aggregator(args.concept_dim, args.nclasses)
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
if args.theta_reg_type == 'unreg':
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ=1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ=2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ=3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecoginzed theta_reg_type')
trainer.train(train_loader,
valid_loader,
epochs=args.epochs,
save_path=model_path)
trainer.plot_losses(save_path=results_path)
# Load Best One
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'))
model = checkpoint['model']
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
model = Wrapper(model)
x = x_test[0, :]
print("Verifying Explanation Setup:")
print("Using model.predict() - ", model.predict(x))
print("Using explanation - ", pred_from_coefs(model.explain(x), x))
results = {}
test_acc = trainer.validate(test_loader, fold='test').data.numpy().reshape(
(1))[0]
results['acc'] = np.float64(test_acc) / 100
print('Test accuracy: {:8.2f}'.format(test_acc))
standard, causal, stability = metrics_self(model, x_test)
results['standard'] = np.float64(standard[0])
results['causal'] = np.float64(causal[0, 0])
results['stability'] = np.float64(stability)
print(results)
with open("out.json", "w") as f:
json.dump(results, f)
def main():
args = parse_args()
args.nclasses = 10
args.theta_dim = args.nclasses
model_path, log_path, results_path = generate_dir_names('mnist', args)
# print("Model path out", model_path)
train_loader, valid_loader, test_loader, train_tds, test_tds = load_mnist_data(
batch_size=args.batch_size, num_workers=args.num_workers)
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = 28 * 28 + int(not args.nobias)
elif args.h_type == 'cnn':
#args.nconcepts += int(not args.nobias)
conceptizer = image_cnn_conceptizer(
28 * 28, args.nconcepts,
args.concept_dim) #, sparsity = sparsity_l)
else:
#args.nconcepts += int(not args.nobias)
conceptizer = image_fcc_conceptizer(
28 * 28, args.nconcepts,
args.concept_dim) #, sparsity = sparsity_l)
parametrizer = image_parametrizer(28 * 28,
args.nconcepts,
args.theta_dim,
only_positive=args.positive_theta)
aggregator = additive_scalar_aggregator(args.concept_dim, args.nclasses)
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
if args.load_model:
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'),
map_location=lambda storage, loc: storage)
checkpoint.keys()
model = checkpoint['model']
if args.theta_reg_type in ['unreg', 'none', None]:
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ=1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ=2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ=3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecoginzed theta_reg_type')
if not args.load_model and args.train:
trainer.train(train_loader,
valid_loader,
epochs=args.epochs,
save_path=model_path)
trainer.plot_losses(save_path=results_path)
else:
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'),
map_location=lambda storage, loc: storage)
checkpoint.keys()
model = checkpoint['model']
trainer = VanillaClassTrainer(model, args)
All_Results = {}
features = None
classes = [str(i) for i in range(10)]
model.eval()
expl = new_wrapper(model,
mode='classification',
input_type='image',
multiclass=True,
feature_names=features,
class_names=classes,
train_data=train_loader,
skip_bias=True,
verbose=False)
## Faithfulness analysis. Generates faithfulness plots, dependency plots and correlation scores.
print("Performing faithfulness analysis...")
correlations = np.array([])
altcorrelations = np.array([])
for i, (inputs, targets) in enumerate(tqdm(test_loader)):
# get the inputs
if args.demo:
if args.nconcepts == 5:
if i != 0:
continue
else:
args.demo = 0 + 1
elif args.nconcepts == 22:
if i != 0:
continue
else:
args.demo = 1 + 1
if args.cuda:
inputs, targets = inputs.cuda(), targets.cuda()
input_var = torch.autograd.Variable(inputs, volatile=True)
target_var = torch.autograd.Variable(targets)
if not args.noplot:
save_path = results_path + '/faithfulness' + str(i) + '/'
if not os.path.isdir(save_path):
os.mkdir(save_path)
else:
save_path = None
corrs, altcorrs = expl.compute_dataset_consistency(
input_var,
targets=target_var,
inputs_are_concepts=False,
save_path=save_path,
demo_mode=args.demo)
correlations = np.append(correlations, corrs)
altcorrelations = np.append(altcorrelations, altcorrs)
average_correlation = np.sum(correlations) / len(correlations)
std_correlation = np.std(correlations)
average_alt_correlation = np.sum(altcorrelations) / len(altcorrelations)
std_alt_correlation = np.std(altcorrelations)
print("Average correlation:", average_correlation)
print("Standard deviation of correlations: ", std_correlation)
print("Average alternative correlation:", average_alt_correlation)
print("Standard deviation of alternative correlations: ",
std_alt_correlation)
print("Generating Faithfulness correlation box plot..")
box_plot_values = [correlations, altcorrelations]
box = plt.boxplot(box_plot_values,
patch_artist=True,
labels=['theta(x)', 'theta(x) h(x)'])
colors = ['blue', 'purple']
for patch, color in zip(box['boxes'], colors):
patch.set_facecolor(color)
plt.savefig(results_path + '/faithfulness_box_plot.png',
format="png",
dpi=300,
verbose=True)
if not args.noplot and args.nconcepts == 5:
print("Generating theta, h and theta*h distribution histograms...")
# Make histograms
plot_distribution_h(test_loader,
expl,
'thetaxhx',
fig=0,
results_path=results_path)
plot_distribution_h(test_loader,
expl,
'thetax',
fig=1,
results_path=results_path)
plot_distribution_h(test_loader,
expl,
'hx',
fig=2,
results_path=results_path)
print("Generating stability metrics...")
# Compute stabilites
noises = np.arange(0, 0.21, 0.02)
dist_dict, dist_dict_2 = {}, {}
for noise in noises:
distances = eval_stability_2(test_loader, expl, noise, False)
distances_2 = eval_stability_2(test_loader, expl, noise, True)
dist_dict[noise] = distances
dist_dict_2[noise] = distances_2
print("Generating stability plot...")
# Plot stability
distances, distances_2, noises = dist_dict, dist_dict_2, noises
means = [np.mean(distances[noise]) for noise in noises]
stds = [np.std(distances[noise]) for noise in noises]
means_min = [means[i] - stds[i] for i in range(len(means))]
means_max = [means[i] + stds[i] for i in range(len(means))]
means_2 = [np.mean(distances_2[noise]) for noise in noises]
stds_2 = [np.std(distances_2[noise]) for noise in noises]
means_min_2 = [means_2[i] - stds_2[i] for i in range(len(means_2))]
means_max_2 = [means_2[i] + stds_2[i] for i in range(len(means_2))]
fig, ax = plt.subplots(1)
ax.plot(noises, means, lw=2, label='theta(x)', color='blue')
ax.plot(noises, means_2, lw=2, label='theta(x)^T h(x)', color='purple')
ax.fill_between(noises,
means_max,
means_min,
facecolor='blue',
alpha=0.3)
ax.fill_between(noises,
means_max_2,
means_min_2,
facecolor='purple',
alpha=0.3)
ax.set_title('Stability')
ax.legend(loc='upper left')
ax.set_xlabel('Added noise')
ax.set_ylabel('Norm of relevance coefficients')
ax.grid()
fig.savefig(results_path + '/stability' + '.png',
format="png",
dpi=300)
if (not args.demo) or (args.demo and args.nconcepts == 5):
concept_grid(model,
test_loader,
cuda=args.cuda,
top_k=10,
save_path=results_path + '/concept_grid.png')
print("Finished")
def main():
args = parse_args()
args.nclasses = 1
args.theta_dim = args.nclasses
args.print_freq = 100
args.epochs = 10
train_loader, valid_loader, test_loader, train, valid, test, data, x_train, x_valid, x_test = load_cancer_data(
)
layer_sizes = (10, 10, 5)
input_dim = 30
# model
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = input_dim + int(not args.nobias)
elif args.h_type == 'fcc':
args.nconcepts += int(not args.nobias)
conceptizer = image_fcc_conceptizer(
11, args.nconcepts, args.concept_dim) #, sparsity = sparsity_l)
else:
raise ValueError('Unrecognized h_type')
args.theta_arch = "dummy"
model_path, log_path, results_path = generate_dir_names('cancer', args)
parametrizer = dfc_parametrizer(input_dim, *layer_sizes, args.nconcepts,
args.theta_dim)
aggregator = additive_scalar_aggregator(args.concept_dim, args.nclasses)
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
if args.theta_reg_type == 'unreg':
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ=1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ=2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ=3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecoginzed theta_reg_type')
trainer.train(train_loader,
valid_loader,
epochs=args.epochs,
save_path=model_path)
trainer.plot_losses(save_path=results_path)
# Load Best One
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'))
model = checkpoint['model']
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
results = {}
test_acc = trainer.validate(test_loader, fold='test').data.numpy().reshape(
(1))[0]
results['acc'] = np.float64(test_acc) / 100
print('Test accuracy: {:8.2f}'.format(test_acc))
wrapper = Wrapper(model)
r_maple = np.float64(metrics_maple(wrapper, x_train, x_valid, x_test))
print('MAPLE', r_maple)
results['maple_pf'] = r_maple[0][0]
results['maple_nf'] = r_maple[1][0]
results['maple_s'] = r_maple[2][0]
r_lime = np.float64(metrics_lime(wrapper, x_train, x_test))
print('LIME', r_lime)
results['lime_pf'] = r_lime[0][0]
results['lime_nf'] = r_lime[1][0]
results['lime_s'] = r_lime[2][0]
print(results)
with open("out.json", "w") as f:
json.dump(results, f)
def main():
args = parse_args()
args.nclasses = 10
args.theta_dim = args.nclasses
model_path, log_path, results_path = generate_dir_names('mnist', args)
train_loader, valid_loader, test_loader, train_tds, test_tds = load_mnist_data(
batch_size=args.batch_size, num_workers=args.num_workers)
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = 28 * 28 + int(not args.nobias)
elif args.h_type == 'cnn':
#args.nconcepts += int(not args.nobias)
conceptizer = image_cnn_conceptizer(
28 * 28, args.nconcepts,
args.concept_dim) #, sparsity = sparsity_l)
else:
#args.nconcepts += int(not args.nobias)
conceptizer = image_fcc_conceptizer(
28 * 28, args.nconcepts,
args.concept_dim) #, sparsity = sparsity_l)
parametrizer = image_parametrizer(28 * 28,
args.nconcepts,
args.theta_dim,
only_positive=args.positive_theta)
aggregator = additive_scalar_aggregator(args.concept_dim, args.nclasses)
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
if args.load_model:
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'),
map_location=lambda storage, loc: storage)
checkpoint.keys()
model = checkpoint['model']
if args.theta_reg_type in ['unreg', 'none', None]:
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ=1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ=2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ=3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecoginzed theta_reg_type')
if not args.load_model and args.train:
trainer.train(train_loader,
valid_loader,
epochs=args.epochs,
save_path=model_path)
trainer.plot_losses(save_path=results_path)
else:
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'),
map_location=lambda storage, loc: storage)
checkpoint.keys()
model = checkpoint['model']
trainer = VanillaClassTrainer(model, args)
trainer.validate(test_loader, fold='test')
All_Results = {}
### 1. Single point lipshiz estimate via black box optim
# All methods tested with BB optim for fair comparison)
features = None
classes = [str(i) for i in range(10)]
model.eval()
expl = new_wrapper(model,
mode='classification',
input_type='image',
multiclass=True,
feature_names=features,
class_names=classes,
train_data=train_loader,
skip_bias=True,
verbose=False)
#### Debug single input
# x = next(iter(train_tds))[0]
# attr = expl(x, show_plot = False)
# pdb.set_trace()
# #### Debug multi input
# x = next(iter(test_loader))[0] # Transformed
# x_raw = test_loader.dataset.test_data[:args.batch_size,:,:]
# attr = expl(x, x_raw = x_raw, show_plot = True)
# #pdb.set_trace()
### Consistency analysis
# for i, (inputs, targets) in enumerate(test_loader):
# # get the inputs
# if model.cuda:
# inputs, targets = inputs.cuda(), targets.cuda()
# input_var = torch.autograd.Variable(inputs, volatile=True)
# save_path = results_path + '/faithfulness' + str(i) + '/'
# if not os.path.isdir(save_path):
# os.mkdir(save_path)
# corrs = expl.compute_dataset_consistency(input_var, inputs_are_concepts = False, save_path = save_path)
# #### Debug argmax plot_theta_stability
if args.h_type == 'input':
x = next(iter(test_tds))[0].numpy()
y = next(iter(test_tds))[0].numpy()
x_raw = (test_tds.test_data[0].float() / 255).numpy()
y_raw = revert_to_raw(x)
att_x = expl(x, show_plot=False)
att_y = expl(y, show_plot=False)
lip = 1
lipschitz_argmax_plot(x_raw, y_raw, att_x, att_y,
lip) # save_path=fpath)
#pdb.set_trace()
### 2. Single example lipschitz estimate with Black Box
do_bb_stability_example = False # Aangepast: was False
if do_bb_stability_example:
print('**** Performing lipschitz estimation for a single point ****')
idx = 0
print('Example index: {}'.format(idx))
#x = train_tds[idx][0].view(1,28,28).numpy()
x = next(iter(test_tds))[0].numpy()
x_raw = (test_tds.test_data[0].float() / 255).numpy()
#x_raw = next(iter(train_tds))[0]
# args.optim = 'gp'
# args.lip_eps = 0.1
# args.lip_calls = 10
Results = {}
lip, argmax = expl.local_lipschitz_estimate(x,
bound_type='box_std',
optim=args.optim,
eps=args.lip_eps,
n_calls=4 * args.lip_calls,
njobs=1,
verbose=2)
#pdb.set_trace()
Results['lip_argmax'] = (x, argmax, lip)
# .reshape(inputs.shape[0], inputs.shape[1], -1)
att = expl(x, None, show_plot=False) #.squeeze()
# .reshape(inputs.shape[0], inputs.shape[1], -1)
att_argmax = expl(argmax, None, show_plot=False) #.squeeze()
#pdb.set_trace()
Argmax_dict = {'lip': lip, 'argmax': argmax, 'x': x}
fpath = os.path.join(results_path, 'argmax_lip_gp_senn.pdf')
if args.h_type == 'input':
lipschitz_argmax_plot(x_raw,
revert_to_raw(argmax),
att,
att_argmax,
lip,
save_path=fpath)
pickle.dump(Argmax_dict,
open(results_path + '/argmax_lip_gp_senn.pkl', "wb"))
pdb.set_trace()
# print(asd.asd)
noise_stability_plots(model,
test_tds,
cuda=args.cuda,
save_path=results_path)
### 3. Local lipschitz estimate over multiple samples with Black BOx Optim
do_bb_stability = False # Aangepast, was: True
if do_bb_stability:
print(
'**** Performing black-box lipschitz estimation over subset of dataset ****'
)
maxpoints = 20
#valid_loader 0 it's shuffled, so it's like doing random choice
mini_test = next(iter(valid_loader))[0][:maxpoints].numpy()
lips = expl.estimate_dataset_lipschitz(mini_test,
n_jobs=-1,
bound_type='box_std',
eps=args.lip_eps,
optim=args.optim,
n_calls=args.lip_calls,
verbose=2)
pdb.set_trace()
Stability_dict = {'lips': lips}
pickle.dump(Stability_dict,
open(results_path + '_stability_blackbox.pkl', "wb"))
All_Results['stability_blackbox'] = lips
# add concept plot
concept_grid(model,
test_loader,
top_k=10,
save_path=results_path + '/concept_grid.pdf')
pickle.dump(All_Results,
open(results_path + '_combined_metrics.pkl',
"wb")) # Aangepast: .pkl was .format(dataname)
def main():
args = parse_args()
args.nclasses = 1
args.theta_dim = args.nclasses
args.print_freq = 100
args.epochs = 10
train_loader, valid_loader, test_loader, train, valid, test, data, feat_names = load_compas_data(
)
layer_sizes = (10, 10, 5)
input_dim = 11
if args.h_type == 'input':
conceptizer = input_conceptizer()
args.nconcepts = 11 + int(not args.nobias)
elif args.h_type == 'fcc':
args.nconcepts += int(not args.nobias)
conceptizer = image_fcc_conceptizer(
11, args.nconcepts, args.concept_dim) #, sparsity = sparsity_l)
else:
raise ValueError('Unrecognized h_type')
model_path, log_path, results_path = generate_dir_names('compas', args)
parametrizer = dfc_parametrizer(input_dim, *layer_sizes, args.nconcepts,
args.theta_dim)
aggregator = additive_scalar_aggregator(args.concept_dim, args.nclasses)
model = GSENN(conceptizer, parametrizer,
aggregator) #, learn_h = args.train_h)
if args.theta_reg_type == 'unreg':
trainer = VanillaClassTrainer(model, args)
elif args.theta_reg_type == 'grad1':
trainer = GradPenaltyTrainer(model, args, typ=1)
elif args.theta_reg_type == 'grad2':
trainer = GradPenaltyTrainer(model, args, typ=2)
elif args.theta_reg_type == 'grad3':
trainer = GradPenaltyTrainer(model, args, typ=3)
elif args.theta_reg_type == 'crosslip':
trainer = CLPenaltyTrainer(model, args)
else:
raise ValueError('Unrecognized theta_reg_type')
trainer.train(train_loader,
valid_loader,
epochs=args.epochs,
save_path=model_path)
trainer.plot_losses(save_path=results_path)
# Load Best One
checkpoint = torch.load(os.path.join(model_path, 'model_best.pth.tar'))
model = checkpoint['model']
results = {}
train_acc = trainer.validate(train_loader, fold='train')
valid_acc = trainer.validate(valid_loader, fold='valid')
test_acc = trainer.validate(test_loader, fold='test')
results['train_accuracy'] = train_acc
results['valid_accuracy'] = valid_acc
results['test_accuracy'] = test_acc
print('Train accuracy: {:8.2f}'.format(train_acc))
print('Valid accuracy: {:8.2f}'.format(valid_acc))
print('Test accuracy: {:8.2f}'.format(test_acc))
#noise_stability_plots(model, test_tds, cuda = args.cuda, save_path = results_path)
lips, argmaxes = sample_local_lipschitz(model,
test,
mode=2,
top_k=10,
max_distance=3)
max_lip = lips.max()
imax = np.unravel_index(np.argmax(lips), lips.shape)[0]
jmax = argmaxes[imax][0][0]
print('Max Lip value: {}, attained for pair ({},{})'.format(
max_lip, imax, jmax))
x = test.tensors[0][imax]
argmax = test.tensors[0][jmax]
pred_x = model(Variable(x.view(1, -1), volatile=True)).data
att_x = model.thetas.data.squeeze().numpy().squeeze()
pred_argmax = model(Variable(argmax.view(1, -1), volatile=True)).data
att_argmax = model.thetas.data.squeeze().numpy().squeeze()
# pdb.set_trace()
results['x_max'] = x
results['x_argmax'] = argmax
results['test_discrete_glip'] = lips
results['test_discrete_glip_argmaxes'] = argmaxes
print('Local g-Lipschitz estimate: {:8.2f}'.format(lips.mean()))
fpath = os.path.join(results_path, 'discrete_lip_gsenn')
ppath = os.path.join(results_path, 'relevance_argmax_gsenn')
pickle.dump(results, open(fpath + '.pkl', "wb")) # FOrmerly model_metrics
print(ppath)
lipschitz_feature_argmax_plot(x,
argmax,
att_x,
att_argmax,
feat_names=feat_names,
digits=2,
figsize=(5, 6),
widths=(2, 3),
save_path=ppath + '.pdf')