def attribution_map(func: Callable[[torch.Tensor], torch.Tensor],
attribution_type: Type,
with_noise: bool,
probe_data: torch.Tensor,
norm_square: bool = False,
**attribution_kwargs) -> torch.Tensor:
"""
Calculate attribution map with given attribution type(algorithm).
Args:
model: pytorch module
attribution_type: attribution algorithm, e.g. IntegratedGradients, InputXGradient, ...
with_noise: whether to add noise tunnel
probe_data: input data to model
device: torch.device("cuda: 0")
attribution_kwargs: other kwargs for attribution method
Return: attribution map
"""
attribution: Attribution = attribution_type(
lambda x: with_norm(func, x, norm_square))
if with_noise:
attribution = NoiseTunnel(attribution)
attr_map = attribution.attribute(inputs=probe_data,
target=None,
**attribution_kwargs)
return attr_map.detach()
def extract_IGNT(self, X_test):
ig = IntegratedGradients(self.net)
ig_nt = NoiseTunnel(ig)
start = time.time()
ig_nt_attr_test = ig_nt.attribute(X_test.to(self.device))
print("temps train", time.time() - start)
return ig_nt_attr_test.detach().cpu().numpy()
def _get_attributions(self, enc_data, metric, n_steps, nt_type,
baseline_type, strides, sliding_window_shapes):
# Get Baseline
baseline = self.get_baseline_img(enc_data[0], baseline_type)
supported_metrics = {}
if metric == 'IG':
self._integrated_gradients = self._integrated_gradients if hasattr(
self, '_integrated_gradients') else IntegratedGradients(
self.model)
return self._integrated_gradients.attribute(enc_data[0],
baseline,
target=enc_data[1],
n_steps=200)
elif metric == 'NT':
self._integrated_gradients = self._integrated_gradients if hasattr(
self, '_integrated_gradients') else IntegratedGradients(
self.model)
self._noise_tunnel = self._noise_tunnel if hasattr(
self, '_noise_tunnel') else NoiseTunnel(
self._integrated_gradients)
return self._noise_tunnel.attribute(enc_data[0].to(
self.dls.device),
n_samples=1,
nt_type=nt_type,
target=enc_data[1])
elif metric == 'Occl':
self._occlusion = self._occlusion if hasattr(
self, '_occlusion') else Occlusion(self.model)
return self._occlusion.attribute(
enc_data[0].to(self.dls.device),
strides=strides,
target=enc_data[1],
sliding_window_shapes=sliding_window_shapes,
baselines=baseline)
class NoiseTunnelCallback(Callback):
"Captum Callback for Resnet Interpretation"
def __init__(self):
pass
def after_fit(self):
self.integrated_gradients = IntegratedGradients(self.model)
self._noise_tunnel= NoiseTunnel(self.integrated_gradients)
def visualize(self,inp_data,cmap_name='custom blue',colors=None,N=256,methods=['original_image','heat_map'],signs=["all", "positive"],nt_type='smoothgrad'):
dl = self.dls.test_dl(L(inp_data),with_labels=True, bs=1)
self.enc_inp,self.enc_preds= dl.one_batch()
dec_data=dl.decode((self.enc_inp,self.enc_preds))
self.dec_img,self.dec_pred=dec_data[0][0],dec_data[1][0]
self.colors = [(0, '#ffffff'),(0.25, '#000000'),(1, '#000000')] if colors is None else colors
attributions_ig_nt = self._noise_tunnel.attribute(self.enc_inp.to(self.dl.device), n_samples=1, nt_type=nt_type, target=self.enc_preds)
default_cmap = LinearSegmentedColormap.from_list(cmap_name,
self.colors, N=N)
_ = viz.visualize_image_attr_multiple(np.transpose(attributions_ig_nt.squeeze().cpu().detach().numpy(), (1,2,0)),
np.transpose(self.dec_img.numpy(), (1,2,0)),
methods,signs,
cmap=default_cmap,
show_colorbar=True,titles=[f'Original Image - ({self.dec_pred})', 'Noise Tunnel'])
def compute_saliency_noise_tunnel(self, img_path, target):
# open image
img, transformed_img, input = self.open_image(img_path)
gradient_saliency = Saliency(self.model)
noise_tunnel = NoiseTunnel(gradient_saliency)
attributions_sa_nt = noise_tunnel.attribute(
input,
n_samples=10,
nt_type='smoothgrad',
# internal_batch_size=8,
target=target)
attributions_sa_nt = np.transpose(
attributions_sa_nt.squeeze().cpu().detach().numpy(), (1, 2, 0))
return attributions_sa_nt
def _attr_noise_tunnel(self, input, pred):
attr_algo = NoiseTunnel(IntegratedGradients(self.model))
default_cmap = LinearSegmentedColormap.from_list('custom blue',
[(0, '#ffffff'),
(0.25, '#000000'),
(1, '#000000')], N=256)
attr_ = attr_algo.attribute(input, n_samples=10, nt_type='smoothgrad_sq', target=pred)
_ = viz.visualize_image_attr_multiple(
np.transpose(attr_.squeeze().cpu().detach().numpy(), (1, 2, 0)),
np.transpose(input.squeeze().cpu().detach().numpy(), (1, 2, 0)),
["original_image", "heat_map"],
["all", "positive"],
cmap=default_cmap,
show_colorbar=True)
def compute_integrated_gradients_noise_tunnel(self, img_path, target):
# open image
img, transformed_img, input = self.open_image(img_path)
integrated_gradients = IntegratedGradients(self.model)
# attributions_ig = integrated_gradients.attribute(input,
# target=target,
# n_steps=200)
noise_tunnel = NoiseTunnel(integrated_gradients)
attributions_ig_nt = noise_tunnel.attribute(input,
n_samples=10,
nt_type='smoothgrad',
internal_batch_size=8,
n_steps=200,
target=target)
attributions_ig_nt = np.transpose(
attributions_ig_nt.squeeze().cpu().detach().numpy(), (1, 2, 0))
return attributions_ig_nt
def PT_SmoothGradient(model,
x,
y_onthot,
multiply_with_input=False,
device='cuda:0',
n_steps=50,
stdevs=0.15,
**kwargs):
input = torch.tensor(x).to(device)
model = model.to(device)
model.eval()
saliency = NoiseTunnel(Saliency(model))
target = torch.tensor(np.argmax(y_onthot, -1)).to(device)
attribution_map = saliency.attribute(input,
n_samples=n_steps,
target=target,
stdevs=stdevs,
abs=False)
if multiply_with_input:
attribution_map *= input
return attribution_map.detach().cpu().numpy()
def main(args):
train_loader, test_loader = data_generator(args.data_dir,1)
for m in range(len(models)):
model_name = "model_{}_NumFeatures_{}".format(models[m],args.NumFeatures)
model_filename = args.model_dir + 'm_' + model_name + '.pt'
pretrained_model = torch.load(open(model_filename, "rb"),map_location=device)
pretrained_model.to(device)
if(args.GradFlag):
Grad = Saliency(pretrained_model)
if(args.IGFlag):
IG = IntegratedGradients(pretrained_model)
if(args.DLFlag):
DL = DeepLift(pretrained_model)
if(args.GSFlag):
GS = GradientShap(pretrained_model)
if(args.DLSFlag):
DLS = DeepLiftShap(pretrained_model)
if(args.SGFlag):
Grad_ = Saliency(pretrained_model)
SG = NoiseTunnel(Grad_)
if(args.ShapleySamplingFlag):
SS = ShapleyValueSampling(pretrained_model)
if(args.GSFlag):
FP = FeaturePermutation(pretrained_model)
if(args.FeatureAblationFlag):
FA = FeatureAblation(pretrained_model)
if(args.OcclusionFlag):
OS = Occlusion(pretrained_model)
timeMask=np.zeros((args.NumTimeSteps, args.NumFeatures),dtype=int)
featureMask=np.zeros((args.NumTimeSteps, args.NumFeatures),dtype=int)
for i in range (args.NumTimeSteps):
timeMask[i,:]=i
for i in range (args.NumTimeSteps):
featureMask[:,i]=i
indexes = [[] for i in range(5,10)]
for i ,(data, target) in enumerate(test_loader):
if(target==5 or target==6 or target==7 or target==8 or target==9):
index=target-5
if(len(indexes[index])<1):
indexes[index].append(i)
for j, index in enumerate(indexes):
print(index)
# indexes = [[21],[17],[84],[9]]
for j, index in enumerate(indexes):
print("Getting Saliency for number", j+1)
for i, (data, target) in enumerate(test_loader):
if(i in index):
labels = target.to(device)
input = data.reshape(-1, args.NumTimeSteps, args.NumFeatures).to(device)
input = Variable(input, volatile=False, requires_grad=True)
baseline_single=torch.Tensor(np.random.random(input.shape)).to(device)
baseline_multiple=torch.Tensor(np.random.random((input.shape[0]*5,input.shape[1],input.shape[2]))).to(device)
inputMask= np.zeros((input.shape))
inputMask[:,:,:]=timeMask
inputMask =torch.Tensor(inputMask).to(device)
mask_single= torch.Tensor(timeMask).to(device)
mask_single=mask_single.reshape(1,args.NumTimeSteps, args.NumFeatures).to(device)
Data=data.reshape(args.NumTimeSteps, args.NumFeatures).data.cpu().numpy()
target_=int(target.data.cpu().numpy()[0])
plotExampleBox(Data,args.Graph_dir+'Sample_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.GradFlag):
attributions = Grad.attribute(input, \
target=labels)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_Grad_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(Grad,input, args.NumFeatures,args.NumTimeSteps, labels,hasBaseline=None)
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_Grad_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.IGFlag):
attributions = IG.attribute(input, \
baselines=baseline_single, \
target=labels)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_IG_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(IG,input, args.NumFeatures,args.NumTimeSteps, labels,hasBaseline=baseline_single)
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_IG_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.DLFlag):
attributions = DL.attribute(input, \
baselines=baseline_single, \
target=labels)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_DL_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(DL,input, args.NumFeatures,args.NumTimeSteps, labels,hasBaseline=baseline_single)
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_DL_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.GSFlag):
attributions = GS.attribute(input, \
baselines=baseline_multiple, \
stdevs=0.09,\
target=labels)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_GS_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(GS,input, args.NumFeatures,args.NumTimeSteps, labels,hasBaseline=baseline_multiple)
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_GS_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.DLSFlag):
attributions = DLS.attribute(input, \
baselines=baseline_multiple, \
target=labels)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_DLS_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(DLS,input, args.NumFeatures,args.NumTimeSteps, labels,hasBaseline=baseline_multiple)
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_DLS_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.SGFlag):
attributions = SG.attribute(input, \
target=labels)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_SG_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(SG,input, args.NumFeatures,args.NumTimeSteps, labels)
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_SG_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.ShapleySamplingFlag):
attributions = SS.attribute(input, \
baselines=baseline_single, \
target=labels,\
feature_mask=inputMask)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_SVS_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(SS,input, args.NumFeatures,args.NumTimeSteps, labels,hasBaseline=baseline_single,hasFeatureMask=inputMask)
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_SVS_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
# if(args.FeaturePermutationFlag):
# attributions = FP.attribute(input, \
# target=labels),
# # perturbations_per_eval= 1,\
# # feature_mask=mask_single)
# saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
# plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_FP',greyScale=True)
if(args.FeatureAblationFlag):
attributions = FA.attribute(input, \
target=labels)
# perturbations_per_eval= input.shape[0],\
# feature_mask=mask_single)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_FA_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(FA,input, args.NumFeatures,args.NumTimeSteps, labels)
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_FA_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.OcclusionFlag):
attributions = OS.attribute(input, \
sliding_window_shapes=(1,int(args.NumFeatures/10)),
target=labels,
baselines=baseline_single)
saliency_=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,attributions)
plotExampleBox(saliency_[0],args.Graph_dir+models[m]+'_FO_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
if(args.TSRFlag):
TSR_attributions = getTwoStepRescaling(OS,input, args.NumFeatures,args.NumTimeSteps, labels,hasBaseline=baseline_single,hasSliding_window_shapes= (1,int(args.NumFeatures/10)))
TSR_saliency=Helper.givenAttGetRescaledSaliency(args.NumTimeSteps, args.NumFeatures,TSR_attributions,isTensor=False)
plotExampleBox(TSR_saliency,args.Graph_dir+models[m]+'_TSR_FO_MNIST_'+str(target_)+'_index_'+str(i+1),greyScale=True)
def measure_model(
model_version,
dataset,
out_folder,
weights_dir,
device,
method=METHODS["gradcam"],
sample_images=50,
step=1,
):
invTrans = get_inverse_normalization_transformation()
data_dir = os.path.join("data")
if model_version == "resnet18":
model = create_resnet18_model(num_of_classes=NUM_OF_CLASSES[dataset])
elif model_version == "resnet50":
model = create_resnet50_model(num_of_classes=NUM_OF_CLASSES[dataset])
elif model_version == "densenet":
model = create_densenet121_model(
num_of_classes=NUM_OF_CLASSES[dataset])
else:
model = create_efficientnetb0_model(
num_of_classes=NUM_OF_CLASSES[dataset])
model.load_state_dict(torch.load(weights_dir))
# print(model)
model.eval()
model.to(device)
test_dataset = CustomDataset(
dataset=dataset,
transformer=get_default_transformation(),
data_type="test",
root_dir=data_dir,
step=step,
)
data_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=1,
shuffle=False,
num_workers=4)
try:
image_ids = random.sample(range(0, test_dataset.__len__()),
sample_images)
except ValueError:
raise ValueError(
f"Image sample number ({sample_images}) exceeded dataset size ({test_dataset.__len__()})."
)
classes_map = test_dataset.classes_map
print(f"Measuring {model_version} on {dataset} dataset, with {method}")
print("-" * 10)
pbar = tqdm(total=test_dataset.__len__(), desc="Model test completion")
multipy_by_inputs = False
if method == METHODS["ig"]:
attr_method = IntegratedGradients(model)
nt_samples = 8
n_perturb_samples = 3
if method == METHODS["saliency"]:
attr_method = Saliency(model)
nt_samples = 8
n_perturb_samples = 10
if method == METHODS["gradcam"]:
if model_version == "efficientnet":
attr_method = GuidedGradCam(model, model._conv_stem)
elif model_version == "densenet":
attr_method = GuidedGradCam(model, model.features.conv0)
else:
attr_method = GuidedGradCam(model, model.conv1)
nt_samples = 8
n_perturb_samples = 10
if method == METHODS["deconv"]:
attr_method = Deconvolution(model)
nt_samples = 8
n_perturb_samples = 10
if method == METHODS["gradshap"]:
attr_method = GradientShap(model)
nt_samples = 8
if model_version == "efficientnet":
n_perturb_samples = 3
elif model_version == "densenet":
n_perturb_samples = 2
else:
n_perturb_samples = 10
if method == METHODS["gbp"]:
attr_method = GuidedBackprop(model)
nt_samples = 8
n_perturb_samples = 10
if method == "lime":
attr_method = Lime(model)
nt_samples = 8
n_perturb_samples = 10
feature_mask = torch.tensor(lime_mask).to(device)
multipy_by_inputs = True
if method == METHODS['ig']:
nt = attr_method
else:
nt = NoiseTunnel(attr_method)
scores = []
@infidelity_perturb_func_decorator(multipy_by_inputs=multipy_by_inputs)
def perturb_fn(inputs):
noise = torch.tensor(np.random.normal(0, 0.003, inputs.shape)).float()
noise = noise.to(device)
return inputs - noise
for input, label in data_loader:
pbar.update(1)
inv_input = invTrans(input)
input = input.to(device)
input.requires_grad = True
output = model(input)
output = F.softmax(output, dim=1)
prediction_score, pred_label_idx = torch.topk(output, 1)
prediction_score = prediction_score.cpu().detach().numpy()[0][0]
pred_label_idx.squeeze_()
if method == METHODS['gradshap']:
baseline = torch.randn(input.shape)
baseline = baseline.to(device)
if method == "lime":
attributions = attr_method.attribute(input, target=1, n_samples=50)
elif method == METHODS['ig']:
attributions = nt.attribute(
input,
target=pred_label_idx,
n_steps=25,
)
elif method == METHODS['gradshap']:
attributions = nt.attribute(input,
target=pred_label_idx,
baselines=baseline)
else:
attributions = nt.attribute(
input,
nt_type="smoothgrad",
nt_samples=nt_samples,
target=pred_label_idx,
)
infid = infidelity(model,
perturb_fn,
input,
attributions,
target=pred_label_idx)
if method == "lime":
sens = sensitivity_max(
attr_method.attribute,
input,
target=pred_label_idx,
n_perturb_samples=1,
n_samples=200,
feature_mask=feature_mask,
)
elif method == METHODS['ig']:
sens = sensitivity_max(
nt.attribute,
input,
target=pred_label_idx,
n_perturb_samples=n_perturb_samples,
n_steps=25,
)
elif method == METHODS['gradshap']:
sens = sensitivity_max(nt.attribute,
input,
target=pred_label_idx,
n_perturb_samples=n_perturb_samples,
baselines=baseline)
else:
sens = sensitivity_max(
nt.attribute,
input,
target=pred_label_idx,
n_perturb_samples=n_perturb_samples,
)
inf_value = infid.cpu().detach().numpy()[0]
sens_value = sens.cpu().detach().numpy()[0]
if pbar.n in image_ids:
attr_data = attributions.squeeze().cpu().detach().numpy()
fig, ax = viz.visualize_image_attr_multiple(
np.transpose(attr_data, (1, 2, 0)),
np.transpose(inv_input.squeeze().cpu().detach().numpy(),
(1, 2, 0)),
["original_image", "heat_map"],
["all", "positive"],
titles=["original_image", "heat_map"],
cmap=default_cmap,
show_colorbar=True,
use_pyplot=False,
fig_size=(8, 6),
)
ax[0].set_xlabel(
f"Infidelity: {'{0:.6f}'.format(inf_value)}\n Sensitivity: {'{0:.6f}'.format(sens_value)}"
)
fig.suptitle(
f"True: {classes_map[str(label.numpy()[0])][0]}, Pred: {classes_map[str(pred_label_idx.item())][0]}\nScore: {'{0:.4f}'.format(prediction_score)}",
fontsize=16,
)
fig.savefig(
os.path.join(
out_folder,
f"{str(pbar.n)}-{classes_map[str(label.numpy()[0])][0]}-{classes_map[str(pred_label_idx.item())][0]}.png",
))
plt.close(fig)
# if pbar.n > 25:
# break
scores.append([inf_value, sens_value])
pbar.close()
np.savetxt(
os.path.join(out_folder, f"{model_version}-{dataset}-{method}.csv"),
np.array(scores),
delimiter=",",
header="infidelity,sensitivity",
)
print(f"Artifacts stored at {out_folder}")
def save_explanation(inputImage: torch.Tensor, modeladapter: torch.nn.Module,
cfg: DictConfig, pred_label_idx: int, pred_label_num: int,
gt_label_num: int, filename: str, filepath: str,
filename_without_ext: str, prediction_score: float):
"""
Return explanation value dict
"""
input_gradients = torch.unsqueeze(inputImage, 0)
integrated_gradients = IntegratedGradients(modeladapter)
default_cmap = LinearSegmentedColormap.from_list('custom blue',
[(0, '#ffffff'),
(0.25, '#000000'),
(1, '#000000')],
N=256)
noise_tunnel = NoiseTunnel(integrated_gradients)
attributions_ig = noise_tunnel.attribute(
input_gradients,
nt_samples=cfg.inference.captum.noise_tunnel.nt_samples,
nt_samples_batch_size=cfg.inference.captum.noise_tunnel.
nt_samples_batch_size,
nt_type=cfg.inference.captum.noise_tunnel.nt_type,
target=pred_label_idx)
# Standard Captum Visualization
figure, plot = viz.visualize_image_attr(np.transpose(
attributions_ig.squeeze().cpu().detach().numpy(), (1, 2, 0)),
method="heat_map",
sign="positive",
cmap=default_cmap,
show_colorbar=True)
dict_col_name = {}
dict_col_name.update({
"pred": pred_label_num,
"GT": gt_label_num,
"predict_score": prediction_score.squeeze_().item(),
"image_path": filename
})
save_path_original = "/PREDcls_" + str(pred_label_num) + "_GTcls_" + str(
gt_label_num) + "_" + filename
save_path_explanation = "/PREDcls_" + str(
pred_label_num) + "_GTcls_" + str(
gt_label_num) + "_" + filename_without_ext + "_explain" + ".png"
if (pred_label_num == gt_label_num):
# path to image
shutil.copy(
filepath,
cfg.inference.captum.correct_explanation_path + save_path_original)
figure.savefig(cfg.inference.captum.correct_explanation_path +
save_path_explanation,
dpi=figure.dpi)
else:
shutil.copy(
filepath,
cfg.inference.captum.error_explanation_path + save_path_original)
figure.savefig(cfg.inference.captum.error_explanation_path +
save_path_explanation,
dpi=figure.dpi)
plt.close()
return dict_col_name
def main(args, DatasetsTypes, DataGenerationTypes, models, device):
for m in range(len(models)):
for x in range(len(DatasetsTypes)):
for y in range(len(DataGenerationTypes)):
if (DataGenerationTypes[y] == None):
args.DataName = DatasetsTypes[x] + "_Box"
else:
args.DataName = DatasetsTypes[
x] + "_" + DataGenerationTypes[y]
Training = np.load(args.data_dir + "SimulatedTrainingData_" +
args.DataName + "_F_" +
str(args.NumFeatures) + "_TS_" +
str(args.NumTimeSteps) + ".npy")
TrainingMetaDataset = np.load(args.data_dir +
"SimulatedTrainingMetaData_" +
args.DataName + "_F_" +
str(args.NumFeatures) + "_TS_" +
str(args.NumTimeSteps) + ".npy")
TrainingLabel = TrainingMetaDataset[:, 0]
Testing = np.load(args.data_dir + "SimulatedTestingData_" +
args.DataName + "_F_" +
str(args.NumFeatures) + "_TS_" +
str(args.NumTimeSteps) + ".npy")
TestingDataset_MetaData = np.load(args.data_dir +
"SimulatedTestingMetaData_" +
args.DataName + "_F_" +
str(args.NumFeatures) +
"_TS_" +
str(args.NumTimeSteps) +
".npy")
TestingLabel = TestingDataset_MetaData[:, 0]
Training = Training.reshape(
Training.shape[0], Training.shape[1] * Training.shape[2])
Testing = Testing.reshape(Testing.shape[0],
Testing.shape[1] * Testing.shape[2])
scaler = MinMaxScaler()
scaler.fit(Training)
Training = scaler.transform(Training)
Testing = scaler.transform(Testing)
TrainingRNN = Training.reshape(Training.shape[0],
args.NumTimeSteps,
args.NumFeatures)
TestingRNN = Testing.reshape(Testing.shape[0],
args.NumTimeSteps,
args.NumFeatures)
train_dataRNN = data_utils.TensorDataset(
torch.from_numpy(TrainingRNN),
torch.from_numpy(TrainingLabel))
train_loaderRNN = data_utils.DataLoader(
train_dataRNN, batch_size=args.batch_size, shuffle=True)
test_dataRNN = data_utils.TensorDataset(
torch.from_numpy(TestingRNN),
torch.from_numpy(TestingLabel))
test_loaderRNN = data_utils.DataLoader(
test_dataRNN, batch_size=args.batch_size, shuffle=False)
modelName = "Simulated"
modelName += args.DataName
saveModelName = "../Models/" + models[m] + "/" + modelName
saveModelBestName = saveModelName + "_BEST.pkl"
pretrained_model = torch.load(saveModelBestName,
map_location=device)
Test_Acc = checkAccuracy(test_loaderRNN, pretrained_model,
args)
print('{} {} model BestAcc {:.4f}'.format(
args.DataName, models[m], Test_Acc))
if (Test_Acc >= 90):
if (args.GradFlag):
rescaledGrad = np.zeros((TestingRNN.shape))
Grad = Saliency(pretrained_model)
if (args.IGFlag):
rescaledIG = np.zeros((TestingRNN.shape))
IG = IntegratedGradients(pretrained_model)
if (args.DLFlag):
rescaledDL = np.zeros((TestingRNN.shape))
DL = DeepLift(pretrained_model)
if (args.GSFlag):
rescaledGS = np.zeros((TestingRNN.shape))
GS = GradientShap(pretrained_model)
if (args.DLSFlag):
rescaledDLS = np.zeros((TestingRNN.shape))
DLS = DeepLiftShap(pretrained_model)
if (args.SGFlag):
rescaledSG = np.zeros((TestingRNN.shape))
Grad_ = Saliency(pretrained_model)
SG = NoiseTunnel(Grad_)
if (args.ShapleySamplingFlag):
rescaledShapleySampling = np.zeros((TestingRNN.shape))
SS = ShapleyValueSampling(pretrained_model)
if (args.GSFlag):
rescaledFeaturePermutation = np.zeros(
(TestingRNN.shape))
FP = FeaturePermutation(pretrained_model)
if (args.FeatureAblationFlag):
rescaledFeatureAblation = np.zeros((TestingRNN.shape))
FA = FeatureAblation(pretrained_model)
if (args.OcclusionFlag):
rescaledOcclusion = np.zeros((TestingRNN.shape))
OS = Occlusion(pretrained_model)
idx = 0
mask = np.zeros((args.NumTimeSteps, args.NumFeatures),
dtype=int)
for i in range(args.NumTimeSteps):
mask[i, :] = i
for i, (samples, labels) in enumerate(test_loaderRNN):
print('[{}/{}] {} {} model accuracy {:.2f}'\
.format(i,len(test_loaderRNN), models[m], args.DataName, Test_Acc))
input = samples.reshape(-1, args.NumTimeSteps,
args.NumFeatures).to(device)
input = Variable(input,
volatile=False,
requires_grad=True)
batch_size = input.shape[0]
baseline_single = torch.from_numpy(
np.random.random(input.shape)).to(device)
baseline_multiple = torch.from_numpy(
np.random.random(
(input.shape[0] * 5, input.shape[1],
input.shape[2]))).to(device)
inputMask = np.zeros((input.shape))
inputMask[:, :, :] = mask
inputMask = torch.from_numpy(inputMask).to(device)
mask_single = torch.from_numpy(mask).to(device)
mask_single = mask_single.reshape(
1, args.NumTimeSteps, args.NumFeatures).to(device)
labels = torch.tensor(labels.int().tolist()).to(device)
if (args.GradFlag):
attributions = Grad.attribute(input, \
target=labels)
rescaledGrad[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.IGFlag):
attributions = IG.attribute(input, \
baselines=baseline_single, \
target=labels)
rescaledIG[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.DLFlag):
attributions = DL.attribute(input, \
baselines=baseline_single, \
target=labels)
rescaledDL[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.GSFlag):
attributions = GS.attribute(input, \
baselines=baseline_multiple, \
stdevs=0.09,\
target=labels)
rescaledGS[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.DLSFlag):
attributions = DLS.attribute(input, \
baselines=baseline_multiple, \
target=labels)
rescaledDLS[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.SGFlag):
attributions = SG.attribute(input, \
target=labels)
rescaledSG[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.ShapleySamplingFlag):
attributions = SS.attribute(input, \
baselines=baseline_single, \
target=labels,\
feature_mask=inputMask)
rescaledShapleySampling[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.FeaturePermutationFlag):
attributions = FP.attribute(input, \
target=labels,
perturbations_per_eval= input.shape[0],\
feature_mask=mask_single)
rescaledFeaturePermutation[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.FeatureAblationFlag):
attributions = FA.attribute(input, \
target=labels)
# perturbations_per_eval= input.shape[0],\
# feature_mask=mask_single)
rescaledFeatureAblation[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
if (args.OcclusionFlag):
attributions = OS.attribute(input, \
sliding_window_shapes=(1,args.NumFeatures),
target=labels,
baselines=baseline_single)
rescaledOcclusion[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
args, attributions)
idx += batch_size
if (args.plot):
index = random.randint(0, TestingRNN.shape[0] - 1)
plotExampleBox(TestingRNN[index, :, :],
args.Saliency_Maps_graphs_dir +
args.DataName + "_" + models[m] +
'_sample',
flip=True)
print("Plotting sample", index)
if (args.GradFlag):
plotExampleBox(rescaledGrad[index, :, :],
args.Saliency_Maps_graphs_dir +
args.DataName + "_" + models[m] +
'_Grad',
greyScale=True,
flip=True)
if (args.IGFlag):
plotExampleBox(rescaledIG[index, :, :],
args.Saliency_Maps_graphs_dir +
args.DataName + "_" + models[m] +
'_IG',
greyScale=True,
flip=True)
if (args.DLFlag):
plotExampleBox(rescaledDL[index, :, :],
args.Saliency_Maps_graphs_dir +
args.DataName + "_" + models[m] +
'_DL',
greyScale=True,
flip=True)
if (args.GSFlag):
plotExampleBox(rescaledGS[index, :, :],
args.Saliency_Maps_graphs_dir +
args.DataName + "_" + models[m] +
'_GS',
greyScale=True,
flip=True)
if (args.DLSFlag):
plotExampleBox(rescaledDLS[index, :, :],
args.Saliency_Maps_graphs_dir +
args.DataName + "_" + models[m] +
'_DLS',
greyScale=True,
flip=True)
if (args.SGFlag):
plotExampleBox(rescaledSG[index, :, :],
args.Saliency_Maps_graphs_dir +
args.DataName + "_" + models[m] +
'_SG',
greyScale=True,
flip=True)
if (args.ShapleySamplingFlag):
plotExampleBox(
rescaledShapleySampling[index, :, :],
args.Saliency_Maps_graphs_dir + args.DataName +
"_" + models[m] + '_ShapleySampling',
greyScale=True,
flip=True)
if (args.FeaturePermutationFlag):
plotExampleBox(
rescaledFeaturePermutation[index, :, :],
args.Saliency_Maps_graphs_dir + args.DataName +
"_" + models[m] + '_FeaturePermutation',
greyScale=True,
flip=True)
if (args.FeatureAblationFlag):
plotExampleBox(
rescaledFeatureAblation[index, :, :],
args.Saliency_Maps_graphs_dir + args.DataName +
"_" + models[m] + '_FeatureAblation',
greyScale=True,
flip=True)
if (args.OcclusionFlag):
plotExampleBox(rescaledOcclusion[index, :, :],
args.Saliency_Maps_graphs_dir +
args.DataName + "_" + models[m] +
'_Occlusion',
greyScale=True,
flip=True)
if (args.save):
if (args.GradFlag):
print("Saving Grad", modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_Grad_rescaled", rescaledGrad)
if (args.IGFlag):
print("Saving IG", modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_IG_rescaled", rescaledIG)
if (args.DLFlag):
print("Saving DL", modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_DL_rescaled", rescaledDL)
if (args.GSFlag):
print("Saving GS", modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_GS_rescaled", rescaledGS)
if (args.DLSFlag):
print("Saving DLS", modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_DLS_rescaled", rescaledDLS)
if (args.SGFlag):
print("Saving SG", modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_SG_rescaled", rescaledSG)
if (args.ShapleySamplingFlag):
print("Saving ShapleySampling",
modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_ShapleySampling_rescaled",
rescaledShapleySampling)
if (args.FeaturePermutationFlag):
print("Saving FeaturePermutation",
modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_FeaturePermutation_rescaled",
rescaledFeaturePermutation)
if (args.FeatureAblationFlag):
print("Saving FeatureAblation",
modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_FeatureAblation_rescaled",
rescaledFeatureAblation)
if (args.OcclusionFlag):
print("Saving Occlusion",
modelName + "_" + models[m])
np.save(
args.Saliency_dir + modelName + "_" +
models[m] + "_Occlusion_rescaled",
rescaledOcclusion)
else:
logging.basicConfig(filename=args.log_file,
level=logging.DEBUG)
logging.debug('{} {} model BestAcc {:.4f}'.format(
args.DataName, models[m], Test_Acc))
if not os.path.exists(args.ignore_list):
with open(args.ignore_list, 'w') as fp:
fp.write(args.DataName + '_' + models[m] + '\n')
else:
with open(args.ignore_list, "a") as fp:
fp.write(args.DataName + '_' + models[m] + '\n')
(0.25, '#000000'),
(1, '#000000')],
N=256)
vis_img = viz.visualize_image_attr(
np.transpose(attributions_ig.squeeze().cpu().detach().numpy(),
(1, 2, 0)),
np.transpose(transformed_img.squeeze().cpu().detach().numpy(),
(1, 2, 0)),
method='heat_map',
cmap=default_cmap,
show_colorbar=True,
sign='positive',
outlier_perc=1)
noise_tunnel = NoiseTunnel(integrated_gradients)
attributions_ig_nt = noise_tunnel.attribute(input,
nt_samples=10,
nt_type='smoothgrad_sq',
target=pred_label_idx,
internal_batch_size=10)
_ = viz.visualize_image_attr_multiple(
np.transpose(attributions_ig_nt.squeeze().cpu().detach().numpy(),
(1, 2, 0)),
np.transpose(transformed_img.squeeze().cpu().detach().numpy(),
(1, 2, 0)), ["original_image", "heat_map"],
["all", "positive"],
cmap=default_cmap,
show_colorbar=True)
def visualize_maps(
model: torch.nn.Module,
inputs: Union[Tuple[torch.Tensor, torch.Tensor]],
labels: torch.Tensor,
title: str,
second_occlusion: Tuple[int, int, int] = (1, 2, 2),
baselines: Tuple[int, int] = (0, 0),
closest: bool = False,
) -> None:
"""
Visualizes the average of the inputs, or the single input, using various different XAI approaches
"""
single = inputs[1].ndim == 2
model.zero_grad()
model.eval()
occ = Occlusion(model)
saliency = Saliency(model)
saliency = NoiseTunnel(saliency)
igrad = IntegratedGradients(model)
igrad_2 = NoiseTunnel(igrad)
# deep_lift = DeepLift(model)
grad_shap = ShapleyValueSampling(model)
output = model(inputs[0], inputs[1])
output = F.softmax(output, dim=-1).argmax(dim=1, keepdim=True)
labels = F.softmax(labels, dim=-1).argmax(dim=1, keepdim=True)
if np.all(labels.cpu().numpy() == 1) and not closest:
return
if True:
targets = labels
else:
targets = output
print(targets)
correct = targets.cpu().numpy() == labels.cpu().numpy()
# if correct:
# return
occ_out = occ.attribute(
inputs,
baselines=baselines,
sliding_window_shapes=((1, 5, 5), second_occlusion),
target=targets,
)
# occ_out2 = occ.attribute(inputs, sliding_window_shapes=((1,20,20), second_occlusion), strides=(8,1), target=targets)
saliency_out = saliency.attribute(inputs,
nt_type="smoothgrad_sq",
n_samples=5,
target=targets,
abs=False)
# igrad_out = igrad.attribute(inputs, target=targets, internal_batch_size=1)
igrad_out = igrad_2.attribute(
inputs,
baselines=baselines,
target=targets,
n_samples=5,
nt_type="smoothgrad_sq",
internal_batch_size=1,
)
# deep_lift_out = deep_lift.attribute(inputs, target=targets)
grad_shap_out = grad_shap.attribute(inputs,
baselines=baselines,
target=targets)
if single:
inputs = convert_to_image(inputs)
occ_out = convert_to_image(occ_out)
saliency_out = convert_to_image(saliency_out)
igrad_out = convert_to_image(igrad_out)
# grad_shap_out = convert_to_image(grad_shap_out)
else:
inputs = convert_to_image_multi(inputs)
occ_out = convert_to_image_multi(occ_out)
saliency_out = convert_to_image_multi(saliency_out)
igrad_out = convert_to_image_multi(igrad_out)
grad_shap_out = convert_to_image_multi(grad_shap_out)
fig, axes = plt.subplots(2, 5)
(fig, axes[0, 0]) = visualization.visualize_image_attr(
occ_out[0][0],
inputs[0][0],
title="Original Image",
method="original_image",
show_colorbar=True,
plt_fig_axis=(fig, axes[0, 0]),
use_pyplot=False,
)
(fig, axes[0, 1]) = visualization.visualize_image_attr(
occ_out[0][0],
None,
sign="all",
title="Occ (5x5)",
show_colorbar=True,
plt_fig_axis=(fig, axes[0, 1]),
use_pyplot=False,
)
(fig, axes[0, 2]) = visualization.visualize_image_attr(
saliency_out[0][0],
None,
sign="all",
title="Saliency",
show_colorbar=True,
plt_fig_axis=(fig, axes[0, 2]),
use_pyplot=False,
)
(fig, axes[0, 3]) = visualization.visualize_image_attr(
igrad_out[0][0],
None,
sign="all",
title="Integrated Grad",
show_colorbar=True,
plt_fig_axis=(fig, axes[0, 3]),
use_pyplot=False,
)
(fig, axes[0, 4]) = visualization.visualize_image_attr(
grad_shap_out[0],
None,
title="GradSHAP",
show_colorbar=True,
plt_fig_axis=(fig, axes[0, 4]),
use_pyplot=False,
)
##### Second Input Labels #########################################################################################
(fig, axes[1, 0]) = visualization.visualize_image_attr(
occ_out[1],
inputs[1],
title="Original Aux",
method="original_image",
show_colorbar=True,
plt_fig_axis=(fig, axes[1, 0]),
use_pyplot=False,
)
(fig, axes[1, 1]) = visualization.visualize_image_attr(
occ_out[1],
None,
sign="all",
title="Occ (1x1)",
show_colorbar=True,
plt_fig_axis=(fig, axes[1, 1]),
use_pyplot=False,
)
(fig, axes[1, 2]) = visualization.visualize_image_attr(
saliency_out[1],
None,
sign="all",
title="Saliency",
show_colorbar=True,
plt_fig_axis=(fig, axes[1, 2]),
use_pyplot=False,
)
(fig, axes[1, 3]) = visualization.visualize_image_attr(
igrad_out[1],
None,
sign="all",
title="Integrated Grad",
show_colorbar=True,
plt_fig_axis=(fig, axes[1, 3]),
use_pyplot=False,
)
(fig, axes[1, 4]) = visualization.visualize_image_attr(
grad_shap_out[1],
None,
title="GradSHAP",
show_colorbar=True,
plt_fig_axis=(fig, axes[1, 4]),
use_pyplot=False,
)
fig.suptitle(
title +
f" Label: {labels.cpu().numpy()} Pred: {targets.cpu().numpy()}")
plt.savefig(
f"{title}_{'single' if single else 'multi'}_{'Failed' if correct else 'Success'}_baseline{baselines[0]}.png",
dpi=300,
)
plt.clf()
plt.cla()
pos_neg_file='data/labels_covid19_posi.tsv',
splits=[0.7, 0.15, 0.15],
replicate_channel=1,
batch_size=1,
random_seed=123,
input_size=224,
mode='ct',
num_workers=0)
model = DenseNetModel(121, 2)
model.model.load_state_dict(
torch.load('interpretability/model_best.pth')['state_dict'])
model = model.eval()
model.model = model.model.to('cuda')
gc = GuidedGradCam(model.model, model.model.features.denseblock3)
noise_tunnel = NoiseTunnel(gc)
for (img, label, sublabel, subject_id, ct_id,
slice_id) in tqdm(covid_loaders.test):
img = img.to('cuda')
output = model.model(img)
prediction_score, predicted_label = torch.max(output, 1)
attributions_ig_nt = noise_tunnel.attribute(img,
nt_samples=10,
nt_type='smoothgrad_sq',
target=predicted_label,
stdevs=0.01)
filename = subject_id[0] + '_' + ct_id[0] + '_' + str(
slice_id.item()) + '.png'
dest = os.path.join('interp_output', subject_id[0], ct_id[0], filename)
if not os.path.exists(os.path.dirname(dest)):
[(0, '#ffffff'),
(0.25, '#000000'),
(1, '#000000')],
N=256)
baseline = 2.6400 # ~maximum in the image, white place
for batch_idx, (input, target) in enumerate(dataloaders['test']):
image_input = input.cpu().data[0].numpy().transpose((1, 2, 0))
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
image_input = std * image_input + mean
image_input = np.clip(image_input, 0, 1)
output = model_ft(input)
prediction_score, pred_label_idx = torch.topk(output, 1)
pred_clname = class_names[pred_label_idx]
integrated_gradients = IntegratedGradients(model_ft)
noise_tunnel = NoiseTunnel(integrated_gradients)
attributions_ig_nt = noise_tunnel.attribute(input,
n_samples=10,
nt_type='smoothgrad_sq',
target=pred_label_idx,
baselines=baseline)
pdf = PdfPages(inputdir + "res/image_intepret_" + str(batch_idx) + ".pdf")
fig, axis = viz.visualize_image_attr_multiple(
np.transpose(attributions_ig_nt.squeeze().cpu().detach().numpy(),
(1, 2, 0)),
image_input, ["original_image", "heat_map"], ["all", "positive"],
cmap=default_cmap,
show_colorbar=True,
titles=[
'Truth: ' + class_names[target] + ' Prediction: ' + pred_clname,
'NoiseTunnel'
def run_saliency_methods(saliency_methods,
pretrained_model,
test_shape,
train_loader,
test_loader,
device,
model_type,
model_name,
saliency_dir,
tsr_graph_dir=None,
tsr_inputs_to_graph=()):
_, num_timesteps, num_features = test_shape
run_grad = "Grad" in saliency_methods
run_grad_tsr = "Grad_TSR" in saliency_methods
run_ig = "IG" in saliency_methods
run_ig_tsr = "IG_TSR" in saliency_methods
run_dl = "DL" in saliency_methods
run_gs = "GS" in saliency_methods
run_dls = "DLS" in saliency_methods
run_dls_tsr = "DLS_TSR" in saliency_methods
run_sg = "SG" in saliency_methods
run_shapley_sampling = "ShapleySampling" in saliency_methods
run_feature_permutation = "FeaturePermutation" in saliency_methods
run_feature_ablation = "FeatureAblation" in saliency_methods
run_occlusion = "Occlusion" in saliency_methods
run_fit = "FIT" in saliency_methods
run_ifit = "IFIT" in saliency_methods
run_wfit = "WFIT" in saliency_methods
run_iwfit = "IWFIT" in saliency_methods
if run_grad or run_grad_tsr:
Grad = Saliency(pretrained_model)
if run_grad:
rescaledGrad = np.zeros(test_shape)
if run_grad_tsr:
rescaledGrad_TSR = np.zeros(test_shape)
if run_ig or run_ig_tsr:
IG = IntegratedGradients(pretrained_model)
if run_ig:
rescaledIG = np.zeros(test_shape)
if run_ig_tsr:
rescaledIG_TSR = np.zeros(test_shape)
if run_dl:
rescaledDL = np.zeros(test_shape)
DL = DeepLift(pretrained_model)
if run_gs:
rescaledGS = np.zeros(test_shape)
GS = GradientShap(pretrained_model)
if run_dls or run_dls_tsr:
DLS = DeepLiftShap(pretrained_model)
if run_dls:
rescaledDLS = np.zeros(test_shape)
if run_dls_tsr:
rescaledDLS_TSR = np.zeros(test_shape)
if run_sg:
rescaledSG = np.zeros(test_shape)
Grad_ = Saliency(pretrained_model)
SG = NoiseTunnel(Grad_)
if run_shapley_sampling:
rescaledShapleySampling = np.zeros(test_shape)
SS = ShapleyValueSampling(pretrained_model)
if run_gs:
rescaledFeaturePermutation = np.zeros(test_shape)
FP = FeaturePermutation(pretrained_model)
if run_feature_ablation:
rescaledFeatureAblation = np.zeros(test_shape)
FA = FeatureAblation(pretrained_model)
if run_occlusion:
rescaledOcclusion = np.zeros(test_shape)
OS = Occlusion(pretrained_model)
if run_fit:
rescaledFIT = np.zeros(test_shape)
FIT = FITExplainer(pretrained_model, ft_dim_last=True)
generator = JointFeatureGenerator(num_features, data='none')
# TODO: Increase epochs
FIT.fit_generator(generator, train_loader, test_loader, n_epochs=300)
if run_ifit:
rescaledIFIT = np.zeros(test_shape)
if run_wfit:
rescaledWFIT = np.zeros(test_shape)
if run_iwfit:
rescaledIWFIT = np.zeros(test_shape)
idx = 0
mask = np.zeros((num_timesteps, num_features), dtype=int)
for i in range(num_timesteps):
mask[i, :] = i
for i, (samples, labels) in enumerate(test_loader):
input = samples.reshape(-1, num_timesteps, num_features).to(device)
input = Variable(input, volatile=False, requires_grad=True)
batch_size = input.shape[0]
baseline_single = torch.from_numpy(np.random.random(
input.shape)).to(device)
baseline_multiple = torch.from_numpy(
np.random.random((input.shape[0] * 5, input.shape[1],
input.shape[2]))).to(device)
inputMask = np.zeros((input.shape))
inputMask[:, :, :] = mask
inputMask = torch.from_numpy(inputMask).to(device)
mask_single = torch.from_numpy(mask).to(device)
mask_single = mask_single.reshape(1, num_timesteps,
num_features).to(device)
labels = torch.tensor(labels.int().tolist()).to(device)
if run_grad:
attributions = Grad.attribute(input, target=labels)
rescaledGrad[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_grad_tsr:
rescaledGrad_TSR[idx:idx + batch_size, :, :] = get_tsr_saliency(
Grad,
input,
labels,
graph_dir=tsr_graph_dir,
graph_name=f'{model_name}_{model_type}_Grad_TSR',
inputs_to_graph=tsr_inputs_to_graph,
cur_batch=i)
if run_ig:
attributions = IG.attribute(input,
baselines=baseline_single,
target=labels)
rescaledIG[idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_ig_tsr:
rescaledIG_TSR[idx:idx + batch_size, :, :] = get_tsr_saliency(
IG,
input,
labels,
baseline=baseline_single,
graph_dir=tsr_graph_dir,
graph_name=f'{model_name}_{model_type}_IG_TSR',
inputs_to_graph=tsr_inputs_to_graph,
cur_batch=i)
if run_dl:
attributions = DL.attribute(input,
baselines=baseline_single,
target=labels)
rescaledDL[idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_gs:
attributions = GS.attribute(input,
baselines=baseline_multiple,
stdevs=0.09,
target=labels)
rescaledGS[idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_dls:
attributions = DLS.attribute(input,
baselines=baseline_multiple,
target=labels)
rescaledDLS[idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_dls_tsr:
rescaledDLS_TSR[idx:idx + batch_size, :, :] = get_tsr_saliency(
DLS,
input,
labels,
baseline=baseline_multiple,
graph_dir=tsr_graph_dir,
graph_name=f'{model_name}_{model_type}_DLS_TSR',
inputs_to_graph=tsr_inputs_to_graph,
cur_batch=i)
if run_sg:
attributions = SG.attribute(input, target=labels)
rescaledSG[idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_shapley_sampling:
attributions = SS.attribute(input,
baselines=baseline_single,
target=labels,
feature_mask=inputMask)
rescaledShapleySampling[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_feature_permutation:
attributions = FP.attribute(input,
target=labels,
perturbations_per_eval=input.shape[0],
feature_mask=mask_single)
rescaledFeaturePermutation[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_feature_ablation:
attributions = FA.attribute(input, target=labels)
# perturbations_per_eval= input.shape[0],\
# feature_mask=mask_single)
rescaledFeatureAblation[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_occlusion:
attributions = OS.attribute(input,
sliding_window_shapes=(1,
num_features),
target=labels,
baselines=baseline_single)
rescaledOcclusion[
idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_fit:
attributions = torch.from_numpy(FIT.attribute(input, labels))
rescaledFIT[idx:idx +
batch_size, :, :] = Helper.givenAttGetRescaledSaliency(
num_timesteps, num_features, attributions)
if run_ifit:
attributions = torch.from_numpy(
inverse_fit_attribute(input,
pretrained_model,
ft_dim_last=True))
rescaledIFIT[idx:idx + batch_size, :, :] = attributions
if run_wfit:
attributions = torch.from_numpy(
wfit_attribute(input,
pretrained_model,
N=test_shape[1],
ft_dim_last=True,
single_label=True))
rescaledWFIT[idx:idx + batch_size, :, :] = attributions
if run_iwfit:
attributions = torch.from_numpy(
wfit_attribute(input,
pretrained_model,
N=test_shape[1],
ft_dim_last=True,
single_label=True,
inverse=True))
rescaledIWFIT[idx:idx + batch_size, :, :] = attributions
idx += batch_size
if run_grad:
print("Saving Grad", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type + "_Grad_rescaled",
rescaledGrad)
if run_grad_tsr:
print("Saving Grad_TSR", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type +
"_Grad_TSR_rescaled", rescaledGrad_TSR)
if run_ig:
print("Saving IG", model_name + "_" + model_type)
np.save(saliency_dir + model_name + "_" + model_type + "_IG_rescaled",
rescaledIG)
if run_ig_tsr:
print("Saving IG_TSR", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type + "_IG_TSR_rescaled",
rescaledIG_TSR)
if run_dl:
print("Saving DL", model_name + "_" + model_type)
np.save(saliency_dir + model_name + "_" + model_type + "_DL_rescaled",
rescaledDL)
if run_gs:
print("Saving GS", model_name + "_" + model_type)
np.save(saliency_dir + model_name + "_" + model_type + "_GS_rescaled",
rescaledGS)
if run_dls:
print("Saving DLS", model_name + "_" + model_type)
np.save(saliency_dir + model_name + "_" + model_type + "_DLS_rescaled",
rescaledDLS)
if run_dls_tsr:
print("Saving DLS_TSR", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type + "_DLS_TSR_rescaled",
rescaledDLS_TSR)
if run_sg:
print("Saving SG", model_name + "_" + model_type)
np.save(saliency_dir + model_name + "_" + model_type + "_SG_rescaled",
rescaledSG)
if run_shapley_sampling:
print("Saving ShapleySampling", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type +
"_ShapleySampling_rescaled", rescaledShapleySampling)
if run_feature_permutation:
print("Saving FeaturePermutation", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type +
"_FeaturePermutation_rescaled", rescaledFeaturePermutation)
if run_feature_ablation:
print("Saving FeatureAblation", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type +
"_FeatureAblation_rescaled", rescaledFeatureAblation)
if run_occlusion:
print("Saving Occlusion", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type +
"_Occlusion_rescaled", rescaledOcclusion)
if run_fit:
print("Saving FIT", model_name + "_" + model_type)
np.save(saliency_dir + model_name + "_" + model_type + "_FIT_rescaled",
rescaledFIT)
if run_ifit:
print("Saving IFIT", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type + "_IFIT_rescaled",
rescaledIFIT)
if run_wfit:
print("Saving WFIT", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type + "_WFIT_rescaled",
rescaledWFIT)
if run_iwfit:
print("Saving IWFIT", model_name + "_" + model_type)
np.save(
saliency_dir + model_name + "_" + model_type + "_IWFIT_rescaled",
rescaledIWFIT)
def after_fit(self):
self.integrated_gradients = IntegratedGradients(self.model)
self._noise_tunnel= NoiseTunnel(self.integrated_gradients)
def __init__(self, model, train_data):
model.eval()
self.explainer = NoiseTunnel(IntegratedGradients(model))
self.model = model
global_vars.set('explainer_sampling_rate', 0.1)
def measure_filter_model(
model_version,
dataset,
out_folder,
weights_dir,
device,
method=METHODS["gradcam"],
sample_images=50,
step=1,
use_infidelity=False,
use_sensitivity=False,
render=False,
ids=None,
):
invTrans = get_inverse_normalization_transformation()
data_dir = os.path.join("data")
if model_version == "resnet18":
model = create_resnet18_model(num_of_classes=NUM_OF_CLASSES[dataset])
elif model_version == "resnet50":
model = create_resnet50_model(num_of_classes=NUM_OF_CLASSES[dataset])
elif model_version == "densenet":
model = create_densenet121_model(
num_of_classes=NUM_OF_CLASSES[dataset])
else:
model = create_efficientnetb0_model(
num_of_classes=NUM_OF_CLASSES[dataset])
model.load_state_dict(torch.load(weights_dir))
# print(model)
model.eval()
model.to(device)
test_dataset = CustomDataset(
dataset=dataset,
transformer=get_default_transformation(),
data_type="test",
root_dir=data_dir,
step=step,
add_filters=True,
ids=ids,
)
data_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=1,
shuffle=False,
num_workers=4)
try:
image_ids = random.sample(range(0, test_dataset.__len__()),
test_dataset.__len__())
except ValueError:
raise ValueError(
f"Image sample number ({test_dataset.__len__()}) exceeded dataset size ({test_dataset.__len__()})."
)
classes_map = test_dataset.classes_map
print(f"Measuring {model_version} on {dataset} dataset, with {method}")
print("-" * 10)
pbar = tqdm(total=test_dataset.__len__(), desc="Model test completion")
multipy_by_inputs = False
if method == METHODS["ig"]:
attr_method = IntegratedGradients(model)
nt_samples = 1
n_perturb_samples = 1
if method == METHODS["saliency"]:
attr_method = Saliency(model)
nt_samples = 8
n_perturb_samples = 2
if method == METHODS["gradcam"]:
if model_version == "efficientnet":
attr_method = GuidedGradCam(model, model._conv_stem)
elif model_version == "densenet":
attr_method = GuidedGradCam(model, model.features.conv0)
else:
attr_method = GuidedGradCam(model, model.conv1)
nt_samples = 8
n_perturb_samples = 2
if method == METHODS["deconv"]:
attr_method = Deconvolution(model)
nt_samples = 8
n_perturb_samples = 2
if method == METHODS["gradshap"]:
attr_method = GradientShap(model)
nt_samples = 8
n_perturb_samples = 2
if method == METHODS["gbp"]:
attr_method = GuidedBackprop(model)
nt_samples = 8
n_perturb_samples = 2
if method == "lime":
attr_method = Lime(model)
nt_samples = 8
n_perturb_samples = 2
feature_mask = torch.tensor(lime_mask).to(device)
multipy_by_inputs = True
if method == METHODS["ig"]:
nt = attr_method
else:
nt = NoiseTunnel(attr_method)
scores = []
@infidelity_perturb_func_decorator(multipy_by_inputs=multipy_by_inputs)
def perturb_fn(inputs):
noise = torch.tensor(np.random.normal(0, 0.003, inputs.shape)).float()
noise = noise.to(device)
return inputs - noise
OUR_FILTERS = [
"none",
"fx_freaky_details 2,10,1,11,0,32,0",
"normalize_local 8,10",
"fx_boost_chroma 90,0,0",
"fx_mighty_details 25,1,25,1,11,0",
"sharpen 300",
]
idx = 0
filter_count = 0
filter_attrs = {filter_name: [] for filter_name in OUR_FILTERS}
predicted_main_class = 0
for input, label in data_loader:
pbar.update(1)
inv_input = invTrans(input)
input = input.to(device)
input.requires_grad = True
output = model(input)
output = F.softmax(output, dim=1)
prediction_score, pred_label_idx = torch.topk(output, 1)
prediction_score = prediction_score.cpu().detach().numpy()[0][0]
pred_label_idx.squeeze_()
if OUR_FILTERS[filter_count] == 'none':
predicted_main_class = pred_label_idx.item()
if method == METHODS["gradshap"]:
baseline = torch.randn(input.shape)
baseline = baseline.to(device)
if method == "lime":
attributions = attr_method.attribute(input, target=1, n_samples=50)
elif method == METHODS["ig"]:
attributions = nt.attribute(
input,
target=predicted_main_class,
n_steps=25,
)
elif method == METHODS["gradshap"]:
attributions = nt.attribute(input,
target=predicted_main_class,
baselines=baseline)
else:
attributions = nt.attribute(
input,
nt_type="smoothgrad",
nt_samples=nt_samples,
target=predicted_main_class,
)
if use_infidelity:
infid = infidelity(model,
perturb_fn,
input,
attributions,
target=predicted_main_class)
inf_value = infid.cpu().detach().numpy()[0]
else:
inf_value = 0
if use_sensitivity:
if method == "lime":
sens = sensitivity_max(
attr_method.attribute,
input,
target=predicted_main_class,
n_perturb_samples=1,
n_samples=200,
feature_mask=feature_mask,
)
elif method == METHODS["ig"]:
sens = sensitivity_max(
nt.attribute,
input,
target=predicted_main_class,
n_perturb_samples=n_perturb_samples,
n_steps=25,
)
elif method == METHODS["gradshap"]:
sens = sensitivity_max(
nt.attribute,
input,
target=predicted_main_class,
n_perturb_samples=n_perturb_samples,
baselines=baseline,
)
else:
sens = sensitivity_max(
nt.attribute,
input,
target=predicted_main_class,
n_perturb_samples=n_perturb_samples,
)
sens_value = sens.cpu().detach().numpy()[0]
else:
sens_value = 0
# filter_name = test_dataset.data.iloc[pbar.n]["filter"].split(" ")[0]
attr_data = attributions.squeeze().cpu().detach().numpy()
if render:
fig, ax = viz.visualize_image_attr_multiple(
np.transpose(attr_data, (1, 2, 0)),
np.transpose(inv_input.squeeze().cpu().detach().numpy(),
(1, 2, 0)),
["original_image", "heat_map"],
["all", "positive"],
titles=["original_image", "heat_map"],
cmap=default_cmap,
show_colorbar=True,
use_pyplot=False,
fig_size=(8, 6),
)
if use_sensitivity or use_infidelity:
ax[0].set_xlabel(
f"Infidelity: {'{0:.6f}'.format(inf_value)}\n Sensitivity: {'{0:.6f}'.format(sens_value)}"
)
fig.suptitle(
f"True: {classes_map[str(label.numpy()[0])][0]}, Pred: {classes_map[str(pred_label_idx.item())][0]}\nScore: {'{0:.4f}'.format(prediction_score)}",
fontsize=16,
)
fig.savefig(
os.path.join(
out_folder,
f"{str(idx)}-{str(filter_count)}-{str(label.numpy()[0])}-{str(OUR_FILTERS[filter_count])}-{classes_map[str(label.numpy()[0])][0]}-{classes_map[str(pred_label_idx.item())][0]}.png",
))
plt.close(fig)
# if pbar.n > 25:
# break
score_for_true_label = output.cpu().detach().numpy(
)[0][predicted_main_class]
filter_attrs[OUR_FILTERS[filter_count]] = [
np.moveaxis(attr_data, 0, -1),
"{0:.8f}".format(score_for_true_label),
]
data_range_for_current_set = MAX_ATT_VALUES[model_version][method][
dataset]
filter_count += 1
if filter_count >= len(OUR_FILTERS):
ssims = []
for rot in OUR_FILTERS:
ssims.append("{0:.8f}".format(
ssim(
filter_attrs["none"][0],
filter_attrs[rot][0],
win_size=11,
data_range=data_range_for_current_set,
multichannel=True,
)))
ssims.append(filter_attrs[rot][1])
scores.append(ssims)
filter_count = 0
predicted_main_class = 0
idx += 1
pbar.close()
indexes = []
for filter_name in OUR_FILTERS:
indexes.append(str(filter_name) + "-ssim")
indexes.append(str(filter_name) + "-score")
np.savetxt(
os.path.join(
out_folder,
f"{model_version}-{dataset}-{method}-ssim-with-range.csv"),
np.array(scores),
delimiter=";",
fmt="%s",
header=";".join([str(rot) for rot in indexes]),
)
print(f"Artifacts stored at {out_folder}")
gf = GaussianFilter(2.0)
ig = IntegratedGradients(net)
attr_ig, delta = attribute_image_features(ig,
input,
n_steps=50,
baselines=gf(input),
return_convergence_delta=True)
attr_ig = np.transpose(attr_ig.squeeze().cpu().detach().numpy(), (1, 2, 0))
print('Approximation delta: ', abs(delta))
# %% [markdown]
# Below we demonstrate how to use integrated gradients and noise tunnel with smoothgrad square option on the test image. Noise tunnel with `smoothgrad square` option adds gaussian noise with a standard deviation of `stdevs=0.2` to the input image `n_samples` times, computes the attributions for `n_samples` images and returns the mean of the squared attributions across `n_samples` images.
# %%
ig = IntegratedGradients(net)
nt = NoiseTunnel(ig)
attr_ig_nt = attribute_image_features(nt,
input,
baselines=input * 0,
nt_type='smoothgrad_sq',
n_samples=100,
stdevs=0.2)
attr_ig_nt = np.transpose(
attr_ig_nt.squeeze(0).cpu().detach().numpy(), (1, 2, 0))
# %% [markdown]
# Applies DeepLift on test image. Deeplift assigns attributions to each input pixel by looking at the differences of output and its reference in terms of the differences of the input from the reference.
# %%
dl = DeepLift(net)
attr_dl = attribute_image_features(dl, input, baselines=input * 0)