def attribute_sampled_shapley_values(text_input_ids: torch.Tensor, target: int,
model: BertForSequenceClassification,
settings: Dict[str, Any],
**kwargs) -> Tuple[np.ndarray, None]:
def forward(model_input):
pred = model(model_input)
return torch.softmax(pred[0], dim=1)
svs = ShapleyValueSampling(forward)
attributions = svs.attribute(inputs=text_input_ids,
target=target,
n_samples=settings.get("n_samples", 25))
scores = attributions.cpu().detach().numpy().flatten()
return scores, None
def get_shap_attributions(text, tech_tv, tech_bb, true_label):
try:
model, tokenizer = get_model(settings.MODEL_NAME, settings.MODEL_STAGE)
except Exception as e:
logger.info(f"wtf is going on here: {e}")
ref_token_id = tokenizer.pad_token_id # A token used for generating token reference
sep_token_id = (
tokenizer.sep_token_id
) # A token used as a separator between question and text and it is also added to the end of the text.
cls_token_id = tokenizer.cls_token_id # A token used for prepending to the concatenated question-text word sequence
bert_string = data_utils.stitch_bert_string("", text, tech_tv, tech_bb)
input_ids, ref_input_ids, _ = construct_input_ref_pair(tokenizer, bert_string, ref_token_id, sep_token_id, cls_token_id)
indices = input_ids[0].detach().tolist()
all_tokens = tokenizer.convert_ids_to_tokens(indices)
pred = model(input_ids)[0]
pred_proba = torch.softmax(pred, dim=1)[0]
model_custom = ModelWrapper(model)
shap = ShapleyValueSampling(model_custom.custom_forward)
attributions = shap.attribute(
inputs=input_ids,
baselines=ref_input_ids,
target=torch.argmax(pred[0]),
)
score_vis = viz.VisualizationDataRecord(
attributions[0, :],
torch.softmax(pred, dim=1)[0][torch.argmax(pred[0]).cpu().numpy().item()],
model.config.id2label[torch.argmax(pred[0]).cpu().numpy().item()],
true_label,
model.config.id2label[torch.argmax(pred[0]).cpu().numpy().item()],
attributions.sum(),
all_tokens,
0,
)
labels = list(model.config.id2label.values())
return score_vis, pred_proba, labels
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 extract_SV(self, X_test):
Sv = ShapleyValueSampling(self.net)
start = time.time()
sv_attr_test = Sv.attribute(X_test.to(self.device))
print("temps train", time.time() - start)
return sv_attr_test.detach().cpu().numpy()
# heloc_checkpoint = "cchvae/classifier-heloc.ckpt"
heloc_checkpoint = "checkpoints/stl-21-04-21-08-11-33/heloc-epoch=196-loss_validate=0.37-top-validate.ckpt"
heloc = pd.read_csv("data/heloc/heloc_dataset_v1_pruned.csv", index_col=False)
# print(heloc)
heloc_c = pd.read_csv("data/heloc/counterfactuals.csv",
header=None,
index_col=False)
amplified_heloc_c = pd.read_csv("data/heloc/augmented_counterfactuals.csv",
index_col=False)
heloc_t = heloc["RiskPerformance"].map(lambda target: 1
if target == "Good" else 0)
heloc["RiskPerformance"] = heloc_t
classifier = SingleTaskLearner.load_from_checkpoint(heloc_checkpoint)
sv = ShapleyValueSampling(classifier)
heloc_shapley_values = []
labels = Labels(heloc.to_numpy())
targets, predictors = split_normalized(labels.labels)
with tqdm(total=len(targets)) as progress_bar:
for predictor in predictors.values:
inp = torch.tensor(predictor, dtype=torch.float32).unsqueeze(0)
attributions = sv.attribute(inp, target=0)
heloc_shapley_values.append(attributions)
progress_bar.update(1)
df = DataFrame(heloc_shapley_values, index=None)
df.to_csv("data/heloc/shapley_heloc_pruned.csv")
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 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()
def generate_saliency(model_path, saliency_path):
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_args = Namespace(**checkpoint['args'])
model_args.batch_size = args.batch_size if args.batch_size != None else \
model_args.batch_size
if args.model == 'transformer':
transformer_config = BertConfig.from_pretrained(
'bert-base-uncased', num_labels=model_args.labels)
modelb = BertForSequenceClassification.from_pretrained(
'bert-base-uncased', config=transformer_config).to(device)
modelb.load_state_dict(checkpoint['model'])
model = BertModelWrapper(modelb)
elif args.model == 'lstm':
model = LSTM_MODEL(tokenizer,
model_args,
n_labels=checkpoint['args']['labels'],
device=device).to(device)
model.load_state_dict(checkpoint['model'])
model.train()
model = ModelWrapper(model)
else:
# model_args.batch_size = 1000
model = CNN_MODEL(tokenizer,
model_args,
n_labels=checkpoint['args']['labels']).to(device)
model.load_state_dict(checkpoint['model'])
model.train()
model = ModelWrapper(model)
ablator = ShapleyValueSampling(model)
coll_call = get_collate_fn(dataset=args.dataset, model=args.model)
collate_fn = partial(coll_call,
tokenizer=tokenizer,
device=device,
return_attention_masks=False,
pad_to_max_length=False)
test = get_dataset(args.dataset_dir, mode=args.split)
test_dl = DataLoader(batch_size=model_args.batch_size,
dataset=test,
shuffle=False,
collate_fn=collate_fn)
# PREDICTIONS
predictions_path = model_path + '.predictions'
if not os.path.exists(predictions_path):
predictions = defaultdict(lambda: [])
for batch in tqdm(test_dl, desc='Running test prediction... '):
logits = model(batch[0])
logits = logits.detach().cpu().numpy().tolist()
predicted = np.argmax(np.array(logits), axis=-1)
predictions['class'] += predicted.tolist()
predictions['logits'] += logits
with open(predictions_path, 'w') as out:
json.dump(predictions, out)
# COMPUTE SALIENCY
saliency_flops = []
with open(saliency_path, 'w') as out_mean:
for batch in tqdm(test_dl, desc='Running Saliency Generation...'):
class_attr_list = defaultdict(lambda: [])
if args.model == 'rnn':
additional = batch[-1]
else:
additional = None
if not args.no_time:
high.start_counters([events.PAPI_FP_OPS])
token_ids = batch[0].detach().cpu().numpy().tolist()
for cls_ in range(args.labels):
attributions = ablator.attribute(
batch[0].float(),
target=cls_,
additional_forward_args=additional)
attributions = attributions.detach().cpu().numpy().tolist()
class_attr_list[cls_] += attributions
if not args.no_time:
x = sum(high.stop_counters())
saliency_flops.append(x / batch[0].shape[0])
for i in range(len(batch[0])):
saliencies = []
for token_i, token_id in enumerate(token_ids[i]):
if token_id == tokenizer.pad_token_id:
continue
token_sal = {'token': tokenizer.ids_to_tokens[token_id]}
for cls_ in range(args.labels):
token_sal[int(
cls_)] = class_attr_list[cls_][i][token_i]
saliencies.append(token_sal)
out_mean.write(json.dumps({'tokens': saliencies}) + '\n')
out_mean.flush()
return saliency_flops
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')