Page MenuHomec4science
Files F77460548

Saliency_Utils.py
No OneTemporary
Actions

File Metadata

Created
Wed, Aug 14, 15:11

Saliency_Utils.py
View Options

# -*- coding: utf-8 -*-
"""
Created on Sat Feb 8 22:10:18 2020
---------------------------------------------------------------------
-- Author: Vigneashwara Pandiyan
---------------------------------------------------------------------
Utils file for saliency
"""
#%%
import torch
import matplotlib.pyplot as plt
from mlxtend.plotting import plot_confusion_matrix
import numpy as np
from Data_Manipulation import *
# device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
import pandas as pd
# print(device)
#%% Loading the data for saliency model
def dataloading_funtion_saliency(folder,window):
# folder='data'
rawfile_1 = str(folder)+'/'+'Channel0_'+str(window)+'.npy'
rawfile_2 = str(folder)+'/'+'Channel1_'+str(window)+'.npy'
rawfile_3 = str(folder)+'/'+'Channel2_'+str(window)+'.npy'
rawfile_4 = str(folder)+'/'+'Channel3_'+str(window)+'.npy'
classfile = str(folder)+'/'+'classspace_'+str(window)+'.npy'
Rawspace_1 = normalize(rawfile_1).astype(np.float64)
Rawspace_2 = normalize(rawfile_2).astype(np.float64)
Rawspace_3 = normalize(rawfile_3).astype(np.float64)
Rawspace_4 = normalize(rawfile_4).astype(np.float64)
classspace= np.load(classfile).astype(np.float64)
trainset,testset=data_batch_prep_saliency(Rawspace_1,Rawspace_2,Rawspace_3,Rawspace_4,classspace)
return testset
#%%
def data_batch_prep_saliency(Rawspace_1,Rawspace_2,Rawspace_3,Rawspace_4,classspace):
sequences_batch =[]
for i in range(len(classspace)):
# print(i)
sequence_features_1 = Rawspace_1[i]
# sequence_features_1 = normalize(sequence_features_1)
sequence_features_2 = Rawspace_2[i]
# sequence_features_2 = normalize(sequence_features_2)
sequence_features_3 = Rawspace_3[i]
# sequence_features_3 = normalize(sequence_features_3)
sequence_features_4 = Rawspace_4[i]
# sequence_features_4 = normalize(sequence_features_4)
label = classspace[i]
sequences_batch.append((sequence_features_1,sequence_features_2,sequence_features_3,sequence_features_4,label))
sequences_batch = Mechanism(sequences_batch)
train, test = train_test_split(sequences_batch, test_size=0.3, random_state=42)
trainset = torch.utils.data.DataLoader(train, batch_size=200, num_workers=0,
shuffle=True)
testset = torch.utils.data.DataLoader(test, batch_size=1, num_workers=0,
shuffle=True)
return trainset, testset
#%%
def compute_saliency_time(input_sample, model, relative=False):
# Prepare the input
input_sample.requires_grad_()
if (torch.cuda.is_available()):
input_sample = input_sample.cuda()
# Collect output from the model
output = model(input_sample)
#print(output.shape)
# Require the gradient
output.requires_grad_()
# Collect the unit responsible for the classification
output_max = output.max(1)[0]
#print(output_max.shape)
# Retain grads
output_max.retain_grad()
input_sample.retain_grad()
# print(output_max)
# Compute the gradients
output_max.backward()
# Collect gradient
grad = input_sample.grad
# slc = (grad - grad.min())/(grad.max()-grad.min())
# Compute abs value
slc = torch.abs(grad)
if relative:
eps = 1e-5
slc = slc/(torch.abs(input_sample)+eps)
saliency = slc.detach()
input_sample.grad = torch.zeros_like(grad)
return saliency
def window_saliency_results(testset,model,device,window):
y_pred = []
y_true = []
saliencies = []
model.eval()
count = 0
# iterate over test data
for batches in testset:
data,output = batches
data,output = data.to(device,dtype=torch.float),output.to(device,dtype=torch.long)
output = output.squeeze()
if count%100==0:
print('Computing prediction for sample: ', count)
prediction = model(data)
prediction = torch.argmax(prediction, dim=1)
prediction = prediction.data.cpu().numpy()
output = output.data.cpu().numpy()
y_true.append(output) # Save Truth
y_pred.append(prediction) # Save Prediction
if count%100==0:
print('Computing saliency for sample: ', count)
saliency = compute_saliency_time(data, model, relative=True)
saliencies.append(saliency)
count += 1
# classes = ('LoF pores', 'Conduction mode', 'Keyhole pores')
# plotname= 'CNN_LSTM_MultivariateNormRel_'+str(window)+'_confusion_matrix'+'.png'
# plt.figure()
# plot_confusion_matrix(y_true, y_pred,classes,plotname)
# torch.save(y_true, 'y_trueNormRel'+str(window))
# torch.save(y_pred, 'y_predNormRel'+str(window))
torch.save(y_true, 'data/y_trueNorm'+str(window))
torch.save(y_pred, 'data/y_predNorm'+str(window))
torch.save(saliencies, 'data/salienciesNorm'+str(window))
return y_true,y_pred,saliencies
def compute_normalize_window(saliences,testset,folder,window):
medianAmpl = []
for datum in testset:
singleDatum, _ = datum
medianAmpl.append(torch.median(torch.abs(singleDatum), dim=2, keepdim=True)[0])
medianAmpl = torch.cat(medianAmpl, dim=0)
medianAmpl.shape
# print(medianAmpl.shape)
# print(saliences.shape)
saliences = saliences/medianAmpl
salienceCat = torch.median(saliences, dim=2, keepdim=False)[0]
salienceCat
salienceCat.shape
title=str(folder)+'/'+'saliencies'+str(window)+'PerCat'
torch.save(salienceCat, title)
return salienceCat

Event Timeline

c4science · Help