How to integrate LIME with PyTorch? - pytorch

Question

Using this mnist image classification model :
%reset -f
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
import torch.utils.data as data_utils
import numpy as np
import matplotlib.pyplot as plt
from sklearn.datasets import make_moons
from matplotlib import pyplot
from pandas import DataFrame
import torchvision.datasets as dset
import os
import torch.nn.functional as F
import time
import random
import pickle
from sklearn.metrics import confusion_matrix
import pandas as pd
import sklearn
trans = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
root = './data'
if not os.path.exists(root):
os.mkdir(root)
train_set = dset.MNIST(root=root, train=True, transform=trans, download=True)
test_set = dset.MNIST(root=root, train=False, transform=trans, download=True)
batch_size = 64
train_loader = torch.utils.data.DataLoader(
dataset=train_set,
batch_size=batch_size,
shuffle=True)
test_loader = torch.utils.data.DataLoader(
dataset=test_set,
batch_size=batch_size,
shuffle=True)
class NeuralNet(nn.Module):
def __init__(self):
super(NeuralNet, self).__init__()
self.fc1 = nn.Linear(28*28, 500)
self.fc2 = nn.Linear(500, 256)
self.fc3 = nn.Linear(256, 2)
def forward(self, x):
x = x.view(-1, 28*28)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
num_epochs = 2
random_sample_size = 200
values_0_or_1 = [t for t in train_set if (int(t[1]) == 0 or int(t[1]) == 1)]
values_0_or_1_testset = [t for t in test_set if (int(t[1]) == 0 or int(t[1]) == 1)]
print(len(values_0_or_1))
print(len(values_0_or_1_testset))
train_loader_subset = torch.utils.data.DataLoader(
dataset=values_0_or_1,
batch_size=batch_size,
shuffle=True)
test_loader_subset = torch.utils.data.DataLoader(
dataset=values_0_or_1_testset,
batch_size=batch_size,
shuffle=False)
train_loader = train_loader_subset
# Hyper-parameters
input_size = 100
hidden_size = 100
num_classes = 2
# learning_rate = 0.00001
learning_rate = .0001
# Device configuration
device = 'cpu'
print_progress_every_n_epochs = 1
model = NeuralNet().to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
N = len(train_loader)
# Train the model
total_step = len(train_loader)
most_recent_prediction = []
test_actual_predicted_dict = {}
rm = random.sample(list(values_0_or_1), random_sample_size)
train_loader_subset = data_utils.DataLoader(rm, batch_size=4)
for epoch in range(num_epochs):
for i, (images, labels) in enumerate(train_loader_subset):
# Move tensors to the configured device
images = images.reshape(-1, 2).to(device)
labels = labels.to(device)
# Forward pass
outputs = model(images)
loss = criterion(outputs, labels)
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (epoch) % print_progress_every_n_epochs == 0:
print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'.format(epoch+1, num_epochs, i+1, total_step, loss.item()))
predicted_test = []
model.eval() # eval mode (batchnorm uses moving mean/variance instead of mini-batch mean/variance)
probs_l = []
predicted_values = []
actual_values = []
labels_l = []
with torch.no_grad():
for images, labels in test_loader_subset:
images = images.to(device)
labels = labels.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
predicted_test.append(predicted.cpu().numpy())
sm = torch.nn.Softmax()
probabilities = sm(outputs)
probs_l.append(probabilities)
labels_l.append(labels.cpu().numpy())
predicted_values.append(np.concatenate(predicted_test).ravel())
actual_values.append(np.concatenate(labels_l).ravel())
if (epoch) % 1 == 0:
print('test accuracy : ', 100 * len((np.where(np.array(predicted_values[0])==(np.array(actual_values[0])))[0])) / len(actual_values[0]))
I'm to attempting to integrate 'Local Interpretable Model-Agnostic Explanations for machine learning classifiers' : https://marcotcr.github.io/lime/
It appears PyTorch support is not enabled as it is not mentioned in doc and following tutorial :
https://marcotcr.github.io/lime/tutorials/Tutorial%20-%20images.html
With my updated code for PyTorch :
from lime import lime_image
import time
explainer = lime_image.LimeImageExplainer()
explanation = explainer.explain_instance(images[0].reshape(28,28), model(images[0]), top_labels=5, hide_color=0, num_samples=1000)
Causes error :
/opt/conda/lib/python3.6/site-packages/skimage/color/colorconv.py in gray2rgb(image, alpha)
830 is_rgb = False
831 is_alpha = False
--> 832 dims = np.squeeze(image).ndim
833
834 if dims == 3:
AttributeError: 'Tensor' object has no attribute 'ndim'
So appears tensorflow object is expected here ?
How to integrate LIME with PyTorch image classification ?

Answer 1

Here's my solution:
Lime expects an image input of type numpy. This is why you get the attribute error and a solution would be to convert the image (from Tensor) to numpy before passing it to the explainer object. Another solution would be to select a specific image with the test_loader_subset and convert it with img = img.numpy().
Secondly, in order to make LIME work with pytorch (or any other framework), you'll need to specify a batch prediction function which outputs the prediction scores of each class for each image. The name of this function (here I've called it batch_predict) is then passed to explainer.explain_instance(img, batch_predict, ...). The batch_predict needs to loop through all images passed to it, convert them to Tensor, make a prediction and finally return the prediction score list (with numpy values). This is how I got it working.
Note also that the images need to have shape (... ,... ,3) or (... ,... ,1) in order to be properly segmented by the default segmentation algorithm. This means that you might have to use np.transpose(img, (...)). You may specify the segmentation algorithm as well if the results are poor.
Finally you'll need to display the LIME image mask on top of the original image. This snippet shows how this may be done:
from skimage.segmentation import mark_boundaries
temp, mask = explanation.get_image_and_mask(explanation.top_labels[0], positive_only=False, num_features=5, hide_rest=False)
img_boundry = mark_boundaries(temp, mask)
plt.imshow(img_boundry)
plt.show()
This notebook is a good reference:
https://github.com/marcotcr/lime/blob/master/doc/notebooks/Tutorial%20-%20images%20-%20Pytorch.ipynb