Introduce explainability into deep learning models.
This repository is related to the Chapter 7 of the thesis: "Human-Centered xAI: Towards Overcoming Interpretation Biases in Biomedical Signal Analysis".
The library versions used here are:
- python=3.8
- pytorch=1.11.0
- cudatoolkit=11.3
- fastai=2.7.9
- fastcore=1.5.24
- ipykernel=6.25.0
- matplotlib=3.5.2
- numpy=1.22.3
- scikit-learn=1.1.0
- seaborn=0.11.2
- torchmetrics=0.7.3
All the required packages could be directly installed within your conda environment by using the file environment.yml through:
conda env create -n <your environment name> -f environment.yml
We provide an example of cat/dog classification from Oxford-IIIT Pet Dataset.
data_path = untar_data(URLs.PETS)
print(data_path.ls())
catblock = MultiCategoryBlock(encoded=True, vocab=['cat', 'dog'])
dblock = DataBlock(
blocks=(ImageBlock(), catblock),
get_items=get_image_files,
splitter=RandomSplitter(valid_pct=0.2, seed=42),
get_y=label_func,
item_tfms=Resize(128),
batch_tfms=[Normalize.from_stats(*imagenet_stats)],
)
With label_func defined as:
def label_func(f):
name = f.name #on veut accéder aux noms uniquement
if name[0].isupper(): #on veut tester la première lettre uniquement donc on applique "isupper" au premier élément de name (name[0])
lab = torch.tensor([1, 0], dtype=torch.float32)
else:
lab = torch.tensor([0, 1], dtype=torch.float32)
return lab
The data loader is therefore created using:
dls = dblock.dataloaders(data_path/"images", bs=16, drop_last=True)
The general architecture is an Adversarial AutoEncoder (AAE), as shown in the following figure:
The model MUST be an autoencoder and additional fully connected layers should be defined for the classifier and the discriminator.
You can modify the provided one in the model.py file.
def __init__(
self,
input_size,
input_channels,
encoding_dims=128,
step_channels=16,
nonlinearity=nn.LeakyReLU(0.2),
classes=2,
gen_train=True
):
super(AAEGen, self).__init__()
self.gen_train = gen_train
self.count_acc = 1
self.classes = classes
self.pool = nn.AdaptiveAvgPool2d(1)
self.flatten = nn.Flatten()
self.dropout = nn.Dropout(p=0.2)#, inplace=True)
# self.linear = nn.Linear(self.encoder.out_channels[-1], 2, bias=True) #2 classes
self.linear = nn.Linear(encoding_dims, self.classes, bias=True) #8 classes
self.bn_lin = nn.BatchNorm1d(num_features=encoding_dims)
self.fc_crit1 = nn.Linear(encoding_dims*2, 64)
self.fc_crit2 = nn.Linear(64, 16)
self.fc_crit3 = nn.Linear(16, 1)
self.bn_crit1 = nn.BatchNorm1d(num_features=64)
self.bn_crit2 = nn.BatchNorm1d(num_features=16)
encoder = [
nn.Sequential(
nn.Conv2d(input_channels, step_channels, 5, 2, 2), nonlinearity
)
]
size = input_size // 2
channels = step_channels
while size > 1:
encoder.append(
nn.Sequential(
nn.Conv2d(channels, channels * 4, 5, 4, 2),
nn.BatchNorm2d(channels * 4),
nonlinearity,
)
)
channels *= 4
size = size // 4
self.encoder = nn.Sequential(*encoder)
self.encoder_fc = nn.Linear(
channels, encoding_dims
) # Can add a Tanh nonlinearity if training is unstable as noise prior is Gaussian
self.decoder_fc = nn.Linear(encoding_dims, step_channels)
decoder = []
size = 1
channels = step_channels
while size < input_size // 2:
decoder.append(
nn.Sequential(
nn.ConvTranspose2d(channels, channels * 4, 5, 4, 2, 3),
nn.BatchNorm2d(channels * 4),
nonlinearity,
)
)
channels *= 4
size *= 4
decoder.append(nn.ConvTranspose2d(channels, input_channels, 5, 2, 2, 1))
self.decoder = nn.Sequential(*decoder)
Then, in the main.py file, you can instantiate it using:
model = AAE(
input_size=128,
input_channels=3,
encoding_dims=128,
classes=2,
)
Run the main.py script in your command prompt
python main.py
The result provided by this repository consists of the final latent space within which you can navigate to identify the most important features for the classification.
In our example, we have intentionnaly biased the dataset by applying rotation of [0,90] degrees to cat images and rotation of [90,180] degrees to dog images.
Here is the added value of using the AAE model instead of a common classifier (Resnet34):
When navigating along the most discriminant dimension, we can clearly observe the effect of the rotation on the classification on the AAE latent space:
While, on the initial Resnet34 model, the effect of the rotation cannot be identified: