troj 0.0.1

TrojAI provides the troj Python convenience package to allow users to integrate TrojAI adversarial protections and robustness metrics seamlessly into their AI development pipelines.

TrojAI Integration Package

TrojAI provides the troj Python convenience package to allow users to integrate TrojAI adversarial protections and robustness metrics seamlessly into their AI development pipelines.

How to develop locally

these commands add the local package to the wheel pip install wheel python setup.py bdist_wheel pip install -e.[dev]

Installation

Run the following to install:

pip install troj

Usage

'''
ImageNet folder style: 
/train
----/class1
----/class2
...
/test
----/class1
----/class2
...etc 

CoCo annotation style:
/images
----image1.jpg
----image2.jpg
...etc
annotation.json

'''



import string
import random
import troj


session = troj.start(...)

PROJECT_ID = 'My Project'


if not session.project_exists(PROJECT_ID):
    print("Creating Project")

    # need to create this function to return labels map
    label_class_map = session.dataset.getLabelsMap()

    # This associates each class name with an ID and visualization color
    # label_class_map = troj.LabelClassMap.from_classnames(LABEL_CLASSES)

    # We do not specify `primary_task`
    # This will save the project ID into client, and when creating a dataset, will call it's own property
    session.create_project(PROJECT_ID, label_class_map)
    DATASET_ID = "New Dataset"
    session.create_dataset(dataset_ID)

if not session.dataset_exists(PROJECT_ID, DATASET_ID):
    print("Creating Dataset")
    session.create_dataset(
        PROJECT_ID,
        DATASET_ID,
        dataset=ds, 
        wait_until_finish=True, # Poll for job completion
        preview_first_frame=True # Preview frame to confirm that things look as expected
    )


# creates the dataframe inside the session object
# stored in session.dataset
session.create_troj_dataset("CIFAR10-Imagenet-Format/test/")

# uses the created dataframe to create a dataloader
# stored in session.dataloader
session.create_troj_dataloader()


import torch
from torch.utils.data import DataLoader

batch_size=64
test_loader = torch.utils.data.DataLoader(session.dataloader,
                                           batch_size=batch_size,
                                           shuffle=True)

input_shape = ()
'''
At this point the data is loaded in, now we need to set up the model and loss function
creating the attack target instantiates an attackable object and stores it with the session
'''
session.create_attack_target(model, loss, input_shape)
session.create_attack_target(model, loss, input_size)

session.run_troj_test(test_loader, loss_function)

Things we need to upload to aws in test loop

1. Images
2. Labels
3. Embeddings
4. Inferences
5. Perturbations