Minor changes to README

Clean up build + environment
cpu tests for adversary

Author: GeoffNN

.travis.yml

@@ -25,6 +25,8 @@ install:
   - pip install cox requests
   - pip install git+https://github.com/RobustBench/robustbench
   - pip install pytest
+  - pip install advertorch
+  - pip install copt
 
 
 script:

README.md

@@ -5,7 +5,7 @@
 
 !! WARNING !! This library is in early development. Its API may change often for the time being.
 
-## Stochastic Constrained Algorithms
+## Stochastic Algorithms
 We define stochastic optimizers in the `chop.stochastic` module. These follow PyTorch Optimizer conventions, similar to the `torch.optim` module.
 
 ### Examples:

chop/stochastic.py

@@ -101,6 +101,8 @@ def __init__(self, params, constraint, lr=.1):
     @property
     @torch.no_grad()
     def certificate(self):
+        """A generator over the current convergence certificate estimate
+        for each optimized parameter."""
         for groups in self.param_groups:
             for p in groups['params']:
                 state = self.state[p]
@@ -157,6 +159,8 @@ def _lmo(u, x):
     @property
     @torch.no_grad()
     def certificate(self):
+        """A generator over the current convergence certificate estimate
+        for each optimized parameter."""
         for groups in self.param_groups:
             for p in groups['params']:
                 state = self.state[p]
@@ -240,8 +244,8 @@ def _lmo(u, x):
     @property
     @torch.no_grad()
     def certificate(self):
-        """For each parameter optimized, gives the current value of the 
-        estimated optimality certificate."""
+        """A generator over the current convergence certificate estimate
+        for each optimized parameter."""
         for group in self.param_groups:
             for p in group['params']:
                 state = self.state[p]

environment.yml

@@ -7,15 +7,9 @@ dependencies:
  - pandas
  - pytorch
  - cudatoolkit
- - jupyter
  - torchvision
- - pylint
- - pytest
- - rope
  - pip
  - pip:
    - tables
    - dill
-   - copt
-   - advertorch
    - easydict

examples/adversarial_robustness/attack_benchmark.py

@@ -1,10 +1,8 @@
-from chop.optim import minimize_frank_wolfe, minimize_pgd, minimize_pgd_madry, minimize_three_split
-from chop.utils import closure
-
 import torch
 from tqdm import tqdm
 
 import chop
+from chop.optim import minimize_frank_wolfe, minimize_pgd, minimize_pgd_madry, minimize_three_split
 from chop.data import load_cifar10
 from chop.adversary import Adversary
 
@@ -70,7 +68,6 @@ def prox(delta, step_size=None):
                                                      step=2. / max_iter,
                                                      max_iter=max_iter)
 
-    delta_split = torch.zeros_like(data, device=data.device)
     _, delta_split = adversary_split.perturb(data, target, model,
                                              criterion,
                                              use_best=False,

examples/training_constrained_net_on_mnist.py

@@ -46,7 +46,7 @@
 # Project model parameters in the constraint set.
 constraint.make_feasible(model)
 
-optimizer = chop.stochastic.MomentumFrankWolfe(model.parameters(), constraint, lr=.3)
+optimizer = chop.stochastic.FrankWolfe(model.parameters(), constraint, lr=lr, momentum=momentum)
 
 # Training loop
 for epoch in range(nb_epochs):

tests/test_adversary.py

@@ -30,8 +30,7 @@ def test_adversary_synthetic_data(algorithm, step_size, p):
     # Setup
     torch.manual_seed(0)
     use_cuda = torch.cuda.is_available()
-    # device = torch.device("cuda" if use_cuda else "cpu")
-    device = torch.device("cpu")
+    device = torch.device("cuda" if use_cuda else "cpu")
 
     data = torch.rand((1, 25, 25))
     target = torch.zeros(1).long()
@@ -93,12 +92,11 @@ def test_adversary_mnist(algorithm, step, p, model_filename):
 
     use_cuda = torch.cuda.is_available()
     device = torch.device("cuda" if use_cuda else "cpu")
-    # device = torch.device("cpu")
+    device = torch.device("cpu")
 
     model = LeNet5()
     model_path = os.path.join(TRAINED_MODEL_PATH, model_filename)
-    model.load_state_dict(torch.load(model_path))
-    model.to(device)
+    model.load_state_dict(torch.load(model_path, map_location=device))
     model.eval()
 
     criterion = nn.CrossEntropyLoss()