Cuda Requirement?

[ElizabethSotomayor]

While I am aware that AB Crown only works with Ubuntu, I am wondering if a NVIDIA GPU is a requirement. I believe this may be possible to enable while using a VM, although my attempts have not been successful. Any guidance on that would be wonderful, but in the meantime I am wondering if there is any way to utilize the tool without the CUDA?

Here is my output when running the command: python abcrown.py --config exp_configs/mnist_6_100.yaml

I receive the same output on my VM which runs Kubuntu 23.04, which I attempted to install CUDA on.

/Users/lizsotomayor/miniconda3/envs/alpha-beta-crown/lib/python3.7/site-packages/onnx/mapping.py:27: DeprecationWarning: `np.object` is a deprecated alias for the builtin `object`. To silence this warning, use `object` by itself. Doing this will not modify any behavior and is safe. 
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  int(TensorProto.STRING): np.dtype(np.object)
Configurations:

general:
  device: cuda
  seed: 100
  conv_mode: patches
  deterministic: false
  double_fp: false
  loss_reduction_func: sum
  record_bounds: false
  sparse_alpha: true
  save_adv_example: false
  precompile_jit: false
  complete_verifier: bab-refine
  enable_incomplete_verification: true
  csv_name: null
  results_file: out.txt
  root_path: ''
model:
  name: mnist_6_100
  path: models/eran/mnist_6_100_nat.pth
  onnx_path: null
  onnx_path_prefix: ''
  cache_onnx_conversion: false
  onnx_quirks: null
  input_shape: null
  onnx_loader: default_onnx_and_vnnlib_loader
  onnx_optimization_flags: none
data:
  start: 0
  end: 10000
  select_instance: null
  num_outputs: 10
  mean: [0.0]
  std: [1.0]
  pkl_path: null
  dataset: MNIST_ERAN_UN
  data_filter_path: null
  data_idx_file: null
specification:
  type: lp
  robustness_type: verified-acc
  norm: .inf
  epsilon: 0.026
  vnnlib_path: null
  vnnlib_path_prefix: ''
solver:
  batch_size: 1024
  min_batch_size_ratio: 0.1
  use_float64_in_last_iteration: false
  early_stop_patience: 10
  start_save_best: 0.5
  bound_prop_method: alpha-crown
  prune_after_crown: false
  crown:
    batch_size: 1000000000
    max_crown_size: 1000000000
  alpha-crown:
    alpha: true
    lr_alpha: 0.1
    iteration: 100
    share_slopes: false
    no_joint_opt: false
    lr_decay: 0.98
    full_conv_alpha: true
  beta-crown:
    lr_alpha: 0.01
    lr_beta: 0.05
    lr_decay: 0.98
    optimizer: adam
    iteration: 20
    beta: true
    beta_warmup: true
    enable_opt_interm_bounds: false
    all_node_split_LP: false
  forward:
    refine: false
    dynamic: false
    max_dim: 10000
  multi_class:
    multi_class_method: allclass_domain
    label_batch_size: 32
    skip_with_refined_bound: true
  mip:
    parallel_solvers: 16
    solver_threads: 1
    refine_neuron_timeout: 15
    refine_neuron_time_percentage: 0.8
    early_stop: true
    adv_warmup: true
    mip_solver: gurobi
bab:
  initial_max_domains: 1
  max_domains: .inf
  decision_thresh: 0
  timeout: 300
  timeout_scale: 1
  override_timeout: null
  get_upper_bound: false
  dfs_percent: 0.0
  pruning_in_iteration: true
  pruning_in_iteration_ratio: 0.2
  sort_targets: false
  batched_domain_list: true
  optimized_intermediate_layers: ''
  interm_transfer: true
  cut:
    enabled: false
    bab_cut: false
    lp_cut: false
    method: null
    lr: 0.01
    lr_decay: 1.0
    iteration: 100
    bab_iteration: -1
    early_stop_patience: -1
    lr_beta: 0.02
    number_cuts: 50
    topk_cuts_in_filter: 100
    batch_size_primal: 100
    max_num: 1000000000
    patches_cut: false
    cplex_cuts: false
    cplex_cuts_wait: 0
    cplex_cuts_revpickup: true
    cut_reference_bounds: true
    fix_intermediate_bounds: false
  branching:
    method: kfsb
    candidates: 3
    reduceop: max
    sb_coeff_thresh: 0.001
    input_split:
      enable: false
      enhanced_bound_prop_method: alpha-crown
      enhanced_branching_method: naive
      enhanced_bound_patience: 100000000.0
      attack_patience: 100000000.0
      adv_check: 0
      sort_domain_interval: -1
  attack:
    enabled: false
    beam_candidates: 8
    beam_depth: 7
    max_dive_fix_ratio: 0.8
    min_local_free_ratio: 0.2
    mip_start_iteration: 5
    mip_timeout: 30.0
    adv_pool_threshold: null
    refined_mip_attacker: false
    refined_batch_size: null
attack:
  pgd_order: after
  pgd_steps: 100
  pgd_restarts: 30
  pgd_early_stop: true
  pgd_lr_decay: 0.99
  pgd_alpha: auto
  pgd_loss_mode: null
  enable_mip_attack: false
  cex_path: ./test_cex.txt
  attack_mode: PGD
  gama_lambda: 10.0
  gama_decay: 0.9
  check_clean: false
  input_split:
    pgd_steps: 100
    pgd_restarts: 30
    pgd_alpha: auto
  input_split_enhanced:
    pgd_steps: 200
    pgd_restarts: 5000000
    pgd_alpha: auto
  input_split_check_adv:
    pgd_steps: 5
    pgd_restarts: 5
    pgd_alpha: auto
debug:
  lp_test: null

Experiments at Wed Sep 20 01:30:43 2023 on Lizs-Air-3.lan
Sequential(
  (0): Flatten()
  (1): Linear(in_features=784, out_features=100, bias=True)
  (2): ReLU()
  (3): Linear(in_features=100, out_features=100, bias=True)
  (4): ReLU()
  (5): Linear(in_features=100, out_features=100, bias=True)
  (6): ReLU()
  (7): Linear(in_features=100, out_features=100, bias=True)
  (8): ReLU()
  (9): Linear(in_features=100, out_features=100, bias=True)
  (10): ReLU()
  (11): Linear(in_features=100, out_features=10, bias=True)
)
/Users/lizsotomayor/alpha-beta-CROWN-main/complete_verifier/utils.py:623: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
  eps_temp = torch.tensor(eps_temp).reshape(1, -1, 1, 1)
############################
Sampled data loaded. No normalization used!
Shape: torch.Size([1000, 1, 28, 28]) torch.Size([1000]) torch.Size([1000])
X range: tensor(1.) tensor(0.) tensor(0.12226457)
Note runnerup label is empty here!
############################
Internal results will be saved to Verified_ret_[mnist_6_100]_start=0_end=10000_iter=20_b=1024_timeout=300_branching=kfsb-max-3_lra-init=0.1_lra=0.01_lrb=0.05_PGD=after_cplex_cuts=False_multiclass=allclass_domain.npy.

 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% idx: 0, vnnlib ID: 0 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Traceback (most recent call last):
  File "abcrown.py", line 647, in <module>
    main()
  File "abcrown.py", line 512, in main
    model_ori = model_ori.to(device)
  File "/Users/lizsotomayor/miniconda3/envs/alpha-beta-crown/lib/python3.7/site-packages/torch/nn/modules/module.py", line 989, in to
    return self._apply(convert)
  File "/Users/lizsotomayor/miniconda3/envs/alpha-beta-crown/lib/python3.7/site-packages/torch/nn/modules/module.py", line 641, in _apply
    module._apply(fn)
  File "/Users/lizsotomayor/miniconda3/envs/alpha-beta-crown/lib/python3.7/site-packages/torch/nn/modules/module.py", line 664, in _apply
    param_applied = fn(param)
  File "/Users/lizsotomayor/miniconda3/envs/alpha-beta-crown/lib/python3.7/site-packages/torch/nn/modules/module.py", line 987, in convert
    return t.to(device, dtype if t.is_floating_point() or t.is_complex() else None, non_blocking)
  File "/Users/lizsotomayor/miniconda3/envs/alpha-beta-crown/lib/python3.7/site-packages/torch/cuda/__init__.py", line 221, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

[shizhouxing]

Hi @ElizabethSotomayor , it looks like the issue is on the pytorch installation and not directly our library. Could you please check how pytorch can be properly installed for your GPU? Your pytorch seems to be a CPU version. You'll need a Pytorch for CUDA.

[shizhouxing]

@ElizabethSotomayor If you mean you want to use the library without any GPU, you may set --device cpu to use CPU only.