refactoring

.gitignore

@@ -5,7 +5,7 @@
 **/mlruns
 **/mlartifacts
 **datasets
-!gnn_toolbox/custom_modules/datasets
+!gnn_toolbox/custom_components/datasets
 *.ckpt
 *.log
 *.json

PyG_to_sci_sparse.py

@@ -305,54 +305,60 @@ def to_symmetric_scipy(adjacency: sp.csr_matrix):
 def accuracy(pred, y, mask):
     return (pred.argmax(-1)[mask] == y[mask]).float().mean()
 # print(data9.is_directed())
-dataset2= PygNodePropPredDataset(name='ogbn-arxiv', root='./datasets', transform=T.ToSparseTensor(remove_edge_index=False))
+# dataset2= PygNodePropPredDataset(name='ogbn-arxiv', root='./datasets', transform=T.ToSparseTensor(remove_edge_index=False))
 dataset1= Planetoid(name = 'cora', root = './datasets')
 data = dataset1[0]
 train=data.train_mask.nonzero().squeeze()
 model1= GCN(in_channels=dataset1.num_features, out_channels=dataset1.num_classes, hidden_channels=32, num_layers=2)
 logits = model1(data.x, data.edge_index)
 # print(accuracy(logits, data.y, train))
 # print(accuracy(logits, data.y, data.train_mask))
-# from torch_geometric.datasets import PPI
-# dataset1 = PPI(root='./datasets', transform=T.ToSparseTensor(remove_edge_index=False))
+from torch_geometric.datasets import PPI, KarateClub
+dataset9 = QM9(root = './datasets', transform=T.ToSparseTensor(remove_edge_index=False))
+data12 = dataset9[0]
+print(data12)
+
 # print('pygnodepred', dataset1)
 # we = dataset1[0]
 # print('we', we)
 
 # dataset2= Planetoid(name = 'citeseer', root = './datasets')
 # dataset2= QM9(root = './datasets')
-data2 = dataset2[0]
-# train_mask=data2.train_mask.nonzero().squeeze()
-# attr, adj, labels, splits, n = prep_graph(dataset2, make_undirected=False)
-# print()
-from torch_geometric.utils import to_undirected, is_undirected
-print('is_undirected', is_undirected(data2.edge_index, num_nodes=data2.num_nodes))
-adj = SparseTensor(row=data2.edge_index[0], col=data2.edge_index[1], value=data2.edge_attr).to('cuda')
-adj2 = SparseTensor(row=data2.edge_index[0], col=data2.edge_index[1], value=data2.edge_weight if data2.edge_weight is not None else torch.ones(data2.edge_index.size(1))).to('cuda')
-print('num_edge',adj.nnz())
-print('what', adj)
-print('what2', adj.t())
-
-row, col, edge_attr = adj.t().coo()
-edge_index = torch.stack([row, col], dim=0)
-print('edge_index3', edge_index)
-
-# row, col, edge_attr = adj.coo()
+# data2 = dataset2[0]
+# # train_mask=data2.train_mask.nonzero().squeeze()
+# # attr, adj, labels, splits, n = prep_graph(dataset2, make_undirected=False)
+# # print()
+# from torch_geometric.utils import to_undirected, is_undirected
+# print('is_undirected', is_undirected(data2.edge_index, num_nodes=data2.num_nodes))
+# adj = SparseTensor(row=data2.edge_index[0], col=data2.edge_index[1], value=data2.edge_attr).to('cuda')
+# adj2 = SparseTensor(row=data2.edge_index[0], col=data2.edge_index[1], value=data2.edge_weight if data2.edge_weight is not None else torch.ones(data2.edge_index.size(1))).to('cuda')
+# print('num_edge',adj.nnz())
+# print('what', adj)
+# print('what2', adj.t())
+
+# row, col, edge_attr = adj.t().coo()
 # edge_index = torch.stack([row, col], dim=0)
-# print('edge_index1', edge_index)
+# print('edge_index3', edge_index)
+# print('edge_attr', edge_attr)
+# # row, col, edge_attr = adj.coo()
+# # edge_index = torch.stack([row, col], dim=0)
+# # print('edge_index1', edge_index)
+
+
+# edge_index = to_undirected(data2.edge_index)
+# print('x',edge_index.shape[1])
+# print('undirected', edge_index)
+# edge_index2 = to_undirected(edge_index)
+# print('undirected', edge_index)
+# print('q',torch.equal(edge_index, edge_index2))
+# print(torch.equal(data2.edge_index, edge_index))
+# print("numm", data2.num_nodes)
 
+# print("numm2", data2.x.shape[0])
+# print("numm2", data2.x.size(0))
 
-edge_index = to_undirected(data2.edge_index)
-print('x',edge_index.shape[1])
-print('undirected', edge_index)
-edge_index2 = to_undirected(edge_index)
-print('undirected', edge_index)
-print('q',torch.equal(edge_index, edge_index2))
-print(torch.equal(data2.edge_index, edge_index))
-print("numm", data2.num_nodes)
+# print(T.ToSparseTensor.__name__)
 
-print("numm2", data2.x.shape[0])
-print("numm2", data2.x.size(0))
 # row, col, edge_attr = adj.coo()
 # cora_edge_index = torch.stack([row, col], dim=0)
 # print('edge_attr', edge_attr)

configs/default_experiment.yaml

@@ -1,30 +1,31 @@
-output_dir: ./output6
+output_dir: ./output8
 # resume_output: False
 # csv_save: True
 # esd: True
 experiment_templates:
-  - name: My experiment
+  - name: Local attack_3
     seed: [0]
     device: cuda
     model:
       - name: GCN
         params:
           hidden_channels: 64
     dataset:
-      name: Cora # ogbn-arxiv 
+      name: [QM9] # ogbn-arxiv 
       root: ./datasets
       make_undirected: true
       # transforms:
+      #   - name: ToUndirected
       #   - name: NormalizeFeatures
         # - name: Constant
         #   params:
         #     value: 0.8
 
     attack:
-      # - scope: global
-      #   name: DICE
-      #   type: poison
-      #   epsilon: [0.5]
+      - scope: 'global'
+        name: DICE
+        type: evasion
+        epsilon: [0.5]
       # - scope: global
       #   name: FGSM
       #   type: poison
@@ -33,10 +34,14 @@ experiment_templates:
       #   name: PGD
       #   type: poison
       #   epsilon: [0.5]
-      - scope: global
-        name: [DICE]
-        type: evasion
-        epsilon: [0.6]
+
+      # - scope: local
+      #   name: [LocalDICE]
+      #   type: evasion
+      #   epsilon: [0.6]
+      #   nodes_topk: 10
+      #   min_node_degree: 2
+
         # params:
         #   block_size: 100000
       # - scope: local

gnn_toolbox/custom_components/attacks/global_attacks/dice.py → custom_components/attacks/global_attacks/dice.py

@@ -5,8 +5,8 @@
 from torch_sparse import SparseTensor
 from tqdm import tqdm
 
-import gnn_toolbox.custom_components.utils as utils
-from gnn_toolbox.custom_components.attacks.base_attack import GlobalAttack
+import custom_components.utils as utils
+from custom_components.attacks.base_attack import GlobalAttack
 from gnn_toolbox.registry import register_global_attack
 
 @register_global_attack("DICE")

gnn_toolbox/custom_components/attacks/global_attacks/greedy_rbcd.py → custom_components/attacks/global_attacks/greedy_rbcd.py

@@ -3,8 +3,8 @@
 import torch_sparse
 from torch_sparse import SparseTensor
 
-from gnn_toolbox.custom_components import utils
-from gnn_toolbox.custom_components.attacks.global_attacks.prbcd import PRBCD
+from custom_components import utils
+from custom_components.attacks.global_attacks.prbcd import PRBCD
 from gnn_toolbox.registry import register_global_attack
 
 @register_global_attack("GreedyRBCD")

gnn_toolbox/custom_components/attacks/global_attacks/new_prbcd.py → custom_components/attacks/global_attacks/new_prbcd.py

@@ -12,7 +12,7 @@
 import scipy.sparse as sp
 
 from torch_geometric.utils import coalesce, to_undirected
-from gnn_toolbox.custom_components.attacks.base_attack import GlobalAttack
+from custom_components.attacks.base_attack import GlobalAttack
 from gnn_toolbox.registry import register_global_attack
 # (predictions, labels, ids/mask) -> Tensor with one element
 LOSS_TYPE = Callable[[Tensor, Tensor, Optional[Tensor]], Tensor]

gnn_toolbox/custom_components/attacks/global_attacks/prbcd.py → custom_components/attacks/global_attacks/prbcd.py

@@ -11,8 +11,8 @@
 from torch_sparse import SparseTensor
 
 # from rgnn_at_scale.models import MODEL_TYPE
-from gnn_toolbox.custom_components import utils
-from gnn_toolbox.custom_components.attacks.base_attack import GlobalAttack
+from custom_components import utils
+from custom_components.attacks.base_attack import GlobalAttack
 from gnn_toolbox.registry import register_global_attack
 
 @register_global_attack("PRBCD")

gnn_toolbox/custom_components/attacks/local_attacks/local_dice.py → custom_components/attacks/local_attacks/local_dice.py

@@ -1,7 +1,7 @@
 import torch
 from torch_sparse import SparseTensor
 
-from gnn_toolbox.custom_components.attacks.base_attack import LocalAttack
+from custom_components.attacks.base_attack import LocalAttack
 from gnn_toolbox.registry import register_local_attack
 
 @register_local_attack("LocalDICE")

gnn_toolbox/custom_components/attacks/local_attacks/local_prbcd.py → custom_components/attacks/local_attacks/local_prbcd.py

@@ -12,8 +12,8 @@
 
 from tqdm import tqdm
 
-from gnn_toolbox.custom_components.attacks.base_attack import LocalAttack
-from gnn_toolbox.custom_components import utils
+from custom_components.attacks.base_attack import LocalAttack
+from custom_components import utils
 from gnn_toolbox.registry import register_local_attack
 
 

gnn_toolbox/custom_components/data_transforms/transforms.py → custom_components/data_transforms/transforms.py

@@ -3,9 +3,11 @@
     AddSelfLoops,
     Constant,
     NormalizeFeatures,
+    ToSparseTensor,
 )
 from gnn_toolbox.registry import register_transform
 
 register_transform("AddSelfLoops", AddSelfLoops)
 register_transform("Constant", Constant)
 register_transform("NormalizeFeatures", NormalizeFeatures)
+register_transform("ToSparseTensor", ToSparseTensor(remove_edge_index=False))

gnn_toolbox/custom_components/models/architecture.py → custom_components/models/architecture.py

@@ -6,12 +6,10 @@
 from torch_geometric.nn.conv.gcn_conv import gcn_norm
 # from gnn_toolbox.custom_modules.models.model
 # import gnn_toolbox
-from gnn_toolbox.registry import register_model, registry
+from gnn_toolbox.registry import register_model
 import os
 from typing import Any, Dict, Union
 
-
-
 @register_model("GCN")
 class GCNWH(nn.Module):
     def __init__(self, in_channels: int, out_channels: int, hidden_channels: int, num_layers: int = 2, dropout: float = 0.5, **kwargs):
@@ -32,7 +30,7 @@ def forward(self, x, edge_index, edge_weight, **kwargs):
 
 
 
-@register_model('gcn2')
+# @register_model('gcn2')
 def GCN_(in_channels, out_channels, hidden_channels):
     return GCN(in_channels, out_channels, hidden_channels)
 
@@ -51,7 +49,7 @@ def reset_parameters(self):
         self.conv1.reset_parameters()
         self.conv2.reset_parameters()
 
-    def forward(self, x, edge_index, edge_weight=None, **kwargs):
+    def forward(self, x, edge_index, edge_weight=None):
         # Normalize edge indices only once:
         if not kwargs.get('skip_norm', False):
             edge_index, edge_weight = self.norm(
@@ -115,7 +113,7 @@ def forward(self, arguments: Tuple[torch.Tensor, torch.Tensor]) -> torch.Tensor:
 
 import collections
 from torch_sparse import coalesce, SparseTensor 
-@register_model('DenseGCN')
+# @register_model('DenseGCN')
 class DenseGCN(nn.Module):
     """Dense two layer GCN for the FGSM attack that requires a gradient towards the adjacency matrix.
     """

gnn_toolbox/custom_components/models/model.py → custom_components/models/model.py

@@ -7,7 +7,7 @@
 from pytorch_lightning import LightningModule
 
 # from gnn_toolbox.old.config_def import cfg
-from gnn_toolbox.custom_components.optimizers.optimizers import register_optimizer
+# from custom_components.optimizers.optimizers import register_optimizer
 from gnn_toolbox.registry import registry
 
 class BaseModel(LightningModule):

gnn_toolbox/custom_components/optimizers/optimizers.py → custom_components/optimizers/optimizers.py

@@ -1,6 +1,6 @@
 from torch.optim import Adam
 from torch.nn.parameter import Parameter
-from ...registry import register_optimizer
+from gnn_toolbox.registry import register_optimizer
 from typing import Iterator, Dict
 
 @register_optimizer('adam')

gnn_toolbox/cmd_args.py

@@ -1,17 +1,27 @@
 import os
 import argparse
 import logging
-
+from custom_components import *
+from gnn_toolbox.registry import registry
 def parse_args() -> argparse.Namespace:
     """Parses the command line arguments."""
-    parser = argparse.ArgumentParser(description='Run experiments from a configuration file.')
-
+    parser = argparse.ArgumentParser(description='GNN Robustness Toolbox: Run experiments to test GNNs against adversarial attacks from configuration YAML file.')
+    # parser = argparse.ArgumentParser(description=f'available models {registry["model"].keys()} \n available global attacks {registry["global_attack"].keys()} \n available local attacks {registry["local_attack"].keys()} \n available datasets {registry["dataset"].keys()} \n available transforms {registry["transform"].keys()} \n available optimizers {registry["optimizer"].keys()} \n available losses {registry["loss"].keys()}')
     parser.add_argument('--cfg', type=str, default= os.path.join('configs', 'default_experiment.yaml'), help='The configuration YAML file path. Default is configs/good_1.yaml.')
 
     parser.add_argument('--log', type=str, default='INFO', help='Logging level. Choose from DEBUG, INFO, WARNING, ERROR, CRITICAL.')
 
+    parser.add_argument('--list-components', action='store_true', help="List registered components and exit")
+
     return parser.parse_args()
 
+def list_registered_components():
+    """
+    List all registered components
+    """
+    for key, value in registry.items():
+        print(f"Registered {key}: {list(value.keys())}")
+
 def logger_setup(logging_level):
     """
     Setup the logger for the experiment. Modified from https://docs.python.org/3/howto/logging.html