Latest code

.gitignore

@@ -3,5 +3,7 @@ dataset/
 logs/
 models/
 embeddings/
+gat/dataset copy/
+gat/ogbn-cora-submission/
 __*
 **/submission_ogbn_cora/*

README.md

@@ -43,8 +43,8 @@ cp ogbn-cora-submission/master.csv /Users/samdatta/miniconda/envs/gmlpipe/lib/py
 **NOTE**: The current version of the code does not aim to reproduce the paper exactly - while we use (inside `adapter.py`, which downloads Cora from DGL and packages it for a submission to OGB) the standard train-val-test split of 140:500:1000, the C&S paper reportedly used a 60:20:20 _random_ split (see Sec. 3).
 
 ### Experiments
-- `python run_experiments.py --dataset cora --method lp` gives val/test accuracies of 0.698/0.707.
-- `python gen_models.py --dataset arxiv --model mlp --use_embeddings` gives val/test accuracies of 76.6600 ± 0.8796/76.9600 ± 1.0384.
+- `python run_experiments.py --dataset cora --method lp` gives val/test accuracies of 0.698/0.707.		->D
+- `python gen_models.py --dataset arxiv --model mlp --use_embeddings` gives val/test accuracies of 76.6600 ±	->D 0.8796/76.9600 ± 1.0384.
 
 **NOTE** Other models (e.g., GAT+C&S) have not been tested yet. Most likely, those pathways will throw errors (the original code has quite a bit of code-duplication - not all pathways were modified to accept Cora). However, fixing them should be routine.
 
@@ -53,11 +53,19 @@ cp ogbn-cora-submission/master.csv /Users/samdatta/miniconda/envs/gmlpipe/lib/py
 ### Label Propagation (0 params):
 ```
 python run_experiments.py --dataset arxiv --method lp
+Valid acc:  0.7018356320681902
+Test acc: 0.683496903483324
+
 
 Valid acc: 0.7013658176448874
 Test acc: 0.6832294302820814
 ```
+```
+python run_experiments.py --dataset cora --method lp		->DONE
 
+Valid acc: 0.968
+Test acc: 0.707
+```
 ### Plain Linear + C&S (5160 params, 52.5% base accuracy)
 ```
 python gen_models.py --dataset arxiv --model plain --epochs 1000    
@@ -67,13 +75,62 @@ Valid acc -> Test acc
 Args []: 73.00 ± 0.01 -> 71.26 ± 0.01
 ```
 
+```
+python gen_models.py --dataset cora --model plain --epochs 1000     D
+python run_experiments.py --dataset cora --method plain		D
+
+All runs:
+Highest Train: 100.0000 ± 0.0000
+Highest Valid: 51.2600 ± 0.3893
+  Final Train: 100.0000 ± 0.0000
+   Final Test: 47.4400 ± 0.5641
+
+
+Valid acc -> Test acc
+Args []: 74.34 ± 0.21 -> 75.17 ± 0.09
+
+
+
+_________________________________________________________________________
+
+python run_experiments.py --dataset cora --method plain_gen_bound		D
+
+Valid acc -> Test acc
+Args []: 71.84 ± 0.67 -> 74.18 ± 0.62
+
+```
+
 ### Linear + C&S (15400 params, 70.11% base accuracy)
 ```
 python gen_models.py --dataset arxiv --model linear --use_embeddings --epochs 1000 
 python run_experiments.py --dataset arxiv --method linear
 
 Valid acc -> Test acc
 Args []: 73.68 ± 0.04 -> 72.22 ± 0.02;
+```
+```
+python gen_models.py --dataset cora --model linear --use_embeddings --epochs 1000 	D
+python run_experiments.py --dataset cora --method linear				D
+
+
+
+All runs:
+Highest Train: 100.0000 ± 0.0000
+Highest Valid: 72.9600 ± 0.4695
+  Final Train: 99.7857 ± 0.3450
+   Final Test: 71.7000 ± 0.3055
+
+Valid acc -> Test acc
+Args []: 75.78 ± 0.06 -> 75.36 ± 0.08
+
+_________________________________________________________________________
+
+python run_experiments.py --dataset cora --method linear_gen_bound		D
+
+Valid acc -> Test acc
+Args []: 74.18 ± 0.82 -> 75.29 ± 0.40
+
+
 ```
 
 ### MLP + C&S (175656 params, 71.44% base accuracy)
@@ -83,17 +140,77 @@ python run_experiments.py --dataset arxiv --method mlp
 
 Valid acc -> Test acc
 Args []: 73.91 ± 0.15 -> 73.12 ± 0.12
+```
+
+```
+python gen_models.py --dataset cora --model mlp --use_embeddings	D
+python run_experiments.py --dataset cora --method mlp	
+
+
+All runs:
+Highest Train: 100.0000 ± 0.0000
+Highest Valid: 76.8800 ± 1.2300
+  Final Train: 100.0000 ± 0.0000
+   Final Test: 76.9000 ± 1.0965
+
+Valid acc -> Test acc
+Args []: 78.50 ± 1.33 -> 79.10 ± 1.28
+
+_________________________________________________________________________
+
+python run_experiments.py --dataset cora --method mlp_gen_bound 		D
+
+Valid acc -> Test acc
+Args []: 78.50 ± 1.33 -> 79.10 ± 1.28
+
+
+
 ```
 
 ### GAT + C&S (1567000 params, 73.56% base accuracy)
 ```
 cd gat && python gat.py --use-norm
+Average epoch time: 21.718203401565553, Test acc: 0.6538896775919182
+Runned 10 times
+Val Accs: [0.6605255209906372, 0.6424376656934796, 0.6559951676230746, 0.6449880868485519, 0.6551562132957481, 0.6409275479042921, 0.6187120373166884, 0.6712976945535085, 0.6283432329943958, 0.6312292358803987]
+Test Accs: [0.667160463345884, 0.6027817212929243, 0.6717075077670103, 0.6384791062280106, 0.6247145237948275, 0.6592597164784066, 0.578585684011275, 0.6555356665226426, 0.6484579141205276, 0.6538896775919182]
+Average val accuracy: 0.6449612403100775 ± 0.015290282129320602
+Average test accuracy: 0.6400571981153427 ± 0.02830663270099123
+[768, 768, 98304, 768, 768, 589824, 120, 120, 92160, 98304, 589824, 92160, 768, 768, 768, 768, 40]
+Number of params: 1567000
+
+
+
 cd .. && python run_experiments.py --dataset arxiv --method gat
+Valid acc -> Test acc
+Args []: 67.27 ± 1.39 -> 66.25 ± 2.90
+
 
 Valid acc -> Test acc
 Args []: 74.84 ± 0.07 -> 73.86 ± 0.14
 ```
 
+```
+cd gat && python gat.py --use-norm --dataset cora
+cd .. && python run_experiments.py --dataset cora --method gat
+
+
+Average epoch time: 0.3995665550231934, Test acc: 0.262
+Runned 10 times
+**************************************************
+Average epoch time: 0.39290672254562375, Test acc: 0.791
+Runned 10 times
+Val Accs: [0.772, 0.77, 0.768, 0.788, 0.78, 0.78, 0.79, 0.746, 0.774, 0.76]
+Test Accs: [0.788, 0.775, 0.771, 0.804, 0.794, 0.802, 0.797, 0.772, 0.772, 0.791]
+Average val accuracy: 0.7727999999999999 ± 0.012432216214336054
+Average test accuracy: 0.7866000000000002 ± 0.012362847568420484
+[768, 768, 1100544, 768, 768, 589824, 21, 21, 16128, 1100544, 589824, 16128, 768, 768, 768, 768, 7]
+Number of params: 3419185
+
+Valid acc -> Test acc
+Args []: 78.18 ± 1.14 -> 80.38 ± 1.15
+```
+
 ### Notes
 As opposed to the paper's results, which only use spectral embeddings, here we use spectral *and* diffusion embeddings, which we find improves Arxiv performance.
 

adapter.py

@@ -1,6 +1,6 @@
 from curses.ascii import isspace
 import torch_geometric.transforms as T
-from ogb.nodeproppred import PygNodePropPredDataset
+from ogb.nodeproppred import PygNodePropPredDataset , DglNodePropPredDataset
 from ogb.io import DatasetSaver
 import numpy as np
 from dgl.data import *
@@ -11,6 +11,9 @@ def dgl_to_ogbn(dataset_name, mapping_path, is_sparse=False):
         dataset = CoraGraphDataset()
         dataset_name = 'ogbn-cora'
 
+        print("#### cora type initial = ",type(dataset))
+
+
     saver = DatasetSaver(dataset_name=dataset_name, is_hetero=False, version=1)
 
     g = dataset[0].to_networkx()
@@ -29,6 +32,16 @@ def dgl_to_ogbn(dataset_name, mapping_path, is_sparse=False):
     split_idx['train'] = np.array(g.nodes)[dataset[0].ndata['train_mask']]
     split_idx['valid'] = np.array(g.nodes)[dataset[0].ndata['val_mask']]
     split_idx['test'] = np.array(g.nodes)[dataset[0].ndata['test_mask']]
+    # split_idx['train'] = np.load("cora_splits/cora_train_split.npy")
+    # split_idx['valid'] = np.load("cora_splits/cora_validate_split.npy")
+    # split_idx['test'] = np.load("cora_splits/cora_test_split.npy")
+
+    print("train len = ",len(split_idx['train']))
+
+    print("valid len = ",len(split_idx['valid']))
+    print("test len = ",len(split_idx['test']))
+    print("train type = ",type(split_idx['train']))
+    print("train  = ",split_idx['train'])
     saver.save_split(split_idx, split_name='std')
 
     saver.copy_mapping_dir(mapping_path)
@@ -46,3 +59,53 @@ def dgl_to_ogbn(dataset_name, mapping_path, is_sparse=False):
     saver.zip()
 
     return pyg_dataset
+
+
+def dgl_to_dgl_ogbn(dataset_name, mapping_path, is_sparse=False):
+    if 'cora' == dataset_name:
+        dataset = CoraGraphDataset()
+        dataset_name = 'ogbn-cora'
+        print("#### cora type initial = ",type(dataset))
+    saver = DatasetSaver(dataset_name=dataset_name, is_hetero=False, version=1)
+
+
+    print("dataset[0] = ",type(dataset[0]))
+    print("dataset[0] = ",dataset[0])
+
+    g = dataset[0].to_networkx()
+    print("cora g = ",g)
+    graph = dict()
+    graph['edge_index'] = np.array(
+        [(u, v) for (u, v, w) in g.edges]).transpose()
+    num_nodes = len(g.nodes)
+    graph['num_nodes'] = num_nodes
+    graph['node_feat'] = np.array(dataset[0].ndata['feat'])
+    saver.save_graph_list([graph])
+
+    saver.save_target_labels(
+        np.array(dataset[0].ndata['label']).reshape((num_nodes, 1)))
+
+    split_idx = dict()
+    split_idx['train'] = np.array(g.nodes)[dataset[0].ndata['train_mask']]
+    split_idx['valid'] = np.array(g.nodes)[dataset[0].ndata['val_mask']]
+    split_idx['test'] = np.array(g.nodes)[dataset[0].ndata['test_mask']]
+    print("train len = ",len(split_idx['train']))
+    print("valid len = ",len(split_idx['valid']))
+    print("test len = ",len(split_idx['test']))
+    saver.save_split(split_idx, split_name='std')
+
+    saver.copy_mapping_dir(mapping_path)
+
+    saver.save_task_info(task_type='classification',
+                         eval_metric='acc', num_classes=dataset.num_classes)
+    meta_dict = saver.get_meta_dict()
+
+    if is_sparse:
+        pyg_dataset = DglNodePropPredDataset(
+            dataset_name, meta_dict=meta_dict, transform=T.ToSparseTensor())
+    else:
+        pyg_dataset = DglNodePropPredDataset(dataset_name, meta_dict=meta_dict)
+
+    saver.zip()
+
+    return pyg_dataset

cora_splits/new_added_train_idx.csv

@@ -0,0 +1,52 @@
+0
+22
+131
+23
+627
+5
+725
+4
+1984
+924
+68
+478
+0
+2081
+643
+2
+6
+187
+860
+1174
+66
+947
+1429
+1422
+491
+902
+1335
+1321
+8
+2036
+1453
+1322
+1
+2105
+51
+1480
+2411
+1975
+1602
+2022
+2104
+1309
+1486
+1331
+1648
+1601
+1937
+1820
+1848
+2066
+1940
+1957