Merge branch 'external_dataset' of https://github.com/snap-stanford/ogb into external_dataset

Author: weihua916

ogb/io/README.md

@@ -5,7 +5,7 @@ The `DatasetSaver` class allows external contributors to prepare their datasets
 Below is the quick example of how to use `DatasetSaver` class, where we focus on graph property prediction datasets.
 Please follow the steps below **in the exact order** to generate final dataset files.
 
-## 0. Constructor
+## 1. Constructor
 Create a constructor of `DatasetSaver`. `dataset_name` needs to follow OGB convention and start from either `ogbn-`, `ogbl-`, or `ogbg-`. `is_hetero` is `True` for heterogeneous graphs, and `version` indicates the dataset version.
 ```python
 from ogb.io import DatasetSaver
@@ -18,13 +18,13 @@ dataset_name = 'ogbg-toy'
 saver = DatasetSaver(dataset_name = dataset_name, is_hetero = False, version = 1)
 ```
 
-## 1. Saving graph list
+## 2. Saving graph list
 
 Create `graph_list`, storing your graph objects, and call `saver.save_graph_list(graph_list)`. 
 
 Graph objects are dictionaries containing the following keys.
 ### Homogeneous graph:
-- `edge_index` (necessary): `numpy.ndarray` of shape `(2, num_edges)`. Includes bidirectional edges explicitly if graphs are undirected.
+- `edge_index` (necessary): `numpy.ndarray` of shape `(2, num_edges)`. Please include bidirectional edges explicitly if graphs are undirected.
 - `num_nodes` (necessary): `int`, denoting the number of nodes in the graph.
 - `node_feat` (optional): `numpy.ndarray` of shape `(num_nodes, node_feat_dim)`.
 - `edge_feat` (optional): `numpy.ndarray` of shape `(num_edges, edge_feat_dim)`. 
@@ -33,7 +33,7 @@ Graph objects are dictionaries containing the following keys.
 - `edge_index_dict` (necessary): Dictionary mapping triplets `(head type, relation type, tail type)` to `edge_index`
 - `num_nodes_dict` (necessary): Dictionary mapping `entity type` to `num_nodes`.
 - `node_feat_dict` (optional): Dictionary mapping `entity type` to `node_feat`.
-- `edge_feat_dict` (optional): Dictionary mapping `entity type` to `edge_feat`.
+- `edge_feat_dict` (optional): Dictionary mapping `(head type, relation type, tail type)` to `edge_feat`.
 
 ```python
 # generate random graphs with node and edge features
@@ -89,17 +89,18 @@ saver.copy_mapping_dir(mapping_path)
 ```
 
 ## 6. Saving task information
-Save task information by calling `saver.save_task_info(task_type = 'classification', eval_metric = 'acc', num_classes = num_classes)`.
+Save task information by calling `saver.save_task_info(task_type, eval_metric, num_classes = num_classes)`.
+`eval_metric` is used to call `Evaluator` (c.f. [here](https://github.com/snap-stanford/ogb/blob/master/ogb/graphproppred/evaluate.py)). You can reuse one of the existing metrics, or you can implement your own by creating a pull request.
 ```python
 saver.save_task_info(task_type = 'classification', eval_metric = 'acc', num_classes = num_classes)
 ```
 
-## 7. Get meta information dictionary
+## 7. Getting meta information dictionary
 ```python
 meta_dict = saver.get_meta_dict()
 ```
 
-## 8. Test the dataset object
+## 8. Testing the dataset object
 Test the OGB dataset object to confirm it is working as you expect. You can similarly test Pytorch Geometric and DGL dataset objects.
 ```python
 from ogb.graphproppred import GraphPropPredDataset
@@ -110,11 +111,11 @@ print(dataset[0])
 print(dataset.get_idx_split())
 ```
 
-## 10. Zip and clean up
+## 9. Zipping and cleaning up
 ```python
 saver.zip()
 saver.cleanup()
 ```
 
-## 11. Send us two files
-In this example, under `submission_ogbg_toy/`, you will find two files `meta_dict.pt` and `toy.zip`. Please send them to us.
+## 10. Sending us two files
+In this example, under `submission_ogbg_toy/`, you will find two files `meta_dict.pt` and `toy.zip`. Please send them to us.