Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

#

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

Simplest version of dumping train config #558

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

Simplest version of dumping train config #558

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
125 changes: 125 additions & 0 deletions examples/kge-transe.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,125 @@
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"id": "b68543d5e71ceeb2",
"metadata": {},
"outputs": [],
"source": [
"import os\n",
"from graphdatascience import GraphDataScience"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "e685a47b61f968ef",
"metadata": {},
"outputs": [],
"source": [
"NEO4J_URI = \"bolt://localhost:7687\"\n",
"NEO4J_DB = \"neo4j\""
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "initial_id",
"metadata": {},
"outputs": [],
"source": [
"if os.environ.get(\"NEO4J_USER\") and os.environ.get(\"NEO4J_PASSWORD\"):\n",
" NEO4J_AUTH = (\n",
" os.environ.get(\"NEO4J_USER\"),\n",
" os.environ.get(\"NEO4J_PASSWORD\"),\n",
" )\n",
"gds = GraphDataScience(NEO4J_URI, auth=NEO4J_AUTH, database=NEO4J_DB)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "a14f06aebe1ed34c",
"metadata": {},
"outputs": [],
"source": [
"_ = gds.run_cypher(\n",
" \"\"\"\n",
" CREATE\n",
" (dan:Person {name: 'Dan'}),\n",
" (annie:Person {name: 'Annie'}),\n",
" (matt:Person {name: 'Matt'}),\n",
" (jeff:Person {name: 'Jeff'}),\n",
" (brie:Person {name: 'Brie'}),\n",
" (elsa:Person {name: 'Elsa'}),\n",
"\n",
" (cookies:Product {name: 'Cookies'}),\n",
" (tomatoes:Product {name: 'Tomatoes'}),\n",
" (cucumber:Product {name: 'Cucumber'}),\n",
" (celery:Product {name: 'Celery'}),\n",
" (kale:Product {name: 'Kale'}),\n",
" (milk:Product {name: 'Milk'}),\n",
" (chocolate:Product {name: 'Chocolate'}),\n",
"\n",
" (dan)-[:BUYS {amount: 1.2}]->(cookies),\n",
" (dan)-[:BUYS {amount: 3.2}]->(milk),\n",
" (dan)-[:BUYS {amount: 2.2}]->(chocolate),\n",
"\n",
" (annie)-[:BUYS {amount: 1.2}]->(cucumber),\n",
" (annie)-[:BUYS {amount: 3.2}]->(milk),\n",
" (annie)-[:BUYS {amount: 3.2}]->(tomatoes),\n",
"\n",
" (matt)-[:BUYS {amount: 3}]->(tomatoes),\n",
" (matt)-[:BUYS {amount: 2}]->(kale),\n",
" (matt)-[:BUYS {amount: 1}]->(cucumber),\n",
"\n",
" (jeff)-[:BUYS {amount: 3}]->(cookies),\n",
" (jeff)-[:BUYS {amount: 2}]->(milk),\n",
"\n",
" (brie)-[:BUYS {amount: 1}]->(tomatoes),\n",
" (brie)-[:BUYS {amount: 2}]->(milk),\n",
" (brie)-[:BUYS {amount: 2}]->(kale),\n",
" (brie)-[:BUYS {amount: 3}]->(cucumber),\n",
" (brie)-[:BUYS {amount: 0.3}]->(celery),\n",
"\n",
" (elsa)-[:BUYS {amount: 3}]->(chocolate),\n",
" (elsa)-[:BUYS {amount: 3}]->(milk)\n",
" \"\"\"\n",
")\n",
"node_projection = [\"Person\", \"Product\"]\n",
"relationship_projection = {\"BUYS\": {\"orientation\": \"UNDIRECTED\", \"properties\": \"amount\"}}\n",
"G, result = gds.graph.project(\"purchases222\", node_projection, relationship_projection)\n",
"print(f\"The projection took {result['projectMillis']} ms\")\n",
"print(f\"Graph '{G.name()}' node count: {G.node_count()}\")\n",
"print(f\"Graph '{G.name()}' node labels: {G.node_labels()}\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "e049480efa34e8ca",
"metadata": {},
"outputs": [],
"source": [
"gds.model.transe.train(\n",
" G,\n",
" proportions=[0.8, 0.1, 0.1],\n",
" embedding_dimension=50,\n",
" batch_size=512,\n",
" epochs=100,\n",
" optimizer=\"Adam\",\n",
" optimizer_kwargs={\"lr\": 0.01, \"weight_decay\": 5e-4},\n",
" # loss\n",
")"
]
}
],
"metadata": {
"language_info": {
"name": "python"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
29 changes: 29 additions & 0 deletions graphdatascience/model/model_proc_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,4 @@
import json
from typing import Any, Dict, List, Optional, Tuple

from pandas import DataFrame, Series
Expand Down Expand Up @@ -45,6 +46,34 @@ def create(
relationship_type_embeddings,
)

@compatible_with("train", min_inclusive=ServerVersion(2, 5, 0))
@client_only_endpoint("gds.model.transe")
def train(self,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm a little bit confused about how the predict-only TransE we've had previously fits into this. They kind of seem like different things. Also I wonder if it's better if we have a gds.kge.train call instead of one dedicated to each particular scoring function

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think that we are adding training step in addition to existing prediction one. Maybe it's worth to have a fully separate API like gds.kge.train or gds.kge.transe.train.

Personally I prefer to mention scoring method in a function name, like gds.kge.transe.train, not gds.kge.train. Because KGE algorithms are different and supposed to catch different relationship properties.

Our API proposal has gds.model.transe.train call, that's why I wrote it that way.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok! Fair enough :) I do still prefer gds.kge.train: I think it makes sense since the embeddings are what are trained, not TransE which is a scoring function. And since all KGE algos are the same (in terms of eg. hyperparameters) except for the scoring function, I think it would make sense to group them in the API for simplicity, sharing the same docs, etc. In that sense scoring function is just another hyperparam I think, and one may even want to use an ensemble of them. I like how pyKEEN designed their API

G: Graph,
proportions: list,
embedding_dimension: int,
batch_size: int,
epochs: int,
optimizer: str,
optimizer_kwargs: dict,
# loss: str
) -> int:
config = {'scoring_function': 'TransE',
'proportions': proportions,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I call this split_ratios in the python runtime. I think "ratio" is a more common term to refer to this (also what pyKEEN use for example)

'embedding_dimension': embedding_dimension,
'num_epochs': epochs,
'graph_name': G.name(),
'batch_size': batch_size,
'optimizer': optimizer,
'optimizer_kwargs': optimizer_kwargs,
# 'loss': loss,
}
config_path = "/tmp/kge-train-config-dump.json"
print('Dumped to ' + config_path)
config_file = open(config_path, "w")

json.dump(config, config_file)
return 0

class ModelProcRunner(ModelResolver):
@client_only_endpoint("gds.model")
Expand Down