Add G-retriever (GNN+LLM) example (#9167)

1. #9462
2. #9480
3. #9481
4. **->** #9167

---

repro:
Latest NVIDIA PyG container
+
`git config --global credential.helper store; huggingface-cli login; cd
/opt/pyg; pip uninstall -y torch-geometric; rm -rf pytorch_geometric;
git clone -b gnn-llm-model-integration
https://github.com/pyg-team/pytorch_geometric.git; cd
/opt/pyg/pytorch_geometric; pip install .; pip install peft datasets
transformers pcst_fast sentencepiece; python3
examples/llm_plus_gnn/g_retriever.py`

old PR: #9154

note: pure cpu is 220x slower than pure GPU using a single Grace Hopper
(for llama-7b)

info:
tried gemma, performs worse in all train/val/test metrics. most likely
needs some tuning, will leave this as future work as part of the
community sprint to try many LLM and GNN combos and tune them. Therefore
keeping the default llama2

the new gemma-v2 is also much worse than llama2

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Akihiro Nitta <[email protected]>
Co-authored-by: rusty1s <[email protected]>

CHANGELOG.md

@@ -8,7 +8,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 ### Added
 
 - Added the `WebQSPDataset` dataset ([#9481](https://github.com/pyg-team/pytorch_geometric/pull/9481))
-- Added the `GRetriever` model ([#9480](https://github.com/pyg-team/pytorch_geometric/pull/9480))
+- Added the `GRetriever` model and an example ([#9480](https://github.com/pyg-team/pytorch_geometric/pull/9480), [#9167](https://github.com/pyg-team/pytorch_geometric/pull/9167))
 - Added the `ClusterPooling` layer ([#9627](https://github.com/pyg-team/pytorch_geometric/pull/9627))
 - Added the `LinkPredMRR` metric ([#9632](https://github.com/pyg-team/pytorch_geometric/pull/9632))
 - Added PyTorch 2.4 support ([#9594](https://github.com/pyg-team/pytorch_geometric/pull/9594))

examples/llm/README.md

@@ -1,5 +1,5 @@
 # Examples for Co-training LLMs and GNNs
 
-| Example | Description |
-| ------- | ----------- |
-|  |  |
+| Example  | Description  |
+| ------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| [`g_retriever.py`](./g_retriever.py) | Example for Retrieval-Augmented Generation (RAG) w/ GNN+LLM by co-training `LLAMA2` with `GAT` for answering questions based on knowledge graph information |

examples/llm/g_retriever.py

@@ -0,0 +1,272 @@
+"""This example implements the G-Retriever model
+(https://arxiv.org/abs/2402.07630) using PyG.
+
+G-Retriever significantly reduces hallucinations by 54% compared to the
+stand-alone LLM baseline.
+
+Requirements:
+`pip install datasets transformers pcst_fast sentencepiece accelerate`
+"""
+import argparse
+import math
+import os.path as osp
+import re
+import time
+
+import pandas as pd
+import torch
+from torch import Tensor
+from torch.nn.utils import clip_grad_norm_
+from tqdm import tqdm
+
+from torch_geometric import seed_everything
+from torch_geometric.datasets import WebQSPDataset
+from torch_geometric.loader import DataLoader
+from torch_geometric.nn.models import GAT, GRetriever
+from torch_geometric.nn.nlp import LLM
+
+
+def compute_metrics(eval_output):
+ df = pd.concat([pd.DataFrame(d) for d in eval_output])
+ all_hit = []
+ all_precision = []
+ all_recall = []
+ all_f1 = []
+
+ for pred, label in zip(df.pred.tolist(), df.label.tolist()):
+ try:
+ pred = pred.split('[/s]')[0].strip().split('|')
+ hit = re.findall(pred[0], label)
+ all_hit.append(len(hit) > 0)
+
+ label = label.split('|')
+ matches = set(pred).intersection(set(label))
+ precision = len(matches) / len(set(label))
+ recall = len(matches) / len(set(pred))
+ if recall + precision == 0:
+ f1 = 0
+ else:
+ f1 = 2 * precision * recall / (precision + recall)
+
+ all_precision.append(precision)
+ all_recall.append(recall)
+ all_f1.append(f1)
+
+ except Exception as e:
+ print(f'Label: {label}')
+ print(f'Pred: {pred}')
+ print(f'Exception: {e}')
+ print('------------------')
+
+ hit = sum(all_hit) / len(all_hit)
+ precision = sum(all_precision) / len(all_precision)
+ recall = sum(all_recall) / len(all_recall)
+ f1 = sum(all_f1) / len(all_f1)
+
+ print(f'Hit: {hit:.4f}')
+ print(f'Precision: {precision:.4f}')
+ print(f'Recall: {recall:.4f}')
+ print(f'F1: {f1:.4f}')
+
+
+def save_params_dict(model, save_path):
+ state_dict = model.state_dict()
+ param_grad_dict = {
+ k: v.requires_grad
+ for (k, v) in model.named_parameters()
+ }
+ for k in list(state_dict.keys()):
+ if k in param_grad_dict.keys() and not param_grad_dict[k]:
+ del state_dict[k] # Delete parameters that do not require gradient
+ torch.save(state_dict, save_path)
+
+
+def load_params_dict(model, save_path):
+ state_dict = torch.load(save_path)
+ model.load_state_dict(state_dict)
+ return model
+
+
+def get_loss(model, batch, model_save_name) -> Tensor:
+ if model_save_name == 'llm':
+ return model(batch.question, batch.label, batch.desc)
+ else:
+ return model(batch.question, batch.x, batch.edge_index, batch.batch,
+ batch.label, batch.edge_attr, batch.desc)
+
+
+def inference_step(model, batch, model_save_name):
+ if model_save_name == 'llm':
+ return model.inference(batch.question, batch.desc)
+ else:
+ return model.inference(batch.question, batch.x, batch.edge_index,
+ batch.batch, batch.edge_attr, batch.desc)
+
+
+def train(
+ num_epochs,
+ hidden_channels,
+ num_gnn_layers,
+ batch_size,
+ eval_batch_size,
+ lr,
+ checkpointing=False,
+ tiny_llama=False,
+):
+ def adjust_learning_rate(param_group, LR, epoch):
+ # Decay the learning rate with half-cycle cosine after warmup
+ min_lr = 5e-6
+ warmup_epochs = 1
+ if epoch < warmup_epochs:
+ lr = LR
+ else:
+ lr = min_lr + (LR - min_lr) * 0.5 * (
+ 1.0 + math.cos(math.pi * (epoch - warmup_epochs) /
+ (num_epochs - warmup_epochs)))
+ param_group['lr'] = lr
+ return lr
+
+ start_time = time.time()
+ path = osp.dirname(osp.realpath(__file__))
+ path = osp.join(path, '..', '..', 'data', 'WebQSPDataset')
+ train_dataset = WebQSPDataset(path, split='train')
+ val_dataset = WebQSPDataset(path, split='val')
+ test_dataset = WebQSPDataset(path, split='test')
+
+ seed_everything(42)
+
+ train_loader = DataLoader(train_dataset, batch_size=batch_size,
+ drop_last=True, pin_memory=True, shuffle=True)
+ val_loader = DataLoader(val_dataset, batch_size=eval_batch_size,
+ drop_last=False, pin_memory=True, shuffle=False)
+ test_loader = DataLoader(test_dataset, batch_size=eval_batch_size,
+ drop_last=False, pin_memory=True, shuffle=False)
+
+ gnn = GAT(
+ in_channels=1024,
+ hidden_channels=hidden_channels,
+ out_channels=1024,
+ num_layers=num_gnn_layers,
+ heads=4,
+ )
+ if tiny_llama:
+ llm = LLM(
+ model_name='TinyLlama/TinyLlama-1.1B-Chat-v0.1',
+ num_params=1,
+ )
+ model = GRetriever(llm=llm, gnn=gnn, mlp_out_channels=2048)
+ else:
+ llm = LLM(model_name='meta-llama/Llama-2-7b-chat-hf', num_params=7)
+ model = GRetriever(llm=llm, gnn=gnn)
+
+ model_save_name = 'gnn_llm' if num_gnn_layers is not None else 'llm'
+ params = [p for _, p in model.named_parameters() if p.requires_grad]
+ optimizer = torch.optim.AdamW([
+ {
+ 'params': params,
+ 'lr': lr,
+ 'weight_decay': 0.05
+ },
+ ], betas=(0.9, 0.95))
+ grad_steps = 2
+
+ best_epoch = 0
+ best_val_loss = float('inf')
+ for epoch in range(num_epochs):
+ model.train()
+ epoch_loss = 0
+ if epoch == 0:
+ print(f"Total Preparation Time: {time.time() - start_time:2f}s")
+ start_time = time.time()
+ print("Training beginning...")
+ epoch_str = f'Epoch: {epoch + 1}|{num_epochs}'
+ loader = tqdm(train_loader, desc=epoch_str)
+ for step, batch in enumerate(loader):
+ optimizer.zero_grad()
+ loss = get_loss(model, batch, model_save_name)
+ loss.backward()
+
+ clip_grad_norm_(optimizer.param_groups[0]['params'], 0.1)
+
+ if (step + 1) % grad_steps == 0:
+ adjust_learning_rate(optimizer.param_groups[0], lr,
+ step / len(train_loader) + epoch)
+
+ optimizer.step()
+ epoch_loss = epoch_loss + float(loss)
+
+ if (step + 1) % grad_steps == 0:
+ lr = optimizer.param_groups[0]['lr']
+ train_loss = epoch_loss / len(train_loader)
+ print(epoch_str + f', Train Loss: {train_loss:4f}')
+
+ val_loss = 0
+ eval_output = []
+ model.eval()
+ with torch.no_grad():
+ for step, batch in enumerate(val_loader):
+ loss = get_loss(model, batch, model_save_name)
+ val_loss += loss.item()
+ val_loss = val_loss / len(val_loader)
+ print(epoch_str + f", Val Loss: {val_loss:4f}")
+ if checkpointing and val_loss < best_val_loss:
+ print("Checkpointing best model...")
+ best_val_loss = val_loss
+ best_epoch = epoch
+ save_params_dict(model, f'{model_save_name}_best_val_loss_ckpt.pt')
+ torch.cuda.empty_cache()
+ torch.cuda.reset_max_memory_allocated()
+
+ if checkpointing and best_epoch != num_epochs - 1:
+ print("Loading best checkpoint...")
+ model = load_params_dict(
+ model,
+ f'{model_save_name}_best_val_loss_ckpt.pt',
+ )
+
+ model.eval()
+ eval_output = []
+ print("Final evaluation...")
+ progress_bar_test = tqdm(range(len(test_loader)))
+ for step, batch in enumerate(test_loader):
+ with torch.no_grad():
+ pred = inference_step(model, batch, model_save_name)
+ eval_data = {
+ 'pred': pred,
+ 'question': batch.question,
+ 'desc': batch.desc,
+ 'label': batch.label
+ }
+ eval_output.append(eval_data)
+ progress_bar_test.update(1)
+
+ compute_metrics(eval_output)
+ print(f"Total Training Time: {time.time() - start_time:2f}s")
+ save_params_dict(model, f'{model_save_name}.pt')
+ torch.save(eval_output, f'{model_save_name}_eval_outs.pt')
+
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser()
+ parser.add_argument('--gnn_hidden_channels', type=int, default=1024)
+ parser.add_argument('--num_gnn_layers', type=int, default=4)
+ parser.add_argument('--lr', type=float, default=1e-5)
+ parser.add_argument('--epochs', type=int, default=2)
+ parser.add_argument('--batch_size', type=int, default=8)
+ parser.add_argument('--eval_batch_size', type=int, default=16)
+ parser.add_argument('--checkpointing', action='store_true')
+ parser.add_argument('--tiny_llama', action='store_true')
+ args = parser.parse_args()
+
+ start_time = time.time()
+ train(
+ args.epochs,
+ args.gnn_hidden_channels,
+ args.num_gnn_layers,
+ args.batch_size,
+ args.eval_batch_size,
+ args.lr,
+ checkpointing=args.checkpointing,
+ tiny_llama=args.tiny_llama,
+ )
+ print(f"Total Time: {time.time() - start_time:2f}s")

torch_geometric/nn/models/g_retriever.py

@@ -3,7 +3,6 @@
 import torch
 from torch import Tensor
 
-from torch_geometric.nn.models import GAT
 from torch_geometric.nn.nlp.llm import BOS, LLM, MAX_NEW_TOKENS
 from torch_geometric.utils import scatter
 
@@ -43,7 +42,6 @@ def __init__(
  llm: LLM,
  gnn: torch.nn.Module,
  use_lora: bool = False,
- gnn_to_use=GAT,
  mlp_out_channels: int = 4096,
  ) -> None:
  super().__init__()
@@ -126,7 +124,15 @@ def forward(
  """
  x = self.encode(x, edge_index, batch, edge_attr)
  x = self.projector(x)
- xs = x.split(x.size(0), dim=0)
+ xs = x.split(1, dim=0)
+
+ # Handle questions without node features:
+ batch_unique = batch.unique()
+ batch_size = len(question)
+ if len(batch_unique) < batch_size:
+ xs = [
+ xs[i] if i in batch_unique else None for i in range(batch_size)
+ ]
 
  (
  inputs_embeds,
@@ -174,7 +180,15 @@ def inference(
  """
  x = self.encode(x, edge_index, batch, edge_attr)
  x = self.projector(x)
- xs = x.split(x.size(0), dim=0)
+ xs = x.split(1, dim=0)
+
+ # Handle questions without node features:
+ batch_unique = batch.unique()
+ batch_size = len(question)
+ if len(batch_unique) < batch_size:
+ xs = [
+ xs[i] if i in batch_unique else None for i in range(batch_size)
+ ]
 
  inputs_embeds, attention_mask, _ = self.llm._get_embeds(
  question, additional_text_context, xs)

torch_geometric/nn/nlp/llm.py

@@ -1,3 +1,4 @@
+import warnings
 from contextlib import nullcontext
 from typing import Any, Dict, List, Optional
 
@@ -85,6 +86,7 @@ def __init__(
  self.word_embedding = self.llm.model.get_input_embeddings()
 
  if 'max_memory' not in kwargs: # Pure CPU:
+ warnings.warn("LLM is being used on CPU, which may be slow")
  self.device = torch.device('cpu')
  self.autocast_context = nullcontext()
  else: