feat: feat: GraphLama contributions - GraphGPT improvements

examples/gcn/debug_edge_index.py

@@ -0,0 +1,21 @@
+import os
+os.environ['PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION'] = 'python'
+
+import sys
+sys.path.insert(0, '/home/zbs2/GammaGL_algo/论文复现/GraphLama/GammaGL/tensorlayerx_patch')
+
+import gammagl
+import tensorlayerx as tlx
+from gammagl.datasets import Planetoid
+from gammagl.utils.check import check_is_numpy
+
+print("Loading Cora dataset...")
+dataset = Planetoid(root='/tmp/cora', name='cora')
+graph = dataset[0]
+
+print(f"Edge index type: {type(graph.edge_index)}")
+print(f"Edge index shape: {graph.edge_index.shape}")
+print(f"Is tensor: {tlx.is_tensor(graph.edge_index)}")
+print(f"Is numpy: {check_is_numpy(graph.edge_index)}")
+print(f"Edge index dtype: {graph.edge_index.dtype}")
+print(f"First few edges: {graph.edge_index[:, :5]}")

examples/gcn/test_gammagl.py

@@ -0,0 +1,23 @@
+import os
+os.environ['PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION'] = 'python'
+
+import sys
+sys.path.insert(0, '/home/zbs2/GammaGL_algo/论文复现/GraphLama/GammaGL/tensorlayerx_patch')
+
+import gammagl
+import torch
+from gammagl.data import Graph
+
+print("GammaGL version:", getattr(gammagl, '__version__', 'Unknown'))
+print("PyTorch version:", torch.__version__)
+
+# Create a simple graph
+edge_index = torch.tensor([[0, 1, 1, 2], [1, 0, 2, 1]], dtype=torch.long)
+x = torch.randn(3, 4)
+graph = Graph(x=x, edge_index=edge_index)
+
+print("Graph created successfully:")
+print(f"  Nodes: {graph.num_nodes}")
+print(f"  Edges: {graph.num_edges}")
+print(f"  Node features shape: {graph.x.shape}")
+print("\nGammaGL basic functionality test passed!")

examples/graphgpt/eval.sh

@@ -1,13 +1,19 @@
-export PYTHONPATH=$(dirname $(dirname $(realpath $0))):$PYTHONPATH
+export PYTHONPATH=/home/zbs2/GammaGL_algo/论文复现/GraphLama/GammaGL
 # to fill in the following path to extract projector for the second tuning stage!
-output_model=/local/yy3/graphgpt/GraphGPT-7B-mix-all # path to the pre-trained model checkpoint
-datapath=/local/yy3/graphgpt/data/eval/cora_test_instruct_std.json # path to the instruction datset
-graph_data_path=/local/yy3/graphgpt/data/graph_data_all.pt # path to the graph data
-res_path=./output_stage_2_cora_nc # path to save the results
+output_model=/home/zbs2/GammaGL_algo/论文复现/GraphLama/GraphGPT_data/GraphGPT
+datapath=/home/zbs2/GammaGL_algo/论文复现/GraphLama/GraphGPT_data/GraphGPT_eval_dataset/cora_test_instruct_std.json
+graph_data_path=/home/zbs2/GammaGL_algo/论文复现/GraphLama/GraphGPT_data/graphdata_all/graph_data_all.pt
+pretrain_graph_model_path=/home/zbs2/GammaGL_algo/论文复现/GraphLama/GraphGPT_data/graph_cli_gt
+res_path=./output_cora_nc # path to save the results
 start_id=0
 end_id=20000 # total number of instructions to test
 num_gpus=1
 
-export CUDA_VISIBLE_DEVICES=2 # specify the GPU id
+export CUDA_VISIBLE_DEVICES=2
 
-python ./examples/graphgpt/graphgpt_trainer.py --model-name ${output_model}  --prompting_file ${datapath} --graph_data_path ${graph_data_path} --output_res_path ${res_path} --start_id ${start_id} --end_id ${end_id} --num_gpus ${num_gpus}
+python /home/zbs2/GammaGL_algo/论文复现/GraphLama/GammaGL/examples/graphgpt/graphgpt_trainer.py --model-name ${output_model}  --prompting_file ${datapath} --graph_data_path ${graph_data_path} --output_res_path ${res_path} --start_id ${start_id} --pretrain_graph_model_path ${pretrain_graph_model_path} --end_id ${end_id}
+
+# TLX 版本运行示例：
+# python /home/zbs2/GammaGL_algo/论文复现/GraphLama/GammaGL/examples/graphgpt/graphgpt_trainer_tlx.py --model-name ${output_model}  --prompting_file ${datapath} --graph_data_path ${graph_data_path} --output_res_path ${res_path} --start_id ${start_id} --pretrain_graph_model_path ${pretrain_graph_model_path} --end_id ${end_id}
+
+# TL_BACKEND="torch" nohup bash examples/graphgpt/eval.sh > log/test_graphgpt.out &

examples/graphgpt/graphgpt_eval.py

@@ -17,8 +17,6 @@
     "pubmed":{"Experimentally induced diabetes": 0, "Type 2 diabetes": 2, "Type 1 diabetes": 1}
 }
 
-
-
 data_list = []
 folder = 'output_stage_2_{}_nc'.format(args.dataset)
 for filename in os.listdir(folder):

examples/graphgpt/graphgpt_trainer.py

@@ -1,20 +1,21 @@
 import argparse
-from transformers import AutoTokenizer, AutoModelForCausalLM
 import torch
 import os
 from gammagl.utils.conversation import conv_templates, SeparatorStyle
 from gammagl.utils.gfm_utils import disable_torch_init, KeywordsStoppingCriteria
 from gammagl.utils.gfm_utils import DEFAULT_G_END_TOKEN, DEFAULT_G_START_TOKEN, DEFAULT_GRAPH_PATCH_TOKEN, DEFAULT_GRAPH_TOKEN, GRAPH_TOKEN_INDEX
 from transformers import CLIPVisionModel, CLIPImageProcessor, StoppingCriteria
 from gammagl.models.graphgpt import *
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import sys
 
 from torch_geometric.data import Data
 import json
 import copy
 from tqdm import tqdm
 import json
 import os.path as osp
-import ray
+
 
 os.environ['TL_BACKEND'] = 'torch'
 
@@ -44,50 +45,21 @@ def load_prompting_file(file_path):
 
 # def prepare_query(instruct_item): 
 
-
-def run_eval(args, num_gpus):
-    # split question file into num_gpus files
+def run_eval(args):
+    # Load and slice prompting file according to start/end
     prompt_file = load_prompting_file(args.prompting_file)
     args.end_id = min(args.end_id, len(prompt_file))
-    prompt_file = prompt_file[args.start_id:args.end_id]
-    chunk_size = len(prompt_file) // num_gpus
-    ans_handles = []
-    split_list = list(range(args.start_id, args.end_id, chunk_size))
-    idx_list = list(range(0, len(prompt_file), chunk_size))
-    if len(split_list) == num_gpus: 
-        split_list.append(args.end_id)
-        idx_list.append(len(prompt_file))
-    elif len(split_list) == num_gpus + 1: 
-        split_list[-1] = args.end_id
-        idx_list[-1] = len(prompt_file)
-    else: 
-        raise ValueError('error in the number of list')
-
-    if osp.exists(args.output_res_path) is False: 
-        os.mkdir(args.output_res_path)
-
-    for idx in range(len(idx_list) - 1):
-        start_idx = idx_list[idx]
-        end_idx = idx_list[idx + 1]
-
-        start_split = split_list[idx]
-        end_split = split_list[idx + 1]
-        ans_handles.append(
-            eval_model.remote(
-                args, prompt_file[start_idx:end_idx], start_split, end_split
-            )
-        )
+    prompt_slice = prompt_file[args.start_id:args.end_id]
 
-    ans_jsons = []
-    for ans_handle in ans_handles:
-        ans_jsons.extend(ray.get(ans_handle))
+    # Ensure output directory exists
+    os.makedirs(args.output_res_path, exist_ok=True)
 
-    # with open(args.output_res_path, "w") as ans_file:
-    #     for line in ans_jsons:
-    #         ans_file.write(json.dumps(line) + "\n")
+    # Directly run evaluation without Ray
+    print('--- evaluating without Ray ---')
+    ans_jsons = eval_model(args, prompt_slice, args.start_id, args.end_id)
+    print('--- done ---')
 
 
-@ray.remote(num_gpus=1)
 @torch.inference_mode()
 def eval_model(args, prompt_file, start_idx, end_idx):
     # load prompting file
@@ -115,7 +87,17 @@ def eval_model(args, prompt_file, start_idx, end_idx):
     # TODO: add graph tower
     # if graph_tower.device.type == 'meta':
     #     print('meta')
-    clip_graph, args_graph= load_model_pretrained(CLIP, model.config.pretrain_graph_model_path)
+    # Resolve graph tower pretrain path (supports absolute or relative to model dir)
+    cfg_pretrain_path = getattr(model.config, "pretrain_graph_model_path", None)
+    if args.pretrain_graph_model_path is not None:
+        pretrain_path = args.pretrain_graph_model_path
+    else:
+        if cfg_pretrain_path is None:
+            raise ValueError("pretrain_graph_model_path not set; please provide --pretrain_graph_model_path")
+        pretrain_path = cfg_pretrain_path if osp.isabs(cfg_pretrain_path) else osp.join(args.model_name, cfg_pretrain_path)
+    assert osp.exists(osp.join(pretrain_path, 'config.json')), f'config.json missing at {pretrain_path}'
+
+    clip_graph, args_graph = load_model_pretrained(CLIP, pretrain_path)
     graph_tower = graph_transformer(args_graph)
     graph_tower = transfer_param_tograph(clip_graph, graph_tower)
 
@@ -208,25 +190,23 @@ def eval_model(args, prompt_file, start_idx, end_idx):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model-name", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-name", type=str, default="/home/zbs2/GammaGL_algo/论文复现/GraphLama/GraphGPT_data/GraphGPT")
     # parser.add_argument("--image-file", type=str, required=True)
     # parser.add_argument("--query", type=str, required=True)
     parser.add_argument("--prompting_file", type=str, default=None)
     parser.add_argument("--conv-mode", type=str, default=None)
     parser.add_argument("--graph_data_path", type=str, default=None)
 
     parser.add_argument("--output_res_path", type=str, default=None)
-    parser.add_argument("--num_gpus", type=int, default=4)
 
     parser.add_argument("--start_id", type=int, default=0)
     parser.add_argument("--end_id", type=int, default=20567)
+    parser.add_argument("--pretrain_graph_model_path", type=str, default=None,
+                        help="Path to graph tower pretrain directory containing config.json (overrides model config)")
 
     args = parser.parse_args()
 
-    # eval_model(args)
-
-    ray.init()
-    run_eval(args, args.num_gpus)
-
+    # Run in single process without Ray
+    run_eval(args)
 
 # protobuf             4.22.3

examples/graphgpt/test.py

@@ -0,0 +1,7 @@
+# import argparse
+import torch
+import os
+from gammagl.utils.conversation import conv_templates, SeparatorStyle
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+print('导入成功')

examples/graphlama/README.md

@@ -0,0 +1,31 @@
+# GraphLAMA: Enabling Efficient Adaptation of Graph Language Models with Limited Annotations
+* Paper link: https://arxiv.org/pdf/2506.21559
+* Author's code repo: https://github.com/BUPT-GAMMA/GraphLAMA.git
+
+# How to Run
+
+* First, follow the [GraphLAMA original repo](https://github.com/BUPT-GAMMA/GraphLAMA.git) to install all required packages where will tell you to follow the instructions to 
+
+* Then download all required datasets and pretrained checkpoints, and fill their path into corresponding values in eval.sh
+
+# Dataset Statics
+| Dataset | # Nodes | # Edges |  # Classes | 
+| :-------: | :-------: | :------: | :------: |
+| Cora | 25,120 | 182,280 | 70 |
+| PubMed | 19,717 | 44,338 | 3 |
+| ogb-arxiv | 169,343 | 1,166,243 | 40 |
+
+# Files Description
+* graphgpt_trainer.py: the trainer of graphgpt, inference stage
+* graphgpt_eval.py: run this to evaluate 
+
+# Results
+```bash
+# run inference
+TL_BACKEND="torch" nohup bash examples/graphgpt/eval.sh > log/test_graphgpt.out &
+# run evaluation
+python examples/graphgpt/graphgpt_eval.py --dataset cora
+```
+| Dataset | Paper | Our(torch) |
+| :-------: | :-------: | :------: |
+| Cora | 0.1501 | 0.1451 | 

examples/graphlama/eval.sh

@@ -0,0 +1,14 @@
+export PYTHONPATH=/home/zbs2/GammaGL_algo/论文复现/GraphLama/GammaGL
+# Fill in the following path to extract projector for the tuning stage following the instructions.
+output_model=/path/to/GraphLama/weights
+datapath=path/to/cora_test_instruct_std.json
+graph_data_path=path/to/graph_data_all.pt
+pretrain_graph_model_path=path/to/graph_cli_gt # remember to download arixv ones
+res_path=./output_cora_nc # path to save the results
+start_id=0
+end_id=20000 # total number of instructions to test
+num_gpus=1
+
+export CUDA_VISIBLE_DEVICES=2 # define the gpu id to use
+
+python graphlama_trainer.py --model-name ${output_model}  --prompting_file ${datapath} --graph_data_path ${graph_data_path} --output_res_path ${res_path} --start_id ${start_id} --pretrain_graph_model_path ${pretrain_graph_model_path} --end_id ${end_id}

examples/graphlama/graphgpt_eval_tlx.py

@@ -0,0 +1,109 @@
+import json
+import os.path as osp
+import os
+import re
+import pandas as pd
+from tqdm import tqdm
+from sklearn.metrics import classification_report
+
+import tensorlayerx as tlx
+
+import argparse
+parser = argparse.ArgumentParser()
+parser.add_argument('--dataset', type=str, default='arxiv')
+args = parser.parse_args()
+
+label_to_idx = {
+    "cora":{"databases, object oriented": 29, "operating systems, memory management": 59, "data structures  algorithms and theory, quantum computing": 24, "artificial intelligence, planning": 13, "artificial intelligence, knowledge representation": 4, "artificial intelligence, data mining": 1, "artificial intelligence, vision and pattern recognition": 17, "artificial intelligence, machine learning, case-based": 5, "artificial intelligence, agents": 0, "artificial intelligence, machine learning, probabilistic methods": 8, "encryption and compression, security": 36, "operating systems, distributed": 57, "human computer interaction, interface design": 46, "artificial intelligence, machine learning, genetic algorithms": 6, "human computer interaction, graphics and virtual reality": 45, "artificial intelligence, machine learning, rule learning": 10, "programming, functional": 63, "programming, object oriented": 67, "encryption and compression, encryption": 35, "databases, performance": 30, "networking, protocols": 54, "data structures  algorithms and theory, randomized": 25, "data structures  algorithms and theory, formal languages": 20, "data structures  algorithms and theory, parallel": 23, "programming, software development": 69, "programming, compiler design": 61, "artificial intelligence, machine learning, theory": 11, "artificial intelligence, machine learning, neural networks": 7, "programming, logic": 66, "databases, relational": 32, "information retrieval, retrieval": 52, "programming, debugging": 62, "networking, wireless": 56, "artificial intelligence, theorem proving": 16, "databases, temporal": 33, "encryption and compression, compression": 34, "information retrieval, filtering": 51, "data structures  algorithms and theory, computational complexity": 18, "programming, garbage collection": 64, "artificial intelligence, machine learning, reinforcement learning": 9, "human computer interaction, multimedia": 47, "hardware and architecture, vlsi": 43, "artificial intelligence, nlp": 12, "hardware and architecture, microprogramming": 42, "operating systems, fault tolerance": 58, "programming, java": 65, "operating systems, realtime": 60, "human computer interaction, cooperative": 44, "artificial intelligence, speech": 15, "databases, deductive": 28, "artificial intelligence, robotics": 14, "data structures  algorithms and theory, logic": 22, "networking, routing": 55, "hardware and architecture, logic design": 40, "hardware and architecture, distributed architectures": 37, "data structures  algorithms and theory, hashing": 21, "programming, semantics": 68, "artificial intelligence, games and search": 3, "databases, concurrency": 27, "data structures  algorithms and theory, sorting": 26, "human computer interaction, wearable computers": 48, "information retrieval, digital library": 49, "artificial intelligence, expert systems": 2, "information retrieval, extraction": 50, "data structures  algorithms and theory, computational geometry": 19, "databases, query evaluation": 31, "networking, internet": 53, "hardware and architecture, memory structures": 41, "hardware and architecture, high performance computing": 38, "hardware and architecture, input output and storage": 39},
+    "pubmed":{"Experimentally induced diabetes": 0, "Type 2 diabetes": 2, "Type 1 diabetes": 1}
+}
+
+data_list = []
+folder = 'output_stage_2_{}_nc'.format(args.dataset)
+for filename in os.listdir(folder):
+    if filename.endswith('.json'):
+        file_path = os.path.join(folder, filename)
+        with open(file_path, 'r') as f:
+            data = json.load(f)
+            data_list.extend(data)
+
+print(data_list[1])
+
+# 加载图数据（原始为 .pt 文件），使用 torch 读取后转为 TLX Tensor
+import torch as th
+graph_data = th.load('/local/yy3/graphgpt/data/graph_data_all.pt')[args.dataset]
+labels_th = graph_data.y
+labels = tlx.convert_to_tensor(labels_th.detach().cpu().numpy(), dtype=tlx.int64)
+
+def cal_map():
+    label_dict = {}
+    if args.dataset == "arxiv":
+        df = pd.read_csv(os.path.expanduser('~/datasets/OGB/ogbn_arxiv/mapping/labelidx2arxivcategeory.csv.gz'), compression='gzip')
+        for index, line in df.iterrows():
+            lb = line['arxiv category'].split(' ')[-1]
+            lb_new = 'cs.' + lb.upper()
+            label_dict[lb_new] = line['label idx']
+    else:
+        label_dict = label_to_idx[args.dataset]
+    return label_dict
+
+class_map = cal_map()
+
+inverse_class_map = {}
+for lb, lb_id in class_map.items():
+    inverse_class_map[lb_id] = lb
+
+
+pattern = r"cs\.[A-Z]{2}"
+
+
+topk = 3
+
+correct = 0
+total = len(data_list)
+
+trues = []
+preds = []
+
+for instruct_item in tqdm(data_list):
+    nid = instruct_item['node_idx']
+    gpt_res = instruct_item['res']
+
+
+    true_y = labels[nid]
+    true_y_val = int(tlx.ops.convert_to_numpy(true_y))
+
+    pred_y = []
+    if args.dataset == "arxiv":
+        matches = list(set(re.findall(pattern, gpt_res))) # pred
+        sorted_matches = sorted(matches, key=lambda x: gpt_res.index(x))
+        for m in sorted_matches:
+            try:
+                pred_y.append(class_map[m])
+            except:
+                pass
+        try:
+            # print(sorted_matches)
+            preds.append(pred_y[0])
+        except:
+            preds.append(-1)
+    else:
+        for lb, lb_id in class_map.items():
+            if lb in gpt_res:
+                pred_y.append(lb_id)
+        try:
+            # print(sorted_matches)
+            preds.append(pred_y[0])
+        except:
+            preds.append(-1)
+    trues.append(true_y_val)
+    res_tmp = 1 if true_y_val in pred_y[:topk] else 0
+    correct = correct + 1 if true_y_val in pred_y[:topk] else correct
+
+acc = correct / total
+
+print("Accuracy:", acc)
+
+report = classification_report(trues, preds, digits=6)
+
+print(report)