Pdf2 model task (#33)

* add needed dependency

* add extract_model_and_implement pipeline

* add merge_file_to_model_dict_to_model_dict

* implement `rdagent\app\model_implementation\eval.py`

* Running benchmark

* refine import

---------

Co-authored-by: Young <afe.young@gmail.com>

rdagent/app/model_extraction_and_implementation/model_extraction_and_implementation.py

@@ -0,0 +1,16 @@
+# %%
+from dotenv import load_dotenv
+from rdagent.components.task_implementation.model_implementation.one_shot import ModelTaskGen
+from rdagent.components.task_implementation.model_implementation.task_extraction import ModelImplementationTaskLoaderFromPDFfiles
+
+
+
+def extract_models_and_implement(report_file_path: str="../test_doc") -> None:
+    factor_tasks = ModelImplementationTaskLoaderFromPDFfiles().load(report_file_path)
+    implementation_result = ModelTaskGen().generate(factor_tasks)
+    return implementation_result
+
+
+import fire
+if __name__ == "__main__":
+    fire.Fire(extract_models_and_implement)

rdagent/app/model_implementation/README.md

@@ -17,8 +17,9 @@ pip3 install torch_geometric
 ## Task Extraction
 From paper to task.
 ```bash
-python rdagent/app/model_implementation/task_extraction.py
+# python rdagent/app/model_implementation/task_extraction.py
 # It may based on rdagent/document_reader/document_reader.py
+python rdagent/components/task_implementation/model_implementation/task_extraction.py ./PaperImpBench/raw_paper/
 ```
 
 ## Complete workflow
@@ -30,10 +31,14 @@ From paper to implementation
 
 ## Paper benchmark
 ```bash
+# TODO: it does not work well now.
 python rdagent/app/model_implementation/eval.py
-
-TODO:
-- Is evaluation reasonable
 ```
 
+TODO:
+- Create reasonable benchmark
+  - with uniform input
+  - manually create task
+- Create reasonable evaluation metrics
+
 ## Evolving

rdagent/app/model_implementation/eval.py

@@ -2,14 +2,18 @@ from pathlib import Path
 
 DIRNAME = Path(__file__).absolute().resolve().parent
 
-from rdagent.model_implementation.benchmark.eval import ModelImpValEval
-from rdagent.model_implementation.one_shot import ModelTaskGen
-from rdagent.model_implementation.task import ModelImpLoader, ModelTaskLoderJson
+from rdagent.components.task_implementation.model_implementation.benchmark.eval import ModelImpValEval
+from rdagent.components.task_implementation.model_implementation.one_shot import ModelTaskGen
+from rdagent.components.task_implementation.model_implementation.task import  ModelImpLoader, ModelTaskLoderJson
 
-mtl = ModelTaskLoderJson("TODO: A Path to json")
+
+bench_folder = DIRNAME.parent.parent / "components" / "task_implementation" / "model_implementation" / "benchmark"
+mtl = ModelTaskLoderJson(str(bench_folder / "model_dict.json"))
 
 task_l = mtl.load()
 
+task_l = [t for t in task_l if t.key == "A-DGN"]  # FIXME: other models does not work well
+
 mtg = ModelTaskGen()
 
 impl_l = mtg.generate(task_l)
@@ -17,12 +21,13 @@ impl_l = mtg.generate(task_l)
 # TODO: Align it with the benchmark framework after @wenjun's refine the evaluation part.
 # Currently, we just handcraft a workflow for fast evaluation.
 
-mil = ModelImpLoader(DIRNAME.parent.parent / "model_implementation" / "benchmark" / "gt_code")
+mil = ModelImpLoader(bench_folder / "gt_code")
 
 mie = ModelImpValEval()
 # Evaluation:
 eval_l = []
 for impl in impl_l:
+    print(impl.target_task)
     gt_impl = mil.load(impl.target_task)
     eval_l.append(mie.evaluate(gt_impl, impl))
 

rdagent/components/task_implementation/model_implementation/benchmark/eval.py

@@ -1,6 +1,6 @@
 # TODO: inherent from the benchmark base class
 import torch
-from rdagent.model_implementation.task import ModelTaskImpl
+from rdagent.components.task_implementation.model_implementation.task import ModelTaskImpl
 
 
 def get_data_conf(init_val):

rdagent/components/task_implementation/model_implementation/benchmark/gt_code/dirgnn.py

@@ -74,6 +74,9 @@ class DirGNNConv(torch.nn.Module):
         return f"{self.__class__.__name__}({self.conv_in}, alpha={self.alpha})"
 
 
+model_cls = DirGNNConv
+
+
 if __name__ == "__main__":
     node_features = torch.load("node_features.pt")
     edge_index = torch.load("edge_index.pt")

rdagent/components/task_implementation/model_implementation/benchmark/gt_code/gpsconv.py

@@ -183,6 +183,9 @@ class GPSConv(torch.nn.Module):
         )
 
 
+model_cls = GPSConv
+
+
 if __name__ == "__main__":
     node_features = torch.load("node_features.pt")
     edge_index = torch.load("edge_index.pt")

rdagent/components/task_implementation/model_implementation/benchmark/gt_code/linkx.py

@@ -166,6 +166,7 @@ class LINKX(torch.nn.Module):
             f"out_channels={self.out_channels})"
         )
 
+model_cls = LINKX
 
 if __name__ == "__main__":
     node_features = torch.load("node_features.pt")

rdagent/components/task_implementation/model_implementation/benchmark/gt_code/pmlp.py

@@ -99,17 +99,14 @@ class PMLP(torch.nn.Module):
         return f"{self.__class__.__name__}({self.in_channels}, " f"{self.out_channels}, num_layers={self.num_layers})"
 
 
+model_cls = PMLP
+
 if __name__ == "__main__":
     node_features = torch.load("node_features.pt")
     edge_index = torch.load("edge_index.pt")
 
     # Model instantiation and forward pass
-    model = PMLP(
-        in_channels=node_features.size(-1),
-        hidden_channels=node_features.size(-1),
-        out_channels=node_features.size(-1),
-        num_layers=1,
-    )
+    model = PMLP(in_channels=node_features.size(-1), hidden_channels=node_features.size(-1), out_channels=node_features.size(-1), num_layers=1)
     output = model(node_features, edge_index)
 
     # Save output to a file

rdagent/components/task_implementation/model_implementation/benchmark/gt_code/visnet.py

@@ -1176,6 +1176,9 @@ class ViSNet(torch.nn.Module):
         return y, None
 
 
+model_cls = ViSNet
+
+
 if __name__ == "__main__":
     node_features = torch.load("node_features.pt")
     edge_index = torch.load("edge_index.pt")

rdagent/components/task_implementation/model_implementation/prompts.yaml

@@ -0,0 +1,8 @@
+extract_model_formulation_system: |-
+    offer description of the proposed model in this paper, write a latex formula with variable of the model. the format should be like "        "Model Name": {
+                "description": "",
+                "formulation": "",
+                "variables": {
+                    "\\hat{y}_u": "The predicted output for node u",
+                }"
+    such format content should be begin with ```json and end with ``` and the content should be in json format.

rdagent/components/task_implementation/model_implementation/task.py

@@ -16,6 +16,7 @@ from rdagent.core.task import (
     TaskLoader,
 )
 from rdagent.utils import get_module_by_module_path
+import json
 
 
 class ModelImplTask(BaseTask):
@@ -42,53 +43,104 @@ class ModelImplTask(BaseTask):
         self.formulation = formulation
         self.variables = variables
         self.key = key
+    def get_information(self):
+        return f"""name: {self.name}
+description: {self.description}
+formulation: {self.formulation}
+variables: {self.variables}
+key: {self.key}
+"""
+    
+    @staticmethod
+    def from_dict(dict):
+        return ModelImplTask(**dict)
+    
+    def __repr__(self) -> str:
+        return f"<{self.__class__.__name__} {self.name}>"
 
 
 class ModelTaskLoderJson(TaskLoader):
+    # def __init__(self, json_uri: str, select_model: Optional[str] = None) -> None:
+    #     super().__init__()
+    #     self.json_uri = json_uri
+    #     self.select_model = 'A-DGN'
+
+    # def load(self, *argT, **kwargs) -> Sequence[ModelImplTask]:
+    #     # json is supposed to be in the format of {model_name: dict{model_data}}
+    #     model_dict = json.load(open(self.json_uri, "r"))
+    #     if self.select_model is not None:
+    #         assert self.select_model in model_dict
+    #         model_name = self.select_model
+    #         model_data = model_dict[self.select_model]
+    #     else:
+    #         model_name, model_data = list(model_dict.items())[0]
+
+    #     model_impl_task = ModelImplTask(
+    #         name=model_name,
+    #         description=model_data["description"],
+    #         formulation=model_data["formulation"],
+    #         variables=model_data["variables"],
+    #         key=model_name
+    #     )
+        
+    #     return [model_impl_task]
+
     def __init__(self, json_uri: str) -> None:
         super().__init__()
-        # TODO: the json should be loaded from URI.
         self.json_uri = json_uri
 
     def load(self, *argT, **kwargs) -> Sequence[ModelImplTask]:
-        # TODO: we should load the tasks from json;
+        # json is supposed to be in the format of {model_name: dict{model_data}}
+        model_dict = json.load(open(self.json_uri, "r"))
 
-        # this version does not align with the right answer
+        # FIXME: the model in the json file is not right due to extraction error
+        #       We should fix them case by case in the future
+        # 
         # formula_info = {
         #     "name": "Anti-Symmetric Deep Graph Network (A-DGN)",
         #     "description": "A framework for stable and non-dissipative DGN design. It ensures long-range information preservation between nodes and prevents gradient vanishing or explosion during training.",
-        #     "formulation": "x_u^{(l)} = x_u^{(l-1)} + \\epsilon \\sigma \\left( W^T x_u^{(l-1)} + \\Phi(X^{(l-1)}, N_u) + b \\right)",
+        #     "formulation": r"\mathbf{x}^{\prime}_i = \mathbf{x}_i + \epsilon \cdot \sigma \left( (\mathbf{W}-\mathbf{W}^T-\gamma \mathbf{I}) \mathbf{x}_i + \Phi(\mathbf{X}, \mathcal{N}_i) + \mathbf{b}\right),",
         #     "variables": {
-        #         "x_u^{(l)}": "The state of node u at layer l",
-        #         "\\epsilon": "The step size in the Euler discretization",
-        #         "\\sigma": "A monotonically non-decreasing activation function",
-        #         "W": "An anti-symmetric weight matrix",
-        #         "X^{(l-1)}": "The node feature matrix at layer l-1",
-        #         "N_u": "The set of neighbors of node u",
-        #         "b": "A bias vector",
+        #         r"\mathbf{x}_i": "The state of node i at previous layer",
+        #         r"\epsilon": "The step size in the Euler discretization",
+        #         r"\sigma": "A monotonically non-decreasing activation function",
+        #         r"\Phi": "A graph convolutional operator",
+        #         r"W": "An anti-symmetric weight matrix",
+        #         r"\mathbf{x}^{\prime}_i": "The node feature matrix at layer l-1",
+        #         r"\mathcal{N}_i": "The set of neighbors of node u",
+        #         r"\mathbf{b}": "A bias vector",
         #     },
         #     "key": "A-DGN",
         # }
-        formula_info = {
-            "name": "Anti-Symmetric Deep Graph Network (A-DGN)",
-            "description": "A framework for stable and non-dissipative DGN design. It ensures long-range information preservation between nodes and prevents gradient vanishing or explosion during training.",
-            "formulation": r"\mathbf{x}^{\prime}_i = \mathbf{x}_i + \epsilon \cdot \sigma \left( (\mathbf{W}-\mathbf{W}^T-\gamma \mathbf{I}) \mathbf{x}_i + \Phi(\mathbf{X}, \mathcal{N}_i) + \mathbf{b}\right),",
-            "variables": {
-                r"\mathbf{x}_i": "The state of node i at previous layer",
-                r"\epsilon": "The step size in the Euler discretization",
-                r"\sigma": "A monotonically non-decreasing activation function",
-                r"\Phi": "A graph convolutional operator",
-                r"W": "An anti-symmetric weight matrix",
-                r"\mathbf{x}^{\prime}_i": "The node feature matrix at layer l-1",
-                r"\mathcal{N}_i": "The set of neighbors of node u",
-                r"\mathbf{b}": "A bias vector",
-            },
-            "key": "A-DGN",
-        }
-        return [ModelImplTask(**formula_info)]
+        model_impl_task_list = []
+        for model_name, model_data in model_dict.items():
+            model_impl_task = ModelImplTask(
+                name=model_name,
+                description=model_data["description"],
+                formulation=model_data["formulation"],
+                variables=model_data["variables"],
+                key=model_data["key"]
+            )
+            model_impl_task_list.append(model_impl_task)
+        return model_impl_task_list
+    
+class ModelImplementationTaskLoaderFromDict(TaskLoader):
+    def load(self, model_dict: dict) -> list:
+        """Load data from a dict."""
+        task_l = []
+        for model_name, model_data in model_dict.items():
+            task = ModelImplTask(
+                name=model_name,
+                description=model_data["description"],
+                formulation=model_data["formulation"],
+                variables=model_data["variables"],
+                key=model_name
+            )
+            task_l.append(task)
+        return task_l
 
 
-class ModelTaskImpl(TaskImplementation):
+class ModelTaskImpl(FBTaskImplementation):
     """
     It is a Pytorch model implementation task;
     All the things are placed in a folder.

rdagent/components/task_implementation/model_implementation/task_extraction.py

@@ -0,0 +1,117 @@
+from __future__ import annotations
+
+import json
+import multiprocessing as mp
+import re
+from pathlib import Path
+from typing import Mapping
+
+import numpy as np
+import pandas as pd
+import tiktoken
+from jinja2 import Template
+from rdagent.core.conf import RD_AGENT_SETTINGS
+from rdagent.core.log import RDAgentLog
+from rdagent.core.prompts import Prompts
+from rdagent.core.task import TaskLoader
+from rdagent.components.document_reader.document_reader import load_and_process_pdfs_by_langchain
+from rdagent.components.task_implementation.model_implementation.task import ModelImplementationTaskLoaderFromDict
+from rdagent.oai.llm_utils import APIBackend, create_embedding_with_multiprocessing
+
+
+document_process_prompts = Prompts(file_path=Path(__file__).parent / "prompts.yaml")
+
+def extract_model_from_doc(doc_content: str) -> dict:
+    """
+    Extract model information from document content.
+
+    Parameters
+    ----------
+    doc_content : str
+        Document content.
+
+    Returns
+    -------
+    dict
+        {factor_name: dict{description, formulation, variables}}
+    """
+    session = APIBackend().build_chat_session(
+        session_system_prompt=document_process_prompts["extract_model_formulation_system"],
+    )
+    current_user_prompt = doc_content
+
+    # Extract model information from document content.
+    model_dict = {}
+
+    for _ in range(10):
+        # try to extract model information from the document content, retry at most 10 times.
+        extract_result_resp = session.build_chat_completion(
+            user_prompt=current_user_prompt,
+            json_mode=False,
+        )
+        re_search_res = re.search(r"```json(.*)```", extract_result_resp, re.S)
+        ret_json_str = re_search_res.group(1) if re_search_res is not None else ""
+        try:
+            ret_dict = json.loads(ret_json_str)
+            parse_success = bool(isinstance(ret_dict, dict))
+        except json.JSONDecodeError:
+            parse_success = False
+        if ret_json_str is None or not parse_success:
+            current_user_prompt = "Your response didn't follow the instruction might be wrong json format. Try again."
+        else:
+            for name, formulation_and_description in ret_dict.items():
+                if name not in model_dict:
+                    model_dict[name] = formulation_and_description
+            if len(model_dict) == 0:
+                current_user_prompt = "No model extracted. Please try again."
+            else:
+                break
+
+
+    RDAgentLog().info(f"已经完成{len(model_dict)}个模型的提取")
+
+    return model_dict
+
+def merge_file_to_model_dict_to_model_dict(
+    file_to_model_dict: dict[str, dict],
+) -> dict:
+    model_dict = {}
+    for file_name in file_to_model_dict:
+        for model_name in file_to_model_dict[file_name]:
+            model_dict.setdefault(model_name, [])
+            model_dict[model_name].append(file_to_model_dict[file_name][model_name])
+    
+    model_dict_simple_deduplication = {}
+    for model_name in model_dict:
+        if len(model_dict[model_name]) > 1:
+            model_dict_simple_deduplication[model_name] = max(
+                model_dict[model_name],
+                key=lambda x: len(x["formulation"]),
+            )
+        else:
+            model_dict_simple_deduplication[model_name] = model_dict[model_name][0]
+    return model_dict_simple_deduplication
+
+
+def extract_model_from_docs(docs_dict):
+    model_dict = {}
+    for doc_name, doc_content in docs_dict.items():
+        model_dict[doc_name] = extract_model_from_doc(doc_content)
+    return model_dict
+
+
+class ModelImplementationTaskLoaderFromPDFfiles(TaskLoader):
+    def load(self, file_or_folder_path: Path) -> dict:
+        docs_dict = load_and_process_pdfs_by_langchain(Path(file_or_folder_path)) # dict{file_path:content}
+        model_dict = extract_model_from_docs(docs_dict) # dict{file_name: dict{model_name: dict{description, formulation, variables}}}
+        model_dict = merge_file_to_model_dict_to_model_dict(model_dict) # dict {model_name: dict{description, formulation, variables}}
+        return ModelImplementationTaskLoaderFromDict().load(model_dict)
+
+def main(path="../test_doc"):
+    doc_dict = load_and_process_pdfs_by_langchain(Path(path))
+    print(doc_dict.keys()) # if you run code like "python -u", the print content will be truncated
+
+
+import fire
+if __name__ == "__main__":
+    fire.Fire(main)

requirements.txt

@@ -5,6 +5,19 @@ typer[all]
 cython
 scipy
 python-Levenshtein
+scikit-learn
+filelock
+
+# PDF related
+pypdf
+azure-core
+azure-ai-formrecognizer
+
+# factor implementations
+tables
+
+# azure identity related
+azure.identity
 
 # CI Fix Tool
 tree-sitter-python
@@ -12,3 +25,5 @@ tree-sitter
 
 # Jupyter related
 jupyter
+
+python-dotenv