[WIP] adding mmbench dev evaluation (#75)

* WIP * Update GPT evaluation model name and sys prompt * 🛠️ Scale accuracy to percentage
EvolvingLMMs-Lab · Apr 4, 2024 · 1ea7a8f · 1ea7a8f
1 parent 70cc773
commit 1ea7a8f
Show file tree

Hide file tree

Showing 10 changed files with 447 additions and 19 deletions.
diff --git a/lmms_eval/tasks/mmbench/cc_utils.py b/lmms_eval/tasks/mmbench/cc_utils.py
@@ -9,7 +9,7 @@
 from lmms_eval.tasks.mmbench.mmbench_evals import MMBench_Evaluator
 from lmms_eval.tasks._task_utils.file_utils import generate_submission_file
 
-with open(Path(__file__).parent / "mmbench_cn.yaml", "r") as f:
+with open(Path(__file__).parent / "mmbench.yaml", "r") as f:
     raw_data = f.readlines()
     safe_data = []
     for i, line in enumerate(raw_data):
@@ -19,7 +19,18 @@
 
     config = yaml.safe_load("".join(safe_data))
 
-mmbench_evaluator = MMBench_Evaluator(sys_prompt=config["metadata"]["sys_prompt"])
+GPT_EVAL_MODEL_NAME = config["metadata"]["gpt_eval_model_name"]
+API_TYPE = os.getenv("API_TYPE", "openai")
+
+if API_TYPE == "openai":
+    API_URL = os.getenv("OPENAI_API_URL", "https://api.openai.com/v1/chat/completions")
+    API_KEY = os.getenv("OPENAI_API_KEY", "YOUR_API_KEY")
+elif API_TYPE == "azure":
+    API_URL = os.getenv("AZURE_ENDPOINT", "https://api.cognitive.microsoft.com/sts/v1.0/issueToken")
+    API_KEY = os.getenv("AZURE_API_KEY", "YOUR_API_KEY")
+
+
+mmbench_evaluator = MMBench_Evaluator(sys_prompt=config["metadata"]["sys_prompt"], API_KEY=API_KEY, API_URL=API_URL, model_version=GPT_EVAL_MODEL_NAME)
 
 
 def mmbench_doc_to_visual(doc):
@@ -52,21 +63,44 @@ def mmbench_cn_cc_doc_to_text(doc, model_specific_prompt_kwargs=None):
 def mmbench_cn_cc_process_results(doc, results):
     model_response = results[0].strip()
     data = {
+        "gpt_eval_score": {
+            "index": doc["index"],
+            "question": doc["question"],
+            "answer": doc["answer"],
+            "prediction": model_response,
+            "source": doc["source"],
+            "category": doc["category"],
+        },
         "submission": {
             "index": doc["index"],
             "question": doc["question"],
             "answer": doc["answer"],
             "prediction": model_response,
             "source": doc["source"],
             "category": doc["category"],
-        }
+        },
     }
     option_candidate = ["A", "B", "C", "D", "E"]
     for c in option_candidate:
         data["submission"][c] = doc.get(c, "nan")
+        data["gpt_eval_score"][c] = doc.get(c, "nan")
     return data
 
 
+def mmbench_cn_cc_aggregate_dev_results_eval(results, args):
+    print(f"============= MMBench-CN(CC) Detailed Results =============")
+    overall_acc, category_acc, l2_category_acc = mmbench_evaluator.eval_result(results, eval_method="openai")
+    file = generate_submission_file("mmbench_cn_cc_results.json", args)
+    details_info = {
+        "overall_acc": overall_acc,
+        "category_acc": category_acc,
+        "l2_category_acc": l2_category_acc,
+    }
+    with open(file, "w") as f:
+        json.dump(details_info, f)
+    return overall_acc * 100
+
+
 def mmbench_cn_cc_aggregate_results(results, args):
     df = pd.DataFrame(results)
     file = generate_submission_file("mmbench_cn_cc_results.xlsx", args)

diff --git a/lmms_eval/tasks/mmbench/cn_utils.py b/lmms_eval/tasks/mmbench/cn_utils.py
@@ -8,8 +8,9 @@
 
 eval_logger = logging.getLogger("lmms-eval")
 from lmms_eval.tasks.mmbench.mmbench_evals import MMBench_Evaluator
+from lmms_eval.tasks._task_utils.file_utils import generate_submission_file
 
-with open(Path(__file__).parent / "mmbench_cn.yaml", "r") as f:
+with open(Path(__file__).parent / "mmbench.yaml", "r") as f:
     raw_data = f.readlines()
     safe_data = []
     for i, line in enumerate(raw_data):
@@ -19,7 +20,18 @@
 
     config = yaml.safe_load("".join(safe_data))
 
-mmbench_evaluator = MMBench_Evaluator(sys_prompt=config["metadata"]["sys_prompt"])
+GPT_EVAL_MODEL_NAME = config["metadata"]["gpt_eval_model_name"]
+API_TYPE = os.getenv("API_TYPE", "openai")
+
+if API_TYPE == "openai":
+    API_URL = os.getenv("OPENAI_API_URL", "https://api.openai.com/v1/chat/completions")
+    API_KEY = os.getenv("OPENAI_API_KEY", "YOUR_API_KEY")
+elif API_TYPE == "azure":
+    API_URL = os.getenv("AZURE_ENDPOINT", "https://api.cognitive.microsoft.com/sts/v1.0/issueToken")
+    API_KEY = os.getenv("AZURE_API_KEY", "YOUR_API_KEY")
+
+
+mmbench_evaluator = MMBench_Evaluator(sys_prompt=config["metadata"]["sys_prompt"], API_KEY=API_KEY, API_URL=API_URL, model_version=GPT_EVAL_MODEL_NAME)
 
 
 def mmbench_doc_to_visual(doc):
@@ -55,6 +67,17 @@ def mmbench_doc_to_text(doc, model_specific_prompt_kwargs=None):
 def mmbench_process_results(doc, results):
     model_response = results[0].strip()
     data = {
+        "gpt_eval_score": {
+            "index": doc["index"],
+            "question": doc["question"],
+            "answer": doc["answer"],
+            "prediction": model_response,
+            "hint": doc["hint"],
+            "source": doc["source"],
+            "split": doc["split"],
+            "category": doc["category"],
+            "L2-category": doc["L2-category"],
+        },
         "submission": {
             "index": doc["index"],
             "question": doc["question"],
@@ -65,14 +88,29 @@ def mmbench_process_results(doc, results):
             "split": doc["split"],
             "category": doc["category"],
             "L2-category": doc["L2-category"],
-        }
+        },
     }
     option_candidate = ["A", "B", "C", "D", "E"]
     for c in option_candidate:
         data["submission"][c] = doc.get(c, "nan")
+        data["gpt_eval_score"][c] = doc.get(c, "nan")
     return data
 
 
+def mmbench_aggregate_dev_results_eval(results, args):
+    print(f"============= MMBench-CN(Dev) Detailed Results =============")
+    overall_acc, category_acc, l2_category_acc = mmbench_evaluator.eval_result(results, eval_method="openai")
+    file = generate_submission_file("mmbench_cn_dev_results.json", args)
+    details_info = {
+        "overall_acc": overall_acc,
+        "category_acc": category_acc,
+        "l2_category_acc": l2_category_acc,
+    }
+    with open(file, "w") as f:
+        json.dump(details_info, f)
+    return overall_acc * 100
+
+
 def mmbench_aggregate_dev_results(results, args):
     df = pd.DataFrame(results)
     excel_write_path = generate_submission_file("mmbench_cn_dev_results.xlsx", args)

diff --git a/lmms_eval/tasks/mmbench/en_utils.py b/lmms_eval/tasks/mmbench/en_utils.py
@@ -9,7 +9,7 @@
 from lmms_eval.tasks.mmbench.mmbench_evals import MMBench_Evaluator
 from lmms_eval.tasks._task_utils.file_utils import generate_submission_file
 
-with open(Path(__file__).parent / "mmbench_en.yaml", "r") as f:
+with open(Path(__file__).parent / "mmbench.yaml", "r") as f:
     raw_data = f.readlines()
     safe_data = []
     for i, line in enumerate(raw_data):
@@ -19,7 +19,18 @@
 
     config = yaml.safe_load("".join(safe_data))
 
-mmbench_evaluator = MMBench_Evaluator(sys_prompt=config["metadata"]["sys_prompt"])
+GPT_EVAL_MODEL_NAME = config["metadata"]["gpt_eval_model_name"]
+API_TYPE = os.getenv("API_TYPE", "openai")
+
+if API_TYPE == "openai":
+    API_URL = os.getenv("OPENAI_API_URL", "https://api.openai.com/v1/chat/completions")
+    API_KEY = os.getenv("OPENAI_API_KEY", "YOUR_API_KEY")
+elif API_TYPE == "azure":
+    API_URL = os.getenv("AZURE_ENDPOINT", "https://api.cognitive.microsoft.com/sts/v1.0/issueToken")
+    API_KEY = os.getenv("AZURE_API_KEY", "YOUR_API_KEY")
+
+
+mmbench_evaluator = MMBench_Evaluator(sys_prompt=config["metadata"]["sys_prompt"], API_KEY=API_KEY, API_URL=API_URL, model_version=GPT_EVAL_MODEL_NAME)
 
 
 def mmbench_doc_to_visual(doc):
@@ -55,6 +66,17 @@ def mmbench_doc_to_text(doc, model_specific_prompt_kwargs=None):
 def mmbench_process_results(doc, results):
     model_response = results[0].strip()
     data = {
+        "gpt_eval_score": {
+            "index": doc["index"],
+            "question": doc["question"],
+            "answer": doc["answer"],
+            "prediction": model_response,
+            "hint": doc["hint"],
+            "source": doc["source"],
+            "split": doc["split"],
+            "category": doc["category"],
+            "L2-category": doc["L2-category"],
+        },
         "submission": {
             "index": doc["index"],
             "question": doc["question"],
@@ -65,15 +87,30 @@ def mmbench_process_results(doc, results):
             "split": doc["split"],
             "category": doc["category"],
             "L2-category": doc["L2-category"],
-        }
+        },
     }
     option_candidate = ["A", "B", "C", "D", "E"]
     for c in option_candidate:
         data["submission"][c] = doc.get(c, "nan")
+        data["gpt_eval_score"][c] = doc.get(c, "nan")
     return data
 
 
-def mmbench_aggregate_dev_results(results, args):
+def mmbench_aggregate_dev_results_eval(results, args):
+    print(f"============= MMBench-EN(Dev) Detailed Results =============")
+    overall_acc, category_acc, l2_category_acc = mmbench_evaluator.eval_result(results, eval_method="openai")
+    file = generate_submission_file("mmbench_en_dev_results.json", args)
+    details_info = {
+        "overall_acc": overall_acc,
+        "category_acc": category_acc,
+        "l2_category_acc": l2_category_acc,
+    }
+    with open(file, "w") as f:
+        json.dump(details_info, f)
+    return overall_acc * 100
+
+
+def mmbench_aggregate_dev_results_submission(results, args):
     df = pd.DataFrame(results)
     excel_write_path = generate_submission_file("mmbench_en_dev_results.xlsx", args)
     with pd.ExcelWriter(excel_write_path) as writer:

diff --git a/lmms_eval/tasks/mmbench/mmbench.yaml b/lmms_eval/tasks/mmbench/mmbench.yaml
@@ -4,4 +4,8 @@ task:
   - mmbench_en_test
   - mmbench_cn_dev
   - mmbench_cn_test
-  - mmbench_cn_cc
+  - mmbench_cn_cc
+metadata:
+  version: 0.0
+  sys_prompt: "There are several options:"
+  gpt_eval_model_name: "gpt-3.5-turbo-0613"
diff --git a/lmms_eval/tasks/mmbench/mmbench_cc.yaml b/lmms_eval/tasks/mmbench/mmbench_cc.yaml
@@ -16,12 +16,14 @@ generation_kwargs:
   do_sample: false
 process_results: !function cc_utils.mmbench_cn_cc_process_results
 metric_list:
+  - metric: gpt_eval_score
+    aggregation: !function cc_utils.mmbench_cn_cc_aggregate_dev_results_eval
+    higher_is_better: true
   - metric: submission
     aggregation: !function cc_utils.mmbench_cn_cc_aggregate_results
 metadata:
   version: 0.0
-  gpt_eval_model_name: "gpt-3.5-turbo"
-  quick_extract: true
+  gpt_eval_model_name: "gpt-3.5-turbo-0613"
 
 model_specific_prompt_kwargs:
   default:

diff --git a/lmms_eval/tasks/mmbench/mmbench_cn.yaml b/lmms_eval/tasks/mmbench/mmbench_cn.yaml
@@ -5,6 +5,5 @@ task:
   - mmbench_cn_cc
 metadata:
   version: 0.0
-  gpt_eval_model_name: "gpt-3.5-turbo"
-  quick_extract: true
-  sys_prompt: "有如下几个选项："
+  gpt_eval_model_name: "gpt-3.5-turbo-0613"
+  sys_prompt: "有如下几个选项："
diff --git a/lmms_eval/tasks/mmbench/mmbench_cn_dev.yaml b/lmms_eval/tasks/mmbench/mmbench_cn_dev.yaml
@@ -1,6 +1,9 @@
 task: "mmbench_cn_dev"
 test_split: "dev"
 metric_list:
+  - metric: gpt_eval_score
+    aggregation: !function cn_utils.mmbench_aggregate_dev_results_eval
+    higher_is_better: true
   - metric: submission
     higher_is_better: true
     aggregation: !function cn_utils.mmbench_aggregate_dev_results

diff --git a/lmms_eval/tasks/mmbench/mmbench_en.yaml b/lmms_eval/tasks/mmbench/mmbench_en.yaml
@@ -5,3 +5,4 @@ task:
 metadata:
   version: 0.0
   sys_prompt: "There are several options:"
+  gpt_eval_model_name: "gpt-3.5-turbo-0613"
diff --git a/lmms_eval/tasks/mmbench/mmbench_en_dev.yaml b/lmms_eval/tasks/mmbench/mmbench_en_dev.yaml
@@ -2,6 +2,9 @@ task: "mmbench_en_dev"
 test_split: dev
 include: _default_template_mmbench_en_yaml
 metric_list:
-  - metric: submission
-    aggregation: !function en_utils.mmbench_aggregate_dev_results
+  - metric: gpt_eval_score
+    aggregation: !function en_utils.mmbench_aggregate_dev_results_eval
     higher_is_better: true
+  - metric: submission
+    aggregation: !function en_utils.mmbench_aggregate_dev_results_submission
+    higher_is_better: true