[Fix, Feat] Solve llava wild issue when task num can't divide, add fu…

…yu ppl (#37)

* Add fuyu ppl, fix llava-bench gather issue

* Add gpt4V

* Black lint

lmms_eval/models/__init__.py

@@ -2,3 +2,4 @@
 from .otterhd import OtterHD
 from .qwen_vl import Qwen_VL
 from .fuyu import Fuyu
+from .gpt4v import GPT4V

lmms_eval/models/fuyu.py

@@ -8,6 +8,10 @@
 from lmms_eval.api.instance import Instance
 from tqdm import tqdm
 
+import warnings
+
+warnings.filterwarnings("ignore")
+
 
 @register_model("fuyu")
 class Fuyu(lmms):
@@ -118,7 +122,40 @@ def _collate(x):
 
     def loglikelihood(self, requests: List[Instance]) -> List[Tuple[float, bool]]:
         # TODO
-        assert False, "We have not implemented this function for llava yet"
+        res = []
+        pbar = tqdm(total=len(requests), disable=(self.rank != 0), desc="Model Responding")
+
+        for contexts, doc_to_target, doc_to_visual, doc_id, task, split in [reg.args for reg in requests]:
+            # encode, pad, and truncate contexts for this batch
+            continuation = doc_to_target(self.task_dict[task][split][doc_id])
+            visuals = [doc_to_visual(self.task_dict[task][split][doc_id])]
+            visuals = self.flatten(visuals)
+            formatted_contexts = [f"{contexts}\n"]
+            formatted_continuation = [f"{contexts}\n{continuation}"]
+            model_inputs = self.processor(text=formatted_continuation, images=visuals, device=self.device)
+            for k, v in model_inputs.items():
+                model_inputs[k] = v.to(self.device, non_blocking=True) if isinstance(v, torch.Tensor) else [vv.to(self.device, non_blocking=True) for vv in v]
+
+            for index in range(len(model_inputs["image_patches"])):
+                model_inputs["image_patches"][index] = model_inputs["image_patches"][index].to(dtype=next(self.model.parameters()).dtype)
+
+            labels = model_inputs["input_ids"].clone()
+            contxt_id = self.processor(text=formatted_contexts, return_tensors="pt")["input_ids"]
+            labels[: len(contxt_id)] = -100
+            with torch.inference_mode():
+                outputs = self.model(**model_inputs, labels=labels)
+            loss = outputs["loss"]
+            # loss = torch.exp(loss)
+            logits = outputs["logits"]
+            greedy_tokens = logits.argmax(dim=-1)
+            cont_toks = model_inputs["input_ids"][:, contxt_id.shape[1] :]  # [1, seq]
+            greedy_tokens = greedy_tokens[:, contxt_id.shape[1] : model_inputs["input_ids"].shape[1]]  # [1, seq]
+            max_equal = (greedy_tokens == cont_toks).all()
+            res.append((float(loss.item()), bool(max_equal)))
+            pbar.update(1)
+
+        pbar.close()
+        return res
 
     def loglikelihood_rolling(self, requests: List[Instance]) -> List[float]:
         # TODO

lmms_eval/models/gpt4v.py

@@ -0,0 +1,108 @@
+from io import BytesIO
+import os
+import base64
+from typing import List, Tuple
+from tqdm import tqdm
+import requests as url_requests
+import time
+import logging
+
+from lmms_eval.api.instance import Instance
+from lmms_eval.api.model import lmms
+from lmms_eval.api.registry import register_model
+from lmms_eval import utils
+
+from PIL import Image
+
+API_TYPE = os.getenv("API_TYPE", "openai")
+NUM_SECONDS_TO_SLEEP = 5
+eval_logger = logging.getLogger("lmms-eval")
+
+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")
+    headers = {
+        "Authorization": f"Bearer {API_KEY}",
+        "Content-Type": "application/json",
+    }
+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")
+    headers = {
+        "api-key": API_KEY,
+        "Content-Type": "application/json",
+    }
+
+
+@register_model("gpt4V")
+class GPT4V(lmms):
+    def __init__(self, **kwargs) -> None:
+        super().__init__()
+
+    # Function to encode the image
+    def encode_image(self, image: Image):
+        output_buffer = BytesIO()
+        image.save(output_buffer, format="JPEG")
+        byte_data = output_buffer.getvalue()
+        base64_str = base64.b64encode(byte_data).decode("utf-8")
+        return base64_str
+
+    def flatten(self, input):
+        new_list = []
+        for i in input:
+            for j in i:
+                new_list.append(j)
+        return new_list
+
+    def generate_until(self, requests) -> List[str]:
+        res = []
+        pbar = tqdm(total=len(requests), disable=(self.rank != 0), desc="Model Responding")
+
+        for contexts, gen_kwargs, doc_to_visual, doc_id, task, split in [reg.args for reg in requests]:
+            # encode, pad, and truncate contexts for this batch
+            visuals = [doc_to_visual(self.task_dict[task][split][doc_id])]
+            visuals = self.flatten(visuals)
+
+            payload = {"model": "gpt-4-vision-preview", "messages": [{"role": "user", "content": []}]}
+            payload["messages"][0]["content"].append({"type": "text", "text": contexts})
+
+            for visual in visuals:
+                img = self.encode_image(visual)
+                payload["messages"][0]["content"].append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{img}"}})
+            if "max_new_tokens" not in gen_kwargs:
+                gen_kwargs["max_new_tokens"] = 1024
+            if "temperature" not in gen_kwargs:
+                gen_kwargs["temperature"] = 0
+            if "top_p" not in gen_kwargs:
+                gen_kwargs["top_p"] = None
+            if "num_beams" not in gen_kwargs:
+                gen_kwargs["num_beams"] = 1
+
+            # payload["max_tokens"] = gen_kwargs["max_new_tokens"]
+            # payload["temperature"] = gen_kwargs["temperature"]
+
+            for attempt in range(5):
+                try:
+                    response = url_requests.post(API_URL, headers=headers, json=payload)
+                    response_data = response.json()
+
+                    content = response_data["choices"][0]["message"]["content"].strip()
+                    break  # If successful, break out of the loop
+
+                except Exception as e:
+                    eval_logger.info(f"Attempt {attempt + 1} failed with error: {str(e)}")
+                    if attempt < 5 - 1:  # If we have retries left, sleep and then continue to next attempt
+                        time.sleep(NUM_SECONDS_TO_SLEEP)
+                    else:  # If this was the last attempt, log and return empty
+                        eval_logger.error(f"All 5 attempts failed. Last error message: {str(e)}")
+                        content = ""
+            res.append(content)
+        return res
+
+    def loglikelihood(self, requests: List[Instance]) -> List[Tuple[float, bool]]:
+        # TODO
+        assert False, "GPT4V not support"
+
+    def loglikelihood_rolling(self, requests: List[Instance]) -> List[float]:
+        # TODO
+        assert False, "GPT4V not support"

lmms_eval/tasks/llava-bench-coco/utils.py

@@ -8,10 +8,13 @@
 import time
 import yaml
 from pathlib import Path
+from copy import deepcopy
 
 eval_logger = logging.getLogger("lmms-eval")
 NUM_SECONDS_TO_SLEEP = 0.5
 
+LLAVA_W_METRICS = ["gpt_eval_llava_conv", "gpt_eval_llava_detail", "gpt_eval_llava_conv"]
+
 rule_dict = json.load(open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "rule.json"), "r"))
 
 with open(Path(__file__).parent / "llava-bench-coco.yaml", "r") as f:
@@ -24,16 +27,30 @@
 
     config = yaml.safe_load("".join(safe_data))
 
-API_URL = os.getenv("OPENAI_API_URL", "https://api.openai.com/v1/chat/completions")
-API_KEY = os.getenv("OPENAI_API_KEY", "YOUR_API_KEY")
 GPT_EVAL_MODEL_NAME = config["metadata"]["gpt_eval_model_name"]
 
+GPT_EVAL_MODEL_NAME = config["metadata"]["gpt_eval_model_name"]
+
+API_TYPE = os.getenv("API_TYPE", "openai")
 
-def get_eval(content: str, max_tokens: int, retries: int = 3):
+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")
     headers = {
         "Authorization": f"Bearer {API_KEY}",
         "Content-Type": "application/json",
     }
+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")
+    headers = {
+        "api-key": API_KEY,
+        "Content-Type": "application/json",
+    }
+
+
+def get_eval(content: str, max_tokens: int, retries: int = 3):
+    global headers
 
     messages = [
         {
@@ -125,7 +142,7 @@ def llava_process_results(doc, result):
         scores = [-1, -1]
 
     metric = f"gpt_eval_llava_{doc.get('category', 'unknown')}"
-    review_dict = {
+    category_review_dict = {
         "question": question,
         "ans1": ans1,
         "ans2": ans2,
@@ -136,8 +153,19 @@ def llava_process_results(doc, result):
         "eval_model": model_name,
     }
 
+    non_category_review_dict = deepcopy(category_review_dict)
+    non_category_review_dict["scores"] = [-999, -999]
+
+    data_dict = {}
+    for m in LLAVA_W_METRICS:
+        if m == metric:
+            data_dict[m] = category_review_dict
+        else:
+            data_dict[m] = non_category_review_dict
+    data_dict["gpt_eval_llava_all"] = category_review_dict
+
     # return {"gpt_eval_llava_all": review_dict}
-    return {metric: review_dict, "gpt_eval_llava_all": review_dict}
+    return data_dict
 
 
 def llava_conv_aggregation(results):
@@ -160,6 +188,8 @@ def llava_aggregation(results, category):
     try:
         scores = []
         for result in results:
+            if -999 in result["scores"]:
+                continue
             scores.append(result["scores"])
 
         stats = np.asarray(scores).mean(0).tolist()

lmms_eval/tasks/llava-in-the-wild/utils.py

@@ -8,10 +8,13 @@
 import time
 import yaml
 from pathlib import Path
+from copy import deepcopy
 
 eval_logger = logging.getLogger("lmms-eval")
 NUM_SECONDS_TO_SLEEP = 0.5
 
+LLAVA_W_METRICS = ["gpt_eval_llava_conv", "gpt_eval_llava_detail", "gpt_eval_llava_conv"]
+
 rule_dict = json.load(open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "rule.json"), "r"))
 
 with open(Path(__file__).parent / "llava-in-the-wild.yaml", "r") as f:
@@ -24,16 +27,28 @@
 
     config = yaml.safe_load("".join(safe_data))
 
-API_URL = os.getenv("OPENAI_API_URL", "https://api.openai.com/v1/chat/completions")
-API_KEY = os.getenv("OPENAI_API_KEY", "YOUR_API_KEY")
 GPT_EVAL_MODEL_NAME = config["metadata"]["gpt_eval_model_name"]
 
+API_TYPE = os.getenv("API_TYPE", "openai")
 
-def get_eval(content: str, max_tokens: int, retries: int = 3):
+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")
     headers = {
         "Authorization": f"Bearer {API_KEY}",
         "Content-Type": "application/json",
     }
+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")
+    headers = {
+        "api-key": API_KEY,
+        "Content-Type": "application/json",
+    }
+
+
+def get_eval(content: str, max_tokens: int, retries: int = 3):
+    global headers
 
     messages = [
         {
@@ -125,7 +140,7 @@ def llava_process_results(doc, result):
         scores = [-1, -1]
 
     metric = f"gpt_eval_llava_{doc.get('category', 'all')}"
-    review_dict = {
+    category_review_dict = {
         "question": question,
         "ans1": ans1,
         "ans2": ans2,
@@ -136,8 +151,19 @@ def llava_process_results(doc, result):
         "eval_model": model_name,
     }
 
+    non_category_review_dict = deepcopy(category_review_dict)
+    non_category_review_dict["scores"] = [-999, -999]
+
+    data_dict = {}
+    for m in LLAVA_W_METRICS:
+        if m == metric:
+            data_dict[m] = category_review_dict
+        else:
+            data_dict[m] = non_category_review_dict
+    data_dict["gpt_eval_llava_all"] = category_review_dict
+
     # return {"gpt_eval_llava_all": review_dict}
-    return {metric: review_dict, "gpt_eval_llava_all": review_dict}
+    return data_dict
 
 
 def llava_conv_aggregation(results):
@@ -160,6 +186,8 @@ def llava_aggregation(results, category):
     try:
         scores = []
         for result in results:
+            if -999 in result["scores"]:
+                continue
             scores.append(result["scores"])
 
         stats = np.asarray(scores).mean(0).tolist()