update eval code to match new src args

Author: i-gao

open_flamingo/eval/eval_datasets.py

@@ -7,6 +7,8 @@
 
 from open_flamingo.eval.classification_utils import IMAGENET_CLASSNAMES
 
+SUPPORTED_TASKS = ["coco", "flickr", "vqav2", "ok_vqa", "vizwiz", "textvqa", "hateful_memes", "imagenet"]
+
 
 class CaptionDataset(Dataset):
     def __init__(

open_flamingo/eval/eval_models/__init__.py

@@ -0,0 +1 @@
+from .eval_model import *

open_flamingo/eval/models/blip.py renamed to open_flamingo/eval/eval_models/blip.py

@@ -13,9 +13,6 @@ class EvalModel(BaseEvalModel):
     """BLIP-2 model evaluation."""
 
     def __init__(self, model_args, init_on_device=False):
-        assert (
-            "processor_path" in model_args and "lm_path" in model_args
-        ), "BLIP-2 requires processor_path, lm_path, and device arguments to be specified"
         super().__init__(model_args, init_on_device)
         with self.init_ctx:
             self.processor = Blip2Processor.from_pretrained(model_args["processor_path"])
@@ -25,6 +22,10 @@ def __init__(self, model_args, init_on_device=False):
             self.tokenizer = self.processor.tokenizer
         self._check_init()
 
+    @property
+    def required_args(self):
+        return ["processor_path", "lm_path"]
+
     def prepare_images(self, batch: List[List[Image.Image]]) -> torch.Tensor:
         batch_images = None
         assert all(
@@ -58,6 +59,7 @@ def prepare_text(
         max_length=2000,
         add_special_tokens=True,
     ):
+        self._validate_text(batch)
         encodings = self.tokenizer(
             batch,
             padding=padding,
@@ -95,39 +97,20 @@ def get_outputs(
 
         return self.tokenizer.batch_decode(outputs, skip_special_tokens=True)
 
-    def get_vqa_prompt(self, question, answer=None) -> str:
-        return (
-            f"Question:{question} Short answer:{answer if answer is not None else ''}"
-        )
-
-    def get_caption_prompt(self, caption=None) -> str:
+    def get_vqav2_prompt(self, question, answer=None) -> str:
+        return f"Question:{question} Short answer:{answer if answer is not None else ''}"
+    
+    def get_ok_vqa_prompt(self, question, answer=None) -> str:
+        return f"Question:{question} Short answer:{answer if answer is not None else ''}"
+    
+    def get_vizwiz_prompt(self, question, answer=None) -> str:
+        return f"Question:{question} Short answer:{answer if answer is not None else ''}"
+    
+    def get_textvqa_prompt(self, question, answer=None) -> str:
+        return f"Question:{question} Short answer:{answer if answer is not None else ''}"
+
+    def get_coco_prompt(self, caption=None) -> str:
+        return f"A photo of {caption if caption is not None else ''}"
+    
+    def get_flickr_prompt(self, caption=None) -> str:
         return f"A photo of {caption if caption is not None else ''}"
-
-    def __call__(
-        self,
-        lang_x: torch.Tensor,
-        vision_x: torch.Tensor,
-        attention_mask: torch.Tensor,
-    ):
-        with self.autocast():
-            outputs = self.model(
-                pixel_values=vision_x,
-                input_ids=lang_x,
-                attention_mask=attention_mask,
-            )
-
-        # remove vision tokens
-        outputs.logits = outputs.logits[:, -lang_x.size(1) :, :]
-        return outputs
-
-    def get_rank_classifications(
-        self,
-        batch_text: List[str],
-        batch_images: List[List[Image.Image]],
-        all_class_names: List[str],
-        use_cache: bool,
-        normalize_length: bool,
-    ):
-        raise NotImplementedError(
-            "BLIP-2 classification-based evaluation not implemented"
-        )

open_flamingo/eval/eval_model.py renamed to open_flamingo/eval/eval_models/eval_model.py

@@ -6,36 +6,75 @@
 import torch
 from contextlib import suppress
 
+SUPPORTED_MODELS = ["open_flamingo", "blip", "idefics"]
+ZERO_SHOT_ONLY_MODELS = ["blip"]
+
+
+def get_eval_model(name, *args, **kwargs):
+    """Return an EvalModel object."""
+    if name == "open_flamingo":
+        from .open_flamingo import EvalModel
+
+        return EvalModel(*args, **kwargs)
+    elif name == "blip":
+        from .blip import EvalModel
+
+        return EvalModel(*args, **kwargs)
+    elif name == "idefics":
+        from .idefics import EvalModel
+
+        return EvalModel(*args, **kwargs)
+    else:
+        raise ValueError(f"Unsupported EvalModel type {name}")
+
+
 class BaseEvalModel(abc.ABC):
     """Base class encapsulating functionality needed to evaluate a model."""
 
-    def __init__(self, model_args: List[str]):
+    def __init__(self, model_args: List[str], init_on_device=False):
         """Initialize model.
 
         Args:
             args: arguments to model. These should be parsed, or if the model
                 has no applicable arguments, an error should be thrown if `args`
                 is non-empty.
         """
+        # check model args
+        assert all(
+            arg in model_args for arg in self.required_args
+        ), f"Missing required args for {self.__class__.__name__}: {self.required_args}"
+        self.lm_name = model_args["lm_path"].split("/")[-1]
 
-    def __init__(self, model_args, init_on_device=False):
-        assert "lm_path" in model_args, "All models require the lm_path argument"
+        # set device and precision
         self.device = (
             model_args["device"]
-            if ("device" in model_args and (type(model_args["device"]) != int or model_args["device"] >= 0))
+            if (
+                "device" in model_args
+                and (type(model_args["device"]) != int or model_args["device"] >= 0)
+            )
             else "cpu"
         )
+        print("Using device:", self.device)
         self.precision = model_args.get("precision", "fp32")
-        self.lm_name = model_args["lm_path"].split("/")[-1]
         self.autocast = get_autocast(self.precision)
         self.cast_dtype = get_cast_dtype(self.precision)
+
+        # initialization context
         if init_on_device:
-            # for deepspeed, must init on device, or likely CPU OOM 
+            # for deepspeed, must init on device, or likely CPU OOM
             import deepspeed
-            self.init_ctx = deepspeed.OnDevice(dtype=self.cast_dtype, device=self.device)
+
+            self.init_ctx = deepspeed.OnDevice(
+                dtype=self.cast_dtype, device=self.device
+            )
         else:
             self.init_ctx = suppress()
 
+    @property
+    def required_args(self):
+        """Return list of required arguments to initialize model."""
+        return ["lm_path"]
+
     def _check_init(self):
         """Finish model initialization."""
         assert hasattr(self, "model"), "Model has not been initialized"
@@ -49,6 +88,7 @@ def init_distributed(self, world_size=None, use_deepspeed=False):
         if use_deepspeed:
             assert "amp" not in self.precision, "Deepspeed does not support amp"
             import deepspeed
+
             self.ds_engine = deepspeed.init_inference(
                 self.model,
                 mp_size=world_size,
@@ -61,12 +101,6 @@ def init_distributed(self, world_size=None, use_deepspeed=False):
         else:
             self.model = DDP(self.model, device_ids=[self.device])
 
-    def set_device(self, device):
-        """Set device for model."""
-        torch.cuda.set_device(device)
-        self.device = torch.device("cuda", device)
-        self.model = self.model.to(device, dtype=self.cast_dtype)
-
     def __call__(
         self,
         lang_x: torch.Tensor,
@@ -76,12 +110,13 @@ def __call__(
         use_cache: bool = False,
     ):
         """
-        Calls the forward function of the model.
-        Special logic to handle the case if past_key_values is not None:
+        Calls the forward function of the model, and returns an object that includes logits.
+        Note: implementations should handle the case if past_key_values is not None:
             then lang_x is assumed to contain the tokens to be generated
             *excluding* the tokens already in past_key_values.
             We then repeatedly call forward, updating the past_key_values.
         """
+        raise NotImplementedError
 
     def prepare_text(
         self,
@@ -92,7 +127,7 @@ def prepare_text(
         add_special_tokens=True,
     ):
         """
-        Prepare text for model.
+        Prepare text for model. Note that padding is always on the left.
 
         Args:
             batch: list of text strings
@@ -101,36 +136,38 @@ def prepare_text(
             max_length: maximum length of the text
 
         Returns:
-            input_ids: tensor of shape (B, T)
-            attention_mask: tensor of shape (B, T)
+            input_ids: tensor of shape (B, T_txt)
+            attention_mask: tensor of shape (B, T_txt)
         """
+        raise NotImplementedError
 
     def prepare_images(self, batch: List[List[Image.Image]]):
         """
         Prepare images for model.
         Args:
             batch: list of lists of PIL images
         Returns:
-            tensor of shape (B, T, *, C, H, W)
+            tensor of shape (B, T_img, F, C, H, W)
         """
+        raise NotImplementedError
 
     def get_outputs(
         self,
         batch_text: List[str],
         batch_images: List[List[Image.Image]],
         **decode_kwargs,
     ) -> List[str]:
-        """Get outputs for a batch of images and text.
+        """Call generate on a batch of images and text.
 
         Args:
-            batch_text: list of text strings, with the text "<image>" in place
-                of any images to be included.
+            batch_text: list of text strings
             batch_images: images to provide to model. Should be a list of lists,
               where each list contains the images for a single example.
 
         Returns:
             List of decoded output strings.
         """
+        raise NotImplementedError
 
     def get_rank_classifications(
         self,
@@ -150,7 +187,29 @@ def get_rank_classifications(
             all_class_names: list of all class names.
             use_cache: whether to cache the context to speed up evaluations.
             normalize_length: whether to normalize logprobs by the length of the
-                class name
+                class name; use with caution, as this can change predictions quite a bit.
         Returns:
             (B, |all_class_names|) tensor containing the logprobs for each class name.
         """
+        raise NotImplementedError
+
+    @property
+    def supported_tasks(self):
+        """
+        Return list of tasks that this model can be evaluated on.
+        Parsed by checking whether the model has a method called `get_{task}_prompt`.
+        """
+        return [
+            task.split("_")[1]
+            for task in dir(self)
+            if task.startswith("get_") and task.endswith("_prompt")
+        ]
+
+    def _validate_text(self, batch_text):
+        """
+        Checks for trailing whitespaces in the text and prints a warning.
+        """
+        if any([x.endswith(" ") for x in batch_text]):
+            print(
+                "Warning: trailing whitespace detected in text. This can cause unexpected behavior."
+            )

open_flamingo/eval/models/idefics.py renamed to open_flamingo/eval/eval_models/idefics.py

@@ -18,16 +18,17 @@ class EvalModel(BaseEvalModel):
     """IDEFICS model evaluation."""
 
     def __init__(self, model_args, init_on_device=False):
-        assert (
-            "lm_path" in model_args and "processor_path" in model_args
-        ), "IDEFICS requires lm_path and lm_tokenizer_path"
         super().__init__(model_args, init_on_device)
         with self.init_ctx:
             self.model = IdeficsForVisionText2Text.from_pretrained(model_args["lm_path"])
             self.processor = AutoProcessor.from_pretrained(model_args["processor_path"])
             self.tokenizer = self.processor.tokenizer
         self._check_init()
 
+    @property
+    def required_args(self):
+        return ["lm_path", "processor_path"]
+
     def prepare_images(self, batch: List[List[Image.Image]]) -> torch.Tensor:
         batch_images = self.processor(batch)["pixel_values"]
         if batch_images is not None:
@@ -44,6 +45,7 @@ def prepare_text(
         max_length=2000,
         add_special_tokens=True,
     ):
+        self._validate_text(batch)
         # check to see if there any <image> without <fake_token_around_image> wrapping it
         for i, text in enumerate(batch):
             if "<image>" in text and "<fake_token_around_image>" not in text:
@@ -88,19 +90,6 @@ def _compute_image_attention_mask(self, batch_tokens: torch.Tensor) -> torch.Ten
             )
         return image_attention_mask
 
-    def get_rank_classifications(
-        self,
-        batch_text: List[str],
-        batch_images: List[List[Image.Image]],
-        all_class_names: List[str],
-        use_cache: bool,
-        normalize_length: bool,
-    ):
-        """
-        Returns a (B, |all_class_names|) tensor containing the logprobs for each class name.
-        """
-        raise NotImplementedError
-
     def get_outputs(
         self,
         batch_text: List[str],
@@ -176,18 +165,26 @@ def __call__(
             past_key_values=past_key_values,
         )
 
-    def get_vqa_prompt(self, question, answer=None) -> str:
+    def get_vqav2_prompt(self, question, answer=None) -> str:
         # TODO: handle prefix prompts
         return f"<image>Question:{question} Answer: {answer if answer is not None else ''}{'<|endofchunk|>' if answer is not None else ''}"
-
-    def get_caption_prompt(self, caption=None) -> str:
+    
+    def get_ok_vqa_prompt(self, question, answer=None) -> str:
         # TODO: handle prefix prompts
-        return f"<image>Caption: {caption if caption is not None else ''}{'<|endofchunk|>' if caption is not None else ''}"
+        return f"<image>Question:{question} Answer: {answer if answer is not None else ''}{'<|endofchunk|>' if answer is not None else ''}"
+    
+    def get_vizwiz_prompt(self, question, answer=None) -> str:
+        # TODO: handle prefix prompts
+        return f"<image>Question:{question} Answer: {answer if answer is not None else ''}{'<|endofchunk|>' if answer is not None else ''}"
 
-    def get_imagenet_prompt(self, label=None) -> str:
+    def get_textvqa_prompt(self, question, answer=None) -> str:
         # TODO: handle prefix prompts
-        return f"<image>Output:{label if label is not None else ''}{'<|endofchunk|>' if label is not None else ''}"
+        return f"<image>Question:{question} Answer: {answer if answer is not None else ''}{'<|endofchunk|>' if answer is not None else ''}"
 
-    def get_hateful_memes_prompt(self, text, label=None) -> str:
+    def get_coco_prompt(self, caption=None) -> str:
+        # TODO: handle prefix prompts
+        return f"<image>Caption: {caption if caption is not None else ''}{'<|endofchunk|>' if caption is not None else ''}"
+    
+    def get_flickr_prompt(self, caption=None) -> str:
         # TODO: handle prefix prompts
-        return f"<image>is an image with: '{text}' written on it. Is it hateful? Answer: {label if label is not None else ''}{'<|endofchunk|>' if label is not None else ''}"
+        return f"<image>Caption: {caption if caption is not None else ''}{'<|endofchunk|>' if caption is not None else ''}"