adapt for transformers processing

image_processing_minicpmv.py

@@ -0,0 +1,405 @@
+from typing import Optional, Union, Dict, Any
+
+import torch
+import math
+import PIL.Image
+import PIL.ImageSequence
+import numpy as np
+import PIL
+from PIL import Image
+
+from transformers.utils import TensorType, requires_backends, is_torch_dtype, is_torch_device
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
+from transformers import AutoImageProcessor
+from transformers.image_transforms import to_channel_dimension_format
+from transformers.image_utils import (
+    ImageInput, 
+    make_list_of_images, 
+    valid_images, 
+    is_torch_tensor, 
+    to_numpy_array, 
+    infer_channel_dimension_format,
+    ChannelDimension
+)
+
+
+def recursive_converter(converter, value):
+    if isinstance(value, list):
+        new_value = []
+        for v in value:
+            new_value += [recursive_converter(converter, v)]
+        return new_value
+    else:
+        return converter(value)
+
+
+class MiniCPMVBatchFeature(BatchFeature):
+    r"""
+    Extend from BatchFeature for supporting various image size
+    """
+    def __init__(self, data: Optional[Dict[str, Any]] = None, tensor_type: Union[None, str, TensorType] = None):
+        super().__init__(data)
+        self.convert_to_tensors(tensor_type=tensor_type)
+
+    def convert_to_tensors(self, tensor_type: Optional[Union[str, TensorType]] = None):
+        if tensor_type is None:
+            return self
+        
+        is_tensor, as_tensor = self._get_is_as_tensor_fns(tensor_type)
+
+        def converter(value):
+            try:
+                if not is_tensor(value):
+                    tensor = as_tensor(value)
+                    return tensor
+            except:  # noqa E722
+                if key == "overflowing_values":
+                    raise ValueError("Unable to create tensor returning overflowing values of different lengths. ")
+                raise ValueError(
+                    "Unable to create tensor, you should probably activate padding "
+                    "with 'padding=True' to have batched tensors with the same length."
+                )
+
+
+        for key, value in self.items():
+            self[key] = recursive_converter(converter, value)
+        return self
+            
+    def to(self, *args, **kwargs) -> "MiniCPMVBatchFeature":
+        requires_backends(self, ["torch"])
+        import torch
+
+        def cast_tensor(v):
+            # check if v is a floating point
+            if torch.is_floating_point(v):
+                # cast and send to device
+                return v.to(*args, **kwargs)
+            elif device is not None:
+                return v.to(device=device)
+            else:
+                return v
+
+        new_data = {}
+        device = kwargs.get("device")
+        # Check if the args are a device or a dtype
+        if device is None and len(args) > 0:
+            # device should be always the first argument
+            arg = args[0]
+            if is_torch_dtype(arg):
+                # The first argument is a dtype
+                pass
+            elif isinstance(arg, str) or is_torch_device(arg) or isinstance(arg, int):
+                device = arg
+            else:
+                # it's something else
+                raise ValueError(f"Attempting to cast a BatchFeature to type {str(arg)}. This is not supported.")
+        # We cast only floating point tensors to avoid issues with tokenizers casting `LongTensor` to `FloatTensor`
+        for k, v in self.items():
+            new_data[k] = recursive_converter(cast_tensor, v)
+        self.data = new_data
+        return self
+
+
+class MiniCPMVImageProcessor(BaseImageProcessor):
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+            self, 
+            max_slice_nums=9,
+            scale_resolution=448,
+            patch_size=14,
+            **kwargs):
+        super().__init__(**kwargs)
+        self.max_slice_nums = max_slice_nums
+        self.scale_resolution = scale_resolution
+        self.patch_size = patch_size
+        self.image_feature_size = kwargs.pop("image_feature_size", 64)
+        self.im_start_token = kwargs.pop("im_start", "<image>")
+        self.im_end_token = kwargs.pop("im_end", "</image>")
+        self.slice_start_token = kwargs.pop("slice_start", "<slice>")
+        self.slice_end_token = kwargs.pop("slice_end", "</slice>")
+        self.unk_token = kwargs.pop("unk", "<unk>")
+        self.mean = np.array(kwargs.pop("norm_mean", [0.5, 0.5, 0.5]))
+        self.std = np.array(kwargs.pop("norm_std", [0.5, 0.5, 0.5]))
+        self.version = kwargs.pop("version", 2.0)
+
+    def ensure_divide(self, length, patch_size):
+        return max(round(length / patch_size) * patch_size, patch_size)
+
+    def find_best_resize(self,
+                         original_size,
+                         scale_resolution,
+                         patch_size,
+                         allow_upscale=False):
+        width, height = original_size
+        if (width * height >
+                scale_resolution * scale_resolution) or allow_upscale:
+            r = width / height
+            height = int(scale_resolution / math.sqrt(r))
+            width = int(height * r)
+        best_width = self.ensure_divide(width, patch_size)
+        best_height = self.ensure_divide(height, patch_size)
+        return (best_width, best_height)
+
+    def get_refine_size(self,
+                        original_size,
+                        grid,
+                        scale_resolution,
+                        patch_size,
+                        allow_upscale=False):
+        width, height = original_size
+        grid_x, grid_y = grid
+
+        refine_width = self.ensure_divide(width, grid_x)
+        refine_height = self.ensure_divide(height, grid_y)
+
+        grid_width = refine_width / grid_x
+        grid_height = refine_height / grid_y
+
+        best_grid_size = self.find_best_resize((grid_width, grid_height),
+                                               scale_resolution,
+                                               patch_size,
+                                               allow_upscale=allow_upscale)
+        refine_size = (best_grid_size[0] * grid_x, best_grid_size[1] * grid_y)
+        return refine_size
+
+    def split_to_patches(self, image, grid):
+        patches = []
+        width, height = image.size
+        grid_x = int(width / grid[0])
+        grid_y = int(height / grid[1])
+        for i in range(0, height, grid_y):
+            images = []
+            for j in range(0, width, grid_x):
+                box = (j, i, j + grid_x, i + grid_y)
+                patch = image.crop(box)
+                images.append(patch)
+            patches.append(images)
+        return patches
+
+    def slice_image(
+        self, image, max_slice_nums=9, scale_resolution=448, patch_size=14, never_split=False
+    ):
+        original_size = image.size
+        original_width, original_height = original_size
+        log_ratio = math.log(original_width / original_height)
+        ratio = original_width * original_height / (scale_resolution * scale_resolution)
+        multiple = min(math.ceil(ratio), max_slice_nums)
+
+        source_image = None
+        best_grid = None
+        patches = []
+
+        if multiple <= 1 or never_split:
+            # dont need to slice, upsample
+            best_size = self.find_best_resize(
+                original_size, scale_resolution, patch_size, allow_upscale=True
+            )
+            source_image = image.resize(best_size, resample=Image.Resampling.BICUBIC)
+        else:
+            candidate_split_grids_nums = []
+            for i in [multiple - 1, multiple, multiple + 1]:
+                if i == 1 or i > max_slice_nums:
+                    continue
+                candidate_split_grids_nums.append(i)
+
+            # source image, down-sampling and ensure divided by patch_size
+            best_resize = self.find_best_resize(original_size, scale_resolution, patch_size)
+            source_image = image.copy().resize(best_resize, resample=Image.Resampling.BICUBIC)
+            candidate_grids = []
+
+            # find best grid
+            for split_grids_nums in candidate_split_grids_nums:
+                m = 1
+                while m <= split_grids_nums:
+                    if split_grids_nums % m == 0:
+                        candidate_grids.append([m, split_grids_nums // m])
+                    m += 1
+
+            best_grid = [1, 1]
+            min_error = float("inf")
+            for grid in candidate_grids:
+                error = abs(log_ratio - math.log(grid[0] / grid[1]))
+                if error < min_error:
+                    best_grid = grid
+                    min_error = error
+
+            refine_size = self.get_refine_size(
+                original_size, best_grid, scale_resolution, patch_size, allow_upscale=True
+            )
+
+            refine_image = image.resize(refine_size, resample=Image.Resampling.BICUBIC)
+            patches = self.split_to_patches(refine_image, best_grid)
+
+        return source_image, patches, best_grid
+
+    def get_grid_placeholder(self, grid):
+        if grid is None:
+            return ""
+        image_placeholder = (
+            self.im_start_token 
+            + self.unk_token * self.image_feature_size
+            + self.im_end_token
+        )
+
+        cols = grid[0]
+        rows = grid[1]
+        slices = []
+        for i in range(rows):
+            lines = []
+            for j in range(cols):
+                lines.append(image_placeholder)
+            slices.append("".join(lines))
+            
+        slice_placeholder = self.slice_start_token + "\n".join(slices) + self.slice_end_token
+        return slice_placeholder
+
+    def get_sliced_images(self, image):
+        slice_images = []
+
+        source_image, patches, sliced_grid = self.slice_image(
+            image,
+            self.max_slice_nums,  # default: 9
+            self.scale_resolution,  # default: 448
+            self.patch_size  # default: 14
+        )
+        slice_images.append(source_image)
+
+        if len(patches) > 0:
+            for i in range(len(patches)):
+                for j in range(len(patches[0])):
+                    slice_images.append(patches[i][j])
+        return slice_images
+
+    def get_sliced_grid(self, image_size):
+        original_width, original_height = image_size
+        log_ratio = math.log(original_width / original_height)
+        ratio = original_width * original_height / (self.scale_resolution * self.scale_resolution)
+        multiple = min(math.ceil(ratio), self.max_slice_nums)
+        if multiple <= 1:
+            return None
+        candidate_split_grids_nums = []
+        for i in [multiple - 1, multiple, multiple + 1]:
+            if i == 1 or i > self.max_slice_nums:
+                continue
+            candidate_split_grids_nums.append(i)
+        
+        candidate_grids = []
+        for split_grids_nums in candidate_split_grids_nums:
+            m = 1
+            while m <= split_grids_nums:
+                if split_grids_nums % m == 0:
+                    candidate_grids.append([m, split_grids_nums // m])
+                m += 1
+
+        best_grid = [1, 1]
+        min_error = float("inf")
+        for grid in candidate_grids:
+            error = abs(log_ratio - math.log(grid[0] / grid[1]))
+            if error < min_error:
+                best_grid = grid
+                min_error = error
+        
+        return best_grid
+
+    def get_slice_image_placeholder(self, image_size):
+        grid = self.get_sliced_grid(image_size=image_size)
+        return (
+            self.im_start_token 
+            + self.unk_token * self.image_feature_size 
+            + self.im_end_token
+        ) + self.get_grid_placeholder(grid=grid) + "\n"
+
+    def to_pil_image(self, image, rescale=None) -> PIL.Image.Image:
+        """
+        Converts `image` to a PIL Image. Optionally rescales it and puts the channel dimension back as the last axis if
+        needed.
+
+        Args:
+            image (`PIL.Image.Image` or `numpy.ndarray` or `torch.Tensor`):
+                The image to convert to the PIL Image format.
+            rescale (`bool`, *optional*):
+                Whether or not to apply the scaling factor (to make pixel values integers between 0 and 255). Will
+                default to `True` if the image type is a floating type, `False` otherwise.
+        """
+        if isinstance(image, PIL.Image.Image):
+            return image
+        if is_torch_tensor(image):
+            image = image.numpy()
+
+        if isinstance(image, np.ndarray):
+            if rescale is None:
+                # rescale default to the array being of floating type.
+                rescale = isinstance(image.flat[0], np.floating)
+            # If the channel as been moved to first dim, we put it back at the end.
+            if image.ndim == 3 and image.shape[0] in [1, 3]:
+                image = image.transpose(1, 2, 0)
+            if rescale:
+                image = image * 255
+            image = image.astype(np.uint8)
+            return PIL.Image.fromarray(image)
+        return image
+
+    def reshape_by_patch(self, image):
+        """
+        :param image: shape [3, H, W]
+        :param patch_size:
+        :return: [3, patch_size, HW/patch_size]
+        """
+        image = torch.from_numpy(image)
+        patch_size = self.patch_size
+        patches = torch.nn.functional.unfold(
+            image,
+            (patch_size, patch_size),
+            stride=(patch_size, patch_size)
+        )
+
+        patches = patches.reshape(image.size(0), patch_size, patch_size, -1)
+        patches = patches.permute(0, 1, 3, 2).reshape(image.size(0), patch_size, -1)
+        return patches.numpy()
+
+    def preprocess(
+            self, 
+            images: ImageInput,
+            do_pad: Optional[bool] = True, # TODO: add pad for MiniCPM-Llama3-V-2_5
+            return_tensors: Optional[Union[str, TensorType]] = None
+        ) -> MiniCPMVBatchFeature:
+        images = make_list_of_images(images)
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+        
+        images = [self.to_pil_image(image).convert("RGB") for image in images]
+        input_data_format = infer_channel_dimension_format(np.array(images[0]))
+
+        new_images = []
+        image_sizes = [image.size for image in images]
+        tgt_sizes = []
+        for image in images:
+            image_patches = self.get_sliced_images(image)
+            image_patches = [to_numpy_array(image).astype(np.float32) / 255 for image in image_patches]
+            image_patches = [
+                self.normalize(image=image, mean=self.mean, std=self.std, input_data_format=input_data_format)
+                    for image in image_patches
+            ]
+            image_patches = [
+                to_channel_dimension_format(image, ChannelDimension.FIRST, input_channel_dim=input_data_format) 
+                    for image in image_patches
+            ]
+            patches_tgt_sizes = [
+                np.array((image.shape[1] // self.patch_size, image.shape[2] // self.patch_size))
+                    for image in image_patches
+            ]
+            patches_tgt_sizes = np.vstack(patches_tgt_sizes)
+
+            new_images += [image_patches]  
+            tgt_sizes += [patches_tgt_sizes]      
+
+        return MiniCPMVBatchFeature(
+            data={"pixel_values": new_images, "image_sizes": image_sizes, "tgt_sizes": tgt_sizes}, tensor_type=return_tensors
+        )
+
+AutoImageProcessor.register("MiniCPMVImageProcessor", MiniCPMVImageProcessor)

modeling_minicpmv.py

@@ -7,7 +7,6 @@ import torchvision
 from PIL import Image
 from timm.data import IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
 from torchvision import transforms
-from transformers import LlamaTokenizer
 
 from .configuration_minicpm import MiniCPMVConfig
 from .modeling_minicpm import MiniCPMForCausalLM, MiniCPMPreTrainedModel
@@ -67,11 +66,23 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
         )
 
     def get_input_embeddings(self):
-        return self.llm.get_input_embeddings()
+        return self.llm.embed_tokens
 
     def set_input_embeddings(self, value):
         self.llm.embed_tokens = value
 
+    def get_output_embeddings(self):
+        return self.llm.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.llm.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.llm = decoder
+
+    def get_decoder(self):
+        return self.llm
+
     def get_vision_embedding(self, pixel_values):
         res = []
         dtype = self.vpm.pos_embed.data.dtype
@@ -118,7 +129,7 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
             cur_vs_hs = vision_hidden_states[i]
             if len(cur_vs_hs) > 0:
                 cur_vllm_emb = vllm_embedding[i]
-                cur_image_bound = data["image_bound"][i]
+                cur_image_bound = data["image_bounds"][i]
                 if len(cur_image_bound) > 0:
                     image_indices = torch.stack(
                         [
@@ -150,59 +161,6 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
             **kwargs
         )
 
-    def _convert_to_tensors(
-        self, tokenizer, input_str, max_inp_length: Optional[int] = None
-    ):
-        if tokenizer.add_bos_token:
-            input_ids = tokenizer.encode(input_str)
-        else:
-            input_ids = [tokenizer.bos_id] + tokenizer.encode(input_str)
-        if max_inp_length is not None:
-            input_ids = input_ids[:max_inp_length]
-        input_ids = torch.tensor(input_ids, dtype=torch.int32)
-
-        image_start_tokens = torch.where(input_ids == tokenizer.im_start_id)[0]
-        # 跳过 im_start
-        image_start_tokens += 1
-        image_end_tokens = torch.where(input_ids == tokenizer.im_end_id)[0]
-        valid_image_nums = max(len(image_start_tokens), len(image_end_tokens))
-        image_bound = torch.hstack(
-            [
-                image_start_tokens[:valid_image_nums].unsqueeze(-1),
-                image_end_tokens[:valid_image_nums].unsqueeze(-1),
-            ]
-        )
-
-        model_input = {}
-        model_input["input_ids"] = input_ids.unsqueeze(0).to(self.device)
-        model_input["image_bound"] = image_bound
-
-        return model_input
-
-    def _process_list(
-        self, tokenizer, data_list: List[str], max_inp_length: Optional[int] = None
-    ):
-        pad_keys = ["input_ids"]
-        input_tensors = []
-        for data in data_list:
-            input_tensors.append(
-                self._convert_to_tensors(tokenizer, data, max_inp_length)
-            )
-        padded = {}
-        for key in pad_keys:
-            padded[key] = pad(input_tensors, key, padding_side="left").to(self.device)
-        padded["image_bound"] = [i["image_bound"] for i in input_tensors]
-        return padded
-
-    def _decode(self, inputs_embeds, tokenizer, **kwargs):
-        output = self.llm.generate(
-            inputs_embeds=inputs_embeds,
-            pad_token_id=0,
-            eos_token_id=tokenizer.eos_token_id,
-            **kwargs
-        )
-        return self._decode_text(output, tokenizer)
-
     def _decode_text(self, result_ids, tokenizer):
         result_text = []
         for result in result_ids:
@@ -214,87 +172,52 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
             result_text.append(tokenizer.decode(result).strip())
         return result_text
 
-    def slice_image(self, image):
-        return slice_image(
-            image,
-            self.config.max_slice_nums,
-            self.config.scale_resolution,
-            self.config.patch_size,
-        )
-
-    def get_slice_image_placeholder(self, image, tokenizer):
-        image_placeholder = (
-            tokenizer.im_start
-            + tokenizer.unk_token * self.config.query_num
-            + tokenizer.im_end
-        )
-
-        slice_images = []
-
-        source_image, patches, best_grid = slice_image(
-            image,
-            self.config.max_slice_nums,
-            self.config.scale_resolution,
-            self.config.patch_size,
+    def _decode(self, inputs_embeds, tokenizer, **kwargs):
+        output = self.llm.generate(
+            inputs_embeds=inputs_embeds,
+            pad_token_id=0,
+            eos_token_id=tokenizer.eos_token_id if tokenizer is not None else kwargs.pop("eos_token_id", 2),
+            **kwargs
         )
-
-        slice_images.append(source_image)
-        final_placeholder = image_placeholder
-
-        if len(patches) > 0:
-            for i in range(len(patches)):
-                for j in range(len(patches[0])):
-                    slice_images.append(patches[i][j])
-
-            final_placeholder += get_grid_placeholder(
-                tokenizer, best_grid, self.config.query_num
-            )
-
-        return slice_images, final_placeholder
+        return output
 
     def generate(
         self,
-        data_list=None,
-        img_list=None,
+        input_ids,
+        pixel_values=None,
+        image_sizes=[],
+        image_bounds=[],
+        tgt_sizes=[],
         tokenizer=None,
-        max_inp_length: Optional[int] = None,
         vision_hidden_states=None,
-        return_vision_hidden_states=False,
         **kwargs
     ):
-
-        assert data_list is not None
-        bs = len(data_list)
+        bs = len(input_ids)
+        img_list = pixel_values
         if img_list == None:
             img_list = [[] for i in range(bs)]
         assert bs == len(img_list)
 
-        model_inputs = self._process_list(tokenizer, data_list, max_inp_length)
-
         if vision_hidden_states is None:
             pixel_values = []
             for i in range(bs):
                 img_inps = []
                 for img in img_list[i]:
-                    img_inps.append(self.transform(img).to(self.device))
-                if img_inps:
-                    pixel_values.append(img_inps)
-                else:
-                    pixel_values.append([])
-            model_inputs["pixel_values"] = pixel_values
-        else:
-            model_inputs["vision_hidden_states"] = vision_hidden_states
-
-        with torch.inference_mode():
-            (
-                model_inputs["inputs_embeds"],
-                vision_hidden_states,
-            ) = self.get_vllm_embedding(model_inputs)
-
-            result = self._decode(model_inputs["inputs_embeds"], tokenizer, **kwargs)
-
-        if return_vision_hidden_states:
-            return result, vision_hidden_states
+                    img_inps.append(img.to(self.device, self.dtype))
+                pixel_values.append(img_inps)
+
+        # with torch.inference_mode():
+        (
+            input_embeds,
+            vision_hidden_states,
+        ) = self.get_vllm_embedding({
+            "input_ids": input_ids,
+            "pixel_values": pixel_values,
+            "image_sizes": image_sizes,
+            "image_bounds": image_bounds,
+            "tgt_sizes": tgt_sizes
+        })
+        result = self._decode(input_embeds, tokenizer, **kwargs)
 
         return result
 
@@ -304,6 +227,7 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
         msgs,
         context,
         tokenizer,
+        processor,
         vision_hidden_states=None,
         max_new_tokens=1024,
         sampling=True,
@@ -313,34 +237,9 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
         if isinstance(msgs, str):
             msgs = json.loads(msgs)
         # msgs to prompt
-        prompt = ""
-        for i, msg in enumerate(msgs):
-            role = msg["role"]
-            content = msg["content"]
-            assert role in ["user", "assistant"]
-            if i == 0:
-                if image is None:
-                    images = []
-                else:
-                    assert role == "user", "The role of first msg should be user"
-                    if self.config.slice_mode:
-                        images, final_placeholder = self.get_slice_image_placeholder(
-                            image, tokenizer
-                        )
-                        content = final_placeholder + "\n" + content
-                    else:
-                        images = [image]
-                        content = (
-                            tokenizer.im_start
-                            + tokenizer.unk_token * self.config.query_num
-                            + tokenizer.im_end
-                            + "\n"
-                            + content
-                        )
-            prompt += "<用户>" if role == "user" else "<AI>"
-            prompt += content
-        prompt += "<AI>"
-        final_input = prompt
+
+        prompt = processor.tokenizer.apply_chat_template(msgs)
+        inputs = processor(prompt, [image], return_tensors="pt").to(self.device)
 
         if sampling:
             generation_config = {
@@ -359,235 +258,18 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
         generation_config.update(
             (k, kwargs[k]) for k in generation_config.keys() & kwargs.keys()
         )
-
         with torch.inference_mode():
-            res, vision_hidden_states = self.generate(
-                data_list=[final_input],
-                max_inp_length=max_inp_length,
-                img_list=[images],
+            res = self.generate(
+                **inputs,
                 tokenizer=tokenizer,
                 max_new_tokens=max_new_tokens,
                 vision_hidden_states=vision_hidden_states,
-                return_vision_hidden_states=True,
-                **generation_config
+                **generation_config,
             )
+            res = self._decode_text(res, tokenizer)
         answer = res[0]
         context = msgs.copy()
         context.append({"role": "assistant", "content": answer})
 
         return answer, context, generation_config
-
-
-class LlamaTokenizerWrapper(LlamaTokenizer):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self.im_start = "<image>"
-        self.im_end = "</image>"
-        self.ref_start = "<ref>"
-        self.ref_end = "</ref>"
-        self.box_start = "<box>"
-        self.box_end = "</box>"
-        self.quad_start = "<quad>"
-        self.quad_end = "</quad>"
-        self.point_start = "<point>"
-        self.point_end = "</point>"
-        self.slice_start = "<slice>"
-        self.slice_end = "</slice>"
-
-    @property
-    def eos_id(self):
-        return self.sp_model.eos_id()
-
-    @property
-    def bos_id(self):
-        return self.sp_model.bos_id()
-
-    @property
-    def unk_id(self):
-        return self.sp_model.unk_id()
-
-    @property
-    def im_start_id(self):
-        return self._convert_token_to_id(self.im_start)
-
-    @property
-    def im_end_id(self):
-        return self._convert_token_to_id(self.im_end)
-
-
-def pad(orig_items, key, max_length=None, padding_value=0, padding_side="left"):
-    items = []
-    if isinstance(orig_items[0][key], list):
-        assert isinstance(orig_items[0][key][0], torch.Tensor)
-        for it in orig_items:
-            for tr in it[key]:
-                items.append({key: tr})
-    else:
-        assert isinstance(orig_items[0][key], torch.Tensor)
-        items = orig_items
-
-    batch_size = len(items)
-    shape = items[0][key].shape
-    dim = len(shape)
-    assert dim <= 3
-    if max_length is None:
-        max_length = 0
-    max_length = max(max_length, max(item[key].shape[-1] for item in items))
-    min_length = min(item[key].shape[-1] for item in items)
-    dtype = items[0][key].dtype
-
-    if dim == 1:
-        return torch.cat([item[key] for item in items], dim=0)
-    elif dim == 2:
-        if max_length == min_length:
-            return torch.cat([item[key] for item in items], dim=0)
-        tensor = torch.zeros((batch_size, max_length), dtype=dtype) + padding_value
-    else:
-        tensor = (
-            torch.zeros((batch_size, max_length, shape[-1]), dtype=dtype)
-            + padding_value
-        )
-
-    for i, item in enumerate(items):
-        if dim == 2:
-            if padding_side == "left":
-                tensor[i, -len(item[key][0]) :] = item[key][0].clone()
-            else:
-                tensor[i, : len(item[key][0])] = item[key][0].clone()
-        elif dim == 3:
-            if padding_side == "left":
-                tensor[i, -len(item[key][0]) :, :] = item[key][0].clone()
-            else:
-                tensor[i, : len(item[key][0]), :] = item[key][0].clone()
-
-    return tensor
-
-
-def slice_image(
-    image, max_slice_nums=9, scale_resolution=448, patch_size=14, never_split=False
-):
-    original_size = image.size
-    original_width, original_height = original_size
-    log_ratio = math.log(original_width / original_height)
-    ratio = original_width * original_height / (scale_resolution * scale_resolution)
-    multiple = min(math.ceil(ratio), max_slice_nums)
-
-    source_image = None
-    best_grid = None
-    patches = []
-
-    if multiple <= 1 or never_split:
-        # dont need to slice, upsample
-        best_size = find_best_resize(
-            original_size, scale_resolution, patch_size, allow_upscale=True
-        )
-        source_image = image.resize(best_size, Image.Resampling.BICUBIC)
-    else:
-        candidate_split_grids_nums = []
-        for i in [multiple - 1, multiple, multiple + 1]:
-            if i == 1 or i > max_slice_nums:
-                continue
-            candidate_split_grids_nums.append(i)
-
-        # source image, down-sampling and ensure divided by patch_size
-        best_resize = find_best_resize(original_size, scale_resolution, patch_size)
-        source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
-        candidate_grids = []
-
-        # find best grid
-        for split_grids_nums in candidate_split_grids_nums:
-            m = 1
-            while m <= split_grids_nums:
-                if split_grids_nums % m == 0:
-                    candidate_grids.append([m, split_grids_nums // m])
-                m += 1
-
-        best_grid = [1, 1]
-        min_error = float("inf")
-        for grid in candidate_grids:
-            error = abs(log_ratio - math.log(grid[0] / grid[1]))
-            if error < min_error:
-                best_grid = grid
-                min_error = error
-
-        refine_size = get_refine_size(
-            original_size, best_grid, scale_resolution, patch_size, allow_upscale=True
-        )
-
-        refine_image = image.resize(refine_size, Image.Resampling.BICUBIC)
-        patches = split_to_patches(refine_image, best_grid)
-
-    return source_image, patches, best_grid
-
-
-def ensure_divide(length, patch_size):
-    return max(round(length / patch_size) * patch_size, patch_size)
-
-
-def find_best_resize(original_size, scale_resolution, patch_size, allow_upscale=False):
-    width, height = original_size
-    if (width * height > scale_resolution * scale_resolution) or allow_upscale:
-        r = width / height
-        height = int(scale_resolution / math.sqrt(r))
-        width = int(height * r)
-    best_width = ensure_divide(width, patch_size)
-    best_height = ensure_divide(height, patch_size)
-    return (best_width, best_height)
-
-
-def get_refine_size(
-    original_size, grid, scale_resolution, patch_size, allow_upscale=False
-):
-    width, height = original_size
-    grid_x, grid_y = grid
-
-    refine_width = ensure_divide(width, grid_x)
-    refine_height = ensure_divide(height, grid_y)
-
-    grid_width = refine_width / grid_x
-    grid_height = refine_height / grid_y
-
-    best_grid_size = find_best_resize(
-        (grid_width, grid_height),
-        scale_resolution,
-        patch_size,
-        allow_upscale=allow_upscale,
-    )
-
-    refine_size = (best_grid_size[0] * grid_x, best_grid_size[1] * grid_y)
-
-    return refine_size
-
-
-def split_to_patches(image, grid):
-    patches = []
-    width, height = image.size
-    grid_x = int(width / grid[0])
-    grid_y = int(height / grid[1])
-
-    for i in range(0, height, grid_y):
-        images = []
-        for j in range(0, width, grid_x):
-            box = (j, i, j + grid_x, i + grid_y)
-            patch = image.crop(box)
-            images.append(patch)
-        patches.append(images)
-
-    return patches
-
-
-def get_grid_placeholder(tokenizer, grid, query_num):
-    image_placeholder = (
-        tokenizer.im_start + tokenizer.unk_token * query_num + tokenizer.im_end
-    )
-
-    cols = grid[0]
-    rows = grid[1]
-    slices = []
-    for i in range(rows):
-        lines = []
-        for j in range(cols):
-            lines.append(image_placeholder)
-        slices.append("".join(lines))
-    slice_placeholder = tokenizer.slice_start + "\n".join(slices) + tokenizer.slice_end
-    return slice_placeholder
+    

preprocessor_config.json

@@ -0,0 +1,19 @@
+{
+    "image_processor_type": "MiniCPMVImageProcessor",
+    "auto_map": {
+        "AutoProcessor": "processing_minicpmv.MiniCPMVProcessor",
+        "AutoImageProcessor": "image_processing_minicpmv.MiniCPMVImageProcessor"
+      },
+    "processor_class": "MiniCPMVProcessor",
+    "max_slice_nums": 9,
+    "scale_resolution": 448,
+    "patch_size": 14,
+    "image_feature_size": 64,
+    "im_start": "<image>",
+    "im_end": "</image>",
+    "slice_start": "<slice>",
+    "slice_end": "</slice>",
+    "unk": "<unk>",
+    "norm_mean": [0.5, 0.5, 0.5],
+    "norm_std": [0.5, 0.5, 0.5]
+}

processing_minicpmv.py

@@ -0,0 +1,195 @@
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Processor class for MiniCPMV.
+"""
+
+from typing import List, Optional, Union
+import torch
+import re
+
+from transformers.image_utils import ImageInput
+from transformers.processing_utils import ProcessorMixin
+from transformers.tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from transformers.utils import TensorType
+
+from .image_processing_minicpmv import MiniCPMVBatchFeature
+
+
+class MiniCPMVProcessor(ProcessorMixin):
+    r"""
+    Constructs a MiniCPMV processor which wraps a MiniCPMV image processor and a MiniCPMV tokenizer into a single processor.
+
+    [`MiniCPMVProcessor`] offers all the functionalities of [`MiniCPMVImageProcessor`] and [`LlamaTokenizerWrapper`]. See the
+    [`~MiniCPMVProcessor.__call__`] and [`~MiniCPMVProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`MiniCPMVImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`LlamaTokenizerWrapper`], *optional*):
+            The tokenizer is a required input.
+    """
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(self, image_processor=None, tokenizer=None):
+        super().__init__(image_processor, tokenizer)
+    
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]],
+        images: ImageInput = None,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length: Optional[int] = None,
+        do_pad: Optional[bool] = True,
+        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
+    ) -> MiniCPMVBatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        LlavaNextImageProcessor's [`~LlavaNextImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
+        of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
+                Select a strategy to pad the returned sequences (according to the model's padding side and padding
+                index) among:
+                - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+                  sequence if provided).
+                - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
+                  acceptable input length for the model if that argument is not provided.
+                - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
+                  lengths).
+            max_length (`int`, *optional*):
+                Maximum length of the returned list and optionally padding length (see above).
+            do_pad (`bool`, *optional*, defaults to self.do_pad):
+                Whether to pad the image. If `True` will pad the images in the batch to the largest image in the batch
+                and create a pixel mask. Padding will be applied to the bottom and right of the image with zeros.
+            truncation (`bool`, *optional*):
+                Activates truncation to cut input sequences longer than `max_length` to `max_length`.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+        if images is not None:
+            image_inputs = self.image_processor(images, do_pad=do_pad, return_tensors=return_tensors)
+        return self._convert_images_texts_to_inputs(image_inputs, text, max_length=max_length)
+    
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        output_ids = args[0]
+        result_text = []
+        for result in output_ids:
+            result = result[result != 0]
+            if result[0] == self.tokenizer.bos_id:
+                result = result[1:]
+            if result[-1] == self.tokenizer.eos_id:
+                result = result[:-1]
+            result_text.append(self.tokenizer.decode(result, *args[1:], **kwargs).strip())
+        return result_text
+        # return self.tokenizer.batch_decode(*args, **kwargs)
+    
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        result = args[0]
+        result = result[result != 0]
+        if result[0] == self.tokenizer.bos_id:
+            result = result[1:]
+        if result[-1] == self.tokenizer.eos_id:
+            result = result[:-1]
+        return self.tokenizer.decode(result, *args[1:], **kwargs).strip()
+
+    def _convert(
+        self, input_str, max_inp_length: Optional[int] = None
+    ):
+        if self.tokenizer.add_bos_token:
+            input_ids = self.tokenizer.encode(input_str)
+        else:
+            input_ids = [self.tokenizer.bos_id] + self.tokenizer.encode(input_str)
+        if max_inp_length is not None:
+            input_ids = input_ids[:max_inp_length]
+        input_ids = torch.tensor(input_ids, dtype=torch.int32)
+
+        image_start_tokens = torch.where(input_ids == self.tokenizer.im_start_id)[0]
+        image_start_tokens += 1
+        image_end_tokens = torch.where(input_ids == self.tokenizer.im_end_id)[0]
+        valid_image_nums = max(len(image_start_tokens), len(image_end_tokens))
+        image_bounds = torch.hstack(
+            [
+                image_start_tokens[:valid_image_nums].unsqueeze(-1),
+                image_end_tokens[:valid_image_nums].unsqueeze(-1),
+            ]
+        )
+        return input_ids.unsqueeze(0), image_bounds
+
+    def _convert_images_texts_to_inputs(self, images, texts, do_pad=False, truncation=None, max_length=None, return_tensors=None):
+        if not len(images):
+            model_inputs = self.tokenizer(texts, return_tensors=return_tensors, padding=do_pad, truncation=truncation, max_length=max_length)
+            return MiniCPMVBatchFeature(data={**model_inputs})
+        
+        pattern = "(<image>./</image>)"
+        images, image_sizes = images["pixel_values"], images["image_sizes"]
+
+        image_tags = re.findall(pattern, texts)
+        assert len(image_tags) <= 1 and len(image_sizes) == 1
+        text_chunks = texts.split(pattern)
+        final_texts = text_chunks[0] + self.image_processor.get_slice_image_placeholder(image_sizes[0]) \
+            + text_chunks[1] + "<AI>"
+        input_ids, image_bounds = self._convert(final_texts, max_length)
+        
+        return MiniCPMVBatchFeature(data={
+            "input_ids": input_ids,
+            "pixel_values": images,
+            "image_sizes": [image_sizes],
+            "image_bounds": [image_bounds]
+        })
+
+    @property
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

tokenization_minicpmv.py

@@ -0,0 +1,61 @@
+import json
+
+from transformers import LlamaTokenizer
+
+
+class MiniCPMVTokenizer(LlamaTokenizer):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.im_start = "<image>"
+        self.im_end = "</image>"
+        self.ref_start = "<ref>"
+        self.ref_end = "</ref>"
+        self.box_start = "<box>"
+        self.box_end = "</box>"
+        self.quad_start = "<quad>"
+        self.quad_end = "</quad>"
+        self.point_start = "<point>"
+        self.point_end = "</point>"
+        self.slice_start = "<slice>"
+        self.slice_end = "</slice>"
+
+    @property
+    def eos_id(self):
+        return self.sp_model.eos_id()
+
+    @property
+    def bos_id(self):
+        return self.sp_model.bos_id()
+
+    @property
+    def unk_id(self):
+        return self.sp_model.unk_id()
+
+    @property
+    def im_start_id(self):
+        return self._convert_token_to_id(self.im_start)
+
+    @property
+    def im_end_id(self):
+        return self._convert_token_to_id(self.im_end)
+
+    def apply_chat_template(self, 
+                            conversation, 
+                            add_image_msg: bool=True):
+        if isinstance(conversation, str):
+            conversation = json.loads(conversation)
+        
+        prompt = ""
+        for i, msg in enumerate(conversation):
+            role = msg["role"]
+            content = msg["content"]
+            assert role in ["user", "assistant"]
+            if i == 0:
+                assert role == "user", "The role of first msg should be user"
+                if add_image_msg is True and "(<image>./</image>)" not in content:
+                    content = "(<image>./</image>)" + content
+            prompt += "<用户>" if role == "user" else "<AI>"
+            prompt += content
+        prompt += "<AI>"
+        return prompt
+    

tokenizer_config.json

@@ -139,7 +139,7 @@
   ],
   "auto_map": {
     "AutoTokenizer": [
-      "modeling_minicpmv.LlamaTokenizerWrapper",
+      "tokenization_minicpmv.MiniCPMVTokenizer",
       null
     ]
   },
@@ -152,7 +152,7 @@
   "padding_side": "right",
   "sp_model_kwargs": {},
   "spaces_between_special_tokens": false,
-  "tokenizer_class": "LlamaTokenizerWrapper",
+  "tokenizer_class": "MiniCPMVTokenizer",
   "truncation_side": "right",
   "unk_token": "<unk>",
   "use_default_system_prompt": false