fount-vlm-nell-02 0.3.11__py3-none-any.whl

mlx_vlm/models/deepseekocr_2/processing_deepseekocr.py

@@ -0,0 +1,624 @@
+"""
+From https://github.com/deepseek-ai/DeepSeek-VL2
+"""
+
+from dataclasses import dataclass
+from typing import Dict, List, Literal, Optional, Tuple
+
+import mlx.core as mx
+import numpy as np
+from PIL import Image, ImageOps
+from transformers import LlamaTokenizerFast
+from transformers.processing_utils import ProcessorMixin
+
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float("inf")
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    # print(f'width: {width}, height: {height}, best_ratio: {best_ratio}')
+    return best_ratio
+
+
+def dynamic_preprocess(
+    image, min_num=2, max_num=9, image_size=640, use_thumbnail=False
+):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j)
+        for n in range(min_num, max_num + 1)
+        for i in range(1, n + 1)
+        for j in range(1, n + 1)
+        if i * j <= max_num and i * j >= min_num
+    )
+    # print(target_ratios)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size
+    )
+
+    # print(target_aspect_ratio)
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size,
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images, target_aspect_ratio
+
+
+class DictOutput(object):
+    def keys(self):
+        return self.__dict__.keys()
+
+    def __getitem__(self, item):
+        if isinstance(item, int):
+            return list(self.__dict__.values())[item]
+        if item not in self.__dict__:
+            raise KeyError(item)
+        return self.__dict__[item]
+
+    def __setitem__(self, key, value):
+        self.__dict__[key] = value
+
+
+@dataclass
+class VLChatProcessorOutput(DictOutput):
+    sft_format: str
+    input_ids: mx.array
+    target_ids: mx.array
+    images: mx.array
+    images_seq_mask: mx.array
+    images_spatial_crop: mx.array
+    num_image_tokens: List[int]
+
+    def __len__(self):
+        return len(self.input_ids)
+
+
+@dataclass
+class BatchCollateOutput(DictOutput):
+    sft_format: List[str]
+    input_ids: mx.array
+    labels: mx.array
+    images: mx.array
+    attention_mask: mx.array
+    images_seq_mask: mx.array
+    images_spatial_crop: mx.array
+    seq_lens: List[int]
+
+
+class ImageTransform:
+    def __init__(
+        self,
+        mean: Optional[Tuple[float, float, float]] = (0.5, 0.5, 0.5),
+        std: Optional[Tuple[float, float, float]] = (0.5, 0.5, 0.5),
+        normalize: bool = True,
+    ):
+        self.mean = mean
+        self.std = std
+        self.normalize = normalize
+
+    def __call__(self, pil_img: Image.Image):
+        # Convert PIL image to numpy array and normalize
+
+        img = mx.array(np.array(pil_img)) / 255.0
+
+        # Transpose from HWC to CHW format
+        img = mx.transpose(img, [2, 0, 1])
+
+        if self.normalize:
+            mean = mx.array(self.mean).reshape(-1, 1, 1)
+            std = mx.array(self.std).reshape(-1, 1, 1)
+            img = (img - mean) / std
+
+        return img
+
+
+class DeepseekOCR2Processor(ProcessorMixin):
+    tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
+    attributes = ["tokenizer"]
+
+    def __init__(
+        self,
+        tokenizer: LlamaTokenizerFast,
+        candidate_resolutions: Tuple[Tuple[int, int]],
+        patch_size: int,
+        downsample_ratio: int,
+        image_mean: Tuple[float, float, float] = (0.5, 0.5, 0.5),
+        image_std: Tuple[float, float, float] = (0.5, 0.5, 0.5),
+        normalize: bool = True,
+        image_token: str = "<image>",
+        pad_token: str = "<｜▁pad▁｜>",
+        add_special_token: bool = False,
+        sft_format: str = "deepseek",
+        mask_prompt: bool = True,
+        ignore_id: int = -100,
+        **kwargs,
+    ):
+        self.candidate_resolutions = candidate_resolutions
+        self.image_size = candidate_resolutions[0][0]
+        self.patch_size = patch_size
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.normalize = normalize
+        self.downsample_ratio = downsample_ratio
+
+        self.image_transform = ImageTransform(
+            mean=image_mean, std=image_std, normalize=normalize
+        )
+        self.tokenizer = tokenizer
+        self.tokenizer.padding_side = "left"
+
+        # Add special tokens
+        if tokenizer.pad_token is None:
+            self.tokenizer.add_special_tokens({"pad_token": pad_token})
+        print(
+            f"Add pad token = ['{pad_token}'] to the tokenizer\n"
+            f"{pad_token}:{tokenizer.encode(pad_token, add_special_tokens=False)[0]}"
+        )
+
+        image_token_id = self.tokenizer.vocab.get(image_token)
+        if image_token_id is None:
+            special_tokens = [image_token]
+            special_tokens_dict = {"additional_special_tokens": special_tokens}
+            self.tokenizer.add_special_tokens(special_tokens_dict)
+        self.image_token_id = self.tokenizer.vocab.get(image_token)
+        print(
+            f"Add image token = ['{image_token}'] to the tokenizer\n"
+            f"{image_token}:{tokenizer.encode(image_token, add_special_tokens=False)[0]}"
+        )
+
+        # Add grounding-related tokens
+        special_tokens = ["<|ref|>", "<|/ref|>", "<|det|>", "<|/det|>", "<|grounding|>"]
+        special_tokens_dict = {"additional_special_tokens": special_tokens}
+        self.tokenizer.add_special_tokens(special_tokens_dict)
+        print("Added grounding-related tokens")
+
+        # Add chat tokens
+        special_tokens = ["<|User|>", "<|Assistant|>"]
+        special_tokens_dict = {"additional_special_tokens": special_tokens}
+        self.tokenizer.add_special_tokens(special_tokens_dict)
+        print("Added chat tokens")
+
+        self.image_token = image_token
+        self.pad_token = pad_token
+        self.add_special_token = add_special_token
+        self.sft_format = sft_format
+        self.mask_prompt = mask_prompt
+        self.ignore_id = ignore_id
+
+        super().__init__(tokenizer, **kwargs)
+
+        # Add chat template
+        self.chat_template = kwargs.pop("chat_template", self.default_chat_template)
+
+    @property
+    def default_chat_template(self):
+        return (
+            "{% for message in messages %}"
+            "{% if message['role'] == 'user' %}"
+            "{% elif message['role'] == 'assistant' %}{% endif %}"
+            "{{message['content']}} "
+            "{% endfor %}"
+            "{% if add_generation_prompt %}{% endif %}"
+        )
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+
+    @property
+    def pad_id(self):
+        return self.tokenizer.pad_token_id
+
+    def encode(self, text: str, bos: bool = True, eos: bool = False):
+        t = self.tokenizer.encode(text, add_special_tokens=False)
+
+        if bos:
+            t = [self.bos_id] + t
+        if eos:
+            t = t + [self.eos_id]
+
+        return t
+
+    def decode(self, t: List[int], **kwargs) -> str:
+        return self.tokenizer.decode(t, **kwargs)
+
+    def process_one(
+        self,
+        prompt: str = None,
+        images: List[Image.Image] = None,
+        inference_mode: bool = True,
+        base_size: int = 1024,
+        image_size: int = 768,
+        cropping: bool = True,
+        min_patches: int = 1,
+        max_patches: int = 6,
+    ):
+
+        sft_format = prompt
+        (
+            tokenized_str,
+            images_list,
+            images_seq_mask,
+            images_spatial_crop,
+            num_image_tokens,
+        ) = self.tokenize_with_images(
+            sft_format,
+            images,
+            base_size=base_size,
+            image_size=image_size,
+            cropping=cropping,
+            min_patches=min_patches,
+            max_patches=max_patches,
+        )
+
+        masked_tokenized_str = []
+        for token_index in tokenized_str:
+            if token_index != self.image_token_id:
+                masked_tokenized_str.append(token_index)
+            else:
+                masked_tokenized_str.append(self.ignore_id)
+
+        input_ids = mx.array(tokenized_str)
+        target_ids = mx.array(masked_tokenized_str)
+        images_seq_mask = mx.array(images_seq_mask)
+
+        # Set ignored indices
+        target_ids = mx.where(
+            (input_ids < 0) | (input_ids == self.image_token_id),
+            self.ignore_id,
+            target_ids,
+        )
+        input_ids = mx.where(input_ids < 0, self.pad_id, input_ids)
+
+        if inference_mode:
+            input_ids = input_ids[:-1]
+            target_ids = target_ids[:-1]
+            images_seq_mask = images_seq_mask[:-1]
+
+        return {
+            "input_ids": input_ids[None, :],
+            "attention_mask": input_ids != self.pad_id,
+            "labels": target_ids,
+            "images": images_list,
+            "images_seq_mask": images_seq_mask[None, ...],
+            "images_spatial_crop": images_spatial_crop,
+            "num_image_tokens": num_image_tokens,
+        }
+
+    def pad_sequence(self, sequences, padding_value):
+        # Get max length of sequences
+        max_len = max(len(seq) for seq in sequences)
+
+        # Pad each sequence to max length
+        padded_seqs = []
+        for seq in sequences:
+            pad_length = max_len - len(seq)
+            if pad_length > 0:
+                padding = mx.full((pad_length,), padding_value)
+                padded_seq = mx.concatenate([seq, padding])
+            else:
+                padded_seq = seq
+            padded_seqs.append(padded_seq)
+
+        return mx.stack(padded_seqs)
+
+    def tokenize_with_images(
+        self,
+        conversation: str,
+        images: List[Image.Image],
+        base_size: int = 1024,
+        image_size: int = 768,
+        cropping: bool = True,
+        min_patches: int = 1,
+        max_patches: int = 6,
+    ):
+        """Tokenize text with <image> tags.
+
+        For DeepSeek-OCR-2 with Qwen2 encoder:
+        - Global view (1024x1024): 256 tokens from Qwen2 encoder
+        - Local patches (768x768): 144 tokens each from Qwen2 encoder
+        - Plus 1 view_separator token
+
+        Dynamic resolution:
+        - Total tokens = (num_patches * 144) + 256 + 1
+        - Default: 0-6 patches at 768x768 + 1 global at 1024x1024
+        """
+        # Token counts for Qwen2 encoder
+        TOKENS_PER_PATCH = 144  # 12x12 SAM features for 768x768
+        TOKENS_PER_GLOBAL = 256  # 16x16 SAM features for 1024x1024
+        TOKENS_VIEW_SEP = 1
+
+        assert conversation.count(self.image_token) == len(
+            images
+        ), f"The number of image tokens in the prompt does not match the number of images: {conversation.count(self.image_token)} != {len(images)}"
+
+        text_splits = conversation.split(self.image_token)
+
+        all_patches_list = []
+        all_global_list = []
+        images_seq_mask = []
+        tokenized_str = []
+        images_spatial_crop = []
+        num_image_tokens_list = []
+
+        for text_sep, image in zip(text_splits, images):
+            # Tokenize the text before this image
+            tokenized_sep = self.encode(text_sep, bos=False, eos=False)
+            tokenized_str += tokenized_sep
+            images_seq_mask += [False] * len(tokenized_sep)
+
+            # Process global view: pad to base_size x base_size (1024x1024)
+            global_view = ImageOps.pad(
+                image,
+                (base_size, base_size),
+                color=tuple(int(x * 255) for x in self.image_transform.mean),
+            )
+            global_tensor = self.image_transform(global_view).astype(mx.bfloat16)
+            all_global_list.append(global_tensor)
+
+            # Process local patches using dynamic resolution
+            if cropping and min_patches > 0:
+                # Use dynamic_preprocess to split image into patches
+                patches, (rows, cols) = dynamic_preprocess(
+                    image,
+                    min_num=min_patches,
+                    max_num=max_patches,
+                    image_size=image_size,  # 768x768 patches
+                    use_thumbnail=False,
+                )
+                num_patches = len(patches)
+
+                # Transform each patch
+                patch_tensors = []
+                for patch in patches:
+                    patch_tensor = self.image_transform(patch).astype(mx.bfloat16)
+                    patch_tensors.append(patch_tensor)
+
+                if patch_tensors:
+                    patches_stacked = mx.stack(patch_tensors, axis=0)
+                    all_patches_list.append(patches_stacked)
+
+                images_spatial_crop.append([rows, cols])
+            else:
+                # No patches, only global view
+                num_patches = 0
+                images_spatial_crop.append([0, 0])
+
+            # Calculate number of image tokens for this image
+            # Order: [local_patches, global_view, view_separator]
+            num_image_tokens = (
+                (num_patches * TOKENS_PER_PATCH) + TOKENS_PER_GLOBAL + TOKENS_VIEW_SEP
+            )
+            num_image_tokens_list.append(num_image_tokens)
+
+            # Add image tokens to sequence
+            tokenized_image = [self.image_token_id] * num_image_tokens
+            tokenized_str += tokenized_image
+            images_seq_mask += [True] * len(tokenized_image)
+
+        # Tokenize the text after the last image
+        tokenized_sep = self.encode(text_splits[-1], bos=False, eos=False)
+        tokenized_str += tokenized_sep
+        images_seq_mask += [False] * len(tokenized_sep)
+
+        # Add the bos token
+        bos_id = 0
+        tokenized_str = [bos_id] + tokenized_str
+        images_seq_mask = [False] + images_seq_mask
+
+        images_seq_mask = mx.array(images_seq_mask)
+
+        # Stack global images
+        if len(all_global_list) == 0:
+            images_ori = mx.zeros((1, 3, base_size, base_size))
+            images_spatial_crop = mx.zeros((1, 2))
+        else:
+            images_ori = mx.stack(all_global_list, axis=0)
+            images_spatial_crop = mx.array(images_spatial_crop)
+
+        # Stack patches (or zeros if no patches)
+        if all_patches_list:
+            # Concatenate all patches from all images
+            images_patches = mx.concatenate(all_patches_list, axis=0)
+        else:
+            images_patches = mx.zeros((1, 3, image_size, image_size))
+
+        assert len(tokenized_str) == len(
+            images_seq_mask
+        ), f"tokenize_with_images func: tokenized_str's length {len(tokenized_str)} is not equal to images_seq_mask's length {len(images_seq_mask)}"
+
+        return (
+            tokenized_str,
+            [images_patches, images_ori],
+            images_seq_mask,
+            images_spatial_crop,
+            num_image_tokens_list[0] if num_image_tokens_list else 257,
+        )
+
+    def __call__(
+        self,
+        *,
+        text: str = None,
+        images: List[Image.Image] = None,
+        inference_mode: bool = True,
+        image_size: int = 768,
+        base_size: int = 1024,
+        cropping: bool = True,
+        min_patches: int = 1,
+        max_patches: int = 6,
+        padding: bool = True,
+        return_tensors: Literal["np", "mx", "pt"] = "mx",
+        **kwargs,
+    ):
+        """Process text and images for DeepSeek-OCR-2.
+
+        Args:
+            text (str or List[str]): the formatted prompt(s)
+            images (List[ImageType]): the list of images (one per prompt for batched inputs)
+            inference_mode (bool): if True, remove the last eos token
+            image_size (int): size of local patches (default 768)
+            base_size (int): size of global view (default 1024)
+            cropping (bool): whether to use dynamic resolution with local patches
+            min_patches (int): minimum number of patches (default 1)
+            max_patches (int): maximum number of patches (default 6)
+
+        Returns:
+            outputs (dict): the output of the processor,
+                - input_ids (mx.array): [batch_size, N + image tokens]
+                - images (List[mx.array]): [patches, global_images]
+                - images_seq_mask (mx.array): mask for image token positions
+                - images_spatial_crop (mx.array): patch grid dimensions
+        """
+
+        # Handle batched inputs (list of prompts with list of images)
+        if isinstance(text, list):
+            if images is None:
+                images = [None] * len(text)
+
+            batch_results = []
+            for i, prompt in enumerate(text):
+                # Each prompt has one image
+                img = [images[i]] if images[i] is not None else None
+                result = self.process_one(
+                    prompt=prompt,
+                    images=img,
+                    inference_mode=inference_mode,
+                    image_size=image_size,
+                    base_size=base_size,
+                    cropping=cropping,
+                    min_patches=min_patches,
+                    max_patches=max_patches,
+                )
+                batch_results.append(result)
+
+            # Collate batch results
+            return self._collate_batch(batch_results, padding=padding)
+
+        # Single input case
+        prepare = self.process_one(
+            prompt=text,
+            images=images,
+            inference_mode=inference_mode,
+            image_size=image_size,
+            base_size=base_size,
+            cropping=cropping,
+            min_patches=min_patches,
+            max_patches=max_patches,
+        )
+
+        return prepare
+
+    def _collate_batch(self, batch_results: List[Dict], padding: bool = True) -> Dict:
+        """Collate multiple processed results into a batch."""
+        if not batch_results:
+            return {}
+
+        batch_size = len(batch_results)
+
+        # Get max sequence length for padding
+        max_seq_len = max(r["input_ids"].shape[1] for r in batch_results)
+
+        # Pad and stack input_ids
+        padded_input_ids = []
+        padded_images_seq_mask = []
+        for r in batch_results:
+            seq_len = r["input_ids"].shape[1]
+            pad_len = max_seq_len - seq_len
+
+            if pad_len > 0:
+                # Pad input_ids on the left
+                input_ids = mx.concatenate(
+                    [
+                        mx.full((1, pad_len), self.pad_id, dtype=r["input_ids"].dtype),
+                        r["input_ids"],
+                    ],
+                    axis=1,
+                )
+                # Pad images_seq_mask on the left with False
+                seq_mask = mx.concatenate(
+                    [mx.zeros((1, pad_len), dtype=mx.bool_), r["images_seq_mask"]],
+                    axis=1,
+                )
+            else:
+                input_ids = r["input_ids"]
+                seq_mask = r["images_seq_mask"]
+
+            padded_input_ids.append(input_ids)
+            padded_images_seq_mask.append(seq_mask)
+
+        # Stack into batch
+        input_ids = mx.concatenate(padded_input_ids, axis=0)
+        images_seq_mask = mx.concatenate(padded_images_seq_mask, axis=0)
+
+        # Combine images: [patches, global_images]
+        all_patches = []
+        all_global_images = []
+        all_spatial_crops = []
+
+        for r in batch_results:
+            patches, global_img = r["images"]
+            # Only add non-zero patches
+            if mx.sum(patches).item() != 0:
+                all_patches.append(patches)
+            all_global_images.append(global_img)
+            all_spatial_crops.append(r["images_spatial_crop"])
+
+        # Stack patches and global images
+        if all_patches:
+            combined_patches = mx.concatenate(all_patches, axis=0)
+        else:
+            combined_patches = mx.zeros((1, 3, 1024, 1024))
+
+        combined_global_images = mx.concatenate(all_global_images, axis=0)
+        combined_spatial_crops = mx.concatenate(all_spatial_crops, axis=0)
+
+        return {
+            "input_ids": input_ids,
+            "attention_mask": input_ids != self.pad_id,
+            "images": [combined_patches, combined_global_images],
+            "images_seq_mask": images_seq_mask,
+            "images_spatial_crop": combined_spatial_crops,
+        }
+
+
+# Install a composable AutoProcessor patch for DeepSeek-OCR-2
+from ..base import install_auto_processor_patch
+
+install_auto_processor_patch("deepseekocr_2", DeepseekOCR2Processor)