nexaai 1.0.19rc6__cp310-cp310-macosx_14_0_universal2.whl → 1.0.19rc7__cp310-cp310-macosx_14_0_universal2.whl

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/internvl_chat/processor.py

@@ -0,0 +1,393 @@
+import json
+from pathlib import Path
+from typing import List, Optional, Tuple, Union
+
+import mlx.core as mx
+import numpy as np
+from PIL import Image
+from transformers import (
+    AutoImageProcessor,
+    AutoProcessor,
+    AutoTokenizer,
+    BatchFeature,
+    PreTrainedTokenizerBase,
+    ProcessorMixin,
+)
+from transformers.image_utils import ImageFeatureExtractionMixin
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+# Constants for image processing (from internvl_chat.py)
+
+IMAGENET_MEAN = np.array([0.485, 0.456, 0.406])
+IMAGENET_STD = np.array([0.229, 0.224, 0.225])
+# chat_template = get_conv_template("internvl2_5")
+chat_template = "{% for message in messages %}{{message['role'].capitalize() + ': '}}{# Render all images first #}{% for content in message['content'] | selectattr('type', 'equalto', 'image') %}{{ '<image>\n' }}{% endfor %}{# Render all text next #}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{{ content['content'] }}{% endfor %}{{'\n'}}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:\n' }}{% endif %}"
+
+IMG_START_TOKEN = "<img>"
+IMG_END_TOKEN = "</img>"
+IMG_CONTEXT_TOKEN = "<IMG_CONTEXT>"
+
+
+def build_transform(input_size):
+    """
+    Builds a transformation pipeline for images.
+
+    Args:
+        input_size (int): The target size for the image (height and width).
+
+    Returns:
+        function: A function that takes a PIL image and returns a normalized mx.array.
+    """
+    mean = mx.array(IMAGENET_MEAN)
+    std = mx.array(IMAGENET_STD)
+
+    def transform(img: Image.Image) -> mx.array:
+        # Ensure image is RGB
+        if img.mode != "RGB":
+            img = img.convert("RGB")
+
+        # Resize using PIL - BICUBIC interpolation is default in Pillow >= 9.1.0 for resize
+        # For older versions, you might need Pillow-SIMD or explicitly set
+        # resampling=Image.BICUBIC if available.
+        img = img.resize((input_size, input_size), resample=Image.Resampling.BICUBIC)
+
+        # Convert PIL image to NumPy array (H, W, C) and scale to [0, 1]
+        img_np = np.array(img).astype(np.float32) / 255.0
+
+        # Convert to MLX array and transpose to (C, H, W)
+        img_mx = mx.array(img_np).transpose(2, 0, 1)
+
+        # Normalize
+        img_mx = (img_mx - mean[:, None, None]) / std[:, None, None]
+
+        return img_mx
+
+    return transform
+
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    """Finds the closest aspect ratio from a list of targets."""
+    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:
+            # Prioritize ratios closer to the original image area if diffs are equal
+            target_area = image_size * image_size * ratio[0] * ratio[1]
+            if abs(area - target_area) < abs(
+                area - image_size * image_size * best_ratio[0] * best_ratio[1]
+            ):
+                best_ratio = ratio
+    return best_ratio
+
+
+def dynamic_preprocess(
+    image: Image.Image, min_num=1, max_num=12, image_size=448, use_thumbnail=False
+):
+    """
+    Preprocesses the image by splitting it into blocks based on the closest aspect ratio.
+
+    Args:
+        image (PIL.Image.Image): Input image.
+        min_num (int): Minimum number of blocks.
+        max_num (int): Maximum number of blocks.
+        image_size (int): Target size for each block.
+        use_thumbnail (bool): Whether to include a thumbnail of the original image.
+
+    Returns:
+        list[PIL.Image.Image]: A list of processed image blocks (as PIL images).
+    """
+    orig_width, orig_height = image.size
+    if orig_width == 0 or orig_height == 0:
+        # Handle potential zero dimensions
+        return []
+    aspect_ratio = orig_width / orig_height
+
+    # Calculate the possible target aspect ratios
+    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 min_num <= i * j <= max_num
+    )
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # Find the closest target aspect ratio
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size
+    )
+
+    # Calculate the target dimensions for resizing
+    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 to fit the target block structure
+    # Using BICUBIC resampling
+    resized_img = image.resize(
+        (target_width, target_height), resample=Image.Resampling.BICUBIC
+    )
+
+    processed_images = []
+    # Crop the resized image into blocks
+    for i in range(blocks):
+        # Calculate crop box for the i-th block
+        row_idx = i // target_aspect_ratio[0]
+        col_idx = i % target_aspect_ratio[0]
+        left = col_idx * image_size
+        top = row_idx * image_size
+        right = (col_idx + 1) * image_size
+        bottom = (row_idx + 1) * image_size
+        box = (left, top, right, bottom)
+
+        # Crop and add the block
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+
+    assert (
+        len(processed_images) == blocks
+    ), f"Expected {blocks} blocks, but got {len(processed_images)}"
+
+    # Add a thumbnail if requested and if the image was split
+    if use_thumbnail and blocks > 1:
+        thumbnail_img = image.resize(
+            (image_size, image_size), resample=Image.Resampling.BICUBIC
+        )
+        processed_images.append(thumbnail_img)
+
+    return processed_images
+
+
+class InternVLImageProcessor(ImageFeatureExtractionMixin):
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: int = 448,  # Default image size from dynamic_preprocess
+        resample=Image.Resampling.BICUBIC,
+        do_center_crop: bool = False,  # Not used in original, but standard HF param
+        crop_size=None,
+        do_rescale: bool = True,  # Original code scales by 1/255.0
+        rescale_factor: float = 1 / 255.0,
+        do_normalize: bool = True,
+        image_mean=IMAGENET_MEAN.tolist(),
+        image_std=IMAGENET_STD.tolist(),
+        do_dynamic_preprocess: bool = True,
+        dynamic_min_num: int = 1,
+        dynamic_max_num: int = 12,
+        dynamic_use_thumbnail: bool = True,
+        **kwargs,
+    ):
+        super().__init__()
+        self.do_resize = (
+            do_resize  # Although dynamic_preprocess handles resizing internally
+        )
+        self.size = size
+        self.resample = resample
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        # Custom dynamic processing params
+        self.do_dynamic_preprocess = do_dynamic_preprocess
+        self.dynamic_min_num = dynamic_min_num
+        self.dynamic_max_num = dynamic_max_num
+        self.dynamic_use_thumbnail = dynamic_use_thumbnail
+
+    def preprocess(
+        self,
+        images: List[Image.Image],
+        do_dynamic_preprocess: Optional[bool] = None,
+        size: Optional[int] = None,
+        # ... other params matching __init__ ...
+        return_tensors: Optional[str] = None,
+        **kwargs,
+    ) -> List[mx.array]:
+
+        do_dynamic_preprocess = (
+            do_dynamic_preprocess
+            if do_dynamic_preprocess is not None
+            else self.do_dynamic_preprocess
+        )
+        size = size if size is not None else self.size
+        # ... handle other overrides ...
+
+        if not isinstance(images, list):
+            images = [images]
+
+        if not all(isinstance(image, Image.Image) for image in images):
+            raise ValueError("Input must be a list of PIL Images.")
+
+        processed_images_batch = []
+        for image in images:
+            # Apply dynamic preprocessing
+            if do_dynamic_preprocess:
+                processed_images = dynamic_preprocess(
+                    image,
+                    min_num=self.dynamic_min_num,
+                    max_num=self.dynamic_max_num,
+                    image_size=size,
+                    use_thumbnail=self.dynamic_use_thumbnail,
+                )
+            else:
+                # Fallback or alternative simpler preprocessing if needed
+                # e.g., simple resize + normalize
+                processed_images = [image.resize((size, size), resample=self.resample)]
+
+            # Create transform function
+            transform = build_transform(input_size=size)
+
+            # Apply transform to each image block and collect arrays
+            pixel_values_list = [transform(img) for img in processed_images]
+
+            # Stack the arrays along a new dimension (batch dimension)
+            pixel_values = mx.stack(pixel_values_list, axis=0)
+
+            processed_images_batch.append(pixel_values)
+
+        # At this point, processed_images_batch contains a list of mx arrays,
+        # each array corresponding to an input image with stacked blocks.
+
+        data = {"pixel_values": mx.array(processed_images_batch)}
+        return BatchFeature(data=data, tensor_type=None)
+
+
+class InternVLChatProcessor(ProcessorMixin):
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "InternVLImageProcessor"
+    tokenizer_class = (
+        "AutoTokenizer",
+        "Qwen2TokenizerFast",
+    )  # Specify possible classes
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        chat_template=chat_template,
+        **kwargs,
+    ):
+        if image_processor is None:
+            image_processor = InternVLImageProcessor(**kwargs)
+        if isinstance(tokenizer, str):
+            # Defaulting to the likely repo ID found earlier
+            tokenizer = AutoTokenizer.from_pretrained(
+                tokenizer, trust_remote_code=True, **kwargs
+            )
+
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+        self.num_image_token = int((448 // 14) ** 2 * (0.5**2))
+
+    def __call__(
+        self,
+        text: Union[str, List[str]] = None,
+        images: List[Image.Image] = None,
+        padding: Union[bool, str] = True,
+        truncation: bool = True,
+        max_length: Optional[int] = None,
+        return_tensors: Optional[str] = "pt",  # Default to PyTorch tensors
+        **kwargs,
+    ):
+        processed_inputs = {}
+        if images is not None:
+            image_features = self.image_processor.preprocess(
+                images, return_tensors=return_tensors, **kwargs
+            )
+            processed_inputs.update(image_features)  # Should contain 'pixel_values'
+
+        if text is not None:
+            queries = []
+
+            if isinstance(text, str):
+                text = [text]
+
+            for idx in range(len(images)):
+                question = text[idx]
+
+                if images is not None and "<image>" not in question:
+                    question = "<image>\n" + question
+
+                num_patches = image_features["pixel_values"][idx].shape[0]
+                image_tokens = (
+                    IMG_START_TOKEN
+                    + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches
+                    + IMG_END_TOKEN
+                )
+                question = question.replace("<image>", image_tokens, 1)
+                queries.append(question)
+
+            self.tokenizer.padding_side = "left"
+            text_inputs = self.tokenizer(
+                queries,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+            processed_inputs.update(text_inputs)  # 'input_ids', 'attention_mask'
+
+        return processed_inputs
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to the tokenizer's batch_decode method.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to the tokenizer's decode method.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    def save_pretrained(self, save_directory, **kwargs):
+        pass
+
+    @staticmethod
+    def from_pretrained(pretrained_model_name_or_path, **kwargs):
+        tokenizer = AutoTokenizer.from_pretrained(
+            pretrained_model_name_or_path, **kwargs
+        )
+        image_processor = InternVLImageProcessor(**kwargs)
+        return InternVLChatProcessor(
+            image_processor=image_processor, tokenizer=tokenizer
+        )
+
+    # Need save_pretrained and from_pretrained
+    # save_pretrained should save both tokenizer and image_processor configs/files
+    # from_pretrained should load both
+
+    # Example:
+    # def save_pretrained(self, save_directory, **kwargs):
+    #     self.tokenizer.save_pretrained(save_directory, **kwargs)
+    #     self.image_processor.save_pretrained(save_directory, **kwargs)
+
+    # def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+    #     tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path, **kwargs)
+    #     image_processor = InternVLImageProcessor.from_pretrained(pretrained_model_name_or_path, **kwargs)
+    #     return cls(image_processor=image_processor, tokenizer=tokenizer)
+
+
+# Registration
+MODEL_TYPE = "internvl_chat"  # Verify this from the model's config.json
+
+AutoImageProcessor.register(
+    MODEL_TYPE, slow_image_processor_class=InternVLImageProcessor
+)
+AutoProcessor.register(MODEL_TYPE, InternVLChatProcessor)
+
+logger.info(f"Registered custom processor classes for model type '{MODEL_TYPE}'.")

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/internvl_chat/vision.py

@@ -0,0 +1,293 @@
+import inspect
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from ..base import interpolate
+
+
+@dataclass
+class VisionConfig:
+    model_type: str
+    hidden_size: int = 1024
+    num_attention_heads: int = 16
+    patch_size: int = 14
+    num_hidden_layers: int = 24
+    intermediate_size: int = 4096
+    image_size: int = 448
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-6
+    drop_path_rate: float = 0.1
+    qkv_bias: bool = True
+    qk_normalization: bool = False
+    norm_type: str = "layer_norm"
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+def check_array_shape(arr):
+    shape = arr.shape
+
+    # Check if the shape has 4 dimensions
+    if len(shape) != 4:
+        return False
+
+    out_channels, kH, KW, _ = shape
+
+    # Check if out_channels is the largest, and kH and KW are the same
+    if (out_channels >= kH) and (out_channels >= KW) and (kH == KW):
+        return True
+    else:
+        return False
+
+
+class Attention(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+
+        if (config.hidden_size % config.num_attention_heads) != 0:
+            raise ValueError(
+                "The input feature dimensions should be divisible by the "
+                f"number of heads ({config.hidden_size} % {config.num_attention_heads}) != 0"
+            )
+
+        self.dims = dims = config.hidden_size
+
+        self.num_heads = config.num_attention_heads
+        head_dim = config.hidden_size // config.num_attention_heads
+        self.scale = head_dim**-0.5
+        self.qkv_bias = config.qkv_bias
+
+        self.qkv = nn.Linear(dims, 3 * dims, bias=config.qkv_bias)
+        self.proj = nn.Linear(dims, dims)
+
+        self.qk_normalization = config.qk_normalization
+
+        if self.qk_normalization:
+            self.q_norm = nn.RMSNorm(dims, eps=config.layer_norm_eps)
+            self.k_norm = nn.RMSNorm(dims, eps=config.layer_norm_eps)
+
+    def __call__(self, x, mask=None):
+        B, L, C = x.shape
+        qkv = self.qkv(x).reshape(B, L, 3, self.num_heads, C // self.num_heads)
+        qkv = qkv.transpose(2, 0, 3, 1, 4)
+        queries, keys, values = (
+            qkv[0],
+            qkv[1],
+            qkv[2],
+        )  # Each has shape (B, groups, N, C//groups)
+
+        if self.qk_normalization:
+            B_, H_, N_, D_ = queries.shape
+            queries = (
+                self.q_norm(queries.transpose(0, 2, 1, 3).flatten(-2, -1))
+                .reshape(B_, N_, H_, D_)
+                .transpose(0, 2, 1, 3)
+            )
+            keys = (
+                self.k_norm(keys.transpose(0, 2, 1, 3).flatten(-2, -1))
+                .reshape(B_, N_, H_, D_)
+                .transpose(0, 2, 1, 3)
+            )
+
+        output = mx.fast.scaled_dot_product_attention(
+            queries, keys, values, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.activation_fn = nn.GELU(approx="precise")
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.fc1(x)
+        x = self.activation_fn(x)
+        x = self.fc2(x)
+        return x
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, config: VisionConfig, drop_path_rate: float = 0.0):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.norm_type = getattr(config, "norm_type", "layer_norm")
+
+        self.attn = Attention(config)
+        self.mlp = MLP(config)
+
+        if self.norm_type == "layer_norm":
+            self.norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+            self.norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        elif self.norm_type == "rms_norm":
+            self.norm1 = nn.RMSNorm(self.embed_dim, eps=config.layer_norm_eps)
+            self.norm2 = nn.RMSNorm(self.embed_dim, eps=config.layer_norm_eps)
+        else:
+            raise ValueError(f"Unsupported normalization type: {self.norm_type}")
+
+        self.ls1 = mx.ones((self.embed_dim,))
+        self.ls2 = mx.ones((self.embed_dim,))
+
+        self.drop_path1 = (
+            nn.Dropout(drop_path_rate) if drop_path_rate > 0.0 else nn.Identity()
+        )
+        self.drop_path2 = (
+            nn.Dropout(drop_path_rate) if drop_path_rate > 0.0 else nn.Identity()
+        )
+
+    def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
+        dtype = x.dtype
+        x = x + self.drop_path1(self.attn(self.norm1(x).astype(dtype)) * self.ls1)
+
+        x = x + self.drop_path2(self.mlp(self.norm2(x).astype(dtype)) * self.ls2)
+
+        return x.astype(dtype)
+
+
+class Encoder(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        dpr = [
+            mx.array(x)
+            for x in np.linspace(0, config.drop_path_rate, config.num_hidden_layers)
+        ]
+        self.layers = [
+            EncoderLayer(config, dpr[i]) for i in range(config.num_hidden_layers)
+        ]
+
+    def __call__(
+        self,
+        x: mx.array,
+        output_hidden_states: Optional[bool] = None,
+        mask: Optional[mx.array] = None,
+    ) -> mx.array:
+        encoder_states = (x,) if output_hidden_states else None
+        h = x
+        for l in self.layers:
+            x = l(x, mask=mask)
+            if output_hidden_states:
+                encoder_states = encoder_states + (x,)
+
+            h = x
+
+        return (h, encoder_states)
+
+
+class VisionEmbeddings(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = mx.random.normal((1, 1, self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=3,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+
+        self.position_embedding = mx.random.normal(
+            (1, self.num_positions, self.embed_dim)
+        )
+
+    def _get_pos_embed(self, pos_embed, H, W):
+        target_dtype = pos_embed.dtype
+        pos_embed = pos_embed.reshape(
+            1,
+            self.image_size // self.patch_size,
+            self.image_size // self.patch_size,
+            -1,
+        ).transpose(0, 3, 1, 2)
+        pos_embed = interpolate(pos_embed, (H, W))
+        pos_embed = (
+            pos_embed.reshape(1, -1, H * W).transpose(0, 2, 1).astype(target_dtype)
+        )
+        return pos_embed
+
+    def __call__(self, x: mx.array) -> mx.array:
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(x).transpose(
+            0, 3, 1, 2
+        )  # shape = [*, channel, width, height]
+        batch_size, _, height, width = patch_embeds.shape
+        patch_embeds = mx.flatten(patch_embeds, start_axis=2).transpose(0, 2, 1)
+        class_embeds = mx.broadcast_to(
+            self.class_embedding, (batch_size, 1, self.embed_dim)
+        ).astype(target_dtype)
+        embeddings = mx.concatenate([class_embeds, patch_embeds], axis=1)
+        position_embedding = mx.concatenate(
+            [
+                self.position_embedding[:, :1, :],
+                self._get_pos_embed(self.position_embedding[:, 1:, :], height, width),
+            ],
+            axis=1,
+        )
+        embeddings = embeddings + position_embedding.astype(target_dtype)
+
+        return embeddings
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        if self.model_type not in ["siglip_vision_model", "intern_vit_6b"]:
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        self.embeddings = VisionEmbeddings(config)
+        self.encoder = Encoder(config)
+
+    def __call__(
+        self,
+        x: mx.array,
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+        x = self.embeddings(x)
+        last_hidden_state, encoder_outputs = self.encoder(
+            x=x, output_hidden_states=output_hidden_states, mask=None
+        )
+        pooler_output = last_hidden_state[:, 0, :]
+        return last_hidden_state, pooler_output, encoder_outputs[1:]
+
+    def sanitize(self, weights):
+        sanitized_weights = {}
+        for k, v in weights.items():
+            if "position_ids" in k:
+                # Remove unused position_ids
+                continue
+            elif "patch_embedding.weight" in k:
+                # PyTorch conv2d weight tensors have shape:
+                #   [out_channels, in_channels, kH, KW]
+                # MLX conv2d expects the weight be of shape:
+                #   [out_channels, kH, KW, in_channels]
+                if check_array_shape(v):
+                    sanitized_weights[k] = v
+                else:
+                    sanitized_weights[k] = v.transpose(0, 2, 3, 1)
+            else:
+                sanitized_weights[k] = v
+
+        return sanitized_weights