Add model files

configuration_sit.py

@@ -0,0 +1,132 @@
+# coding=utf-8
+# Copyright 2022 Facebook AI and The HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+""" ViT SiT model configuration"""
+
+from transformers import PretrainedConfig
+from transformers import logging
+
+
+logger = logging.get_logger(__name__)
+
+VIT_SiT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "erow/vit-SiT-base": "https://huggingface.co/erow/SiT/resolve/main/config.json",    
+}
+
+
+class ViTSiTConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ViTSiTModel`]. It is used to instantiate an ViT
+    SiT model according to the specified arguments, defining the model architecture. Instantiating a configuration with
+    the defaults will yield a similar configuration to that of the ViT
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        decoder_num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the decoder.
+        decoder_hidden_size (`int`, *optional*, defaults to 512):
+            Dimensionality of the decoder.
+        decoder_num_hidden_layers (`int`, *optional*, defaults to 8):
+            Number of hidden layers in the decoder.
+        decoder_intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the decoder.
+        mask_ratio (`float`, *optional*, defaults to 0.75):
+            The ratio of the number of masked tokens in the input sequence.
+        norm_pix_loss (`bool`, *optional*, defaults to `False`):
+            Whether or not to train with normalized pixels (see Table 3 in the paper). Using normalized pixels improved
+            representation quality in the experiments of the authors.
+
+    Example:
+
+    ```python
+    >>> from transformers import ViTSiTConfig, ViTSiTModel
+
+    >>> # Initializing a ViT SiT vit-SiT-base style configuration
+    >>> configuration = ViTSiTConfig()
+
+    >>> # Initializing a model (with random weights) from the vit-SiT-base style configuration
+    >>> model = ViTSiTModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "vit_sit"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        out_dim = 256,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        image_size=224,
+        patch_size=16,
+        num_channels=3,
+        qkv_bias=True,
+        mask_ratio=0.75,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.out_dim = out_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.qkv_bias = qkv_bias
+        self.mask_ratio = mask_ratio

modeling_sit.py

@@ -0,0 +1,921 @@
+# coding=utf-8
+# Copyright 2022 Facebook AI and The HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+""" PyTorch ViT SiT (masked autoencoder) model."""
+
+
+import collections.abc
+import math
+from copy import deepcopy
+from dataclasses import dataclass
+from typing import Optional, Set, Tuple, Union
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutput
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
+from transformers.utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from configuration_sit import ViTSiTConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "ViTSiTConfig"
+_CHECKPOINT_FOR_DOC = "erow/vit-sit-base"
+
+VIT_SiT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "erow/vit-sit-base",
+    # See all ViTSiT models at https://huggingface.co/models?filter=vit_sit
+]
+
+
+@dataclass
+class ViTSiTModelOutput(ModelOutput):
+    """
+    Class for ViTSiTModel's outputs, with potential hidden states and attentions.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
+            Tensor indicating which patches are masked (1) and which are not (0).
+        ids_restore (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Tensor containing the original index of the (shuffled) masked patches.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer
+            plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
+            the self-attention heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    noise: torch.LongTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+
+@dataclass
+class ViTSiTForPreTrainingOutput(ModelOutput):
+    """
+    Class for ViTSiTForPreTraining's outputs, with potential hidden states and attentions.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`):
+            Pixel reconstruction loss.
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, patch_size ** 2 * num_channels)`):
+            Pixel reconstruction logits.
+        mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
+            Tensor indicating which patches are masked (1) and which are not (0).
+        ids_restore (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Tensor containing the original index of the (shuffled) masked patches.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer
+            plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
+            the self-attention heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    noise: torch.LongTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+def get_2d_sincos_pos_embed(embed_dim, grid_size, add_cls_token=False):
+    """
+    Create 2D sin/cos positional embeddings.
+
+    Args:
+        embed_dim (`int`):
+            Embedding dimension.
+        grid_size (`int`):
+            The grid height and width.
+        add_cls_token (`bool`, *optional*, defaults to `False`):
+            Whether or not to add a classification (CLS) token.
+
+    Returns:
+        (`torch.FloatTensor` of shape (grid_size*grid_size, embed_dim) or (1+grid_size*grid_size, embed_dim): the
+        position embeddings (with or without classification token)
+    """
+    grid_h = np.arange(grid_size, dtype=np.float32)
+    grid_w = np.arange(grid_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_size, grid_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if add_cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be even")
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)
+    """
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be even")
+
+    omega = np.arange(embed_dim // 2, dtype=float)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+
+class ViTSiTEmbeddings(nn.Module):
+    """
+    Construct the CLS token, position and patch embeddings.
+
+    """
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.patch_embeddings = ViTSiTPatchEmbeddings(config)
+        self.num_patches = self.patch_embeddings.num_patches
+        # fixed sin-cos embedding
+        self.position_embeddings = nn.Parameter(
+            torch.zeros(1, self.num_patches + 1, config.hidden_size), requires_grad=False
+        )
+        self.config = config
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        # initialize (and freeze) position embeddings by sin-cos embedding
+        pos_embed = get_2d_sincos_pos_embed(
+            self.position_embeddings.shape[-1], int(self.patch_embeddings.num_patches**0.5), add_cls_token=True
+        )
+        self.position_embeddings.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
+
+        # initialize patch_embeddings like nn.Linear (instead of nn.Conv2d)
+        w = self.patch_embeddings.projection.weight.data
+        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
+
+        # timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.)
+        torch.nn.init.normal_(self.cls_token, std=self.config.initializer_range)
+
+
+    def forward(self, pixel_values, noise=None):
+        batch_size, num_channels, height, width = pixel_values.shape
+        embeddings = self.patch_embeddings(pixel_values)
+
+        # add position embeddings w/o cls token
+        embeddings = embeddings + self.position_embeddings[:, 1:, :]
+
+        # append cls token
+        cls_token = self.cls_token + self.position_embeddings[:, :1, :]
+        cls_tokens = cls_token.expand(embeddings.shape[0], -1, -1)
+        embeddings = torch.cat((cls_tokens, embeddings), dim=1)
+
+        return embeddings
+
+class ViTSiTPatchEmbeddings(nn.Module):
+    """
+    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
+    `hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a
+    Transformer.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        image_size, patch_size = config.image_size, config.patch_size
+        num_channels, hidden_size = config.num_channels, config.hidden_size
+        image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size)
+        patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.num_patches = num_patches
+
+        self.projection = nn.Conv2d(num_channels, hidden_size, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, pixel_values):
+        batch_size, num_channels, height, width = pixel_values.shape
+        if num_channels != self.num_channels:
+            raise ValueError(
+                "Make sure that the channel dimension of the pixel values match with the one set in the configuration."
+            )
+        if height != self.image_size[0] or width != self.image_size[1]:
+            raise ValueError(
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size[0]}*{self.image_size[1]})."
+            )
+        x = self.projection(pixel_values).flatten(2).transpose(1, 2)
+        return x
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfAttention ViT->ViTSiT
+class ViTSiTSelfAttention(nn.Module):
+    def __init__(self, config: ViTSiTConfig) -> None:
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size {config.hidden_size,} is not a multiple of the number of attention "
+                f"heads {config.num_attention_heads}."
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self, hidden_states, head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        mixed_query_layer = self.query(hidden_states)
+
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfOutput with ViT->ViTSiT
+class ViTSiTSelfOutput(nn.Module):
+    """
+    The residual connection is defined in ViTSiTLayer instead of here (as is the case with other models), due to the
+    layernorm applied before each block.
+    """
+
+    def __init__(self, config: ViTSiTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTAttention with ViT->ViTSiT
+class ViTSiTAttention(nn.Module):
+    def __init__(self, config: ViTSiTConfig) -> None:
+        super().__init__()
+        self.attention = ViTSiTSelfAttention(config)
+        self.output = ViTSiTSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads: Set[int]) -> None:
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.attention.num_attention_heads, self.attention.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.attention.query = prune_linear_layer(self.attention.query, index)
+        self.attention.key = prune_linear_layer(self.attention.key, index)
+        self.attention.value = prune_linear_layer(self.attention.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.attention.num_attention_heads = self.attention.num_attention_heads - len(heads)
+        self.attention.all_head_size = self.attention.attention_head_size * self.attention.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        self_outputs = self.attention(hidden_states, head_mask, output_attentions)
+
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTIntermediate ViT->ViTSiT
+class ViTSiTIntermediate(nn.Module):
+    def __init__(self, config: ViTSiTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTOutput ViT->ViTSiT
+class ViTSiTOutput(nn.Module):
+    def __init__(self, config: ViTSiTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        hidden_states = hidden_states + input_tensor
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTLayer with ViT->ViTSiT
+class ViTSiTLayer(nn.Module):
+    """This corresponds to the Block class in the timm implementation."""
+
+    def __init__(self, config: ViTSiTConfig) -> None:
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = ViTSiTAttention(config)
+        self.intermediate = ViTSiTIntermediate(config)
+        self.output = ViTSiTOutput(config)
+        self.layernorm_before = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.layernorm_after = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        self_attention_outputs = self.attention(
+            self.layernorm_before(hidden_states),  # in ViTSiT, layernorm is applied before self-attention
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        # first residual connection
+        hidden_states = attention_output + hidden_states
+
+        # in ViTSiT, layernorm is also applied after self-attention
+        layer_output = self.layernorm_after(hidden_states)
+        layer_output = self.intermediate(layer_output)
+
+        # second residual connection is done here
+        layer_output = self.output(layer_output, hidden_states)
+
+        outputs = (layer_output,) + outputs
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTEncoder with ViT->ViTSiT
+class ViTSiTEncoder(nn.Module):
+    def __init__(self, config: ViTSiTConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([ViTSiTLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ) -> Union[tuple, BaseModelOutput]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    layer_module.__call__,
+                    hidden_states,
+                    layer_head_mask,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = layer_module(hidden_states, layer_head_mask, output_attentions)
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class ViTSiTPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ViTSiTConfig
+    base_model_prefix = "vit"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+VIT_SiT_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it
+    as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`ViTSiTConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+VIT_SiT_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`ViTImageProcessor.__call__`]
+            for details.
+
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare ViTSiT Model transformer outputting raw hidden-states without any specific head on top.",
+    VIT_SiT_START_DOCSTRING,
+)
+class ViTSiTModel(ViTSiTPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = ViTSiTEmbeddings(config)
+        self.encoder = ViTSiTEncoder(config)
+
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.patch_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(VIT_SiT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=ViTSiTModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        noise: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, ViTSiTModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, ViTSiTModel
+        >>> from PIL import Image
+        >>> import requests
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> image_processor = AutoImageProcessor.from_pretrained("erow/vit-sit-base")
+        >>> model = ViTSiTModel.from_pretrained("erow/vit-sit-base")
+
+        >>> inputs = image_processor(images=image, return_tensors="pt")
+        >>> outputs = model(**inputs)
+        >>> last_hidden_states = outputs.last_hidden_state
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(pixel_values, noise=noise)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        sequence_output = self.layernorm(sequence_output)
+
+        if not return_dict:
+            return (sequence_output, ) + encoder_outputs[1:]
+
+        return ViTSiTModelOutput(
+            last_hidden_state=sequence_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class CLSHead(nn.Module):
+    def __init__(self, in_dim, bottleneck_dim, nlayers=3, hidden_dim=4096):
+        super().__init__()
+        nlayers = max(nlayers, 1)
+        if nlayers == 1:
+            self.mlp = nn.Linear(in_dim, bottleneck_dim)
+        else:
+            layers = [nn.Linear(in_dim, hidden_dim)]
+            layers.append(nn.BatchNorm1d(hidden_dim))
+            layers.append(nn.ReLU(inplace=True))
+            for _ in range(nlayers - 2):
+                layers.append(nn.Linear(hidden_dim, hidden_dim))
+                layers.append(nn.BatchNorm1d(hidden_dim))
+                layers.append(nn.GELU())
+            layers.append(nn.Linear(hidden_dim, bottleneck_dim))
+            layers.append(nn.BatchNorm1d(bottleneck_dim, affine=False))
+            
+            self.mlp = nn.Sequential(*layers)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        x = self.mlp(x)
+        return x
+
+class RECHead(nn.Module):
+    def __init__(self, in_dim, in_chans=3, patch_size=16):
+        super().__init__()
+
+        layers = [nn.Linear(in_dim, in_dim)]
+        layers.append(nn.GELU())
+        layers.append(nn.Linear(in_dim, in_dim))
+        layers.append(nn.GELU())
+        layers.append(nn.Linear(in_dim, in_dim))
+        layers.append(nn.GELU())
+
+        self.mlp = nn.Sequential(*layers)
+        self.apply(self._init_weights)
+        
+        self.convTrans = nn.ConvTranspose2d(in_dim, in_chans, kernel_size=(patch_size, patch_size), 
+                                                stride=(patch_size, patch_size))
+
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            torch.nn.init.normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        x = self.mlp(x)
+        
+        x_rec = x.transpose(1, 2)
+        out_sz = tuple( (  int(math.sqrt(x_rec.size()[2]))  ,   int(math.sqrt(x_rec.size()[2])) ) )
+        x_rec = self.convTrans(x_rec.unflatten(2, out_sz))
+        return x_rec
+
+
+@add_start_docstrings(
+    """The ViTSiT Model transformer with the decoder on top for self-supervised pre-training.
+
+    <Tip>
+
+    Note that we provide a script to pre-train this model on custom data in our [examples
+    directory](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-pretraining).
+
+    </Tip>
+
+    """,
+    VIT_SiT_START_DOCSTRING,
+)
+class ViTSiTForPreTraining(ViTSiTPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.vit = ViTSiTModel(config)
+        self.head_recons = RECHead(config.hidden_size, config.num_channels, config.patch_size)
+        self.head = CLSHead(config.hidden_size, config.out_dim)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.vit.embeddings.patch_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def patchify(self, pixel_values):
+        """
+        Args:
+            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+                Pixel values.
+
+        Returns:
+            `torch.FloatTensor` of shape `(batch_size, num_patches, patch_size**2 * num_channels)`:
+                Patchified pixel values.
+        """
+        patch_size, num_channels = self.config.patch_size, self.config.num_channels
+        # sanity checks
+        if (pixel_values.shape[2] != pixel_values.shape[3]) or (pixel_values.shape[2] % patch_size != 0):
+            raise ValueError("Make sure the pixel values have a squared size that is divisible by the patch size")
+        if pixel_values.shape[1] != num_channels:
+            raise ValueError(
+                "Make sure the number of channels of the pixel values is equal to the one set in the configuration"
+            )
+
+        # patchify
+        batch_size = pixel_values.shape[0]
+        num_patches_one_direction = pixel_values.shape[2] // patch_size
+        patchified_pixel_values = pixel_values.reshape(
+            batch_size, num_channels, num_patches_one_direction, patch_size, num_patches_one_direction, patch_size
+        )
+        patchified_pixel_values = torch.einsum("nchpwq->nhwpqc", patchified_pixel_values)
+        patchified_pixel_values = patchified_pixel_values.reshape(
+            batch_size, num_patches_one_direction * num_patches_one_direction, patch_size**2 * num_channels
+        )
+        return patchified_pixel_values
+
+    def unpatchify(self, patchified_pixel_values):
+        """
+        Args:
+            patchified_pixel_values (`torch.FloatTensor` of shape `(batch_size, num_patches, patch_size**2 * num_channels)`:
+                Patchified pixel values.
+
+        Returns:
+            `torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`:
+                Pixel values.
+        """
+        patch_size, num_channels = self.config.patch_size, self.config.num_channels
+        num_patches_one_direction = int(patchified_pixel_values.shape[1] ** 0.5)
+        # sanity check
+        if num_patches_one_direction**2 != patchified_pixel_values.shape[1]:
+            raise ValueError("Make sure that the number of patches can be squared")
+
+        # unpatchify
+        batch_size = patchified_pixel_values.shape[0]
+        patchified_pixel_values = patchified_pixel_values.reshape(
+            batch_size,
+            num_patches_one_direction,
+            num_patches_one_direction,
+            patch_size,
+            patch_size,
+            num_channels,
+        )
+        patchified_pixel_values = torch.einsum("nhwpqc->nchpwq", patchified_pixel_values)
+        pixel_values = patchified_pixel_values.reshape(
+            batch_size,
+            num_channels,
+            num_patches_one_direction * patch_size,
+            num_patches_one_direction * patch_size,
+        )
+        return pixel_values
+
+    def forward_loss(self, pixel_values, pred, mask):
+        """
+        Args:
+            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+                Pixel values.
+            pred (`torch.FloatTensor` of shape `(batch_size, num_patches, patch_size**2 * num_channels)`:
+                Predicted pixel values.
+            mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
+                Tensor indicating which patches are masked (1) and which are not (0).
+
+        Returns:
+            `torch.FloatTensor`: Pixel reconstruction loss.
+        """
+        target = pixel_values
+        pred = self.unpatchify(pred)
+        
+        loss = (pred - target) ** 2
+
+        loss = (loss * mask).sum() / mask.sum()  # mean loss on removed patches
+        return loss
+
+    @add_start_docstrings_to_model_forward(VIT_SiT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=ViTSiTForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        noise: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, ViTSiTForPreTrainingOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, ViTSiTForPreTraining
+        >>> from PIL import Image
+        >>> import requests
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> image_processor = AutoImageProcessor.from_pretrained("erow/vit-sit-base")
+        >>> model = ViTSiTForPreTraining.from_pretrained("erow/vit-sit-base")
+
+        >>> inputs = image_processor(images=image, return_tensors="pt")
+        >>> outputs = model(**inputs)
+        >>> loss = outputs.loss
+        >>> mask = outputs.mask
+        >>> ids_restore = outputs.ids_restore
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.vit(
+            pixel_values,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        latent = outputs.last_hidden_state
+
+        logits = self.decoder(latent) # shape (batch_size, num_patches, patch_size*patch_size*num_channels)
+
+        loss = self.forward_loss(pixel_values, logits, noise)
+
+        if not return_dict:
+            output = (logits, ) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return ViTSiTForPreTrainingOutput(
+            loss=loss,
+            logits=logits,
+            noise=noise,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+if __name__=="__main__":
+   
+    # Initializing a ViT MAE vit-mae-base style configuration
+    configuration = ViTSiTConfig()
+
+    # Initializing a model (with random weights) from the vit-mae-base style configuration
+    model = ViTSiTModel(configuration)
+
+    # Accessing the model configuration
+    configuration = model.config
+    
+    x = torch.randn(1, 3, 224, 224)
+    output = model(x)