languagebind 0.1.0__py3-none-any.whl

languagebind/image/modeling_image.py

@@ -0,0 +1,1031 @@
+import math
+from typing import Optional, Tuple, Union
+
+import torch
+from einops import rearrange
+from peft import LoraConfig, get_peft_model
+from torch import nn
+from torch.nn import functional as F
+from transformers import PreTrainedModel, add_start_docstrings
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from transformers.models.clip.modeling_clip import CLIPMLP, CLIPAttention, CLIPTextEmbeddings, CLIPVisionEmbeddings, \
+    CLIPVisionModelWithProjection, CLIPTextModelWithProjection, CLIPOutput
+from languagebind._compat import _expand_mask, clip_loss
+from transformers.utils import add_start_docstrings_to_model_forward, replace_return_docstrings
+
+from .configuration_image import LanguageBindImageConfig, CLIPVisionConfig, CLIPTextConfig
+
+
+
+class PatchDropout(nn.Module):
+    """
+    https://arxiv.org/abs/2212.00794
+    """
+
+    def __init__(self, prob, exclude_first_token=True):
+        super().__init__()
+        assert 0 <= prob < 1.
+        self.prob = prob
+        self.exclude_first_token = exclude_first_token  # exclude CLS token
+
+    def forward(self, x, B, T):
+        if not self.training or self.prob == 0.:
+            return x
+
+        if self.exclude_first_token:
+            cls_tokens, x = x[:, :1], x[:, 1:]
+        else:
+            cls_tokens = torch.jit.annotate(torch.Tensor, x[:, :1])
+
+        batch = x.size()[0]
+        num_tokens = x.size()[1]
+
+        batch_indices = torch.arange(batch)
+        batch_indices = batch_indices[..., None]
+
+        keep_prob = 1 - self.prob
+        num_patches_keep = max(1, int(num_tokens * keep_prob))
+
+        if T == 1:
+            rand = torch.randn(batch, num_tokens)
+            patch_indices_keep = rand.topk(num_patches_keep, dim=-1).indices
+        else:
+            rand = torch.randn(B, num_tokens)
+            patch_indices_keep = rand.topk(num_patches_keep, dim=-1).indices
+            patch_indices_keep = patch_indices_keep.unsqueeze(1).repeat(1, T, 1)
+            patch_indices_keep = rearrange(patch_indices_keep, 'b t n -> (b t) n')
+
+
+        x = x[batch_indices, patch_indices_keep]
+
+        if self.exclude_first_token:
+            x = torch.cat((cls_tokens, x), dim=1)
+
+        return x
+
+class CLIPEncoderLayer(nn.Module):
+    def __init__(self, config: LanguageBindImageConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = CLIPAttention(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = CLIPMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+        self.add_time_attn = config.add_time_attn
+        if self.add_time_attn:
+            self.t = config.num_frames
+            self.temporal_embedding = nn.Parameter(torch.zeros(1, config.num_frames, config.hidden_size))
+            nn.init.normal_(self.temporal_embedding, std=config.hidden_size ** -0.5)
+
+            self.embed_dim = config.hidden_size
+            self.temporal_attn = CLIPAttention(config)
+            self.temporal_layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+            self.temporal_mlp = CLIPMLP(config)
+            self.temporal_layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        causal_attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+
+
+        if self.add_time_attn:
+            bt, n, d = hidden_states.shape
+            t = self.t
+
+            # time embed
+            if t != 1:
+                n = hidden_states.shape[1]
+                hidden_states = rearrange(hidden_states, '(b t) n d -> (b n) t d', t=t)
+                hidden_states = hidden_states + self.temporal_embedding[:, :t, :]
+                hidden_states = rearrange(hidden_states, '(b n) t d -> (b t) n d', n=n)
+
+            # time attn
+            residual = hidden_states
+            hidden_states = rearrange(hidden_states, '(b t) n d -> (b n) t d', t=t)
+            # hidden_states = self.layer_norm1(hidden_states)  # share layernorm
+            hidden_states = self.temporal_layer_norm1(hidden_states)
+            hidden_states, attn_weights = self.temporal_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                causal_attention_mask=causal_attention_mask,
+                output_attentions=output_attentions,
+            )
+            hidden_states = residual + rearrange(hidden_states, '(b n) t d -> (b t) n d', n=n)
+
+            residual = hidden_states
+            hidden_states = rearrange(hidden_states, '(b t) n d -> (b n) t d', t=t)
+            # hidden_states = self.layer_norm2(hidden_states)  # share layernorm
+            hidden_states = self.temporal_layer_norm2(hidden_states)
+            hidden_states = self.temporal_mlp(hidden_states)
+            hidden_states = residual + rearrange(hidden_states, '(b n) t d -> (b t) n d', n=n)
+
+        # spatial attn
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+
+
+
+
+
+
+
+class CLIPPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = LanguageBindImageConfig
+    base_model_prefix = "clip"
+    supports_gradient_checkpointing = True
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_factor
+        if isinstance(module, CLIPTextEmbeddings):
+            module.token_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+            module.position_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+        elif isinstance(module, CLIPVisionEmbeddings):
+            factor = self.config.initializer_factor
+            nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
+            nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
+            nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
+        elif isinstance(module, CLIPAttention):
+            factor = self.config.initializer_factor
+            in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            out_proj_std = (module.embed_dim**-0.5) * factor
+            nn.init.normal_(module.q_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.k_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.v_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.out_proj.weight, std=out_proj_std)
+        elif isinstance(module, CLIPMLP):
+            factor = self.config.initializer_factor
+            in_proj_std = (
+                (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            )
+            fc_std = (2 * module.config.hidden_size) ** -0.5 * factor
+            nn.init.normal_(module.fc1.weight, std=fc_std)
+            nn.init.normal_(module.fc2.weight, std=in_proj_std)
+        elif isinstance(module, LanguageBindImage):
+            nn.init.normal_(
+                module.text_projection.weight,
+                std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+            nn.init.normal_(
+                module.visual_projection.weight,
+                std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+        elif isinstance(module, CLIPVisionModelWithProjection):
+            nn.init.normal_(
+                module.visual_projection.weight,
+                std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
+            )
+        elif isinstance(module, CLIPTextModelWithProjection):
+            nn.init.normal_(
+                module.text_projection.weight,
+                std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
+            )
+
+        if isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, CLIPEncoder):
+            module.gradient_checkpointing = value
+
+
+CLIP_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also 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 ([`CLIPConfig`]): 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.
+"""
+
+CLIP_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        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.
+"""
+
+CLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        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.
+"""
+
+CLIP_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        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.
+"""
+
+
+class CLIPEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`CLIPEncoderLayer`].
+
+    Args:
+        config: CLIPConfig
+    """
+
+    def __init__(self, config: LanguageBindImageConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([CLIPEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Causal mask for the text model. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            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.
+        """
+        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
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(encoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    causal_attention_mask,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    causal_attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+class CLIPTextTransformer(nn.Module):
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+        self.embeddings = CLIPTextEmbeddings(config)
+        self.encoder = CLIPEncoder(config)
+        self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        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 input_ids is None:
+            raise ValueError("You have to specify input_ids")
+
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+
+        hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
+
+        # CLIP's text model uses causal mask, prepare it here.
+        # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
+        causal_attention_mask = _make_causal_mask(input_shape, hidden_states.dtype, device=hidden_states.device)
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.final_layer_norm(last_hidden_state)
+
+        # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
+        pooled_output = last_hidden_state[
+            torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+            input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
+        ]
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """The text model from CLIP without any head or projection on top.""",
+    CLIP_START_DOCSTRING,
+)
+class CLIPTextModel(CLIPPreTrainedModel):
+    config_class = CLIPTextConfig
+
+    _no_split_modules = ["CLIPEncoderLayer"]
+
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__(config)
+        self.text_model = CLIPTextTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, value):
+        self.text_model.embeddings.token_embedding = value
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPTextModel
+
+        >>> model = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        return self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+class CLIPVisionTransformer(nn.Module):
+    def __init__(self, config: CLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = CLIPVisionEmbeddings(config)
+        self.patch_dropout = PatchDropout(config.force_patch_dropout)
+        self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self.encoder = CLIPEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        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")
+        ######################################
+        if len(pixel_values.shape) == 7:
+            b_new, pair_new, T, bs_new, channel_new, h_new, w_new = pixel_values.shape
+            # print(pixel_values.shape)
+            B = b_new * pair_new * bs_new
+            pixel_values = pixel_values.reshape(B*T, channel_new, h_new, w_new)
+
+        elif len(pixel_values.shape) == 5:
+            B, _, T, _, _ = pixel_values.shape
+            # print(pixel_values.shape)
+            pixel_values = rearrange(pixel_values, 'b c t h w -> (b t) c h w')
+        else:
+            # print(pixel_values.shape)
+            B, _, _, _ = pixel_values.shape
+            T = 1
+        ###########################
+        hidden_states = self.embeddings(pixel_values)
+
+        hidden_states = self.patch_dropout(hidden_states, B, T)  ##############################################
+
+        hidden_states = self.pre_layrnorm(hidden_states)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        pooled_output = last_hidden_state[:, 0, :]
+        pooled_output = self.post_layernorm(pooled_output)
+
+        pooled_output = pooled_output.reshape(B, T, -1).mean(1) ################################
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """The vision model from CLIP without any head or projection on top.""",
+    CLIP_START_DOCSTRING,
+)
+class CLIPVisionModel(CLIPPreTrainedModel):
+    config_class = CLIPVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: CLIPVisionConfig):
+        super().__init__(config)
+        self.vision_model = CLIPVisionTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPVisionModel
+
+        >>> model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled CLS states
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        return self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+@add_start_docstrings(CLIP_START_DOCSTRING)
+class LanguageBindImage(CLIPPreTrainedModel):
+    config_class = LanguageBindImageConfig
+
+    def __init__(self, config: LanguageBindImageConfig):
+        super().__init__(config)
+
+        if not isinstance(config.text_config, CLIPTextConfig):
+            raise ValueError(
+                "config.text_config is expected to be of type CLIPTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.vision_config, CLIPVisionConfig):
+            raise ValueError(
+                "config.vision_config is expected to be of type CLIPVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+
+        text_config = config.text_config
+        vision_config = config.vision_config
+        self.add_time_attn = vision_config.add_time_attn
+        self.lora_r = vision_config.lora_r
+        self.lora_alpha = vision_config.lora_alpha
+        self.lora_dropout = vision_config.lora_dropout
+
+        self.projection_dim = config.projection_dim
+        self.text_embed_dim = text_config.hidden_size
+        self.vision_embed_dim = vision_config.hidden_size
+
+        self.text_model = CLIPTextTransformer(text_config)
+        self.vision_model = CLIPVisionTransformer(vision_config)
+
+        self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False)
+        self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False)
+        self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value))
+
+        # Initialize weights and apply final processing
+        self.post_init()
+        self.convert_to_lora()
+        self.resize_pos(self.vision_model.embeddings, vision_config)
+
+    def convert_to_lora(self):
+        if self.lora_r == 0:
+            return
+        if self.add_time_attn:
+            target_modules = ["temporal_attn.k_proj", "temporal_attn.v_proj",
+                              "temporal_attn.q_proj", "temporal_attn.out_proj",
+                              "temporal_mlp.fc1", "temporal_mlp.fc2"]
+        else:
+            target_modules = ["k_proj", "v_proj", "q_proj", "out_proj"]
+        config = LoraConfig(
+            r=self.lora_r,  # 16
+            lora_alpha=self.lora_alpha,  # 16
+            target_modules=target_modules,  # self_attn.out_proj
+            lora_dropout=self.lora_dropout,  # 0.1
+            bias="none",
+            modules_to_save=[],
+        )
+        self.vision_model.encoder.is_gradient_checkpointing = False
+        self.vision_model.encoder = get_peft_model(self.vision_model.encoder, config)
+
+    def resize_pos(self, m, vision_config):
+        # convert embedding
+        if vision_config.num_mel_bins!=0 and vision_config.target_length!=0:
+            m.image_size = [vision_config.num_mel_bins, vision_config.target_length]
+        m.config.image_size = [m.image_size, m.image_size] if isinstance(m.image_size, int) else m.image_size
+        # pos resize
+        old_pos_embed_state_dict = m.position_embedding.state_dict()
+        old_pos_embed = old_pos_embed_state_dict['weight']
+        dtype = old_pos_embed.dtype
+        grid_size = [m.config.image_size[0] // m.patch_size, m.config.image_size[1] // m.patch_size]
+        extra_tokens = 1  # FIXME detect different token configs (ie no class token, or more)
+        new_seq_len = grid_size[0] * grid_size[1] + extra_tokens
+        if new_seq_len == old_pos_embed.shape[0]:
+            # m.to(args.device)
+            return
+
+        m.num_patches = grid_size[0] * grid_size[1]
+        m.num_positions = m.num_patches + 1
+        m.register_buffer("position_ids", torch.arange(m.num_positions).expand((1, -1)))
+        new_position_embedding = nn.Embedding(m.num_positions, m.embed_dim)
+
+        if extra_tokens:
+            pos_emb_tok, pos_emb_img = old_pos_embed[:extra_tokens], old_pos_embed[extra_tokens:]
+        else:
+            pos_emb_tok, pos_emb_img = None, old_pos_embed
+        old_grid_size = [int(math.sqrt(len(pos_emb_img)))] * 2
+
+        # if is_master(args):
+        #     logging.info('Resizing position embedding grid-size from %s to %s', old_grid_size, grid_size)
+        pos_emb_img = pos_emb_img.reshape(1, old_grid_size[0], old_grid_size[1], -1).permute(0, 3, 1, 2)
+        pos_emb_img = F.interpolate(
+            pos_emb_img,
+            size=grid_size,
+            mode='bicubic',
+            antialias=True,
+            align_corners=False,
+        )
+        pos_emb_img = pos_emb_img.permute(0, 2, 3, 1).reshape(1, grid_size[0] * grid_size[1], -1)[0]
+        if pos_emb_tok is not None:
+            new_pos_embed = torch.cat([pos_emb_tok, pos_emb_img], dim=0)
+        else:
+            new_pos_embed = pos_emb_img
+        old_pos_embed_state_dict['weight'] = new_pos_embed.to(dtype)
+        m.position_embedding = new_position_embedding
+        m.position_embedding.load_state_dict(old_pos_embed_state_dict)
+
+        # m.to(args.device)
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    def get_text_features(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
+            applying the projection layer to the pooled output of [`CLIPTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        # Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
+        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
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = text_outputs[1]
+        text_features = self.text_projection(pooled_output)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    def get_image_features(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
+            applying the projection layer to the pooled output of [`CLIPVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+        # Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
+        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
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = vision_outputs[1]  # pooled_output
+        image_features = self.visual_projection(pooled_output)
+
+        return image_features
+
+    @add_start_docstrings_to_model_forward(CLIP_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CLIPOutput, config_class=LanguageBindImageConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CLIPOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(
+        ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
+        ... )
+
+        >>> outputs = model(**inputs)
+        >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+        >>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
+        ```"""
+        # Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
+        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
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[1]
+        image_embeds = self.visual_projection(image_embeds)
+
+        text_embeds = text_outputs[1]
+        text_embeds = self.text_projection(text_embeds)
+
+        # normalized features
+        image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
+        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
+        logits_per_image = logits_per_text.t()
+
+        loss = None
+        if return_loss:
+            loss = clip_loss(logits_per_text)
+
+        if not return_dict:
+            output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return CLIPOutput(
+            loss=loss,
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+        )