o hSu@sdZddlZddlmZmZmZmZddlZddlZddlm Z ddl m Z m Z ddl mZddlmZmZmZdd lmZdd lmZmZdd lmZmZmZmZd d lmZee Z!dZ"dZ#dgZ$Gddde j%Z&Gddde j%Z'Gddde j%Z(Gddde j%Z)Gddde j%Z*Gddde j%Z+Gddde j%Z,Gddde j%Z-Gd d!d!e j%Z.Gd"d#d#e j%Z/Gd$d%d%eZ0d&Z1d'Z2ed(e1Gd)d*d*e0Z3ed+e1Gd,d-d-e0Z4dS).z PyTorch ViViT model.N)OptionalSetTupleUnion)nn)CrossEntropyLossMSELoss)ACT2FN)BaseModelOutputBaseModelOutputWithPoolingImageClassifierOutput)PreTrainedModel) find_pruneable_heads_and_indicesprune_linear_layer)add_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings) VivitConfigzgoogle/vivit-b-16x2-kinetics400rc(eZdZdZfddZddZZS)VivitTubeletEmbeddingsa Construct Vivit Tubelet embeddings. This module turns a batch of videos of shape (batch_size, num_frames, num_channels, height, width) into a tensor of shape (batch_size, seq_len, hidden_size) to be consumed by a Transformer encoder. The seq_len (the number of patches) equals (number of frames // tubelet_size[0]) * (height // tubelet_size[1]) * (width // tubelet_size[2]). cs|t|j|_|j|_|j|_|j|jd|j|jd|j|jd|_|j|_t j |j |j|j|jd|_ dS)Nrr) kernel_sizestride) super__init__ num_frames image_size tubelet_size patch_size num_patches hidden_size embed_dimrConv3d num_channels projectionselfconfig __class__^/var/www/html/ai/venv/lib/python3.10/site-packages/transformers/models/vivit/modeling_vivit.pyr8s  zVivitTubeletEmbeddings.__init__c Cs|j\}}}}}||jks||jkr%td|d|d|jd|jd |ddddd }||}||ddd}|S) NzInput image size (*z) doesn't match model (z).rrrr )shaper ValueErrorpermuter'flatten transpose)r) pixel_values batch_sizerr&heightwidthxr-r-r.forwardHs  zVivitTubeletEmbeddings.forward__name__ __module__ __qualname____doc__rr; __classcell__r-r-r+r.r-s rcr)VivitEmbeddingsz Vivit Embeddings. Creates embeddings from a video using VivitTubeletEmbeddings, adds CLS token and positional embeddings. csdtttdd|j|_t||_ ttd|j j d|j|_ t |j |_||_dS)Nr)rrr Parametertorchzerosr# cls_tokenrpatch_embeddingsr"position_embeddingsDropouthidden_dropout_probdropoutr*r(r+r-r.r_s   zVivitEmbeddings.__init__cCsP|jd}||}|j|ddg}tj||fdd}||j}||}|SNrrdim)r1rGrFtilerDcatrHrK)r)r6r7 embeddings cls_tokensr-r-r.r;ks    zVivitEmbeddings.forwardr<r-r-r+r.rBXs  rBc sveZdZdeddffdd ZdejdejfddZ dd eejd e de e ejejfe ejffd d Z Z S)VivitSelfAttentionr*returnNcst|j|jdkr t|ds td|jfd|jd|j|_t|j|j|_|j|j|_t j |j|j|j d|_ t j |j|j|j d|_ t j |j|j|j d|_t |j|_dS)Nrembedding_sizezThe hidden size z4 is not a multiple of the number of attention heads .)bias)rrr#num_attention_headshasattrr2intattention_head_size all_head_sizerLinearqkv_biasquerykeyvaluerIattention_probs_dropout_probrKr(r+r-r.r}s  zVivitSelfAttention.__init__r:cCs6|dd|j|jf}||}|ddddS)Nrrrr )sizerXr[viewr3)r)r: new_x_shaper-r-r.transpose_for_scoress z'VivitSelfAttention.transpose_for_scoresF head_maskoutput_attentionsc Cs||}|||}|||}||}t||dd}|t|j }t j j |dd} | | } |durA| |} t| |} | dddd} | dd|jf} | | } |rj| | f} | S| f} | S)NrcrMrrrr )r_rgr`rarDmatmulr5mathsqrtr[r functionalsoftmaxrKr3 contiguousrdr\re) r) hidden_statesrhrimixed_query_layer key_layer value_layer query_layerattention_scoresattention_probs context_layernew_context_layer_shapeoutputsr-r-r.r;s$      zVivitSelfAttention.forwardNF)r=r>r?rrrDTensorrgrboolrrr;rAr-r-r+r.rS|srScsFeZdZdZdeddffdd Zdejdejdejfd d ZZ S) VivitSelfOutputz The residual connection is defined in VivitLayer instead of here (as is the case with other models), due to the layernorm applied before each block. r*rTNcs.tt|j|j|_t|j|_dSN) rrrr]r#denserIrJrKr(r+r-r.r zVivitSelfOutput.__init__rq input_tensorcCs||}||}|SrrrKr)rqrr-r-r.r;s  zVivitSelfOutput.forward) r=r>r?r@rrrDr|r;rAr-r-r+r.r~s$r~c s~eZdZdeddffdd ZdeeddfddZ  dd ej d e ej d e de e ej ej fe ej ffd dZZS)VivitAttentionr*rTNcs*tt||_t||_t|_dSr)rrrS attentionr~outputset pruned_headsr(r+r-r.rs    zVivitAttention.__init__headscCst|dkrdSt||jj|jj|j\}}t|jj||j_t|jj||j_t|jj ||j_ t|j j |dd|j _ |jjt||j_|jj|jj|j_ |j ||_dSrL)lenrrrXr[rrr_r`rarrr\union)r)rindexr-r-r. prune_headss zVivitAttention.prune_headsFrqrhricCs4||||}||d|}|f|dd}|S)Nrr)rr)r)rqrhri self_outputsattention_outputrzr-r-r.r;szVivitAttention.forwardr{)r=r>r?rrrrZrrDr|rr}rrr;rAr-r-r+r.rsrc$eZdZfddZddZZS)VivitIntermediatecsRtt|j|j|_t|j|_ t |j t r#t |j |_dS|j |_dSr)rrrr]r#intermediate_sizerrIrJrK isinstance hidden_actstrr intermediate_act_fnr(r+r-r.rs   zVivitIntermediate.__init__cCs"||}||}||}|Sr)rrrK)r)rqr-r-r.r;s   zVivitIntermediate.forwardr=r>r?rr;rAr-r-r+r.rs  rcr) VivitOutputcs.tt|j|j|_t|j|_ dSr) rrrr]rr#rrIrJrKr(r+r-r.rrzVivitOutput.__init__cCs ||}||}||}|Srrrr-r-r.r; s  zVivitOutput.forwardrr-r-r+r.r rcs*eZdZdZfddZdddZZS) VivitLayerzNThis corresponds to the EncoderBlock class in the scenic/vivit implementation.csbt|j|_d|_t||_t||_t||_ t j |j |j d|_t j |j |j d|_dS)Nreps)rrchunk_size_feed_forward seq_len_dimrrr intermediaterrr LayerNormr#layer_norm_epslayernorm_beforelayernorm_afterr(r+r-r.rs    zVivitLayer.__init__NFcCs`|j||||d}|d}|dd}||}||}||}|||}|f|}|S)N)rirr)rrrrr)r)rqrhriself_attention_outputsrrz layer_outputr-r-r.r;"s     zVivitLayer.forwardr{r<r-r-r+r.rs  rcs.eZdZfddZ    dddZZS) VivitEncodercs:t|_tfddtjD|_d|_dS)Ncsg|]}tqSr-)r).0_r*r-r. @sz)VivitEncoder.__init__..F) rrr*r ModuleListrangenum_hidden_layerslayergradient_checkpointingr(r+rr.r=s   zVivitEncoder.__init__NFTc Cs|rdnd}|r dnd}t|jD]8\}} |r||f}|dur$||nd} |jr6|jr6|| j|| |} n| || |} | d}|rI|| df}q|rQ||f}|s_tdd|||fDSt|||dS)Nr-rrcss|] }|dur|VqdSrr-)rvr-r-r. gsz'VivitEncoder.forward..)last_hidden_staterq attentions) enumeraterrtraining_gradient_checkpointing_func__call__tupler ) r)rqrhrioutput_hidden_states return_dictall_hidden_statesall_self_attentionsi layer_modulelayer_head_mask layer_outputsr-r-r.r;Cs6      zVivitEncoder.forward)NFFTrr-r-r+r.r<s  rcr) VivitPoolercs*tt|j|j|_t|_dSr)rrrr]r#rTanh activationr(r+r-r.rps zVivitPooler.__init__cCs(|dddf}||}||}|S)Nr)rr)r)rqfirst_token_tensor pooled_outputr-r-r.r;us  zVivitPooler.forwardrr-r-r+r.rorrc@s(eZdZdZeZdZdZdZddZ dS)VivitPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. vivitr6TcCst|tjtjfr#|jjjd|jjd|j dur!|j j dSdSt|tj rF|jjjd|jjd|j durD|jj|j  dSdSt|tj r[|j j |jjddSt|tjrm|jjd|jjddSdS)zInitialize the weightsg)meanstdNg?)rrr]r%weightdatanormal_r*initializer_rangerWzero_ Embedding padding_idxrfill_rC)r)moduler-r-r. _init_weightss       z"VivitPreTrainedModel._init_weightsN) r=r>r?r@r config_classbase_model_prefixmain_input_namesupports_gradient_checkpointingrr-r-r-r.r~s raG 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 ([`VivitConfig`]): 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. a\ Args: pixel_values (`torch.FloatTensor` of shape `(batch_size, num_frames, num_channels, height, width)`): Pixel values. Pixel values can be obtained using [`VivitImageProcessor`]. See [`VivitImageProcessor.preprocess`] 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. z_The bare ViViT Transformer model outputting raw hidden-states without any specific head on top.cseZdZdfdd ZddZddZeeee e d dd e e j d e e j d e ed e ede edeee j e ff ddZZS) VivitModelTcsXt|||_t||_t||_tj|j |j d|_ |r#t |nd|_ |dS)Nr)rrr*rBrQrencoderrrr#r layernormrpooler post_init)r)r*add_pooling_layerr+r-r.rs    zVivitModel.__init__cCs|jjSr)rQrG)r)r-r-r.get_input_embeddingsszVivitModel.get_input_embeddingscCs*|D]\}}|jj|j|qdS)z Prunes heads of the model. Args: heads_to_prune: dict of {layer_num: list of heads to prune in this layer} N)itemsrrrr)r)heads_to_prunerrr-r-r. _prune_headsszVivitModel._prune_heads output_typerNr6rhrirrrTc Cs|dur|n|jj}|dur|n|jj}|dur|n|jj}|dur&td|||jj}||}|j|||||d}|d}| |}|j durP| |nd} |s^|| f|ddSt || |j |j dS)a Returns: Examples: ```python >>> import av >>> import numpy as np >>> from transformers import VivitImageProcessor, VivitModel >>> from huggingface_hub import hf_hub_download >>> np.random.seed(0) >>> def read_video_pyav(container, indices): ... ''' ... Decode the video with PyAV decoder. ... Args: ... container (`av.container.input.InputContainer`): PyAV container. ... indices (`List[int]`): List of frame indices to decode. ... Returns: ... result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3). ... ''' ... frames = [] ... container.seek(0) ... start_index = indices[0] ... end_index = indices[-1] ... for i, frame in enumerate(container.decode(video=0)): ... if i > end_index: ... break ... if i >= start_index and i in indices: ... frames.append(frame) ... return np.stack([x.to_ndarray(format="rgb24") for x in frames]) >>> def sample_frame_indices(clip_len, frame_sample_rate, seg_len): ... ''' ... Sample a given number of frame indices from the video. ... Args: ... clip_len (`int`): Total number of frames to sample. ... frame_sample_rate (`int`): Sample every n-th frame. ... seg_len (`int`): Maximum allowed index of sample's last frame. ... Returns: ... indices (`List[int]`): List of sampled frame indices ... ''' ... converted_len = int(clip_len * frame_sample_rate) ... end_idx = np.random.randint(converted_len, seg_len) ... start_idx = end_idx - converted_len ... indices = np.linspace(start_idx, end_idx, num=clip_len) ... indices = np.clip(indices, start_idx, end_idx - 1).astype(np.int64) ... return indices >>> # video clip consists of 300 frames (10 seconds at 30 FPS) >>> file_path = hf_hub_download( ... repo_id="nielsr/video-demo", filename="eating_spaghetti.mp4", repo_type="dataset" ... ) >>> container = av.open(file_path) >>> # sample 32 frames >>> indices = sample_frame_indices(clip_len=32, frame_sample_rate=1, seg_len=container.streams.video[0].frames) >>> video = read_video_pyav(container=container, indices=indices) >>> image_processor = VivitImageProcessor.from_pretrained("google/vivit-b-16x2-kinetics400") >>> model = VivitModel.from_pretrained("google/vivit-b-16x2-kinetics400") >>> # prepare video for the model >>> inputs = image_processor(list(video), return_tensors="pt") >>> # forward pass >>> outputs = model(**inputs) >>> last_hidden_states = outputs.last_hidden_state >>> list(last_hidden_states.shape) [1, 3137, 768] ```Nz You have to specify pixel_valuesrhrirrrr)r pooler_outputrqr)r*riruse_return_dictr2 get_head_maskrrQrrrr rqr) r)r6rhrirrembedding_outputencoder_outputssequence_outputrr-r-r.r;s4V  zVivitModel.forward)T)NNNNN)r=r>r?rrrrVIVIT_INPUTS_DOCSTRINGrr _CONFIG_FOR_DOCrrD FloatTensorr}rrr;rAr-r-r+r.rs0  rzViViT Transformer model with a video classification head on top (a linear layer on top of the final hidden state of the [CLS] token) e.g. for Kinetics-400.cseZdZfddZeeeeed      dde e j de e j de e j de e d e e d e e d eee j effd d ZZS)VivitForVideoClassificationcsRt||j|_t|dd|_|jdkrt|j|jnt|_ | dS)NF)rr) rr num_labelsrrrr]r#Identity classifierrr(r+r-r.r]s $ z$VivitForVideoClassification.__init__rNr6rhlabelsrirrrTc Cs|dur|n|jj}|j|||||d}|d}||dddddf} d} |durP|jdkr@t} | | d|d} nt} | | d|j|d} |sf| f|dd} | durd| f| S| St| | |j |j dS)a( labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the image classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Returns: Examples: ```python >>> import av >>> import numpy as np >>> import torch >>> from transformers import VivitImageProcessor, VivitForVideoClassification >>> from huggingface_hub import hf_hub_download >>> np.random.seed(0) >>> def read_video_pyav(container, indices): ... ''' ... Decode the video with PyAV decoder. ... Args: ... container (`av.container.input.InputContainer`): PyAV container. ... indices (`List[int]`): List of frame indices to decode. ... Returns: ... result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3). ... ''' ... frames = [] ... container.seek(0) ... start_index = indices[0] ... end_index = indices[-1] ... for i, frame in enumerate(container.decode(video=0)): ... if i > end_index: ... break ... if i >= start_index and i in indices: ... frames.append(frame) ... return np.stack([x.to_ndarray(format="rgb24") for x in frames]) >>> def sample_frame_indices(clip_len, frame_sample_rate, seg_len): ... ''' ... Sample a given number of frame indices from the video. ... Args: ... clip_len (`int`): Total number of frames to sample. ... frame_sample_rate (`int`): Sample every n-th frame. ... seg_len (`int`): Maximum allowed index of sample's last frame. ... Returns: ... indices (`List[int]`): List of sampled frame indices ... ''' ... converted_len = int(clip_len * frame_sample_rate) ... end_idx = np.random.randint(converted_len, seg_len) ... start_idx = end_idx - converted_len ... indices = np.linspace(start_idx, end_idx, num=clip_len) ... indices = np.clip(indices, start_idx, end_idx - 1).astype(np.int64) ... return indices >>> # video clip consists of 300 frames (10 seconds at 30 FPS) >>> file_path = hf_hub_download( ... repo_id="nielsr/video-demo", filename="eating_spaghetti.mp4", repo_type="dataset" ... ) >>> container = av.open(file_path) >>> # sample 32 frames >>> indices = sample_frame_indices(clip_len=32, frame_sample_rate=4, seg_len=container.streams.video[0].frames) >>> video = read_video_pyav(container=container, indices=indices) >>> image_processor = VivitImageProcessor.from_pretrained("google/vivit-b-16x2-kinetics400") >>> model = VivitForVideoClassification.from_pretrained("google/vivit-b-16x2-kinetics400") >>> inputs = image_processor(list(video), return_tensors="pt") >>> with torch.no_grad(): ... outputs = model(**inputs) ... logits = outputs.logits >>> # model predicts one of the 400 Kinetics-400 classes >>> predicted_label = logits.argmax(-1).item() >>> print(model.config.id2label[predicted_label]) LABEL_116 ```Nrrrrcr)losslogitsrqr) r*rrrrrrerr rqr) r)r6rhrrirrrzrrrloss_fctrr-r-r.r;is4_ z#VivitForVideoClassification.forward)NNNNNN)r=r>r?rrrrr rrrDr LongTensorr}rrr;rAr-r-r+r.rWs2  r)5r@rltypingrrrrrDtorch.utils.checkpointrtorch.nnrr activationsr modeling_outputsr r r modeling_utilsr pytorch_utilsrrutilsrrrrconfiguration_vivitr get_loggerr=logger_CHECKPOINT_FOR_DOCr#VIVIT_PRETRAINED_MODEL_ARCHIVE_LISTModulerrBrSr~rrrrrrrVIVIT_START_DOCSTRINGrrrr-r-r-r.sT     +$=''3