o hL@sRdZddlZddlZddlmZddlmZddlm Z m Z m Z m Z ddl ZddlZddlZddlmZddlmZmZmZdd lmZdd lmZmZdd lmZdd lmZmZdd l m!Z!m"Z"m#Z#m$Z$m%Z%ddl&m'Z'm(Z(ddl)m*Z*e$+e,Z-dZ.dZ/dgZ0eGddde!Z1eGddde!Z2ddZ3Gdddej4Z5Gdddej4Z6Gdddej4Z7Gdd d ej4Z8Gd!d"d"ej4Z9Gd#d$d$ej4Z:Gd%d&d&ej4Z;Gd'd(d(ej4Zd-Z?d.Z@e"d/e?Gd0d1d1e>ZAGd2d3d3ej4ZBe"d4e?Gd5d6d6e>ZCe"d7e?Gd8d9d9e>ZDdS):z- PyTorch VideoMAE (masked autoencoder) model.N)deepcopy) dataclass)OptionalSetTupleUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss)ACT2FN)BaseModelOutputImageClassifierOutput)PreTrainedModel) find_pruneable_heads_and_indicesprune_linear_layer) ModelOutputadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings)IMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD)VideoMAEConfigrzMCG-NJU/videomae-basec@sLeZdZUdZdZejed<dZe e ejed<dZ e e ejed<dS)VideoMAEDecoderOutputaO Class for VideoMAEDecoder's outputs, with potential hidden states and attentions. Args: logits (`torch.FloatTensor` of shape `(batch_size, patch_size ** 2 * num_channels)`): Pixel reconstruction logits. 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. Nlogits hidden_states attentions) __name__ __module__ __qualname____doc__rtorch FloatTensor__annotations__rrrrr'r'd/var/www/html/ai/venv/lib/python3.10/site-packages/transformers/models/videomae/modeling_videomae.pyr8s rc@s^eZdZUdZdZeejed<dZ ejed<dZ ee ejed<dZ ee ejed<dS)VideoMAEForPreTrainingOutputa Class for VideoMAEForPreTraining'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, patch_size ** 2 * num_channels)`): Pixel reconstruction logits. 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. Nlossrrr) r r!r"r#r*rr$r%r&rrrrr'r'r'r(r)Os r)csfddtfddt|D}t|dddddf|dddddf<t|dddddf|dddddf<t|dS) z Sinusoid position encoding tablecsfddtDS)Nc s(g|]}tdd|dqS)i')nppower).0hid_j)d_hidpositionr'r( ps(zOget_sinusoid_encoding_table..get_position_angle_vec..)ranger1)r0r4r(get_position_angle_vecosz;get_sinusoid_encoding_table..get_position_angle_veccsg|]}|qSr'r')r.pos_i)r5r'r(r2rz/get_sinusoid_encoding_table..Nrr+r)r,arrayr3sincosr$r% unsqueeze) n_positionr0sinusoid_tabler')r0r5r(get_sinusoid_encoding_tableks ..r>c(eZdZdZfddZddZZS)VideoMAEEmbeddingsz7 Construct the patch and position embeddings. cs8tt||_|jj|_t|j|j|_||_dSN) super__init__VideoMAEPatchEmbeddingspatch_embeddings num_patchesr> hidden_sizeposition_embeddingsconfigselfrI __class__r'r(rCs    zVideoMAEEmbeddings.__init__cCsZ||}||j||j}|dur+|j\}}}||}||d|}|S)N) rErHtype_astodeviceclonedetachshapereshape)rK pixel_valuesbool_masked_pos embeddings batch_size_ num_channelsr'r'r(forwards    zVideoMAEEmbeddings.forwardr r!r"r#rCr\ __classcell__r'r'rLr(r@ys  r@cr?)rDaw Video to Patch Embedding. 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) * (height // patch_size) * (width // patch_size). c st|j}|j}|j}|j}|j}|j}t|t j j r |n||f}t|t j j r-|n||f}||_||_t ||_|d|d|d|d||j}||_||_ tj|||j|d|df|j|d|dfd|_dS)Nrr) in_channels out_channels kernel_sizestride)rBrC image_size patch_sizer[rG num_frames tubelet_size isinstance collectionsabcIterableintrFrConv3d projection) rKrIrcrdr[rGrerfrFrLr'r(rCs,  ( z VideoMAEPatchEmbeddings.__init__c Cs|j\}}}}}||jkrtd||jdks||jdkr6td|d|d|jdd|jdd |dddd d }||ddd}|S) NzeMake sure that the channel dimension of the pixel values match with the one set in the configuration.rrzInput image size (*z) doesn't match model (z).r+r )rTr[ ValueErrorrcpermutermflatten transpose)rKrVrYrer[heightwidthrXr'r'r(r\s (zVideoMAEPatchEmbeddings.forwardr]r'r'rLr(rDs rDc sveZdZdeddffdd ZdejdejfddZ dd eejd e de e ejejfe ejffd d Z Z S)VideoMAESelfAttentionrIreturnNcst|j|jdkr t|ds td|jfd|jd|j|_t|j|j|_|j|j|_t j |j|jdd|_ t j |j|jdd|_ t j |j|jdd|_ |jrmt t|j|_t t|j|_nd|_d|_t |j|_dS)Nrembedding_sizezThe hidden size z4 is not a multiple of the number of attention heads .Fbias)rBrCrGnum_attention_headshasattrrprkattention_head_size all_head_sizerLinearquerykeyvalueqkv_bias Parameterr$zerosq_biasv_biasDropoutattention_probs_dropout_probdropoutrJrLr'r(rCs&  zVideoMAESelfAttention.__init__xcCs6|dd|j|jf}||}|ddddS)NrNrr+rr )sizer|r~viewrq)rKr new_x_shaper'r'r(transpose_for_scoress z*VideoMAESelfAttention.transpose_for_scoresF head_maskoutput_attentionscCs,|jdur tj|jddnd}tjj||jj|d}tjj||j j|jd}tjj||j j|jd}| |}| |} | |} t | | dd} | t|j} tjj| dd} || } |durh| |} t | | } | ddd d } | dd|jf}| |} |r| | f}|S| f}|S) NF) requires_grad)inputweightr{rNdimrr+rr )rr$ zeros_likerr functionallinearrrrrrmatmulrsmathsqrtr~softmaxrrq contiguousrrr)rKrrrk_biaskeysvaluesqueries key_layer value_layer query_layerattention_scoresattention_probs context_layernew_context_layer_shapeoutputsr'r'r(r\s*       zVideoMAESelfAttention.forwardNF)r r!r"rrCr$Tensorrrboolrrr\r^r'r'rLr(rvsrvcsFeZdZdZdeddffdd Zdejdejdejfd d ZZ S) VideoMAESelfOutputz The residual connection is defined in VideoMAELayer instead of here (as is the case with other models), due to the layernorm applied before each block. rIrwNcs.tt|j|j|_t|j|_dSrA) rBrCrrrGdenserhidden_dropout_probrrJrLr'r(rC zVideoMAESelfOutput.__init__r input_tensorcC||}||}|SrArrrKrrr'r'r(r\   zVideoMAESelfOutput.forward) r r!r"r#rrCr$rr\r^r'r'rLr(rs$rc s~eZdZdeddffdd ZdeeddfddZ  dd ej d e ej d e de e ej ej fe ej ffd dZZS)VideoMAEAttentionrIrwNcs*tt||_t||_t|_dSrA)rBrCrv attentionroutputset pruned_headsrJrLr'r(rC)s    zVideoMAEAttention.__init__headscCst|dkrdSt||jj|jj|j\}}t|jj||j_t|jj||j_t|jj ||j_ t|j j |dd|j _ |jjt||j_|jj|jj|j_ |j ||_dS)Nrrr)lenrrr|r~rrrrrrrrunion)rKrindexr'r'r( prune_heads/s zVideoMAEAttention.prune_headsFrrrcCs4||||}||d|}|f|dd}|S)Nrr)rr)rKrrr self_outputsattention_outputrr'r'r(r\AszVideoMAEAttention.forwardr)r r!r"rrCrrkrr$rrrrrr\r^r'r'rLr(r(srcs<eZdZdeddffdd ZdejdejfddZZS) VideoMAEIntermediaterIrwNcsDtt|j|j|_t|jt rt |j|_ dS|j|_ dSrA) rBrCrrrGintermediate_sizerrg hidden_actstrr intermediate_act_fnrJrLr'r(rCQs   zVideoMAEIntermediate.__init__rcCrrA)rr)rKrr'r'r(r\YrzVideoMAEIntermediate.forward r r!r"rrCr$rr\r^r'r'rLr(rPsrcsBeZdZdeddffdd Zdejdejdejfdd ZZS) VideoMAEOutputrIrwNcs.tt|j|j|_t|j|_ dSrA) rBrCrrrrGrrrrrJrLr'r(rCbrzVideoMAEOutput.__init__rrcCs ||}||}||}|SrArrr'r'r(r\gs  zVideoMAEOutput.forwardrr'r'rLr(ras$rc sleZdZdZdeddffdd Z  d dejd eejd e de e ejejfe ejffd d Z Z S) VideoMAELayerz?This corresponds to the Block class in the timm implementation.rIrwNcsbt|j|_d|_t||_t||_t||_ t j |j |j d|_t j |j |j d|_dS)Nreps)rBrCchunk_size_feed_forward seq_len_dimrrr intermediaterrr LayerNormrGlayer_norm_epslayernorm_beforelayernorm_afterrJrLr'r(rCts    zVideoMAELayer.__init__FrrrcCs`|j||||d}|d}|dd}||}||}||}|||}|f|}|S)N)rrr)rrrrr)rKrrrself_attention_outputsrr layer_outputr'r'r(r\~s     zVideoMAELayer.forwardr)r r!r"r#rrCr$rrrrrr\r^r'r'rLr(rqs rcsbeZdZdeddffdd Z    ddejd eejd ed ed ede e e ff d dZ Z S)VideoMAEEncoderrIrwNcs:t|_tfddtjD|_d|_dS)Ncg|]}tqSr'rr.rZrIr'r(r2r7z,VideoMAEEncoder.__init__..F) rBrCrIr ModuleListr3num_hidden_layerslayergradient_checkpointingrJrLrr(rCs   zVideoMAEEncoder.__init__FTrrroutput_hidden_states return_dictc Cs|rdnd}|r dnd}t|jD]8\}} |r||f}|dur$||nd} |jr6|jr6|| j|| |} n| || |} | d}|rI|| df}q|rQ||f}|s_tdd|||fDSt|||dS)Nr'rrcs|] }|dur|VqdSrAr'r.vr'r'r( z*VideoMAEEncoder.forward..last_hidden_staterr) enumeraterrtraining_gradient_checkpointing_func__call__tupler) rKrrrrrall_hidden_statesall_self_attentionsi layer_modulelayer_head_mask layer_outputsr'r'r(r\s6      zVideoMAEEncoder.forward)NFFT)r r!r"rrCr$rrrrrrr\r^r'r'rLr(rs&  rc@s(eZdZdZeZdZdZdZddZ dS)VideoMAEPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. videomaerVTcCstt|tjtjfr#|jjjd|jjd|j dur!|j j dSdSt|tj r8|j j |jj ddSdS)zInitialize the weightsg)meanstdNg?) rgrrrlrdatanormal_rIinitializer_ranger{zero_rfill_)rKmoduler'r'r( _init_weightss   z%VideoMAEPreTrainedModel._init_weightsN) r r!r"r#r config_classbase_model_prefixmain_input_namesupports_gradient_checkpointingrr'r'r'r(rs raJ 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 ([`VideoMAEConfig`]): 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 [`AutoImageProcessor`]. See [`VideoMAEImageProcessor.__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. zbThe bare VideoMAE Model transformer outputting raw hidden-states without any specific head on top.cseZdZfddZddZddZeeee e d     dd e j d e e jd e e jd e ed e ede edeee ffddZZS) VideoMAEModelcsTt|||_t||_t||_|jrd|_n t j |j |j d|_| dS)Nr)rBrCrIr@rXrencoderuse_mean_pooling layernormrrrGr post_initrJrLr'r(rCs    zVideoMAEModel.__init__cCs|jjSrA)rXrE)rKr'r'r(get_input_embeddingssz"VideoMAEModel.get_input_embeddingscCs*|D]\}}|jj|j|qdS)z Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel N)itemsrrrr)rKheads_to_prunerrr'r'r( _prune_heads szVideoMAEModel._prune_heads output_typerNrVrWrrrrrwc Cs|dur|n|jj}|dur|n|jj}|dur|n|jj}|||jj}|||}|j|||||d}|d} |jdurD|| } |sO| f|ddSt | |j |j dS)a bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, sequence_length)`, *optional*): Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). Each video in the batch must have the same number of masked patches. If `None`, then all patches are considered. Sequence length is `(num_frames // tubelet_size) * (image_size // patch_size) ** 2`. Returns: Examples: ```python >>> import av >>> import numpy as np >>> from transformers import AutoImageProcessor, VideoMAEModel >>> 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 16 frames >>> indices = sample_frame_indices(clip_len=16, frame_sample_rate=1, seg_len=container.streams.video[0].frames) >>> video = read_video_pyav(container, indices) >>> image_processor = AutoImageProcessor.from_pretrained("MCG-NJU/videomae-base") >>> model = VideoMAEModel.from_pretrained("MCG-NJU/videomae-base") >>> # 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, 1568, 768] ```Nrrrrrrr) rIrruse_return_dict get_head_maskrrXrrrrr) rKrVrWrrrrembedding_outputencoder_outputssequence_outputr'r'r(r\(s.\   zVideoMAEModel.forward)NNNNN)r r!r"rCr r rVIDEOMAE_INPUTS_DOCSTRINGrr_CONFIG_FOR_DOCr$r%r BoolTensorrrrrr\r^r'r'rLr(r s4   rcs,eZdZfddZ   dddZZS)VideoMAEDecodercst|j|j|jd}t||j_|j_ |j _ |j _ tfddt|jD|_t|j|_|dkrFt|j|nt|_d|_||_dS)Nr+crr'rrdecoder_configr'r(r2r7z,VideoMAEDecoder.__init__..rF)rBrCr[rfrdrdecoder_hidden_sizerGdecoder_num_hidden_layersrdecoder_num_attention_headsr|decoder_intermediate_sizerrrr3decoder_layersrnormrIdentityheadrrI)rKrIrFdecoder_num_labelsrLrr(rCs  zVideoMAEDecoder.__init__FTc Cs|rdnd}|r dnd}t|jD]/\}} |r||f}|jr,|jr,|| j|d|} n| |d|d} | d}|r@|| df}q|rH||f}|dkrW|dd| df}||}||} |sotdd| ||fDSt | ||dS)Nr')rrrrcsrrAr'rr'r'r(rrz*VideoMAEDecoder.forward..)rrr) rrrrrrr r"rr) rKrreturn_token_numrrrrrrrrrr'r'r(r\s4      zVideoMAEDecoder.forward)FFT)r r!r"rCr\r^r'r'rLr(rs  rzXThe VideoMAE Model transformer with the decoder on top for self-supervised pre-training.cs~eZdZfddZeeeeed    dde j de j de e j de ed e ed e ed eeeffd d ZZS)VideoMAEForPreTrainingcs~t|||_t||_tj|j|jdd|_ t t dd|j|_ t|jjj|j|_t||jjjd|_|dS)NFrzr)rF)rBrCrIrrrrrGrencoder_to_decoderrr$r mask_tokenr>rXrFrHrdecoderrrJrLr'r(rCs    zVideoMAEForPreTraining.__init__r NrVrWrrrrrwc# Csj|dur|n|jj}|j||||||d}|d}||}|j\} } } |dur,td|j| dd|} | |j  } | | | d| } | | | d| }tj|| |j|gdd}|||jd}|j}d}t |jjdkr|}n2|j }|j}ttj ||d ddddddf}ttj ||d ddddddf}|||}|j\} }} }}|jj|jj}}|jjr*|| |||| ||||||}|ddd d d d dd}|| |||||||||| }||jddd|j dddd!d}|| |||||||||| }nB|jjdkr5td|| |||| ||||||}|ddd d d d dd}|| |||||||||| }|j\} }} || | d| } Wdn 1swYt"}!|!|| }|s|f|dd}"|dur|f|"S|"St#|||j$|j%dS)a bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, sequence_length)`): Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). Each video in the batch must have the same number of masked patches. Sequence length is `(num_frames // tubelet_size) * (image_size // patch_size) ** 2`. Returns: Examples: ```python >>> from transformers import AutoImageProcessor, VideoMAEForPreTraining >>> import numpy as np >>> import torch >>> num_frames = 16 >>> video = list(np.random.randint(0, 256, (num_frames, 3, 224, 224))) >>> image_processor = AutoImageProcessor.from_pretrained("MCG-NJU/videomae-base") >>> model = VideoMAEForPreTraining.from_pretrained("MCG-NJU/videomae-base") >>> pixel_values = image_processor(video, return_tensors="pt").pixel_values >>> num_patches_per_frame = (model.config.image_size // model.config.patch_size) ** 2 >>> seq_length = (num_frames // model.config.tubelet_size) * num_patches_per_frame >>> bool_masked_pos = torch.randint(0, 2, (1, seq_length)).bool() >>> outputs = model(pixel_values, bool_masked_pos=bool_masked_pos) >>> loss = outputs.loss ```N)rWrrrrrz!One must provided a boolean mask rNrrr )rQdtyperor+rT)rkeepdim)runbiasedr-gư>zQCan't unnormalize non-RGB images. Consider setting config.norm_pix_loss to False.r*rrr)&rIrrr&rTrprHexpandrOrPrQrRrSrUr$catr'r(rno_gradr[r) as_tensorrrrfrd norm_pix_lossrrqrrvarrr r)rr)#rKrVrWrrrrrrrYseq_lenr[expanded_position_embeddingspos_emb_visible pos_emb_maskx_fulldecoder_outputsrr*framesrQr)rrtimertrurfrd frames_norm videos_patchrZlabelsloss_fctrr'r'r(r\s(    &&      J zVideoMAEForPreTraining.forward)NNNN)r r!r"rCrrrr)rr$r%rrrrrrr\r^r'r'rLr(r%s.   r%zVideoMAE Model transformer with a video classification head on top (a linear layer on top of the average pooled hidden states of all tokens) e.g. for ImageNet.cseZdZfddZeeeeed      dde e j de e j de e j de e d e e d e e d e eeffd d ZZS)VideoMAEForVideoClassificationcsft||j|_t||_|jrt|jnd|_ |jdkr(t |j|jnt |_ | dS)Nr)rBrC num_labelsrrrrrrGfc_normrr! classifierrrJrLr'r(rCs  $ z'VideoMAEForVideoClassification.__init__r NrVrr@rrrrwc Cs|dur|n|jj}|j|||||d}|d}|jdur&||d}n|dddf}||} d} |dur|jjdurc|jdkrId|j_n|jdkr_|jt j ksZ|jt j kr_d|j_nd|j_|jjdkrt } |jdkr{| | | } n+| | |} n%|jjdkrt} | | d|j|d} n|jjdkrt} | | |} |s| f|dd} | dur| f| S| St| | |j|jd S) a3 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 torch >>> import numpy as np >>> from transformers import AutoImageProcessor, VideoMAEForVideoClassification >>> 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 16 frames >>> indices = sample_frame_indices(clip_len=16, frame_sample_rate=1, seg_len=container.streams.video[0].frames) >>> video = read_video_pyav(container, indices) >>> image_processor = AutoImageProcessor.from_pretrained("MCG-NJU/videomae-base-finetuned-kinetics") >>> model = VideoMAEForVideoClassification.from_pretrained("MCG-NJU/videomae-base-finetuned-kinetics") >>> 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]) eating spaghetti ```Nrrr regressionsingle_label_classificationmulti_label_classificationrNr/)rIrrrDrrE problem_typerCr)r$longrkr squeezer rr rrr) rKrVrr@rrrrrrr*rArr'r'r(r\sR_     "       z&VideoMAEForVideoClassification.forward)NNNNNN)r r!r"rCrrrrrrr$rrrrr\r^r'r'rLr(rBs2   rB)Er#collections.abcrhrcopyr dataclassesrtypingrrrrnumpyr,r$torch.utils.checkpointrtorch.nnr r r activationsr modeling_outputsrrmodeling_utilsr pytorch_utilsrrutilsrrrrrutils.constantsrrconfiguration_videomaer get_loggerr loggerr_CHECKPOINT_FOR_DOC&VIDEOMAE_PRETRAINED_MODEL_ARCHIVE_LISTrr)r>Moduler@rDrvrrrrrrrVIDEOMAE_START_DOCSTRINGrrrr%rBr'r'r'r(st        5G(+3 D5