UhdZddlZddlZddlmZddlmZmZmZm Z m Z ddl Z ddl m cmZddl m Z ddlmZddlmZmZmZdd lmZdd lmZmZmZmZdd lmZmZm Z m!Z!d d l"m#Z#m$Z$m%Z%e jLe'Z(dZ)dZ*dZ+dWdZ,de jZde jZfdZ.eGddeZ/eGddeZ0eGddeZ1Gdde jdZ3Gdde jdZ4Gdd e jdZ5Gd!d"e jdZ6Gd#d$e jdZ7Gd%d&e jdZ8Gd'd(e jdZ9Gd)d*e jdZ:Gd+d,e jdZ;Gd-d.e jdZ<Gd/d0e jdZ=Gd1d2e jdZ>Gd3d4e jdZ?Gd5d6e jdZ@Gd7d8e jdZAGd9d:e jdZBd;eAiZCGd<d=e jdZDGd>d?e jdZEGd@dAe jdZFGdBdCe jdZGGdDdEe jdZHGdFdGe jdZIeGdHdIeZJGdJdKeJZKedLMGdNdOeJZLeGdPdQeJZMeGdRdSeJZNeGdTdUeJZOgdVZPy)XzPyTorch CLAP model.N) dataclass)AnyListOptionalTupleUnion)nn)ACT2FN))BaseModelOutputWithPastAndCrossAttentionsBaseModelOutputWithPooling,BaseModelOutputWithPoolingAndCrossAttentions)PreTrainedModel)apply_chunking_to_forward find_pruneable_heads_and_indicesmeshgridprune_linear_layer) ModelOutputauto_docstringlogging torch_int)ClapAudioConfig ClapConfigClapTextConfigc|j\}}}|dddddddfjdd|d}|j|||z|}|S)ae Interpolate data in time domain. This is used to compensate the resolution reduction in downsampling of a CNN. Args: hidden_states (`torch.FloatTensor` of shape (batch_size, time_length, classes_num)): Input hidden states ratio (`int`): The ratio of the length of the output to the length of the input. Nr)shaperepeatreshape) hidden_statesratio batch_size time_length classes_num upsampleds x/var/www/catia.catastroantioquia-mas.com/valormas/lib/python3.12/site-packages/transformers/models/clap/modeling_clap.py interpolater'*sX.;-@-@*ZkaD!m,33Aq%CI!!*kE.A;OI c|j\}}}}|j|||z|||z||}|jddddddjjd|||}|S)aR Returns the resized hidden states. The output shape should be `(batch_size * num_windows, window_size, window_size, num_channels)` Args: hidden_states (`torch.FloatTensor` of shape `(batch_size, height, width, num_channels)`): Input hidden states window_size (`int`): Window size rrr rviewpermute contiguous)r window_sizer"heightwidth num_channelswindowss r&window_partitionr7;s}/<.A.A+J|!&&Fk);8Lk[gM##Aq!Q15@@BGGKYdfrsG Nr(c|jd}|jd||z||z|||}|jddddddjjd|||}|S)a Merges windows to produce higher resolution features. Args: windows (`torch.FloatTensor` of shape `(num_windows * batch_size, window_size, window_size, num_channels)`): Input windows window_size (`int`): Window size height (`int`): Height of the resized audio width (`int`): Width of the resized audio r-rrr r*r+r,r.)r6r2r3r4r5s r&window_reverser9Psn==$Lll2v4e{6JKYdfrsGooaAq!Q/::<AA"feUabG Nr(c|j|j}tj|dj ||z|z}|j |zS)a Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols are ignored. This is modified from fairseq's `utils.make_positions`. Args: x: torch.Tensor x: Returns: torch.Tensor rdim)neinttorchcumsumtype_aslong) input_ids padding_idxpast_key_values_lengthmaskincremental_indicess r&"create_position_ids_from_input_idsrHdsW << $ ( ( *D <<!4<