PhBdZddlZddlZddlmZddlmZmZmZm Z ddl Z ddl m cm Zddl m Z ddlmZddlmZdd lmZmZmZdd lmZmZdd lmZmZmZmZdd lm Z m!Z!m"Z"m#Z#m$Z$m%Z%d dl&m'Z'm(Z(m)Z)e$jTe+Z,dZ-dZ.dZ/dbdZ0de jbde jbfdZ2ee!dGdde Z3ee!dGdde Z4ee!Gdde Z5Gdd e jlZ7Gd!d"e jlZ8Gd#d$e jlZ9Gd%d&e jlZ:Gd'd(e jlZ;Gd)d*e jlZ<Gd+d,e jlZ=Gd-d.e jlZ>Gd/d0e jlZ?Gd1d2eZ@Gd3d4e jlZAGd5d6e jlZBGd7d8e jlZCGd9d:e jlZD dcd;e jlde jbd?ee jbd@eEdAeEdBee jbfdCZFGdDdEe jlZGGdFdGe jlZHGdHdIe jlZIGdJdKe jlZJGdLdMe jlZKGdNdOeZLGdPdQe jlZMGdRdSe jlZNe!GdTdUeZOGdVdWeOZPe!dXGdYdZeOZQe!Gd[d\eOZRe!Gd]d^eOZSe!Gd_d`eOZTgdaZUy)dzPyTorch CLAP model.N) dataclass)AnyCallableOptionalUnion)nn)ACT2FN)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithPooling,BaseModelOutputWithPoolingAndCrossAttentions)ALL_ATTENTION_FUNCTIONSPreTrainedModel)apply_chunking_to_forward find_pruneable_heads_and_indicesmeshgridprune_linear_layer) ModelOutputauto_docstringcan_return_tuplefilter_out_non_signature_kwargslogging 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 `/var/www/html/saasai/venv/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 r viewpermute contiguous)r# window_sizer%heightwidth num_channelswindowss r)window_partitionr:<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 r0rrr r-r.r/r1)r9r5r6r7r8s r)window_reverser<Qsn==$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_idsrKesW << $ ( ( *D <<!4<