Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,774 Bytes
7c34c28 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 |
from typing import Dict, List, Optional
import torch
import torch.nn as nn
from transformers import ClapModel, ClapProcessor
from multi_token.model_utils import MultiTaskType
from multi_token.data_tools import load_audio
from multi_token.modalities.base_modality import Modality
from multi_token.modalities.projectors import (
build_mlp_vector_projector, build_mt_vector_projector, MultiTaskModel
)
import json
OUTPUT_EMB_SIZE = 512
class CLAPAudioModule(nn.Module):
def __init__(self, model_name_or_path: str):
super().__init__()
self.model_name_or_path = model_name_or_path
self.model = None
self.processor = None
self.load_model()
def load_model(self):
self.model = ClapModel.from_pretrained(self.model_name_or_path)
self.processor = ClapProcessor.from_pretrained(self.model_name_or_path)
self.model.requires_grad_(False)
@torch.no_grad()
def forward(self, audios) -> torch.Tensor:
embs = []
for audio_features in audios:
features = self.model.get_audio_features(
input_features=audio_features["input_features"].to(torch.float32),
is_longer=audio_features["is_longer"],
)
embs.append(features)
embs = torch.stack(embs)
return embs.view(-1, 1, OUTPUT_EMB_SIZE)
@property
def dtype(self):
return self.model.dtype
@property
def device(self):
return self.model.device
class CLAPAudioModality(Modality):
def __init__(
self,
model_name_or_path: str = "laion/clap-htsat-fused",
num_projector_layers: int = 2,
num_tokens_output: int = 10,
use_multi_task: int = MultiTaskType.NO_MULTI_TASK,
tasks_config: str = None
):
self.model_name_or_path = model_name_or_path
self.module = CLAPAudioModule(model_name_or_path=self.model_name_or_path)
self.num_projector_layers = num_projector_layers
self.num_tokens_output = num_tokens_output
self.dtype = torch.float32
self.use_multi_task = use_multi_task
self.tasks = None
if self.use_multi_task != MultiTaskType.NO_MULTI_TASK:
with open(tasks_config, 'r') as f:
self.tasks = json.load(f)
print("Tasks :", self.tasks)
def build_projector(self, lm_hidden_size: int) -> nn.Module:
if self.use_multi_task == MultiTaskType.PROJECTED_MULTI_TASK:
return MultiTaskModel(OUTPUT_EMB_SIZE, self.tasks)
elif self.use_multi_task == MultiTaskType.SIMPLE_MULTI_TASK:
return build_mt_vector_projector(
# return build_mlp_vector_projector(
input_hidden_size=OUTPUT_EMB_SIZE,
lm_hidden_size=lm_hidden_size,
# num_layers=self.num_projector_layers,
# num_tokens=self.num_tokens_output,
# )
tasks = self.tasks
)
# )["llm_projector"]
else:
return build_mlp_vector_projector(
input_hidden_size=OUTPUT_EMB_SIZE,
lm_hidden_size=lm_hidden_size,
num_layers=self.num_projector_layers,
num_tokens=self.num_tokens_output,
)
@property
def name(self) -> str:
return "audio_clap"
@property
def token(self) -> str:
return "<sound>"
@property
def data_key(self) -> str:
return "sounds"
@property
def token_width(self) -> int:
return self.num_tokens_output
def to(self, dtype: torch.dtype, device: torch.device) -> "CLAPAudioModality":
self.dtype = dtype
self.module.to(device=device)
return self
def preprocess_rows(self, rows: List[Dict]) -> List[Optional[Dict]]:
row_values = []
for row in rows:
audios = []
for audio_dict in row[self.data_key]:
audio_dict = load_audio(
audio_dict,
target_sampling_rate=self.module.processor.feature_extractor.sampling_rate,
)
audio_processed = self.module.processor(
audios=audio_dict["array"],
return_tensors="pt",
sampling_rate=audio_dict["sampling_rate"],
)
audios.append(audio_processed)
row_values.append(audios)
return row_values
@torch.no_grad()
def forward(self, encoded_values: List[torch.Tensor]) -> List[torch.Tensor]:
audio_features = []
for audio_batch in encoded_values:
audio_features.append(self.module.forward(audio_batch).to(dtype=self.dtype))
return audio_features
|