mispeech/ced-small

1# coding=utf-82# Copyright 2023 Xiaomi Corporation and The HuggingFace Inc. team. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15""" CED model configuration"""16 17import csv18from transformers import PretrainedConfig19from transformers.utils import logging20from transformers.utils.hub import cached_file21 22logger = logging.get_logger(__name__)23 24CED_PRETRAINED_CONFIG_ARCHIVE_MAP = {25    "mispeech/ced-tiny": "https://huggingface.co/mispeech/ced-tiny/resolve/main/config.json",26}27 28 29class CedConfig(PretrainedConfig):30    model_type = "ced"31 32    r"""33    Configuration class for the CED model.34 35    Args:36        name (str, optional, *optional*):37            Name of the pre-defined configuration. Can be "ced-tiny", "ced-mini", "ced-small" or "ced-base".38        attn_drop_rate (float, *optional*, defaults to 0.0):39            Dropout probability for attention weights. Default to 0.0.40        depth (int, *optional*, defaults to 12): Number of transformer layers. Default to 12.41        drop_path_rate (float, *optional*, defaults to 0.0): Drop path is taken from timm. Default to 0.0.42        drop_rate (float, *optional*, defaults to 0.0):43            Dropout probability for input embeddings. Default to 0.0.44        embed_dim (int, *optional*, defaults to 768):45            Dimensionality of the audio patch embeddings. Default to 768.46        eval_avg (str, *optional*, defaults to `"mean"`):47            Type of pooling to use for evaluation. Can be "mean", "token", "dm" or "logit". Default to "mean".48        mlp_ratio (float, *optional*, defaults to 4.0):49            Ratio of hidden size in the feedforward layer to the embedding size. Default to 4.0.50        num_heads (int, *optional*, defaults to 12): Number of attention heads. Default to 12.51        outputdim (int, *optional*, defaults to 527): Dimensionality of the output. Default to 527.52        patch_size (int, *optional*, defaults to 16): Size of the patches. Default to 16.53        patch_stride (int, *optional*, defaults to 16): Stride of the patches. Default to 16.54        pooling (str, *optional*, defaults to `"mean"`):55            Type of pooling to use for the output. Can be "mean", "token", "dm" or "logit". Default to "mean".56        qkv_bias (bool, *optional*, defaults to `True`):57            Whether to include bias terms in the query, key and value projections. Default to True.58        target_length (int, *optional*, defaults to 1012): Frames of an audio chunk. Default to 1012.59    """60 61    def __init__(62        self,63        name=None,64        attn_drop_rate=0.0,65        depth=12,66        drop_path_rate=0.0,67        drop_rate=0.0,68        embed_dim=768,69        eval_avg="mean",70        mlp_ratio=4.0,71        num_heads=12,72        outputdim=527,73        patch_size=16,74        patch_stride=16,75        pooling="mean",76        qkv_bias=True,77        target_length=1012,78        **kwargs,79    ):80        r"""81        TODO: Add docstring82        """83 84        super().__init__(**kwargs)85 86        if name == "ced-tiny":87            embed_dim = 19288            num_heads = 389        elif name == "ced-mini":90            embed_dim = 25691            num_heads = 492        elif name == "ced-small":93            embed_dim = 38494            num_heads = 695        elif name == "ced-base":96            embed_dim = 76897            num_heads = 1298        else:99            logger.info("No model name specified for CedConfig, use default settings.")100 101        assert pooling in ("mean", "token", "dm", "logit")102        self.name = name103        self.attn_drop_rate = attn_drop_rate104        self.center = kwargs.get("center", True)105        self.depth = depth106        self.drop_path_rate = drop_path_rate107        self.drop_rate = drop_rate108        self.embed_dim = embed_dim109        self.eval_avg = eval_avg110        self.f_max = kwargs.get("f_max", 8000)111        self.f_min = kwargs.get("f_min", 0)112        self.hop_size = kwargs.get("hop_size", 160)113        self.mlp_ratio = mlp_ratio114        self.n_fft = kwargs.get("n_fft", 512)115        self.n_mels = kwargs.get("n_mels", 64)116        self.n_mels = kwargs.get("n_mels", 64)117        self.num_heads = num_heads118        self.outputdim = outputdim119        self.pad_last = kwargs.get("pad_last", True)120        self.patch_size = patch_size121        self.patch_stride = patch_stride122        self.pooling = pooling123        self.qkv_bias = qkv_bias124        self.target_length = target_length125        self.win_size = kwargs.get("win_size", 512)126        self.loss = "BCE"127 128        if self.outputdim == 527:129            with open(cached_file("topel/ConvNeXt-Tiny-AT", "class_labels_indices.csv"), "r") as f:130                 reader = csv.reader(f)                                                                                                                                                                 131                 next(reader)  # skip header                                                                                                                                                            132                 self.id2label = {}                                                                                                                                                                     133                 for row in reader:                                                                                                                                                                     134                     idx = int(row[0])                                                                                                                                                                  135                     label = row[2]                                                                                                                                                                     136                     if label not in self.id2label.values():                                                                                                                                            137                         self.id2label[idx] = label138            self.label2id = {v: k for k, v in self.id2label.items()}139        else:140            self.id2label = None141            self.label2id = None142