"""Defines interfaces for simple inference with pretrained models Authors: * Aku Rouhe 2021 * Peter Plantinga 2021 * Loren Lugosch 2020 * Mirco Ravanelli 2020 * Titouan Parcollet 2021 * Abdel Heba 2021 * Andreas Nautsch 2022, 2023 * Pooneh Mousavi 2023 * Sylvain de Langen 2023 * Adel Moumen 2023 * Pradnya Kandarkar 2023 """ import sys import warnings from types import SimpleNamespace import torch import torchaudio from hyperpyyaml import load_hyperpyyaml from torch.nn import DataParallel as DP from torch.nn import SyncBatchNorm from torch.nn.parallel import DistributedDataParallel as DDP from speechbrain.dataio.batch import PaddedBatch, PaddedData from speechbrain.dataio.preprocess import AudioNormalizer from speechbrain.utils.data_pipeline import DataPipeline from speechbrain.utils.data_utils import split_path from speechbrain.utils.distributed import run_on_main from speechbrain.utils.fetching import LocalStrategy, fetch from speechbrain.utils.logger import get_logger from speechbrain.utils.superpowers import import_from_path logger = get_logger(__name__) def foreign_class( source, hparams_file="hyperparams.yaml", pymodule_file="custom.py", classname="CustomInterface", overrides={}, overrides_must_match=True, savedir=None, use_auth_token=False, download_only=False, huggingface_cache_dir=None, local_strategy: LocalStrategy = LocalStrategy.SYMLINK, **kwargs, ): """Fetch and load an interface from an outside source The source can be a location on the filesystem or online/huggingface The pymodule file should contain a class with the given classname. An instance of that class is returned. The idea is to have a custom Pretrained subclass in the file. The pymodule file is also added to the python path before the Hyperparams YAML file is loaded, so it can contain any custom implementations that are needed. The hyperparams file should contain a "modules" key, which is a dictionary of torch modules used for computation. The hyperparams file should contain a "pretrainer" key, which is a speechbrain.utils.parameter_transfer.Pretrainer Arguments --------- source : str or Path or FetchSource The location to use for finding the model. See ``speechbrain.utils.fetching.fetch`` for details. hparams_file : str The name of the hyperparameters file to use for constructing the modules necessary for inference. Must contain two keys: "modules" and "pretrainer", as described. pymodule_file : str The name of the Python file that should be fetched. classname : str The name of the Class, of which an instance is created and returned overrides : dict Any changes to make to the hparams file when it is loaded. overrides_must_match : bool Whether an error will be thrown when an override does not match a corresponding key in the yaml_stream. savedir : str or Path Where to put the pretraining material. If not given, just use cache. use_auth_token : bool (default: False) If true Huggingface's auth_token will be used to load private models from the HuggingFace Hub, default is False because the majority of models are public. download_only : bool (default: False) If true, class and instance creation is skipped. huggingface_cache_dir : str Path to HuggingFace cache; if None -> "~/.cache/huggingface" (default: None) local_strategy : speechbrain.utils.fetching.LocalStrategy The fetching strategy to use, which controls the behavior of remote file fetching with regards to symlinking and copying. See :func:`speechbrain.utils.fetching.fetch` for further details. **kwargs : dict Arguments to forward to class constructor. Returns ------- object An instance of a class with the given classname from the given pymodule file. """ hparams_local_path = fetch( filename=hparams_file, source=source, savedir=savedir, overwrite=False, save_filename=None, use_auth_token=use_auth_token, revision=None, huggingface_cache_dir=huggingface_cache_dir, local_strategy=local_strategy, ) pymodule_local_path = fetch( filename=pymodule_file, source=source, savedir=savedir, overwrite=False, save_filename=None, use_auth_token=use_auth_token, revision=None, huggingface_cache_dir=huggingface_cache_dir, local_strategy=local_strategy, ) sys.path.append(str(pymodule_local_path.parent)) # Load the modules: with open(hparams_local_path, encoding="utf-8") as fin: hparams = load_hyperpyyaml(fin, overrides, overrides_must_match) hparams["savedir"] = savedir # Pretraining: pretrainer = hparams["pretrainer"] pretrainer.set_collect_in(savedir) # For distributed setups, have this here: run_on_main( pretrainer.collect_files, kwargs={ "default_source": source, "use_auth_token": use_auth_token, "local_strategy": local_strategy, }, ) # Load on the CPU. Later the params can be moved elsewhere by specifying if not download_only: # run_opts={"device": ...} pretrainer.load_collected() # Import class and create instance module = import_from_path(pymodule_local_path) cls = getattr(module, classname) return cls(modules=hparams["modules"], hparams=hparams, **kwargs) class Pretrained(torch.nn.Module): """Takes a trained model and makes predictions on new data. This is a base class which handles some common boilerplate. It intentionally has an interface similar to ``Brain`` - these base classes handle similar things. Subclasses of Pretrained should implement the actual logic of how the pretrained system runs, and add methods with descriptive names (e.g. transcribe_file() for ASR). Pretrained is a torch.nn.Module so that methods like .to() or .eval() can work. Subclasses should provide a suitable forward() implementation: by convention, it should be a method that takes a batch of audio signals and runs the full model (as applicable). Arguments --------- modules : dict of str:torch.nn.Module pairs The Torch modules that make up the learned system. These can be treated in special ways (put on the right device, frozen, etc.). These are available as attributes under ``self.mods``, like self.mods.model(x) hparams : dict Each key:value pair should consist of a string key and a hyperparameter that is used within the overridden methods. These will be accessible via an ``hparams`` attribute, using "dot" notation: e.g., self.hparams.model(x). run_opts : dict Options parsed from command line. See ``speechbrain.parse_arguments()``. List that are supported here: * device * data_parallel_count * data_parallel_backend * distributed_launch * distributed_backend * jit * jit_module_keys * compule * compile_module_keys * compile_mode * compile_using_fullgraph * compile_using_dynamic_shape_tracing freeze_params : bool To freeze (requires_grad=False) parameters or not. Normally in inference you want to freeze the params. Also calls .eval() on all modules. """ HPARAMS_NEEDED = [] MODULES_NEEDED = [] def __init__( self, modules=None, hparams=None, run_opts=None, freeze_params=True ): super().__init__() # Arguments passed via the run opts dictionary. Set a limited # number of these, since some don't apply to inference. run_opt_defaults = { "device": "cpu", "data_parallel_count": -1, "data_parallel_backend": False, "distributed_launch": False, "distributed_backend": "nccl", "jit": False, "jit_module_keys": None, "compile": False, "compile_module_keys": None, "compile_mode": "reduce-overhead", "compile_using_fullgraph": False, "compile_using_dynamic_shape_tracing": False, } for arg, default in run_opt_defaults.items(): if run_opts is not None and arg in run_opts: setattr(self, arg, run_opts[arg]) else: # If any arg from run_opt_defaults exist in hparams and # not in command line args "run_opts" if hparams is not None and arg in hparams: setattr(self, arg, hparams[arg]) else: setattr(self, arg, default) # Put modules on the right device, accessible with dot notation self.mods = torch.nn.ModuleDict(modules) for module in self.mods.values(): if module is not None: module.to(self.device) # Check MODULES_NEEDED and HPARAMS_NEEDED and # make hyperparams available with dot notation if self.HPARAMS_NEEDED and hparams is None: raise ValueError("Need to provide hparams dict.") if hparams is not None: # Also first check that all required params are found: for hp in self.HPARAMS_NEEDED: if hp not in hparams: raise ValueError(f"Need hparams['{hp}']") self.hparams = SimpleNamespace(**hparams) # Prepare modules for computation, e.g. jit self._prepare_modules(freeze_params) # Audio normalization self.audio_normalizer = hparams.get( "audio_normalizer", AudioNormalizer() ) def _prepare_modules(self, freeze_params): """Prepare modules for computation, e.g. jit. Arguments --------- freeze_params : bool Whether to freeze the parameters and call ``eval()``. """ # Make jit-able self._compile() self._wrap_distributed() # If we don't want to backprop, freeze the pretrained parameters if freeze_params: self.mods.eval() for p in self.mods.parameters(): p.requires_grad = False def load_audio(self, path, savedir=None): """Load an audio file with this model's input spec When using a speech model, it is important to use the same type of data, as was used to train the model. This means for example using the same sampling rate and number of channels. It is, however, possible to convert a file from a higher sampling rate to a lower one (downsampling). Similarly, it is simple to downmix a stereo file to mono. The path can be a local path, a web url, or a link to a huggingface repo. """ source, fl = split_path(path) path = fetch( fl, source=source, savedir=savedir, local_strategy=LocalStrategy.SYMLINK, ) signal, sr = torchaudio.load(str(path), channels_first=False) signal = signal.to(self.device) return self.audio_normalizer(signal, sr) def _compile(self): """Compile requested modules with either JIT or TorchInductor.""" compile_available = hasattr(torch, "compile") if not compile_available and self.compile_module_keys is not None: raise ValueError( "'compile_module_keys' specified, but this install of PyTorch " "seems to be too old to support it." ) # Modules to compile with torch.compile compile_module_keys = set() if self.compile: if self.compile_module_keys is None: compile_module_keys = set(self.mods) else: compile_module_keys = set(self.compile_module_keys) logger.warning( "--compile and --compile_module_keys are both specified. " "Only modules specified in --compile_module_keys will be compiled." ) # Modules to compile with jit jit_module_keys = set() if self.jit: if self.jit_module_keys is None: jit_module_keys = set(self.mods) else: jit_module_keys = set(self.jit_module_keys) logger.warning( "--jit and --jit_module_keys are both specified. " "Only modules specified in --jit_module_keys will be compiled." ) # find missing keys for name in compile_module_keys | jit_module_keys: if name not in self.mods: raise ValueError( f"module {name} is not defined in your hparams file." ) # try 'torch.compile', remove successful compiles from JIT list for name in compile_module_keys: try: module = torch.compile( self.mods[name], mode=self.compile_mode, fullgraph=self.compile_using_fullgraph, dynamic=self.compile_using_dynamic_shape_tracing, ) except Exception as e: logger.warning( f"'{name}' in 'compile_module_keys' failed to compile " f"and will be skipped (may fallback onto JIT, if " f"specified): {e}" ) continue self.mods[name] = module.to(self.device) jit_module_keys.discard(name) for name in jit_module_keys: module = torch.jit.script(self.mods[name]) self.mods[name] = module.to(self.device) def _compile_jit(self): warnings.warn("'_compile_jit' is deprecated; use '_compile' instead") self._compile() def _wrap_distributed(self): """Wrap modules with distributed wrapper when requested.""" if not self.distributed_launch and not self.data_parallel_backend: return elif self.distributed_launch: for name, module in self.mods.items(): if any(p.requires_grad for p in module.parameters()): # for ddp, all module must run on same GPU module = SyncBatchNorm.convert_sync_batchnorm(module) module = DDP(module, device_ids=[self.device]) self.mods[name] = module else: # data_parallel_backend for name, module in self.mods.items(): if any(p.requires_grad for p in module.parameters()): # if distributed_count = -1 then use all gpus # otherwise, specify the set of gpu to use if self.data_parallel_count == -1: module = DP(module) else: module = DP( module, [i for i in range(self.data_parallel_count)] ) self.mods[name] = module @classmethod def from_hparams( cls, source, hparams_file="hyperparams.yaml", pymodule_file="custom.py", overrides={}, savedir=None, use_auth_token=False, revision=None, download_only=False, huggingface_cache_dir=None, overrides_must_match=True, local_strategy: LocalStrategy = LocalStrategy.SYMLINK, **kwargs, ): """Fetch and load based from outside source based on HyperPyYAML file The source can be a location on the filesystem or online/huggingface You can use the pymodule_file to include any custom implementations that are needed: if that file exists, then its location is added to sys.path before Hyperparams YAML is loaded, so it can be referenced in the YAML. The hyperparams file should contain a "modules" key, which is a dictionary of torch modules used for computation. The hyperparams file should contain a "pretrainer" key, which is a speechbrain.utils.parameter_transfer.Pretrainer Arguments --------- source : str The location to use for finding the model. See ``speechbrain.utils.fetching.fetch`` for details. hparams_file : str The name of the hyperparameters file to use for constructing the modules necessary for inference. Must contain two keys: "modules" and "pretrainer", as described. pymodule_file : str A Python file can be fetched. This allows any custom implementations to be included. The file's location is added to sys.path before the hyperparams YAML file is loaded, so it can be referenced in YAML. This is optional, but has a default: "custom.py". If the default file is not found, this is simply ignored, but if you give a different filename, then this will raise in case the file is not found. overrides : dict Any changes to make to the hparams file when it is loaded. savedir : str or Path Where to put the pretraining material. If not given, just use cache. use_auth_token : bool (default: False) If true Huggingface's auth_token will be used to load private models from the HuggingFace Hub, default is False because the majority of models are public. revision : str The model revision corresponding to the HuggingFace Hub model revision. This is particularly useful if you wish to pin your code to a particular version of a model hosted at HuggingFace. download_only : bool (default: False) If true, class and instance creation is skipped. huggingface_cache_dir : str Path to HuggingFace cache; if None -> "~/.cache/huggingface" (default: None) overrides_must_match : bool Whether the overrides must match the parameters already in the file. local_strategy : LocalStrategy, optional Which strategy to use to deal with files locally. (default: `LocalStrategy.SYMLINK`) **kwargs : dict Arguments to forward to class constructor. Returns ------- Instance of cls """ hparams_local_path = fetch( filename=hparams_file, source=source, savedir=savedir, overwrite=False, save_filename=None, use_auth_token=use_auth_token, revision=revision, huggingface_cache_dir=huggingface_cache_dir, local_strategy=local_strategy, ) try: pymodule_local_path = fetch( filename=pymodule_file, source=source, savedir=savedir, overwrite=False, save_filename=None, use_auth_token=use_auth_token, revision=revision, huggingface_cache_dir=huggingface_cache_dir, local_strategy=local_strategy, ) sys.path.append(str(pymodule_local_path.parent)) except ValueError: if pymodule_file == "custom.py": # The optional custom Python module file did not exist # and had the default name pass else: # Custom Python module file not found, but some other # filename than the default was given. raise # Load the modules: with open(hparams_local_path, encoding="utf-8") as fin: hparams = load_hyperpyyaml( fin, overrides, overrides_must_match=overrides_must_match ) # add savedir to hparams hparams["savedir"] = savedir # Pretraining: pretrainer = hparams.get("pretrainer", None) if pretrainer is not None: pretrainer.set_collect_in(savedir) # For distributed setups, have this here: run_on_main( pretrainer.collect_files, kwargs={ "default_source": source, "use_auth_token": use_auth_token, "local_strategy": local_strategy, }, ) # Load on the CPU. Later the params can be moved elsewhere by specifying if not download_only: # run_opts={"device": ...} pretrainer.load_collected() # Now return the system return cls(hparams["modules"], hparams, **kwargs) else: return cls(hparams["modules"], hparams, **kwargs) class EncodeDecodePipelineMixin: """ A mixin for pretrained models that makes it possible to specify an encoding pipeline and a decoding pipeline """ def create_pipelines(self): """ Initializes the encode and decode pipeline """ self._run_init_steps(self.hparams.encode_pipeline) self._run_init_steps(self.hparams.decode_pipeline) self.encode_pipeline = DataPipeline( static_data_keys=self.INPUT_STATIC_KEYS, dynamic_items=self.hparams.encode_pipeline["steps"], output_keys=self.hparams.encode_pipeline["output_keys"], ) self.decode_pipeline = DataPipeline( static_data_keys=self.hparams.model_output_keys, dynamic_items=self.hparams.decode_pipeline["steps"], output_keys=self.OUTPUT_KEYS, ) def _run_init_steps(self, pipeline_definition): """Encode/decode pipelines may include initialization steps, such as filling text encoders with tokens. Calling this method will run them, if defined""" steps = pipeline_definition.get("init", []) for step in steps: step_func = step.get("func") if not step_func or not callable(step_func): raise ValueError("Invalid pipeline init definition") step_func() def _run_pipeline(self, pipeline, input, batch): if batch: output = pipeline(input) else: output = [pipeline(item) for item in input] return output def _get_encode_pipeline_input(self, input): return input if self.batch_inputs else self._itemize(input) def _get_decode_pipeline_input(self, model_output): model_output_keys = getattr(self.hparams, "model_output_keys", None) pipeline_input = model_output if len(model_output_keys) == 1: pipeline_input = (pipeline_input,) # The input to a pipeline is a dictionary. If model_output_keys # is provided, the output of the model is assumed to be a collection # (e.g. a list or a tuple). if model_output_keys: pipeline_input = dict(zip(model_output_keys, pipeline_input)) # By default, the pipeline will be applied to in batch mode # to the entire model input if not self.batch_outputs: pipeline_input = self._itemize(pipeline_input) return pipeline_input def _itemize(self, pipeline_input): first_item = next(iter(pipeline_input.values())) keys, values = pipeline_input.keys(), pipeline_input.values() batch_length = len(first_item) return [ dict(zip(keys, [value[idx] for value in values])) for idx in range(batch_length) ] def to_dict(self, data): """ Converts padded batches to dictionaries, leaves other data types as is Arguments --------- data: object a dictionary or a padded batch Returns ------- results: dict the dictionary """ if isinstance(data, PaddedBatch): data = { key: self._get_value(data, key) for key in self.hparams.encode_pipeline["output_keys"] } return data def _get_value(self, data, key): """ Retrieves the value associated with the specified key, dereferencing .data where applicable Arguments --------- data: PaddedBatch a padded batch key: str the key Returns ------- result: object the result """ value = getattr(data, key) if not self.input_use_padded_data and isinstance(value, PaddedData): value = value.data return value @property def batch_inputs(self): """ Determines whether the input pipeline operates on batches or individual examples (true means batched) Returns ------- batch_inputs: bool """ return self.hparams.encode_pipeline.get("batch", True) @property def input_use_padded_data(self): """ If turned on, raw PaddedData instances will be passed to the model. If turned off, only .data will be used Returns ------- result: bool whether padded data is used as is """ return self.hparams.encode_pipeline.get("use_padded_data", False) @property def batch_outputs(self): """ Determines whether the output pipeline operates on batches or individual examples (true means batched) Returns ------- batch_outputs: bool """ return self.hparams.decode_pipeline.get("batch", True) def _collate(self, data): if not self.batch_inputs: collate_fn = getattr(self.hparams, "collate_fn", PaddedBatch) data = collate_fn(data) return data def encode_input(self, input): """ Encodes the inputs using the pipeline Arguments --------- input: dict the raw inputs Returns ------- results: object """ pipeline_input = self._get_encode_pipeline_input(input) model_input = self._run_pipeline( pipeline=self.encode_pipeline, input=pipeline_input, batch=self.batch_inputs, ) model_input = self._collate(model_input) if hasattr(model_input, "to"): model_input = model_input.to(self.device) return self.to_dict(model_input) def decode_output(self, output): """ Decodes the raw model outputs Arguments --------- output: tuple raw model outputs Returns ------- result: dict or list the output of the pipeline """ pipeline_input = self._get_decode_pipeline_input(output) return self._run_pipeline( pipeline=self.decode_pipeline, input=pipeline_input, batch=self.batch_outputs, )