From 363c3f1f826f6d12e961252f8885a8867d752365 Mon Sep 17 00:00:00 2001 From: Yuekai Zhang Date: Thu, 7 Sep 2023 15:20:00 +0800 Subject: [PATCH] update finetuning codes --- egs/aishell/ASR/run.sh | 4 + egs/aishell/ASR/seamlessm4t/asr_datamodule.py | 1 + .../ASR/seamlessm4t/label_smoothing.py | 1 + egs/aishell/ASR/seamlessm4t/model.py | 117 ++ egs/aishell/ASR/seamlessm4t/optim.py | 1173 +++++++++++++++ egs/aishell/ASR/seamlessm4t/train2.py | 1254 +++++++++++++++++ icefall/dist.py | 2 +- 7 files changed, 2551 insertions(+), 1 deletion(-) create mode 100644 egs/aishell/ASR/run.sh create mode 120000 egs/aishell/ASR/seamlessm4t/asr_datamodule.py create mode 120000 egs/aishell/ASR/seamlessm4t/label_smoothing.py create mode 100644 egs/aishell/ASR/seamlessm4t/model.py create mode 100644 egs/aishell/ASR/seamlessm4t/optim.py create mode 100644 egs/aishell/ASR/seamlessm4t/train2.py diff --git a/egs/aishell/ASR/run.sh b/egs/aishell/ASR/run.sh new file mode 100644 index 000000000..8c0770b05 --- /dev/null +++ b/egs/aishell/ASR/run.sh @@ -0,0 +1,4 @@ + +export CUDA_VISIBLE_DEVICES="2,3" +export PYTHONPATH=$PYTHONPATH:/mnt/samsung-t7/yuekai/asr/icefall +torchrun --nproc-per-node 2 seamlessm4t/train2.py --use-fp16 1 --max-duration 20 diff --git a/egs/aishell/ASR/seamlessm4t/asr_datamodule.py b/egs/aishell/ASR/seamlessm4t/asr_datamodule.py new file mode 120000 index 000000000..fa1b8cca3 --- /dev/null +++ b/egs/aishell/ASR/seamlessm4t/asr_datamodule.py @@ -0,0 +1 @@ +../tdnn_lstm_ctc/asr_datamodule.py \ No newline at end of file diff --git a/egs/aishell/ASR/seamlessm4t/label_smoothing.py b/egs/aishell/ASR/seamlessm4t/label_smoothing.py new file mode 120000 index 000000000..e9d239fff --- /dev/null +++ b/egs/aishell/ASR/seamlessm4t/label_smoothing.py @@ -0,0 +1 @@ +../../../librispeech/ASR/conformer_ctc/label_smoothing.py \ No newline at end of file diff --git a/egs/aishell/ASR/seamlessm4t/model.py b/egs/aishell/ASR/seamlessm4t/model.py new file mode 100644 index 000000000..18e5996bb --- /dev/null +++ b/egs/aishell/ASR/seamlessm4t/model.py @@ -0,0 +1,117 @@ +import torch +from seamless_communication.models.inference import Translator +from seamless_communication.models.unity import ( + UnitTokenizer, + UnitYModel, + load_unity_model, + load_unity_text_tokenizer, + load_unity_unit_tokenizer, +) +from fairseq2.generation import ( + Seq2SeqGenerator, + SequenceGeneratorOptions, + SequenceGeneratorOutput, + SequenceToTextGenerator, + SequenceToTextOutput, +) +from seamless_communication.models.unity.model import UnitYModel, UnitYX2TModel + +import torchaudio +import torchaudio.compliance.kaldi as ta_kaldi +audio_file="/mnt/samsung-t7/yuekai/asr/Triton-ASR-Client/datasets/mini_en/wav/1089-134686-0001.wav" +src_lang="cmn" + +audio_file="/mnt/samsung-t7/yuekai/asr/Triton-ASR-Client/datasets/mini_zh/wav/long.wav" +src_lang="eng" +target_lang = "cmn" + +audio_input = torchaudio.load(audio_file)[0] +feature = ta_kaldi.fbank(audio_input, num_mel_bins=80) +# feature shape is (T, F), convert it to (B, T, F), source_seq_lens tracks T +source_seqs = feature.unsqueeze(0) +source_seq_lens = torch.tensor([feature.shape[0]]) + +# Initialize a Translator object with a multitask model, vocoder on the GPU. + + +# translator = Translator("seamlessM4T_medium", vocoder_name_or_card="vocoder_36langs", device=torch.device("cuda:2"), dtype=torch.float16) + +# transcribed_text, _, _ = translator.predict(audio_file, "asr", src_lang) + +# print(transcribed_text) + + +model_name_or_card = "seamlessM4T_medium" +device = torch.device("cuda:3") + +# cast source_seq_lens, source_seqs to device, dtype to torch.float16 +source_seq_lens = source_seq_lens.to(device=device, dtype=torch.float16) +source_seqs = source_seqs.to(device=device, dtype=torch.float16) + + + +dtype = torch.float16 +model = load_unity_model(model_name_or_card, device=device, dtype=dtype) +model.eval() +text_tokenizer = load_unity_text_tokenizer(model_name_or_card) +print(text_tokenizer.model.eos_idx, text_tokenizer.model.pad_idx) +text_tokenizer_encoder = text_tokenizer.create_encoder(lang=target_lang, mode="target") +text_tokenizer_decoder = text_tokenizer.create_decoder() +# print attritbut of text_tokenizer_encoder + +print(text_tokenizer_encoder("")) +print(text_tokenizer_decoder(torch.tensor([3,45]))) +exit(0) + + + +# def decode( +# self, +# seqs: Tensor, +# seq_lens: Optional[Tensor], +# encoder_output: Tensor, +# encoder_padding_mask: Optional[Tensor], +# state_bag: Optional[IncrementalStateBag] = None, +# ) -> Tuple[Tensor, Optional[Tensor]]: +# seqs, padding_mask = self.text_decoder_frontend(seqs, seq_lens, state_bag) + +# return self.text_decoder( # type: ignore[no-any-return] +# seqs, padding_mask, encoder_output, encoder_padding_mask, state_bag +# ) + +# def decoding(model, feature): +# seqs, padding_mask = model.speech_encoder_frontend(seqs, seq_lens) +# speech_encoder(seqs, padding_mask) + +# decoder_output, decoder_padding_mask = self.decode( +# batch.target_seqs, +# batch.target_seq_lens, +# encoder_output, +# encoder_padding_mask, +# ) + +# text_logits = model.final_project(decoder_output, decoder_padding_mask) + +text_max_len_a = 1 +text_max_len_b = 200 + +text_opts = SequenceGeneratorOptions( + beam_size=5, soft_max_seq_len=(text_max_len_a, text_max_len_b) +) + +s2t_model = UnitYX2TModel( + encoder_frontend=model.speech_encoder_frontend, + encoder=model.speech_encoder, + decoder_frontend=model.text_decoder_frontend, + decoder=model.text_decoder, + final_proj=model.final_proj, + pad_idx=model.pad_idx, +) +s2t_generator = SequenceToTextGenerator( + s2t_model, text_tokenizer, target_lang, text_opts +) + +text_output = s2t_generator.generate_ex(source_seqs, source_seq_lens) +sentence = text_output.sentences[0] +print(sentence, type(sentence)) +sentence = sentence.bytes().decode("utf-8") diff --git a/egs/aishell/ASR/seamlessm4t/optim.py b/egs/aishell/ASR/seamlessm4t/optim.py new file mode 100644 index 000000000..abfb2092c --- /dev/null +++ b/egs/aishell/ASR/seamlessm4t/optim.py @@ -0,0 +1,1173 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import contextlib +import logging +import random +from collections import defaultdict +from typing import Dict, List, Optional, Tuple, Union + +import torch +from lhotse.utils import fix_random_seed +from torch import Tensor +from torch.optim import Optimizer + + +class BatchedOptimizer(Optimizer): + """ + This class adds to class Optimizer the capability to optimize parameters in batches: + it will stack the parameters and their grads for you so the optimizer can work + on tensors with an extra leading dimension. This is intended for speed with GPUs, + as it reduces the number of kernels launched in the optimizer. + + Args: + params: + """ + + def __init__(self, params, defaults): + super(BatchedOptimizer, self).__init__(params, defaults) + + @contextlib.contextmanager + def batched_params(self, param_group, group_params_names): + """ + This function returns (technically, yields) a list of + of tuples (p, state), where + p is a `fake` parameter that is stacked (over axis 0) from real parameters + that share the same shape, and its gradient is also stacked; + `state` is the state corresponding to this batch of parameters + (it will be physically located in the "state" for one of the real + parameters, the last one that has any particular shape and dtype). + + This function is decorated as a context manager so that it can + write parameters back to their "real" locations. + + The idea is, instead of doing: + + for p in group["params"]: + state = self.state[p] + ... + + you can do: + + with self.batched_params(group["params"]) as batches: + for p, state, p_names in batches: + ... + + + Args: + group: a parameter group, which is a list of parameters; should be + one of self.param_groups. + group_params_names: name for each parameter in group, + which is List[str]. + """ + batches = defaultdict( + list + ) # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter + batches_names = defaultdict( + list + ) # `batches` maps from tuple (dtype_as_str,*shape) to list of str + + assert len(param_group) == len(group_params_names) + for p, named_p in zip(param_group, group_params_names): + key = (str(p.dtype), *p.shape) + batches[key].append(p) + batches_names[key].append(named_p) + + batches_names_keys = list(batches_names.keys()) + sorted_idx = sorted( + range(len(batches_names)), key=lambda i: batches_names_keys[i] + ) + batches_names = [batches_names[batches_names_keys[idx]] for idx in sorted_idx] + batches = [batches[batches_names_keys[idx]] for idx in sorted_idx] + + stacked_params_dict = dict() + + # turn batches into a list, in deterministic order. + # tuples will contain tuples of (stacked_param, state, stacked_params_names), + # one for each batch in `batches`. + tuples = [] + + for batch, batch_names in zip(batches, batches_names): + p = batch[0] + # we arbitrarily store the state in the + # state corresponding to the 1st parameter in the + # group. class Optimizer will take care of saving/loading state. + state = self.state[p] + p_stacked = torch.stack(batch) + grad = torch.stack( + [torch.zeros_like(p) if p.grad is None else p.grad for p in batch] + ) + p_stacked.grad = grad + stacked_params_dict[key] = p_stacked + tuples.append((p_stacked, state, batch_names)) + + yield tuples # <-- calling code will do the actual optimization here! + + for ((stacked_params, _state, _names), batch) in zip(tuples, batches): + for i, p in enumerate(batch): # batch is list of Parameter + p.copy_(stacked_params[i]) + + +class ScaledAdam(BatchedOptimizer): + """ + Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update + proportional to the norm of that parameter; and also learn the scale of the parameter, + in log space, subject to upper and lower limits (as if we had factored each parameter as + param = underlying_param * log_scale.exp()) + + + Args: + params: The parameters or param_groups to optimize (like other Optimizer subclasses) + Unlike common optimizers, which accept model.parameters() or groups of parameters(), + this optimizer could accept model.named_parameters() or groups of named_parameters(). + See comments of function _get_names_of_parameters for its 4 possible cases. + lr: The learning rate. We will typically use a learning rate schedule that starts + at 0.03 and decreases over time, i.e. much higher than other common + optimizers. + clipping_scale: (e.g. 2.0) + A scale for gradient-clipping: if specified, the normalized gradients + over the whole model will be clipped to have 2-norm equal to + `clipping_scale` times the median 2-norm over the most recent period + of `clipping_update_period` minibatches. By "normalized gradients", + we mean after multiplying by the rms parameter value for this tensor + [for non-scalars]; this is appropriate because our update is scaled + by this quantity. + betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad. + Must satisfy 0 < beta <= beta2 < 1. + scalar_lr_scale: A scaling factor on the learning rate, that we use to update the + scale of each parameter tensor and scalar parameters of the mode.. + If each parameter were decomposed + as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale + would be a the scaling factor on the learning rate of p_scale. + eps: A general-purpose epsilon to prevent division by zero + param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of + learning the scale on the parameters (we'll constrain the rms of each non-scalar + parameter tensor to be >= this value) + param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of + learning the scale on the parameters (we'll constrain the rms of each non-scalar + parameter tensor to be <= this value) + scalar_max: Maximum absolute value for scalar parameters (applicable if your + model has any parameters with numel() == 1). + size_update_period: The periodicity, in steps, with which we update the size (scale) + of the parameter tensor. This is provided to save a little time + in the update. + clipping_update_period: if clipping_scale is specified, this is the period + """ + + def __init__( + self, + params, + lr=3e-02, + clipping_scale=None, + betas=(0.9, 0.98), + scalar_lr_scale=0.1, + eps=1.0e-08, + param_min_rms=1.0e-05, + param_max_rms=3.0, + scalar_max=10.0, + size_update_period=4, + clipping_update_period=100, + ): + + defaults = dict( + lr=lr, + clipping_scale=clipping_scale, + betas=betas, + scalar_lr_scale=scalar_lr_scale, + eps=eps, + param_min_rms=param_min_rms, + param_max_rms=param_max_rms, + scalar_max=scalar_max, + size_update_period=size_update_period, + clipping_update_period=clipping_update_period, + ) + + # If params only contains parameters or group of parameters, + # i.e when parameter names are not given, + # this flag will be set to False in funciton _get_names_of_parameters. + self.show_dominant_parameters = True + param_groups, parameters_names = self._get_names_of_parameters(params) + super(ScaledAdam, self).__init__(param_groups, defaults) + assert len(self.param_groups) == len(parameters_names) + self.parameters_names = parameters_names + + def _get_names_of_parameters( + self, params_or_named_params + ) -> Tuple[List[Dict], List[List[str]]]: + """ + Args: + params_or_named_params: according to the way ScaledAdam is initialized in train.py, + this argument could be one of following 4 cases, + case 1, a generator of parameter, e.g.: + optimizer = ScaledAdam(model.parameters(), lr=params.base_lr, clipping_scale=3.0) + + case 2, a list of parameter groups with different config, e.g.: + model_param_groups = [ + {'params': model.encoder.parameters(), 'lr': 0.05}, + {'params': model.decoder.parameters(), 'lr': 0.01}, + {'params': model.joiner.parameters(), 'lr': 0.03}, + ] + optimizer = ScaledAdam(model_param_groups, lr=params.base_lr, clipping_scale=3.0) + + case 3, a generator of named_parameter, e.g.: + optimizer = ScaledAdam(model.named_parameters(), lr=params.base_lr, clipping_scale=3.0) + + case 4, a list of named_parameter groups with different config, e.g.: + model_named_param_groups = [ + {'named_params': model.encoder.named_parameters(), 'lr': 0.05}, + {'named_params': model.decoder.named_parameters(), 'lr': 0.01}, + {'named_params': model.joiner.named_parameters(), 'lr': 0.03}, + ] + optimizer = ScaledAdam(model_named_param_groups, lr=params.base_lr, clipping_scale=3.0) + + For case 1 and case 2, input params is used to initialize the underlying torch.optimizer. + For case 3 and case 4, firstly, names and params are extracted from input named_params, + then, these extracted params are used to initialize the underlying torch.optimizer, + and these extracted names are mainly used by function + `_show_gradient_dominating_parameter` + + Returns: + Returns a tuple containing 2 elements: + - `param_groups` with type List[Dict], each Dict element is a parameter group. + An example of `param_groups` could be: + [ + {'params': `one iterable of Parameter`, 'lr': 0.05}, + {'params': `another iterable of Parameter`, 'lr': 0.08}, + {'params': `a third iterable of Parameter`, 'lr': 0.1}, + ] + - `param_gruops_names` with type List[List[str]], + each `List[str]` is for a group['params'] in param_groups, + and each `str` is the name of a parameter. + A dummy name "foo" is related to each parameter, + if input are params without names, i.e. case 1 or case 2. + """ + # variable naming convention in this function: + # p is short for param. + # np is short for named_param. + # p_or_np is short for param_or_named_param. + # cur is short for current. + # group is a dict, e.g. {'params': iterable of parameter, 'lr': 0.05, other fields}. + # groups is a List[group] + + iterable_or_groups = list(params_or_named_params) + if len(iterable_or_groups) == 0: + raise ValueError("optimizer got an empty parameter list") + + # The first value of returned tuple. A list of dicts containing at + # least 'params' as a key. + param_groups = [] + + # The second value of returned tuple, + # a List[List[str]], each sub-List is for a group. + param_groups_names = [] + + if not isinstance(iterable_or_groups[0], dict): + # case 1 or case 3, + # the input is an iterable of parameter or named parameter. + param_iterable_cur_group = [] + param_names_cur_group = [] + for p_or_np in iterable_or_groups: + if isinstance(p_or_np, tuple): + # case 3 + name, param = p_or_np + else: + # case 1 + assert isinstance(p_or_np, torch.Tensor) + param = p_or_np + # Assign a dummy name as a placeholder + name = "foo" + self.show_dominant_parameters = False + param_iterable_cur_group.append(param) + param_names_cur_group.append(name) + param_groups.append({"params": param_iterable_cur_group}) + param_groups_names.append(param_names_cur_group) + else: + # case 2 or case 4 + # the input is groups of parameter or named parameter. + for cur_group in iterable_or_groups: + assert "named_params" in cur_group + name_list = [ x[0] for x in cur_group["named_params"] ] + p_list = [ x[1] for x in cur_group["named_params"] ] + del cur_group["named_params"] + cur_group["params"] = p_list + param_groups.append(cur_group) + param_groups_names.append(name_list) + + return param_groups, param_groups_names + + def __setstate__(self, state): + super(ScaledAdam, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + batch = True + + for group, group_params_names in zip(self.param_groups, self.parameters_names): + + with self.batched_params(group["params"], group_params_names) as batches: + + # batches is list of pairs (stacked_param, state). stacked_param is like + # a regular parameter, and will have a .grad, but the 1st dim corresponds to + # a stacking dim, it is not a real dim. + + if ( + len(batches[0][1]) == 0 + ): # if len(first state) == 0: not yet initialized + clipping_scale = 1 + else: + clipping_scale = self._get_clipping_scale(group, batches) + + for p, state, _ in batches: + # Perform optimization step. + # grad is not going to be None, we handled that when creating the batches. + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "ScaledAdam optimizer does not support sparse gradients" + ) + # State initialization + if len(state) == 0: + self._init_state(group, p, state) + + self._step_one_batch(group, p, state, clipping_scale) + + return loss + + def _init_state(self, group: dict, p: Tensor, state: dict): + """ + Initializes state dict for parameter 'p'. Assumes that dim 0 of tensor p + is actually the batch dimension, corresponding to batched-together + parameters of a given shape. + + + Args: + group: Dict to look up configuration values. + p: The parameter that we are initializing the state for + state: Dict from string to whatever state we are initializing + """ + size_update_period = group["size_update_period"] + + state["step"] = 0 + + kwargs = {"device": p.device, "dtype": p.dtype} + + # 'delta' implements conventional momentum. There are + # several different kinds of update going on, so rather than + # compute "exp_avg" like in Adam, we store and decay a + # parameter-change "delta", which combines all forms of + # update. this is equivalent to how it's done in Adam, + # except for the first few steps. + state["delta"] = torch.zeros_like(p, memory_format=torch.preserve_format) + + batch_size = p.shape[0] + numel = p.numel() // batch_size + + if numel > 1: + # "param_rms" just periodically records the scalar root-mean-square value of + # the parameter tensor. + # it has a shape like (batch_size, 1, 1, 1, 1) + param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt() + state["param_rms"] = param_rms + + state["scale_exp_avg_sq"] = torch.zeros_like(param_rms) + state["scale_grads"] = torch.zeros( + size_update_period, *param_rms.shape, **kwargs + ) + + # exp_avg_sq is the weighted sum of scaled gradients. as in Adam. + state["exp_avg_sq"] = torch.zeros_like(p, memory_format=torch.preserve_format) + + def _get_clipping_scale( + self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]] + ) -> float: + """ + Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients + by this amount before applying the rest of the update. + + Args: + group: the parameter group, an item in self.param_groups + tuples: a list of tuples of (param, state, param_names) + where param is a batched set of parameters, + with a .grad (1st dim is batch dim) + and state is the state-dict where optimization parameters are kept. + param_names is a List[str] while each str is name for a parameter + in batched set of parameters "param". + """ + assert len(tuples) >= 1 + clipping_scale = group["clipping_scale"] + (first_p, first_state, _) = tuples[0] + step = first_state["step"] + if clipping_scale is None or step == 0: + # no clipping. return early on step == 0 because the other + # parameters' state won't have been initialized yet. + return 1.0 + clipping_update_period = group["clipping_update_period"] + + tot_sumsq = torch.tensor(0.0, device=first_p.device) + for (p, state, param_names) in tuples: + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "ScaledAdam optimizer does not support sparse gradients" + ) + if p.numel() == p.shape[0]: # a batch of scalars + tot_sumsq += (grad**2).sum() # sum() to change shape [1] to [] + else: + tot_sumsq += ((grad * state["param_rms"]) ** 2).sum() + + tot_norm = tot_sumsq.sqrt() + if "model_norms" not in first_state: + first_state["model_norms"] = torch.zeros( + clipping_update_period, device=p.device + ) + first_state["model_norms"][step % clipping_update_period] = tot_norm + + if step % clipping_update_period == 0: + # Print some stats. + # We don't reach here if step == 0 because we would have returned + # above. + sorted_norms = first_state["model_norms"].sort()[0].to("cpu") + quartiles = [] + for n in range(0, 5): + index = min( + clipping_update_period - 1, (clipping_update_period // 4) * n + ) + quartiles.append(sorted_norms[index].item()) + + median = quartiles[2] + threshold = clipping_scale * median + first_state["model_norm_threshold"] = threshold + percent_clipped = ( + first_state["num_clipped"] * 100.0 / clipping_update_period + if "num_clipped" in first_state + else 0.0 + ) + first_state["num_clipped"] = 0 + quartiles = " ".join(["%.3e" % x for x in quartiles]) + logging.info( + f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, " + f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}" + ) + + if step < clipping_update_period: + return 1.0 # We have not yet estimated a norm to clip to. + else: + try: + model_norm_threshold = first_state["model_norm_threshold"] + except KeyError: + logging.info( + "Warning: model_norm_threshold not in state: possibly " + "you changed config when restarting, adding clipping_scale option?" + ) + return 1.0 + ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item()) + if ans < 1.0: + first_state["num_clipped"] += 1 + if ans < 0.1: + logging.warn( + f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}" + ) + if self.show_dominant_parameters: + assert p.shape[0] == len(param_names) + self._show_gradient_dominating_parameter(tuples, tot_sumsq) + return ans + + def _show_gradient_dominating_parameter( + self, tuples: List[Tuple[Tensor, dict, List[str]]], tot_sumsq: Tensor + ): + """ + Show information of parameter which dominates tot_sumsq. + + Args: + tuples: a list of tuples of (param, state, param_names) + where param is a batched set of parameters, + with a .grad (1st dim is batch dim) + and state is the state-dict where optimization parameters are kept. + param_names is a List[str] while each str is name for a parameter + in batched set of parameters "param". + tot_sumsq: sumsq of all parameters. Though it's could be calculated + from tuples, we still pass it to save some time. + """ + all_sumsq_orig = {} + for (p, state, batch_param_names) in tuples: + # p is a stacked batch parameters. + batch_grad = p.grad + if p.numel() == p.shape[0]: # a batch of scalars + batch_sumsq_orig = batch_grad**2 + # Dummy values used by following `zip` statement. + batch_rms_orig = torch.ones(p.shape[0]) + else: + batch_rms_orig = state["param_rms"] + batch_sumsq_orig = ((batch_grad * batch_rms_orig) ** 2).sum( + dim=list(range(1, batch_grad.ndim)) + ) + + for name, sumsq_orig, rms, grad in zip( + batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad + ): + + proportion_orig = sumsq_orig / tot_sumsq + all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad) + + assert torch.isclose( + sum([value[0] for value in all_sumsq_orig.values()]).cpu(), + torch.tensor(1.0), + ) + sorted_by_proportion = { + k: v + for k, v in sorted( + all_sumsq_orig.items(), key=lambda item: item[1][0], reverse=True + ) + } + dominant_param_name = next(iter(sorted_by_proportion)) + ( + dominant_proportion, + dominant_sumsq, + dominant_rms, + dominant_grad, + ) = sorted_by_proportion[dominant_param_name] + logging.info( + f"Parameter dominating tot_sumsq {dominant_param_name}" + f" with proportion {dominant_proportion:.2f}," + f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)" + f"={dominant_sumsq:.3e}," + f" grad_sumsq={(dominant_grad**2).sum():.3e}," + f" orig_rms_sq={(dominant_rms**2).item():.3e}" + ) + + def _step_one_batch( + self, group: dict, p: Tensor, state: dict, clipping_scale: float + ): + """ + Do the step for one parameter, which is actually going to be a batch of + `real` parameters, with dim 0 as the batch dim. + Args: + group: dict to look up configuration values + p: parameter to update (actually multiple parameters stacked together + as a batch) + state: state-dict for p, to look up the optimizer state + """ + lr = group["lr"] + size_update_period = group["size_update_period"] + beta1 = group["betas"][0] + + grad = p.grad + if clipping_scale != 1.0: + grad = grad * clipping_scale + step = state["step"] + delta = state["delta"] + + delta.mul_(beta1) + batch_size = p.shape[0] + numel = p.numel() // batch_size + if numel > 1: + # Update the size/scale of p, and set param_rms + scale_grads = state["scale_grads"] + scale_grads[step % size_update_period] = (p * grad).sum( + dim=list(range(1, p.ndim)), keepdim=True + ) + if step % size_update_period == size_update_period - 1: + param_rms = state["param_rms"] # shape: (batch_size, 1, 1, ..) + param_rms.copy_( + (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt() + ) + if step > 0: + # self._size_update() learns the overall scale on the + # parameter, by shrinking or expanding it. + self._size_update(group, scale_grads, p, state) + + if numel == 1: + # For parameters with 1 element we just use regular Adam. + # Updates delta. + self._step_scalar(group, p, state) + else: + self._step(group, p, state) + + state["step"] = step + 1 + + def _size_update( + self, group: dict, scale_grads: Tensor, p: Tensor, state: dict + ) -> None: + """ + Called only where p.numel() > 1, this updates the scale of the parameter. + If we imagine: p = underlying_param * scale.exp(), and we are doing + gradient descent on underlying param and on scale, this function does the update + on `scale`. + + Args: + group: dict to look up configuration values + scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing + grads w.r.t. the scales. + p: The parameter to update + state: The state-dict of p + """ + + param_rms = state["param_rms"] + beta1, beta2 = group["betas"] + size_lr = group["lr"] * group["scalar_lr_scale"] + param_min_rms = group["param_min_rms"] + param_max_rms = group["param_max_rms"] + eps = group["eps"] + step = state["step"] + batch_size = p.shape[0] + + size_update_period = scale_grads.shape[0] + # correct beta2 for the size update period: we will have + # faster decay at this level. + beta2_corr = beta2**size_update_period + + scale_exp_avg_sq = state["scale_exp_avg_sq"] # shape: (batch_size, 1, 1, ..) + scale_exp_avg_sq.mul_(beta2_corr).add_( + (scale_grads**2).mean(dim=0), # mean over dim `size_update_period` + alpha=1 - beta2_corr, + ) # shape is (batch_size, 1, 1, ...) + + # The 1st time we reach here is when size_step == 1. + size_step = (step + 1) // size_update_period + bias_correction2 = 1 - beta2_corr**size_step + # we don't bother with bias_correction1; this will help prevent divergence + # at the start of training. + + denom = scale_exp_avg_sq.sqrt() + eps + + scale_step = ( + -size_lr * (bias_correction2**0.5) * scale_grads.sum(dim=0) / denom + ) + + is_too_small = param_rms < param_min_rms + + # when the param gets too small, just don't shrink it any further. + scale_step.masked_fill_(is_too_small, 0.0) + + # and ensure the parameter rms after update never exceeds param_max_rms. + # We have to look at the trained model for parameters at or around the + # param_max_rms, because sometimes they can indicate a problem with the + # topology or settings. + scale_step = torch.minimum(scale_step, + (param_max_rms - param_rms) / param_rms) + + delta = state["delta"] + # the factor of (1-beta1) relates to momentum. + delta.add_(p * scale_step, alpha=(1 - beta1)) + + def _step(self, group: dict, p: Tensor, state: dict): + """ + This function does the core update of self.step(), in the case where the members of + the batch have more than 1 element. + + Args: + group: A dict which will be used to look up configuration values + p: The parameter to be updated + grad: The grad of p + state: The state-dict corresponding to parameter p + + This function modifies p. + """ + grad = p.grad + lr = group["lr"] + beta1, beta2 = group["betas"] + eps = group["eps"] + param_min_rms = group["param_min_rms"] + step = state["step"] + + exp_avg_sq = state["exp_avg_sq"] + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2)) + + this_step = state["step"] - (state["zero_step"] if "zero_step" in state else 0) + bias_correction2 = 1 - beta2 ** (this_step + 1) + if bias_correction2 < 0.99: + # note: not in-place. + exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2) + + denom = exp_avg_sq.sqrt() + denom += eps + grad = grad / denom + + alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms) + + delta = state["delta"] + delta.add_(grad * alpha) + p.add_(delta) + + def _step_scalar(self, group: dict, p: Tensor, state: dict): + """ + A simplified form of the core update for scalar tensors, where we cannot get a good + estimate of the parameter rms. + """ + beta1, beta2 = group["betas"] + scalar_max = group["scalar_max"] + eps = group["eps"] + lr = group["lr"] * group["scalar_lr_scale"] + grad = p.grad + + exp_avg_sq = state["exp_avg_sq"] # shape: (batch_size,) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + + # bias_correction2 is like in Adam. Don't bother with bias_correction1; + # slower update at the start will help stability anyway. + bias_correction2 = 1 - beta2 ** (state["step"] + 1) + denom = (exp_avg_sq / bias_correction2).sqrt() + eps + + delta = state["delta"] + delta.add_(grad / denom, alpha=-lr * (1 - beta1)) + p.clamp_(min=-scalar_max, max=scalar_max) + p.add_(delta) + + +class LRScheduler(object): + """ + Base-class for learning rate schedulers where the learning-rate depends on both the + batch and the epoch. + """ + + def __init__(self, optimizer: Optimizer, verbose: bool = False): + # Attach optimizer + if not isinstance(optimizer, Optimizer): + raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__)) + self.optimizer = optimizer + self.verbose = verbose + + for group in optimizer.param_groups: + group.setdefault("base_lr", group["lr"]) + + self.base_lrs = [group["base_lr"] for group in optimizer.param_groups] + + self.epoch = 0 + self.batch = 0 + + def state_dict(self): + """Returns the state of the scheduler as a :class:`dict`. + + It contains an entry for every variable in self.__dict__ which + is not the optimizer. + """ + return { + "base_lrs": self.base_lrs, + "epoch": self.epoch, + "batch": self.batch, + } + + def load_state_dict(self, state_dict): + """Loads the schedulers state. + + Args: + state_dict (dict): scheduler state. Should be an object returned + from a call to :meth:`state_dict`. + """ + self.__dict__.update(state_dict) + + def get_last_lr(self) -> List[float]: + """Return last computed learning rate by current scheduler. Will be a list of float.""" + return self._last_lr + + def get_lr(self): + # Compute list of learning rates from self.epoch and self.batch and + # self.base_lrs; this must be overloaded by the user. + # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ] + raise NotImplementedError + + def step_batch(self, batch: Optional[int] = None) -> None: + # Step the batch index, or just set it. If `batch` is specified, it + # must be the batch index from the start of training, i.e. summed over + # all epochs. + # You can call this in any order; if you don't provide 'batch', it should + # of course be called once per batch. + if batch is not None: + self.batch = batch + else: + self.batch = self.batch + 1 + self._set_lrs() + + def step_epoch(self, epoch: Optional[int] = None): + # Step the epoch index, or just set it. If you provide the 'epoch' arg, + # you should call this at the start of the epoch; if you don't provide the 'epoch' + # arg, you should call it at the end of the epoch. + if epoch is not None: + self.epoch = epoch + else: + self.epoch = self.epoch + 1 + self._set_lrs() + + def _set_lrs(self): + values = self.get_lr() + assert len(values) == len(self.optimizer.param_groups) + + for i, data in enumerate(zip(self.optimizer.param_groups, values)): + param_group, lr = data + param_group["lr"] = lr + self.print_lr(self.verbose, i, lr) + self._last_lr = [group["lr"] for group in self.optimizer.param_groups] + + def print_lr(self, is_verbose, group, lr): + """Display the current learning rate.""" + if is_verbose: + logging.info( + f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate" + f" of group {group} to {lr:.4e}." + ) + + +class Eden(LRScheduler): + """ + Eden scheduler. + The basic formula (before warmup) is: + lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 * + (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup + where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches + and then stays constant at 1. + + + E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam + + Args: + optimizer: the optimizer to change the learning rates on + lr_batches: the number of batches after which we start significantly + decreasing the learning rate, suggest 5000. + lr_epochs: the number of epochs after which we start significantly + decreasing the learning rate, suggest 6 if you plan to do e.g. + 20 to 40 epochs, but may need smaller number if dataset is huge + and you will do few epochs. + """ + + def __init__( + self, + optimizer: Optimizer, + lr_batches: Union[int, float], + lr_epochs: Union[int, float], + warmup_batches: Union[int, float] = 500.0, + warmup_start: float = 0.5, + verbose: bool = False, + ): + super(Eden, self).__init__(optimizer, verbose) + self.lr_batches = lr_batches + self.lr_epochs = lr_epochs + self.warmup_batches = warmup_batches + + assert 0.0 <= warmup_start <= 1.0, warmup_start + self.warmup_start = warmup_start + + def get_lr(self): + factor = ( + (self.batch**2 + self.lr_batches**2) / self.lr_batches**2 + ) ** -0.25 * ( + ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25 + ) + warmup_factor = ( + 1.0 + if self.batch >= self.warmup_batches + else self.warmup_start + (1.0 - self.warmup_start) * (self.batch / self.warmup_batches) + # else 0.5 + 0.5 * (self.batch / self.warmup_batches) + ) + + return [x * factor * warmup_factor for x in self.base_lrs] + + +def _test_eden(): + m = torch.nn.Linear(100, 100) + optim = ScaledAdam(m.parameters(), lr=0.03) + + scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True) + + for epoch in range(10): + scheduler.step_epoch(epoch) # sets epoch to `epoch` + + for step in range(20): + x = torch.randn(200, 100).detach() + x.requires_grad = True + y = m(x) + dy = torch.randn(200, 100).detach() + f = (y * dy).sum() + f.backward() + + optim.step() + scheduler.step_batch() + optim.zero_grad() + + logging.info(f"last lr = {scheduler.get_last_lr()}") + logging.info(f"state dict = {scheduler.state_dict()}") + + +# This is included mostly as a baseline for ScaledAdam. +class Eve(Optimizer): + """ + Implements Eve algorithm. This is a modified version of AdamW with a special + way of setting the weight-decay / shrinkage-factor, which is designed to make the + rms of the parameters approach a particular target_rms (default: 0.1). This is + for use with networks with 'scaled' versions of modules (see scaling.py), which + will be close to invariant to the absolute scale on the parameter matrix. + + The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_. + The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_. + Eve is unpublished so far. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay coefficient (default: 3e-4; + this value means that the weight would decay significantly after + about 3k minibatches. Is not multiplied by learning rate, but + is conditional on RMS-value of parameter being > target_rms. + target_rms (float, optional): target root-mean-square value of + parameters, if they fall below this we will stop applying weight decay. + + + .. _Adam: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _Decoupled Weight Decay Regularization: + https://arxiv.org/abs/1711.05101 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.98), + eps=1e-8, + weight_decay=1e-3, + target_rms=0.1, + ): + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0 <= weight_decay <= 0.1: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + if not 0 < target_rms <= 10.0: + raise ValueError("Invalid target_rms value: {}".format(target_rms)) + defaults = dict( + lr=lr, + betas=betas, + eps=eps, + weight_decay=weight_decay, + target_rms=target_rms, + ) + super(Eve, self).__init__(params, defaults) + + def __setstate__(self, state): + super(Eve, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + + # Perform optimization step + grad = p.grad + if grad.is_sparse: + raise RuntimeError("AdamW does not support sparse gradients") + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + + exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"] + + beta1, beta2 = group["betas"] + + state["step"] += 1 + bias_correction1 = 1 - beta1 ** state["step"] + bias_correction2 = 1 - beta2 ** state["step"] + + # Decay the first and second moment running average coefficient + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_( + group["eps"] + ) + + step_size = group["lr"] / bias_correction1 + target_rms = group["target_rms"] + weight_decay = group["weight_decay"] + + if p.numel() > 1: + # avoid applying this weight-decay on "scaling factors" + # (which are scalar). + is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5)) + p.mul_(1 - (weight_decay * is_above_target_rms)) + + p.addcdiv_(exp_avg, denom, value=-step_size) + + if random.random() < 0.0005: + step = (exp_avg / denom) * step_size + logging.info( + f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}" + ) + + return loss + + +def _test_scaled_adam(hidden_dim: int): + import timeit + + from scaling import ScaledLinear + + E = 100 + B = 4 + T = 2 + logging.info("in test_eve_cain") + # device = torch.device('cuda') + device = torch.device("cpu") + dtype = torch.float32 + + fix_random_seed(42) + # these input_magnitudes and output_magnitudes are to test that + # Abel is working as we expect and is able to adjust scales of + # different dims differently. + input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp() + output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp() + + for iter in [1, 0]: + fix_random_seed(42) + Linear = torch.nn.Linear if iter == 0 else ScaledLinear + + m = torch.nn.Sequential( + Linear(E, hidden_dim), + torch.nn.PReLU(), + Linear(hidden_dim, hidden_dim), + torch.nn.PReLU(), + Linear(hidden_dim, E), + ).to(device) + + train_pairs = [ + ( + 100.0 + * torch.randn(B, T, E, device=device, dtype=dtype) + * input_magnitudes, + torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes, + ) + for _ in range(20) + ] + + if iter == 0: + optim = Eve(m.parameters(), lr=0.003) + elif iter == 1: + optim = ScaledAdam(m.parameters(), lr=0.03, clipping_scale=2.0) + scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False) + + start = timeit.default_timer() + avg_loss = 0.0 + for epoch in range(180): + scheduler.step_epoch() + # if epoch == 100 and iter in [2,3]: + # optim.reset_speedup() # check it doesn't crash. + + # if epoch == 130: + # opts = diagnostics.TensorDiagnosticOptions( + # 2 ** 22 + # ) # allow 4 megabytes per sub-module + # diagnostic = diagnostics.attach_diagnostics(m, opts) + + for n, (x, y) in enumerate(train_pairs): + y_out = m(x) + loss = ((y_out - y) ** 2).mean() * 100.0 + if epoch == 0 and n == 0: + avg_loss = loss.item() + else: + avg_loss = 0.98 * avg_loss + 0.02 * loss.item() + if n == 0 and epoch % 5 == 0: + # norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item() + # norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item() + # norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item() + # norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item() + # scale1 = '%.2e' % (m[0].weight_scale.exp().item()) + # scale1b = '%.2e' % (m[0].bias_scale.exp().item()) + # scale2 = '%.2e' % (m[2].weight_scale.exp().item()) + # scale2b = '%.2e' % (m[2].bias_scale.exp().item()) + lr = scheduler.get_last_lr()[0] + logging.info( + f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}" + ) # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b} + loss.log().backward() + optim.step() + optim.zero_grad() + scheduler.step_batch() + + # diagnostic.print_diagnostics() + + stop = timeit.default_timer() + logging.info(f"Iter={iter}, Time taken: {stop - start}") + + logging.info(f"last lr = {scheduler.get_last_lr()}") + # logging.info("state dict = ", scheduler.state_dict()) + # logging.info("optim state_dict = ", optim.state_dict()) + logging.info(f"input_magnitudes = {input_magnitudes}") + logging.info(f"output_magnitudes = {output_magnitudes}") + + +if __name__ == "__main__": + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + logging.getLogger().setLevel(logging.INFO) + import subprocess + + s = subprocess.check_output( + "git status -uno .; git log -1; git diff HEAD .", shell=True + ) + logging.info(s) + import sys + + if len(sys.argv) > 1: + hidden_dim = int(sys.argv[1]) + else: + hidden_dim = 200 + + _test_scaled_adam(hidden_dim) + _test_eden() diff --git a/egs/aishell/ASR/seamlessm4t/train2.py b/egs/aishell/ASR/seamlessm4t/train2.py new file mode 100644 index 000000000..4802473c9 --- /dev/null +++ b/egs/aishell/ASR/seamlessm4t/train2.py @@ -0,0 +1,1254 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +./prepare.sh + +If you use --datatang-prob=0, then you don't need to run the above script. + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7/exp \ + --full-libri 1 \ + --max-duration 550 +""" + + +import argparse +import copy +import logging +import random +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from typing import List +#from aishell import AIShell +#from asr_datamodule import AsrDataModule +from asr_datamodule import AishellAsrDataModule +#from decoder import Decoder +#from joiner import Joiner +from lhotse import CutSet, load_manifest +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +#from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.nn.functional import pad as pad_tensor +from torch.utils.tensorboard import SummaryWriter +#from zipformer import Zipformer + +from icefall import diagnostics +#from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist, get_world_size, get_rank, get_local_rank +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + MetricsTracker, + filter_uneven_sized_batch, + setup_logger, + str2bool, +) + +from seamless_communication.models.unity import ( + UnitTokenizer, + UnitYModel, + load_unity_model, + load_unity_text_tokenizer, + load_unity_unit_tokenizer, +) +from fairseq2.generation import ( + Seq2SeqGenerator, + SequenceGeneratorOptions, + SequenceGeneratorOutput, + SequenceToTextGenerator, + SequenceToTextOutput, +) +from fairseq2.data.text import ( + SentencePieceDecoder, + SentencePieceEncoder, + SentencePieceModel, + TextTokenDecoder, + TextTokenEncoder, + TextTokenizer, + vocabulary_from_sentencepiece, +) + +from label_smoothing import LabelSmoothingLoss + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=6, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=1, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "frame_shift_ms": 10.0, + "allowed_excess_duration_ratio": 0.1, + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 100, + "env_info": get_env_info(), + } + ) + + return params + + +# def get_transducer_model(params: AttributeDict) -> nn.Module: +# encoder = get_encoder_model(params) +# decoder = get_decoder_model(params) +# joiner = get_joiner_model(params) + +# model = Transducer( +# encoder=encoder, +# decoder=decoder, +# joiner=joiner, +# encoder_dim=int(params.encoder_dims.split(",")[-1]), +# decoder_dim=params.decoder_dim, +# joiner_dim=params.joiner_dim, +# vocab_size=params.vocab_size, +# ) +# return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + text_tokenizer_encoder: SentencePieceEncoder, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + # For the uneven-sized batch, the total duration after padding would possibly + # cause OOM. Hence, for each batch, which is sorted descendingly by length, + # we simply drop the last few shortest samples, so that the retained total frames + # (after padding) would not exceed `allowed_max_frames`: + # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`, + # where `max_frames = max_duration * 1000 // frame_shift_ms`. + # We set allowed_excess_duration_ratio=0.1. + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + def _batch_tensors(tensors: List[Tensor], pad_value: Any) -> Tensor: + padding_size = max(tensor.shape[0] for tensor in tensors) + dims = len(tensors[0].shape) + padded_tensors = [] + for tensor in tensors: + padding = [0] * 2 * dims + padding[-1] = padding_size - tensor.shape[0] + padded_tensors.append(pad_tensor(tensor, padding, "constant", pad_value)) + return torch.stack([tensor for tensor in padded_tensors], dim=0) + + max_frames = params.max_duration * 1000 // params.frame_shift_ms + allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio)) + batch = filter_uneven_sized_batch(batch, allowed_max_frames) + + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + text_tokens_list = [text_tokenizer_encoder(text) for text in texts] + prev_outputs_tokens = _batch_tensors( + [tokens[:-1] for tokens in text_tokens_list], pad_value=params.pad_idx + ) + target_tokens = _batch_tensors( + [tokens[1:] for tokens in text_tokens_list], pad_value=params.pad_idx + ) + target_lengths = torch.LongTensor( + [tokens.shape[0] - 1 for tokens in text_tokens_list] + ) + decoder_criterion = LabelSmoothingLoss(ignore_index=params.pad_idx, label_smoothing=0.1, reduction="sum") + ignore_prefix_size = 1 # ignroe the lang code prediction + + with torch.set_grad_enabled(is_training): + speech_encoder_out, speech_encoder_padding_mask = model.encode_speech( + seqs=feature, + seq_lens=feature_lens, + ) + #assert batch.speech_to_text.prev_output_tokens is not None + text_decoder_out, text_decoder_padding_mask = model.decode( + seqs=prev_outputs_tokens.to(device), + seq_lens=target_lengths.to(device), + encoder_output=speech_encoder_out, + encoder_padding_mask=speech_encoder_padding_mask, + ) + text_logits = model.final_proj(text_decoder_out) + text_logits = text_logits[:, ignore_prefix_size:, :] + target_tokens = target_tokens[:, ignore_prefix_size:] + loss = decoder_criterion(text_logits, target_tokens.to(device)) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + text_tokenizer_encoder: SentencePieceEncoder, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + text_tokenizer_encoder=text_tokenizer_encoder, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + text_tokenizer_encoder: SentencePieceEncoder, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(train_dl): + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + text_tokenizer_encoder=text_tokenizer_encoder, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + text_tokenizer_encoder=text_tokenizer_encoder, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + # rank = get_rank() + # world_size = get_world_size() + # setup_dist(rank, world_size, use_ddp_launch=True) + setup_dist(use_ddp_launch=True) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + + + + logging.info("About to create model") + model_name_or_card = "seamlessM4T_medium" + lang = "cmn" + model = load_unity_model(model_name_or_card, device="cpu", dtype=torch.float32) + del model.t2u_model + del model.text_encoder + del model.text_encoder_frontend + # print(vars(model)) + # exit(0) + text_tokenizer = load_unity_text_tokenizer(model_name_or_card) + text_tokenizer_encoder = SentencePieceEncoder( + text_tokenizer.model, + prefix_tokens=["", f"__{lang}__"], + suffix_tokens=[""], + ) + #params.eos_idx = text_tokenizer.model.eos_idx + params.pad_idx = text_tokenizer.model.pad_idx + logging.info(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + #parameters_names = [] + #parameters_names.append( + # [name_param_pair[0] for name_param_pair in model.named_parameters()] + #) + # optimizer = ScaledAdam( + # model.parameters(), + # lr=params.base_lr, + # clipping_scale=2.0, + # parameters_names=parameters_names, + # ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + ) + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 12.0: + logging.warning( + f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + # T = ((c.num_frames - 7) // 2 + 1) // 2 + # tokens = sp.encode(c.supervisions[0].text, out_type=str) + + # if T < len(tokens): + # logging.warning( + # f"Exclude cut with ID {c.id} from training. " + # f"Number of frames (before subsampling): {c.num_frames}. " + # f"Number of frames (after subsampling): {T}. " + # f"Text: {c.supervisions[0].text}. " + # f"Tokens: {tokens}. " + # f"Number of tokens: {len(tokens)}" + # ) + # return False + + return True + + #aishell = AIShell(manifest_dir=args.manifest_dir) + #train_cuts = aishell.train_cuts() + #asr_datamodule = AishellAsrDataModule(args) + + aishell = AishellAsrDataModule(args) + # train_cuts = asr_datamodule.train_cuts() + # train_cuts = train_cuts.filter(remove_short_and_long_utt) + + # if args.enable_musan: + # cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz") + # else: + # cuts_musan = None + + + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + # train_dl = asr_datamodule.train_dataloaders( + # train_cuts, + # on_the_fly_feats=False, + # cuts_musan=cuts_musan, + # sampler_state_dict=sampler_state_dict, + # ) + + # valid_cuts = aishell.valid_cuts() + # valid_dl = asr_datamodule.valid_dataloaders(valid_cuts) + train_dl = aishell.train_dataloaders(aishell.train_cuts()) + valid_dl = aishell.valid_dataloaders(aishell.valid_cuts()) + # if not params.print_diagnostics: + # scan_pessimistic_batches_for_oom( + # model=model, + # train_dl=train_dl, + # optimizer=optimizer, + # graph_compiler=graph_compiler, + # params=params, + # ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + logging.info(f"start training from epoch {params.start_epoch}") + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + text_tokenizer_encoder=text_tokenizer_encoder, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + # y = graph_compiler.texts_to_ids(supervisions["text"]) + # num_tokens = sum(len(i) for i in y) + # logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + params: AttributeDict, + text_tokenizer_encoder: SentencePieceEncoder, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + text_tokenizer_encoder=text_tokenizer_encoder, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + AishellAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = get_world_size() + rank = get_rank() + assert world_size >= 1 + + run(rank=rank, world_size=world_size, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/icefall/dist.py b/icefall/dist.py index 922f31a2f..ee76e994a 100644 --- a/icefall/dist.py +++ b/icefall/dist.py @@ -22,7 +22,7 @@ from torch import distributed as dist def setup_dist( - rank, world_size, master_port=None, use_ddp_launch=False, master_addr=None + rank=None, world_size=None, master_port=None, use_ddp_launch=False, master_addr=None ): """ rank and world_size are used only if use_ddp_launch is False.