Fairseq学习日记：注定麻烦的旅程

现在开头：Fairseq是一个正在快速迭代的产品，而且是开源的！

这不是表扬，这意味着三件事情：

1.他没有文档！所有框架代码都没有任何注释，包括函数docstring都没有

2.他没有经过有效测试，估计是抢时间吧！即使是官网Readme里的例子也是无法跑起来的！

3.他是一个框架，而且是一个非常不Pythonic的框架，充斥着inline/包装器/莫名其妙的语法。

4.他大量使用类的静态方法和全局函数，这是一种破坏对象化的机制，实在是太难理解了。

虽然这四点决定他真的对不住Facebook的金字招牌，但是作为一个学习者，总要把他运行起来，那么开始这场针对 FaceBOOK派“全新老爷车”的修车之旅吧！

 

通过调试Fairseq我深刻的感到痛苦：各种莫名其妙的跳转，你永远不会知道在什么时候和什么地方发生跳转。

如果想利用Fairseq的机制提供管线封装，那么肯定会很累！

Fairseq的代码质量非常低，模型/模型配置参数/模型管线的参数对应的变量名都是model，任务/任务名称/任务的参数集合的变量名都是task，具体的含义需要调试上下文获得，并且不能保证一个函数的两次重入都是一样的含义。

以上批评意见的时间是1.0版，20201123

Step1 安装

略。。。网上教程很多的，反正你要自己动手

Fairseq

apex

下载预训练模型

下载训练数据 Lirispeech

 

Step2 跑wave2vec

先用官网的例子跑一个看看：

import torch
from fairseq.models.wav2vec import Wav2VecModel

cp = torch.load('/home/**/Documents/Research/fairseq/model/wav2vec_vox_960h_pl.pt.zip')
model = Wav2VecModel.build_model(cp['args'], task=None)
model.load_state_dict(cp['model'])
model.eval()

wav_input_16khz = torch.randn(1,10000)
z = model.feature_extractor(wav_input_16khz)
c = model.feature_aggregator(z)

你猜会咋样呢？

>>> Exception "AttributeError"
'Namespace' object has no attribute 'prediction_steps'
File: /home/**/Documents/workspace/fairseq/fairseq/models/wav2vec/wav2vec.py, Line: 175
Exception "unhandled AttributeError"
'Namespace' object has no attribute 'prediction_steps'
File: /home/**/Documents/workspace/fairseq/fairseq/models/wav2vec/wav2vec.py, Line: 175

哈哈，很神奇吧，其实也没啥就是单纯的代码质量太差！

@register_model("wav2vec", dataclass=Wav2VecConfig)
class Wav2VecModel(BaseFairseqModel):
    @classmethod
    def build_model(cls, cfg: Wav2VecConfig, task: FairseqTask):
        """Build a new model instance."""

        model = Wav2VecModel(cfg)
        logger.info(model)
        return model

    def __init__(self, cfg: Wav2VecConfig):
        super().__init__()

        self.prediction_steps = cfg.prediction_steps
        offset = cfg.offset

        if cfg.activation == "relu":
            activation = nn.ReLU()
        elif cfg.activation == "gelu":
            activation = nn.GELU()
        else:
            raise Exception("unknown activation " + cfg.activation)

调试可以看到，cfg这个参数根本就不是Wav2VecConfig类型，这个args的类型是<class 'argparse.Namespace'>,内容是：

Namespace(activation_dropout=0.1, adam_betas='(0.9, 0.98)', adam_eps=1e-08, all_gather_list_size=16384, 
apply_mask=True, arch='wav2vec_ctc', attention_dropout=0.0, best_checkpoint_metric='wer', 
bf16=False, bpe=None, broadcast_buffers=False, bucket_cap_mb=25, checkpoint_suffix='', clip_norm=0.0, 
cpu=False, criterion='wav2vec', curriculum=0, data='/private/home/qiantong/w2v/data/train_36s_short_960h_rescore', 
data_buffer_size=10, dataset_impl=None, ddp_backend='no_c10d', decay_steps=250000, device_id=0, disable_validation=False, 
distributed_backend='nccl', distributed_init_method='tcp://learnfair1679:56443', distributed_no_spawn=True, 
distributed_num_procs=None, distributed_port=56443, distributed_rank=0, distributed_world_size=24, 
distributed_wrapper='DDP', 
dropout=0.0, dropout_input=0, 
empty_cache_freq=0, enable_padding=False, fast_stat_sync=False, feature_grad_mult=0.0, final_dropout=0.0, final_lr_scale=0.05, find_unused_parameters=False, 
finetune_from_model=None, fix_batches_to_gpus=False,
 fixed_validation_seed=None, fp16=True, fp16_init_scale=128, fp16_no_flatten_grads=False, fp16_scale_tolerance=0.0, fp16_scale_window=None, freeze_finetune_updates=10000, 
hold_steps=210000, init_lr_scale=0.01, keep_best_checkpoints=-1, keep_interval_updates=-1, 
keep_last_epochs=-1, labels='ltr', layerdrop=0.1, localsgd_frequency=3, log_format='json', log_interval=500, lr=[3e-05], lr_scheduler='tri_stage', 
mask_channel_length=64, mask_channel_other=0.0, mask_channel_prob=0.1, mask_channel_selection='static',
 mask_length=10, mask_other=0.0, mask_prob=0.1, mask_selection='static', max_epoch=0, max_sample_size=None, 
max_sentences=None, max_sentences_valid=None, max_tokens=1280000, max_tokens_valid=1280000, max_update=500000, maximize_best_checkpoint_metric=False, 
memory_efficient_bf16=False, memory_efficient_fp16=False, min_loss_scale=0.0001, 
min_lr=-1, min_sample_size=None, model_parallel_size=1, no_epoch_checkpoints=True, no_last_checkpoints=False, 
no_mask_channel_overlap=False, no_mask_overlap=False, no_pretrained_weights=False, no_progress_bar=False, no_save=False, 
no_save_optimizer_state=False, no_seed_provided=False, normalize=True, nprocs_per_node=8, num_workers=4, 
optimizer='adam', optimizer_overrides='{}', patience=-1, pipeline_balance=None, pipeline_checkpoint='never', 
pipeline_chunks=None, pipeline_devices=None, pipeline_model_parallel=False, profile=False, quantization_config_path=None, 
remove_bpe='letter', required_batch_size_multiple=8, required_seq_len_multiple=1, reset_dataloader=False, 
reset_lr_scheduler=False, reset_meters=False, reset_optimizer=False, restore_file='checkpoint_last.pt', 
sample_rate=16000, save_dir='/checkpoint/michaelauli/asr/pseudolbl/960h_vox.fp16.u500000.savg.nrm.ltr.m_static.mstd0.mask10.mprob0.1.ld0.1.mc_static.mcstd0.maskc64.mcprob0.1.fgm0.0.ffu10000.lr3e-05.warmup40000.hld210000.dec250000.frs0.05.fd0.0.drop0.0.ad0.1.atd0.0.ms1280000.sd1337.uf1.ngpu24', 
save_interval=1, save_interval_updates=0, scoring='bleu', seed=1337, sentence_avg=True, skip_invalid_size_inputs_valid_test=False, slowmo_algorithm='LocalSGD', 
slowmo_momentum=None, stop_time_hours=0, task='audio_pretraining', tensorboard_logdir='', 
threshold_loss_scale=None, tokenizer=None, tpu=False, train_subset='train', update_freq=[1], use_bmuf=False, 
use_old_adam=False, user_dir=None, valid_subset='dev_other', validate_after_updates=10000, 
validate_interval=1, validate_interval_updates=0, w2v_args=Namespace(activation_dropout=0.0, activation_fn='gelu', adam_betas='(0.9,0.98)', 
adam_eps=1e-06, arch='wav2vec2', attention_dropout=0.1, attention_type='default', augment=False, 
best_checkpoint_metric='loss', bpe=None, bucket_cap_mb=25, centroids=None, clip_norm=25, codebook_negatives=0, combine_banks=False, combine_dataset=False, conv_bias=True, 
conv_feature_layers='[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] * 2', conv_logs=True, conv_pos=128, conv_pos_groups=16, conv_pos_layers=1, cpu=False, criterion='wav2vec', 
cross_sample_negatives=0, curriculum=0, data='/private/home/abaevski/data/librivox/no_silence', dataset_impl=None, ddp_backend='c10d', debug=False, device_id=0, disable_validation=False, 
distributed_backend='nccl', distributed_init_method='tcp://learnfair1331:55498', distributed_no_spawn=True, distributed_port=55498, distributed_rank=0, distributed_world_size=128, 
div_drop_percent=0, div_pen_threshold=None, dropout=0.0, dropout_features=0.1, dropout_input=0.1, duplicate_negatives=0, empty_cache_freq=0, enable_padding=False, 
encode_padded_indiv=False, encoder_attention_heads=16, encoder_embed_dim=1024, encoder_ffn_embed_dim=4096, encoder_layerdrop=0.0, encoder_layers=24, encoder_normalize_before=True, 
encoder_optimizer_params=None, encoder_schedule=0, end_learning_rate=0.0, extractor_mode='layer_norm', extractor_model=None, extractor_norm_location='default', 
fast_stat_sync=False, feature_glu=False, feature_grad_mult=1.0, feature_noise=0, feature_noise_last=0, features_pen=True, final_dim=768, find_unused_parameters=True, finetune_extractor=True, fix_batches_to_gpus=False, fixed_validation_seed=None, force_anneal=None, fp16=True, fp16_init_scale=128, fp16_no_flatten_grads=False, fp16_scale_tolerance=0.0, fp16_scale_window=None, group_norm_features=False, group_norm_groups=512, 
gumbel_noise_gain=1, infomax=True, input_noise=0.0, keep_interval_updates=1, keep_last_epochs=-1, label_smoothing=0.0, labels=None, latent_groups=2, latent_temp='(2.0,0.1,0.999995)', 
latent_var_banks=2, latent_vars=320, layer_norm_after=-1, layer_norm_before=0, layer_norm_features=True, layer_norm_first=True, layer_norm_repr=True, lazy_load_labels=False, 
log_format='json', log_interval=100, logit_temp=0.1, loss_weights=None, lr=[0.005], lr_scheduler='polynomial_decay', mask_min_space=1, mask_multiple_length=10, 
mask_prob=0.65, mask_same_channels=False, mask_same_timesteps=False, mask_selection='static', 
mask_stdev=0.0, masking_schedule=0, max_epoch=0, max_positions=8000, max_pred_length=0, max_sample_size=320000, max_sentences=None, max_sentences_valid=None, max_tokens=1200000, max_tokens_valid=1200000, max_update=1000000, 
maximize_best_checkpoint_metric=False, memory_efficient_fp16=False, min_loss_scale=0.0001, min_lr=-1, min_sample_size=32000, mlp_mi=768, negatives_from_everywhere=False, new_emb_pen=True, new_logit_pen=False, no_bert_init=False, 
no_epoch_checkpoints=True, no_last_checkpoints=False, no_mask_channel_overlap=False, no_mask_overlap=False, no_norm_after=None, no_progress_bar=False, no_save=False, no_save_optimizer_state=False, no_token_positional_embeddings=True, noise_type='gaussian', norm_init_weight=1.0, normalize=True, num_negatives=100, num_workers=6, optimizer='adam', optimizer_overrides='{}', 
penalize_transformer=False, penalty_coeff='[0,0,0.1,0]', penalty_temp=1.0, pooler_activation_fn='tanh', pooler_dropout=0.0, power=1.0, pre_norm=False, predict_everything=False, predictor_grad_mult=1.0, preemp=False, project_quantized=True, 
quantize_input=False, quantize_targets=True, quantized=False, quantizer_chance=0.0, quantizer_grad_mult=1.0, quantizer_init=True, quantizer_init_gain=1.0, quantizer_init_normal=True, relative_positional_embeddings=0, required_batch_size_multiple=8, rescale_sample_size=False, 
reset_dataloader=False, reset_lr_scheduler&#