我还是在想复现transformer原文的实验，于是按这个查到了一些tips（很有用！）

How to reproduce the result of WMT14 en-de on transformer BASE model? · Issue #346 · facebookresearch/fairseq

 发现fairseq这个库似乎很好用，有预训练好的模型，我就开心的跟着官网教程准备试试看。于是一下午都在配环境和报错~~~~耶耶耶~~~~~~~~~~~~~~（疯了别管我了~~~~

fairseq/examples/translation at main · facebookresearch/fairseq

代码：

import torch

# List available models

torch.hub.list('pytorch/fairseq')  # [..., 'transformer.wmt16.en-de', ... ]

第一次报错：

 RuntimeError Traceback (most recent call last) Cell In[2], line 4 1 import torch 3 # List available models ----> 4 torch.hub.list('pytorch/fairseq') File E:\miniconda\Lib\site-packages\torch\hub.py:430, in list(github, force_reload, skip_validation, trust_repo, verbose) 428 with _add_to_sys_path(repo_dir): 429 hubconf_path = os.path.join(repo_dir, MODULE_HUBCONF) --> 430 hub_module = _import_module(MODULE_HUBCONF, hubconf_path) 432 # We take functions starts with '_' as internal helper functions 433 entrypoints = [f for f in dir(hub_module) if callable(getattr(hub_module, f)) and not f.startswith('_')] File E:\miniconda\Lib\site-packages\torch\hub.py:106, in _import_module(name, path) 104 module = importlib.util.module_from_spec(spec) 105 assert isinstance(spec.loader, Loader) --> 106 spec.loader.exec_module(module) 107 return module File <frozen importlib._bootstrap_external>:940, in exec_module(self, module) File <frozen importlib._bootstrap>:241, in _call_with_frames_removed(f, *args, **kwds) File ~/.cache\torch\hub\pytorch_fairseq_main\hubconf.py:35 33 missing_deps.append(dep) 34 if len(missing_deps) > 0: ---> 35 raise RuntimeError("Missing dependencies: {}".format(", ".join(missing_deps))) 38 # only do fairseq imports after checking for dependencies 39 from fairseq.hub_utils import ( # noqa; noqa 40 BPEHubInterface as bpe, 41 TokenizerHubInterface as tokenizer, 42 ) RuntimeError: Missing dependencies: hydra-core, omegaconf

 okfine我来安装这两个库（巨坑！！根本没有那么简单）

第二次报错：

Using cache found in C:\Users\hp/.cache\torch\hub\pytorch_fairseq_main

---------------------------------------------------------------------------

ValueError                                Traceback (most recent call last)

Cell In[4], line 4

      1 import torch

      3 # List available models

----> 4 torch.hub.list('pytorch/fairseq')

File E:\miniconda\Lib\site-packages\torch\hub.py:430, in list(github, force_reload, skip_validation, trust_repo, verbose)

    428 with _add_to_sys_path(repo_dir):

    429     hubconf_path = os.path.join(repo_dir, MODULE_HUBCONF)

--> 430     hub_module = _import_module(MODULE_HUBCONF, hubconf_path)

    432 # We take functions starts with '_' as internal helper functions

    433 entrypoints = [f for f in dir(hub_module) if callable(getattr(hub_module, f)) and not f.startswith('_')]

File E:\miniconda\Lib\site-packages\torch\hub.py:106, in _import_module(name, path)

    104 module = importlib.util.module_from_spec(spec)

    105 assert isinstance(spec.loader, Loader)

--> 106 spec.loader.exec_module(module)

    107 return module

File <frozen importlib._bootstrap_external>:940, in exec_module(self, module)

File <frozen importlib._bootstrap>:241, in _call_with_frames_removed(f, *args, **kwds)

File ~/.cache\torch\hub\pytorch_fairseq_main\hubconf.py:39

     35     raise RuntimeError("Missing dependencies: {}".format(", ".join(missing_deps)))

     38 # only do fairseq imports after checking for dependencies

---> 39 from fairseq.hub_utils import (  # noqa; noqa

     40     BPEHubInterface as bpe,

     41     TokenizerHubInterface as tokenizer,

     42 )

     43 from fairseq.models import MODEL_REGISTRY  # noqa

     46 # torch.hub doesn't build Cython components, so if they are not found then try

     47 # to build them here

File ~/.cache\torch\hub\pytorch_fairseq_main\fairseq\__init__.py:20

     17 __all__ = ["pdb"]

     19 # backwards compatibility to support `from fairseq.X import Y`

---> 20 from fairseq.distributed import utils as distributed_utils

     21 from fairseq.logging import meters, metrics, progress_bar  # noqa

     23 sys.modules["fairseq.distributed_utils"] = distributed_utils

File ~/.cache\torch\hub\pytorch_fairseq_main\fairseq\distributed\__init__.py:7

      1 # Copyright (c) Facebook, Inc. and its affiliates.

      2 #

      3 # This source code is licensed under the MIT license found in the

      4 # LICENSE file in the root directory of this source tree.

      6 from .distributed_timeout_wrapper import DistributedTimeoutWrapper

----> 7 from .fully_sharded_data_parallel import (

      8     fsdp_enable_wrap,

      9     fsdp_wrap,

     10     FullyShardedDataParallel,

     11 )

     12 from .legacy_distributed_data_parallel import LegacyDistributedDataParallel

     13 from .module_proxy_wrapper import ModuleProxyWrapper

File ~/.cache\torch\hub\pytorch_fairseq_main\fairseq\distributed\fully_sharded_data_parallel.py:10

      7 from typing import Optional

      9 import torch

---> 10 from fairseq.dataclass.configs import DistributedTrainingConfig

     11 from fairseq.distributed import utils as dist_utils

     14 try:

File ~/.cache\torch\hub\pytorch_fairseq_main\fairseq\dataclass\__init__.py:6

      1 # Copyright (c) Facebook, Inc. and its affiliates.

      2 #

      3 # This source code is licensed under the MIT license found in the

      4 # LICENSE file in the root directory of this source tree.

----> 6 from .configs import FairseqDataclass

      7 from .constants import ChoiceEnum

     10 __all__ = [

     11     "FairseqDataclass",

     12     "ChoiceEnum",

     13 ]

File ~/.cache\torch\hub\pytorch_fairseq_main\fairseq\dataclass\configs.py:1127

   1118     ema_update_freq: int = field(

   1119         default=1, metadata={"help": "Do EMA update every this many model updates"}

   1120     )

   1121     ema_fp32: bool = field(

   1122         default=False,

   1123         metadata={"help": "If true, store EMA model in fp32 even if model is in fp16"},

   1124     )

-> 1127 @dataclass

   1128 class FairseqConfig(FairseqDataclass):

   1129     common: CommonConfig = CommonConfig()

   1130     common_eval: CommonEvalConfig = CommonEvalConfig()

File E:\miniconda\Lib\dataclasses.py:1230, in dataclass(cls, init, repr, eq, order, unsafe_hash, frozen, match_args, kw_only, slots, weakref_slot)

   1227     return wrap

   1229 # We're called as @dataclass without parens.

-> 1230 return wrap(cls)

File E:\miniconda\Lib\dataclasses.py:1220, in dataclass.<locals>.wrap(cls)

   1219 def wrap(cls):

-> 1220     return _process_class(cls, init, repr, eq, order, unsafe_hash,

   1221                           frozen, match_args, kw_only, slots,

   1222                           weakref_slot)

File E:\miniconda\Lib\dataclasses.py:958, in _process_class(cls, init, repr, eq, order, unsafe_hash, frozen, match_args, kw_only, slots, weakref_slot)

    955         kw_only = True

    956     else:

    957         # Otherwise it's a field of some type.

--> 958         cls_fields.append(_get_field(cls, name, type, kw_only))

    960 for f in cls_fields:

    961     fields[f.name] = f

File E:\miniconda\Lib\dataclasses.py:815, in _get_field(cls, a_name, a_type, default_kw_only)

    811 # For real fields, disallow mutable defaults.  Use unhashable as a proxy

    812 # indicator for mutability.  Read the __hash__ attribute from the class,

    813 # not the instance.

    814 if f._field_type is _FIELD and f.default.__class__.__hash__ is None:

--> 815     raise ValueError(f'mutable default {type(f.default)} for field '

    816                      f'{f.name} is not allowed: use default_factory')

    818 return f

ValueError: mutable default <class 'fairseq.dataclass.configs.CommonConfig'> for field common is not allowed: use default_factory

 这里是因为fairseq很久没有更新，在python3.8之后的版本中这个默认值用不了

fairseq在py3.9以后版本上的安装 - 酱_油 - 博客园

我创建了一个虚拟环境，想再试一下能不能行，然而。。。。。

第3-N次报错

全部都是包不兼容的问题。python之间的包互相依赖的关系很严重，我第一次感受到hh。

并且不要安装python3.11。（一年前的自己你听到了吗？！！！！）如果在3.8-3.10的版本应该是不会出错的（崩溃。。）

ValueError: mutable default <class 'fairseq.dataclass.configs.CommonConfig'> for field common is not allowed: use default_factory · Issue #12 · RVC-Project/Retrieval-based-Voice-Conversion

github上也有一些解答，我试了还是不行。

pip uninstall fairseq
pip install git+https://github.com/Tps-F/fairseq.git@main

第N+1次报错： 

这里貌似是pip+git的方式不稳定造成的。 这里我觉得应该能成功，又参考了pip install git+https:xxxx安装失败解决方法_pip install git+报错-CSDN博客

我git clone的是这个，这里的fairseq应该是改好了的One-sixth/fairseq: Facebook AI Research Sequence-to-Sequence Toolkit written in Python. 

我运行以下命令又遇到vc没有安装的错误：

pip install --editable ./

第N+2次报错： 

 安装 Microsoft Visual C++ Build Tools-CSDN博客

又重新安装hhh（苦笑

啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊这是成功了吗？？！！！！！

有了！！待我再去试一下代码~~~~ 

不管怎么样折腾一下午是应该装好了！！参考了很多经验和帖子！！！感谢！！！！

二编 18：36

还有最后一步啊啊啊啊啊啊啊啊啊啊啊要配置环境变量

在anaconda安装fairseq，并在Jupyter notebook使用_miniconda 安装 fairseq-CSDN博客

请看这篇！最后一步啊啊啊啊啊啊啊感恩！！！！ 

可以在jupyter notebook上运行了哇啊啊啊啊啊啊啊啊啊啊好开心！！！！