Deprecate outdated features (NVIDIA#1506)

* add warnings

update README

verify deprecated warning

deprecate apex.RNN which is never a thing

Signed-off-by: Masaki Kozuki <[email protected]>

README.md

@@ -12,6 +12,8 @@ The intent of Apex is to make up-to-date utilities available to users as quickly
 
 ## 1. Amp:  Automatic Mixed Precision
 
+**Deprecated. Use [PyTorch AMP](https://pytorch.org/docs/stable/amp.html)**
+
 `apex.amp` is a tool to enable mixed precision training by changing only 3 lines of your script.
 Users can easily experiment with different pure and mixed precision training modes by supplying
 different flags to `amp.initialize`.
@@ -29,6 +31,8 @@ different flags to `amp.initialize`.
 
 ## 2. Distributed Training
 
+**`apex.parallel.DistributedDataParallel` is deprecated. Use [`torch.nn.parallel.DistributedDataParallel`](https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html?highlight=distributeddataparallel#torch.nn.parallel.DistributedDataParallel)**
+
 `apex.parallel.DistributedDataParallel` is a module wrapper, similar to
 `torch.nn.parallel.DistributedDataParallel`.  It enables convenient multiprocess distributed training,
 optimized for NVIDIA's NCCL communication library.
@@ -44,6 +48,8 @@ shows use of `apex.parallel.DistributedDataParallel` along with `apex.amp`.
 
 ### Synchronized Batch Normalization
 
+**Deprecated. Use [`torch.nn.SyncBatchNorm`](https://pytorch.org/docs/stable/generated/torch.nn.SyncBatchNorm.html)**
+
 `apex.parallel.SyncBatchNorm` extends `torch.nn.modules.batchnorm._BatchNorm` to
 support synchronized BN.
 It allreduces stats across processes during multiprocess (DistributedDataParallel) training.

apex/RNN/README.md

@@ -1 +1,3 @@
+**This module will be removed by the end of February 2023**
+
 Under construction...

apex/RNN/models.py

@@ -2,6 +2,7 @@
 
 from torch.nn._functions.rnn import LSTMCell, RNNReLUCell, RNNTanhCell, GRUCell
 
+from apex import deprecated_warning
 from .RNNBackend import bidirectionalRNN, stackedRNN, RNNCell
 from .cells import mLSTMRNNCell, mLSTMCell
 
@@ -10,6 +11,7 @@ def toRNNBackend(inputRNN, num_layers, bidirectional=False, dropout = 0):
     :class:`toRNNBackend`
     """
 
+    deprecated_warning("`apex.RNN` is deprecated and will be removed by the end of February 2023.")
     if bidirectional:
         return bidirectionalRNN(inputRNN, num_layers, dropout = dropout)
     else:
@@ -43,7 +45,7 @@ def Tanh(input_size, hidden_size, num_layers, bias=True, batch_first=False, drop
     """
     inputRNN = RNNCell(1, input_size, hidden_size, RNNTanhCell, 1, bias, output_size)
     return toRNNBackend(inputRNN, num_layers, bidirectional, dropout=dropout)
-        
+
 def mLSTM(input_size, hidden_size, num_layers, bias=True, batch_first=False, dropout=0, bidirectional=False, output_size = None):
     """
     :class:`mLSTM`

apex/__init__.py

@@ -5,8 +5,12 @@
 import torch
 
 
+__all__ = ["amp", "fp16_utils", "optimizers", "normalization", "transformer"]
+
+
 if torch.distributed.is_available():
     from . import parallel
+    __all__.append("parallel")
 
 from . import amp
 from . import fp16_utils
@@ -49,3 +53,16 @@ def check_cudnn_version_and_warn(global_option: str, required_cudnn_version: int
         )
         return False
     return True
+
+
+class DeprecatedFeatureWarning(FutureWarning):
+    pass
+
+
+def deprecated_warning(msg: str) -> None:
+    if (
+        not torch.distributed.is_available
+        or not torch.distributed.is_initialized()
+        or (torch.distributed.is_initialized() and torch.distributed.get_rank() == 0)
+    ):
+        warnings.warn(msg, DeprecatedFeatureWarning)

apex/amp/amp.py

@@ -1,14 +1,14 @@
+import functools
+import itertools
+
+import torch
+
 from . import compat, rnn_compat, utils, wrap
 from .handle import AmpHandle, NoOpHandle
 from .lists import functional_overrides, torch_overrides, tensor_overrides
 from ._amp_state import _amp_state
 from .frontend import *
 
-import functools
-import itertools
-
-import torch
-
 
 _DECORATOR_HANDLE = None
 _USER_CAST_REGISTRY = set()
@@ -28,16 +28,22 @@ def wrapper(*args, **kwargs):
 
 # Decorator form
 def half_function(fn):
+    from apex import deprecated_warning
+    deprecated_warning("apex.amp is deprecated and will be removed by the end of February 2023. Use [PyTorch AMP](https://pytorch.org/docs/stable/amp.html)")
     wrap_fn = functools.partial(wrap.make_cast_wrapper, try_caching=True)
     return _decorator_helper(fn, utils.maybe_half, wrap_fn)
 
 
 def float_function(fn):
+    from apex import deprecated_warning
+    deprecated_warning("apex.amp is deprecated and will be removed by the end of February 2023. Use [PyTorch AMP](https://pytorch.org/docs/stable/amp.html)")
     wrap_fn = functools.partial(wrap.make_cast_wrapper, try_caching=False)
     return _decorator_helper(fn, utils.maybe_float, wrap_fn)
 
 
 def promote_function(fn):
+    from apex import deprecated_warning
+    deprecated_warning("apex.amp is deprecated and will be removed by the end of February 2023. Use [PyTorch AMP](https://pytorch.org/docs/stable/amp.html)")
     wrap_fn = functools.partial(wrap.make_promote_wrapper)
     return _decorator_helper(fn, utils.maybe_float, wrap_fn)
 

apex/amp/frontend.py

@@ -1,7 +1,9 @@
+from collections import OrderedDict
+
 import torch
+
 from ._initialize import _initialize
 from ._amp_state import _amp_state, warn_or_err, maybe_print
-from collections import OrderedDict
 
 
 class Properties(object):
@@ -305,6 +307,8 @@ def initialize(
     .. _`let us know`:
         https://github.com/NVIDIA/apex/issues
     """
+    from apex import deprecated_warning
+    deprecated_warning("apex.amp is deprecated and will be removed by the end of February 2023. Use [PyTorch AMP](https://pytorch.org/docs/stable/amp.html)")
     _amp_state.opt_properties = Properties()
     _amp_state.verbosity = verbosity
 
@@ -377,7 +381,7 @@ def load_state_dict(state_dict):
             len(state_dict), len(_amp_state.loss_scalers)))
 
     state_dict = state_dict.copy()
-    
+
     nb_loss_scalers = len(_amp_state.loss_scalers)
     unexpected_keys = []
     # Initialize idx outside, since unexpected_keys will increase it if enumerate is used