feat(mge/amp): add convert_format module for NHWC training

GitOrigin-RevId: 1b41e1042c

imperative/python/megengine/amp/__init__.py

@@ -2,6 +2,7 @@ import mprop
 
 from ..core.tensor.amp import *
 from .autocast import autocast
+from .convert_format import convert_module_format, convert_tensor_format
 from .grad_scaler import GradScaler
 
 mprop.init()

imperative/python/megengine/amp/autocast.py

@@ -1,5 +1,6 @@
 import functools
 
+from ..core import _config
 from ..core.tensor import amp
 
 
@@ -50,24 +51,37 @@ class autocast:
         self._origin_high = None
         self._origin_low = None
 
+        self._origin_configs = None
+
     def __enter__(self):
         self._origin_enabled = amp._enabled
-        self._origin_high = amp._get_amp_high_prec_dtype()
-        self._origin_low = amp._get_amp_low_prec_dtype()
         amp._enabled = self.enabled
         amp._set_amp_dtype_autocast(self.enabled)
+        if not self.enabled:
+            return
+
+        self._origin_high = amp._get_amp_high_prec_dtype()
+        self._origin_low = amp._get_amp_low_prec_dtype()
         amp._set_amp_high_prec_dtype(self.high_prec_dtype)
         amp._set_amp_low_prec_dtype(self.low_prec_dtype)
 
+        self._origin_configs = _config._reset_execution_config(compute_mode="float32")
+
     def __exit__(self, *args):
         amp._enabled = self._origin_enabled
         amp._set_amp_dtype_autocast(self._origin_enabled)
+        if not self.enabled:
+            return
         amp._set_amp_high_prec_dtype(self._origin_high)
         amp._set_amp_low_prec_dtype(self._origin_low)
 
+        _config._reset_execution_config(*self._origin_configs)
+
     def __call__(self, func):
         @functools.wraps(func)
         def wrapper(*args, **kwargs):
+            if not self.enabled:
+                return func(*args, **kwargs)
             with self:
                 return func(*args, **kwargs)
 

imperative/python/megengine/amp/convert_format.py

@@ -0,0 +1,45 @@
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+from copy import deepcopy
+
+from .. import functional as F
+from ..module import Module
+from ..tensor import Tensor
+
+
+def _is_nchw_format(param: Tensor):
+    # TODO: use better condition
+    return (len(param.shape) == 4 or len(param.shape) == 5) and param.format != "nhwc"
+
+
+def convert_tensor_format(x: Tensor, inplace: bool = True):
+    """Convert NCHW Tensor to NHWC Tensor."""
+    if x.ndim == 4:
+        pattern = (0, 2, 3, 1)
+    elif x.ndim == 5:
+        pattern = (0, 1, 3, 4, 2)
+    else:
+        raise ValueError("Unsupport tensor ndim {}".format(x.ndim))
+    # TODO: use initialization from tensor after fixing format setting
+    if inplace:
+        x[...] = Tensor(x.numpy().transpose(*pattern), format="nhwc")
+    else:
+        x = Tensor(x.numpy().transpose(*pattern), format="nhwc")
+    return x
+
+
+def convert_module_format(module: Module, inplace: bool = True):
+    """Convert NCHW Module to NHWC Module."""
+    if not inplace:
+        module = deepcopy(module)
+
+    for name, param in module.named_tensors():
+        if _is_nchw_format(param):
+            # hostvalue should still be valid, so no d2h cost.
+            convert_tensor_format(param, inplace=True)
+    return module

imperative/python/megengine/autodiff/grad_manager.py

@@ -1,7 +1,13 @@
 import weakref
 from typing import Callable, Iterable, List, Union
 
-from ..core._imperative_rt.core2 import pop_scope, push_scope, set_option
+from ..core._imperative_rt.core2 import (
+    get_auto_format_convert,
+    pop_scope,
+    push_scope,
+    set_auto_format_convert,
+    set_option,
+)
 from ..core.autodiff.grad import Grad
 from ..core.tensor.dtype import is_differentible_dtype
 from ..logger import get_logger
@@ -253,6 +259,8 @@ class GradManager:
         """
         push_scope("backward")
         set_option("record_computing_path", 0)
+        _origin_auto_format = get_auto_format_convert()
+        set_auto_format_convert(False)
         from ..functional import ones_like
 
         global backwarding_grad_manager
@@ -296,6 +304,7 @@ class GradManager:
             self.release()
             backwarding_grad_manager = cache
             set_option("record_computing_path", 1)
+            set_auto_format_convert(_origin_auto_format)
             pop_scope("backward")
 
     def record(self):

imperative/python/megengine/core/_config.py

@@ -10,8 +10,10 @@ from ._imperative_rt.core2 import (
     set_option,
 )
 
+# use "default" to distinguish it from None in _reset_execution_config
 __compute_mode = "default"
 __conv_format = "default"
+__bn_format = "default"
 _benchmark_kernel = False
 _deterministic_kernel = False
 
@@ -22,6 +24,8 @@ __all__ = [
     "disable_memory_forwarding",
     "_compute_mode",
     "_conv_format",
+    "_bn_format",
+    "_auto_format_convert",
     "_override",
 ]
 
@@ -32,6 +36,7 @@ def benchmark_kernel(mod):
     which means use heuristic to choose the fastest algorithm.
 
     Examples:
+
         .. code-block::
 
            import megengine as mge
@@ -55,6 +60,7 @@ def deterministic_kernel(mod):
     which means the algorithm is not reproducible.
 
     Examples:
+
         .. code-block::
 
            import megengine as mge
@@ -75,6 +81,7 @@ def async_level(mod) -> int:
     which means both device and user side errors are async.
 
     Examples:
+
         .. code-block::
 
            import megengine as mge
@@ -110,16 +117,17 @@ def disable_memory_forwarding(mod, disable: bool):
 
 @property
 def _compute_mode(mod):
-    r"""Get or set the precision of intermediate results. The default option is "default",
-    which means that no special requirements will be placed on.  When set to 'float32', it
-    would be used for accumulator and intermediate result, but only effective when input and 
+    r"""Get or set the precision of intermediate results for conv, matmul. The default
+    option is None and will fallback to "default". When set to "float32", it will
+    trigger mixed precision computation on TensorCore, but only effective when input and
     output are of float16 dtype.
 
     Examples:
+
         .. code-block::
 
            import megengine as mge
-           mge.config._compute_mode = "default"
+           mge.config._compute_mode = "float32"
     """
     return __compute_mode
 
@@ -132,7 +140,7 @@ def _compute_mode(mod, _compute_mode: str):
 
 @property
 def _conv_format(mod):
-    r"""Get or set convolution data/filter/output layout format. The default option is "default",
+    r"""Get or set convolution data/filter/output layout format. The default option is None,
     which means that no special format will be placed on. There are all layout definitions
 
     ``NCHW`` layout: ``{N, C, H, W}``
@@ -145,6 +153,7 @@ def _conv_format(mod):
     ``NCHW64`` layout: ``{N, C/64, H, W, 64}``
 
     Examples:
+
         .. code-block::
 
            import megengine as mge
@@ -159,12 +168,35 @@ def _conv_format(mod, format: str):
     __conv_format = format
 
 
+@property
+def _bn_format(mod):
+    r"""Get or set batchnorm param layout format. The default option is None and will
+    fallback to "dim_1c11" which corresponds to NCHW format. When set to "dim_111c",
+    param format of batchnorm will be changed to NHWC.
+
+    Examples:
+
+        .. code-block::
+
+           import megengine as mge
+           mge.config._bn_format = "dim_111c"
+    """
+    return __bn_format
+
+
+@_bn_format.setter
+def _bn_format(mod, format: str):
+    global __bn_format
+    __bn_format = format
+
+
 @property
 def _auto_format_convert(mod):
     r"""Automatically convert indexing params' order for NCHW Tensor to NHWC order.
     The default value is False, which means no convert.
 
     Examples:
+
         .. code-block::
 
            import megengine as mge
@@ -184,15 +216,17 @@ def _reset_execution_config(
     async_level=None,
     compute_mode=None,
     conv_format=None,
+    bn_format=None,
     auto_format_convert=None,
 ):
-    global _benchmark_kernel, _deterministic_kernel, __compute_mode, __conv_format
+    global _benchmark_kernel, _deterministic_kernel, __compute_mode, __conv_format, __bn_format
     orig_flags = (
         _benchmark_kernel,
         _deterministic_kernel,
         get_option("async_level"),
         __compute_mode,
         __conv_format,
+        __bn_format,
         get_auto_format_convert(),
     )
     if benchmark_kernel is not None:
@@ -205,6 +239,8 @@ def _reset_execution_config(
         __compute_mode = compute_mode
     if conv_format is not None:
         __conv_format = conv_format
+    if bn_format is not None:
+        __bn_format = bn_format
     if auto_format_convert is not None:
         set_auto_format_convert(auto_format_convert)
 
@@ -218,12 +254,14 @@ def _override(
     async_level=None,
     compute_mode=None,
     conv_format=None,
+    bn_format=None,
     auto_format_convert=None,
 ):
     r"""A context manager that users can opt in by attaching the decorator to set 
     the config of the global variable.
 
     Examples:
+
         .. code-block::
 
            import megengine as mge
@@ -234,6 +272,7 @@ def _override(
                 async_level=2,
                 compute_mode="float32",
                 conv_format="NHWC",
+                bn_format="dim_111c",
                 auto_format_convert=True,
             )
            def train():
@@ -244,6 +283,7 @@ def _override(
         async_level,
         compute_mode,
         conv_format,
+        bn_format,
         auto_format_convert,
     )
     try:
@@ -254,4 +294,4 @@ def _override(
 
 
 def _get_actual_op_param(function_param, config_param):
-    return function_param if config_param == "default" else config_param
+    return function_param if config_param is "default" else config_param

imperative/python/megengine/core/tensor/amp.py

@@ -10,13 +10,19 @@ from .._imperative_rt.core2 import (
 _enabled = False
 _set_amp_dtype_autocast(_enabled)
 
+__all__ = [
+    "enabled",
+    "high_prec_dtype",
+    "low_prec_dtype",
+]
+
 
 @property
 def enabled(mod):
     r"""Get or set amp autocast mode enabled or not.
-    
+
     Examples:
-    
+
         .. code-block::
 
            import megengine as mge
@@ -36,9 +42,9 @@ def enabled(mod, enabled: bool):
 def high_prec_dtype(mod):
     r"""Get or set amp autocast mode's higher precision dtype. It will change the
     target dtype in tensor casting for better precision. Default: float32.
-    
+
     Examples:
-    
+
         .. code-block::
 
            import megengine as mge
@@ -56,9 +62,9 @@ def high_prec_dtype(mod, dtype: str):
 def low_prec_dtype(mod):
     r"""Get or set amp autocast mode's lower precision dtype. It will change the
     target dtype in tensor casting for better speed and memory. Default: float16.
-    
+
     Examples:
-    
+
         .. code-block::
 
            import megengine as mge

imperative/python/megengine/core/tensor/array_method.py

@@ -63,6 +63,7 @@ def _matmul(
     assert dim1 > 0 and dim2 > 0
     maxdim = dim1 if dim1 > dim2 else dim2
     compute_mode = _config._get_actual_op_param(compute_mode, _config.__compute_mode)
+
     if dim1 == 1 and dim2 == 1:  # dispatch to Dot
         (result,) = apply(builtin.Dot(), inp1, inp2)
         return result

imperative/python/megengine/functional/nn.py

@@ -441,7 +441,6 @@ def deformable_conv2d(
         or conv_mode.name == "CROSS_CORRELATION"
     )
     if amp._enabled:
-        compute_mode = "float32"
         inp, weight, offset, mask, bias = cast_tensors(inp, weight, offset, mask, bias)
     else:
         offset = offset.astype("float32")
@@ -1182,7 +1181,6 @@ def batch_norm(
     momentum: float = 0.9,
     eps: float = 1e-5,
     inplace: bool = True,
-    compute_mode="default",
     param_dim="dim_1c11"
 ):
     r"""Applies batch normalization to the input.

imperative/python/megengine/module/batchnorm.py

@@ -19,7 +19,6 @@ class _BatchNorm(Module):
         affine=True,
         track_running_stats=True,
         freeze=False,
-        compute_mode="default",
         param_dim="dim_1c11",
         **kwargs
     ):
@@ -31,7 +30,6 @@ class _BatchNorm(Module):
         self.track_running_stats = track_running_stats
         self._track_running_stats_saved = track_running_stats
         self.freeze = freeze
-        self.compute_mode = compute_mode
         self.param_dim = param_dim
         if self.freeze:
             assert (
@@ -106,7 +104,6 @@ class _BatchNorm(Module):
             or ((self.running_mean is None) and (self.running_var is None)),
             momentum=exponential_average_factor,
             eps=self.eps,
-            compute_mode=self.compute_mode,
             param_dim=self.param_dim,
         )
 

imperative/python/megengine/optimizer/optimizer.py

@@ -8,7 +8,13 @@ from typing import Union
 
 import numpy as np
 
-from ..core._imperative_rt.core2 import pop_scope, push_scope, set_option
+from ..core._imperative_rt.core2 import (
+    get_auto_format_convert,
+    pop_scope,
+    push_scope,
+    set_auto_format_convert,
+    set_option,
+)
 from ..core.tensor.utils import set_convert_inputs
 from ..tensor import Parameter, Tensor
 from ..utils.deprecation import deprecated
@@ -90,7 +96,7 @@ class Optimizer(metaclass=ABCMeta):
                     "optimizer can only optimize Parameters, but one of the params is "
                     + str(type(param))
                 )
-            param._reset(Tensor(param.numpy(), no_cache=True))
+            param._reset(Tensor(param.numpy(), no_cache=True, format=param.format))
 
         for name, default in self._defaults.items():
             if default is required and name not in param_group:
@@ -139,6 +145,8 @@ class Optimizer(metaclass=ABCMeta):
         # set the globle state `_enable_convert_inputs` to `False` to disable
         # the `convert_inputs` for param updates
         set_option("record_computing_path", 0)
+        _origin_auto_format = get_auto_format_convert()
+        set_auto_format_convert(False)
         if self._disable_type_convert:
             backup = set_convert_inputs(False)
         for group in self.param_groups:
@@ -155,6 +163,7 @@ class Optimizer(metaclass=ABCMeta):
             # restore the globle state `_enable_convert_inputs`
             set_convert_inputs(backup)
         set_option("record_computing_path", 1)
+        set_auto_format_convert(_origin_auto_format)
         return self
 
     @deprecated(version="1.0", reason="use clear_grad instead")

imperative/python/test/unit/amp/test_convert_format.py

@@ -0,0 +1,44 @@
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+import numpy as np
+import pytest
+
+import megengine.functional as F
+import megengine.module as M
+from megengine import Parameter, Tensor, amp, tensor
+
+
+class MyModule(M.Module):
+    class InnerModule(M.Module):
+        def __init__(self):
+            super().__init__()
+            self.bn = M.BatchNorm2d(4)
+
+        def forward(self, x):
+            return self.bn(x)
+
+    def __init__(self):
+        super().__init__()
+        self.i = self.InnerModule()
+        self.conv = M.Conv2d(4, 4, 4, groups=2)
+        self.bn = M.BatchNorm2d(4)
+        self.param = Parameter(np.ones((1, 3, 1, 1), dtype=np.float32))
+        self.buff = Tensor(np.ones((1, 3, 1, 1), dtype=np.float32))
+
+    def forward(self, x):
+        x = self.i(x)
+        x = self.bn(x)
+        return x
+
+
+@pytest.mark.parametrize("is_inplace", [False, True])
+def test_convert_module(is_inplace):
+    m = MyModule()
+    m = amp.convert_module_format(m, is_inplace)
+    for name, param in m.named_tensors():
+        assert param.format == "nhwc"

imperative/python/test/unit/core/test_formatted_tensor.py

@@ -8,14 +8,27 @@ from megengine.autodiff import GradManager
 
 
 def test_basic():
-    a = tensor(np.arange(0, 24).reshape((1, 2, 3, 4)), dtype="float32", format="nhwc")
+    data = np.arange(0, 24).reshape((1, 2, 3, 4))
+    # init from numpy
+    a = tensor(data, format="nhwc")
     assert a.format == "nhwc"
+
+    # init from tensor
     b = tensor(a)
     assert b.format == "nhwc"
-    # TODO: fix Tensor init bug for another Tensor
+
+    # TODO: init from tensor with new format
     # c = tensor(a, format="nchw")
     # assert c.format == "nchw"
 
+    # TODO: reset from numpy
+    # b[...] = data
+    # assert b.format == "nhwc"
+
+    # reset from tensor
+    b[...] = tensor(data, format="nchw")
+    assert b.format == "nchw"
+
 
 def _compare_nchw_nhwc(data, func):
     x1 = tensor(data, format="nchw")
@@ -23,7 +36,7 @@ def _compare_nchw_nhwc(data, func):
     out1 = func(x1)
     with mge.config._override(auto_format_convert=True):
         out2 = func(x2)
-    np.testing.assert_equal(out1, out2)
+    np.testing.assert_almost_equal(out1, out2, decimal=5)
 
 
 def test_dimshuffle():
@@ -296,8 +309,10 @@ def test_backward():
     with gm:
         with mge.config._override(auto_format_convert=True, conv_format="NHWC"):
             x = F.conv2d(x, w, b)
+            # TODO: fix manually convert to NHWC, usually used in detection head
+            # x = x.transpose(0, 2, 3, 1).reshape(1, 18, 2)
             gm.backward(x)
-            # TODO: backward grad has no format yet
+            # backward grad has no format
             np.testing.assert_equal(
                 w.grad.numpy(),
                 np.array([66, 210, 66, 210, 66, 210]).reshape((3, 1, 1, 2)),

imperative/python/test/unit/functional/test_functional.py

@@ -921,12 +921,7 @@ def test_batchnorm2d_autocast():
 
     amp.enabled = False
     expected = F.batch_norm(
-        inp.astype("float16"),
-        weight=weight,
-        bias=bias,
-        training=True,
-        inplace=False,
-        compute_mode="float32",
+        inp.astype("float16"), weight=weight, bias=bias, training=True, inplace=False,
     )
     assert out.dtype == np.float16
     assert expected.dtype == np.float16