You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
478 lines
19 KiB
478 lines
19 KiB
4 years ago
|
""" EfficientNet, MobileNetV3, etc Builder
|
||
|
|
||
|
Assembles EfficieNet and related network feature blocks from string definitions.
|
||
|
Handles stride, dilation calculations, and selects feature extraction points.
|
||
|
|
||
3 years ago
|
Hacked together by / Copyright 2019, Ross Wightman
|
||
4 years ago
|
"""
|
||
|
|
||
5 years ago
|
import logging
|
||
|
import math
|
||
|
import re
|
||
|
from copy import deepcopy
|
||
4 years ago
|
from functools import partial
|
||
5 years ago
|
|
||
|
import torch.nn as nn
|
||
5 years ago
|
|
||
2 years ago
|
from ._efficientnet_blocks import *
|
||
|
from timm.layers import CondConv2d, get_condconv_initializer, get_act_layer, get_attn, make_divisible
|
||
5 years ago
|
|
||
4 years ago
|
__all__ = ["EfficientNetBuilder", "decode_arch_def", "efficientnet_init_weights",
|
||
|
'resolve_bn_args', 'resolve_act_layer', 'round_channels', 'BN_MOMENTUM_TF_DEFAULT', 'BN_EPS_TF_DEFAULT']
|
||
5 years ago
|
|
||
4 years ago
|
_logger = logging.getLogger(__name__)
|
||
5 years ago
|
|
||
|
|
||
4 years ago
|
_DEBUG_BUILDER = False
|
||
|
|
||
|
# Defaults used for Google/Tensorflow training of mobile networks /w RMSprop as per
|
||
|
# papers and TF reference implementations. PT momentum equiv for TF decay is (1 - TF decay)
|
||
|
# NOTE: momentum varies btw .99 and .9997 depending on source
|
||
|
# .99 in official TF TPU impl
|
||
|
# .9997 (/w .999 in search space) for paper
|
||
|
BN_MOMENTUM_TF_DEFAULT = 1 - 0.99
|
||
|
BN_EPS_TF_DEFAULT = 1e-3
|
||
|
_BN_ARGS_TF = dict(momentum=BN_MOMENTUM_TF_DEFAULT, eps=BN_EPS_TF_DEFAULT)
|
||
|
|
||
|
|
||
|
def get_bn_args_tf():
|
||
|
return _BN_ARGS_TF.copy()
|
||
|
|
||
|
|
||
|
def resolve_bn_args(kwargs):
|
||
3 years ago
|
bn_args = {}
|
||
4 years ago
|
bn_momentum = kwargs.pop('bn_momentum', None)
|
||
|
if bn_momentum is not None:
|
||
|
bn_args['momentum'] = bn_momentum
|
||
|
bn_eps = kwargs.pop('bn_eps', None)
|
||
|
if bn_eps is not None:
|
||
|
bn_args['eps'] = bn_eps
|
||
|
return bn_args
|
||
|
|
||
|
|
||
|
def resolve_act_layer(kwargs, default='relu'):
|
||
4 years ago
|
return get_act_layer(kwargs.pop('act_layer', default))
|
||
4 years ago
|
|
||
|
|
||
|
def round_channels(channels, multiplier=1.0, divisor=8, channel_min=None, round_limit=0.9):
|
||
|
"""Round number of filters based on depth multiplier."""
|
||
|
if not multiplier:
|
||
|
return channels
|
||
|
return make_divisible(channels * multiplier, divisor, channel_min, round_limit=round_limit)
|
||
|
|
||
|
|
||
4 years ago
|
def _log_info_if(msg, condition):
|
||
|
if condition:
|
||
4 years ago
|
_logger.info(msg)
|
||
4 years ago
|
|
||
|
|
||
5 years ago
|
def _parse_ksize(ss):
|
||
|
if ss.isdigit():
|
||
|
return int(ss)
|
||
|
else:
|
||
|
return [int(k) for k in ss.split('.')]
|
||
|
|
||
|
|
||
|
def _decode_block_str(block_str):
|
||
|
""" Decode block definition string
|
||
|
|
||
|
Gets a list of block arg (dicts) through a string notation of arguments.
|
||
|
E.g. ir_r2_k3_s2_e1_i32_o16_se0.25_noskip
|
||
|
|
||
|
All args can exist in any order with the exception of the leading string which
|
||
|
is assumed to indicate the block type.
|
||
|
|
||
|
leading string - block type (
|
||
|
ir = InvertedResidual, ds = DepthwiseSep, dsa = DeptwhiseSep with pw act, cn = ConvBnAct)
|
||
|
r - number of repeat blocks,
|
||
|
k - kernel size,
|
||
|
s - strides (1-9),
|
||
|
e - expansion ratio,
|
||
|
c - output channels,
|
||
|
se - squeeze/excitation ratio
|
||
|
n - activation fn ('re', 'r6', 'hs', or 'sw')
|
||
|
Args:
|
||
|
block_str: a string representation of block arguments.
|
||
|
Returns:
|
||
|
A list of block args (dicts)
|
||
|
Raises:
|
||
|
ValueError: if the string def not properly specified (TODO)
|
||
|
"""
|
||
|
assert isinstance(block_str, str)
|
||
|
ops = block_str.split('_')
|
||
|
block_type = ops[0] # take the block type off the front
|
||
|
ops = ops[1:]
|
||
|
options = {}
|
||
4 years ago
|
skip = None
|
||
5 years ago
|
for op in ops:
|
||
|
# string options being checked on individual basis, combine if they grow
|
||
|
if op == 'noskip':
|
||
4 years ago
|
skip = False # force no skip connection
|
||
|
elif op == 'skip':
|
||
|
skip = True # force a skip connection
|
||
5 years ago
|
elif op.startswith('n'):
|
||
|
# activation fn
|
||
|
key = op[0]
|
||
|
v = op[1:]
|
||
|
if v == 're':
|
||
5 years ago
|
value = get_act_layer('relu')
|
||
5 years ago
|
elif v == 'r6':
|
||
5 years ago
|
value = get_act_layer('relu6')
|
||
5 years ago
|
elif v == 'hs':
|
||
5 years ago
|
value = get_act_layer('hard_swish')
|
||
5 years ago
|
elif v == 'sw':
|
||
4 years ago
|
value = get_act_layer('swish') # aka SiLU
|
||
|
elif v == 'mi':
|
||
|
value = get_act_layer('mish')
|
||
5 years ago
|
else:
|
||
|
continue
|
||
|
options[key] = value
|
||
|
else:
|
||
|
# all numeric options
|
||
|
splits = re.split(r'(\d.*)', op)
|
||
|
if len(splits) >= 2:
|
||
|
key, value = splits[:2]
|
||
|
options[key] = value
|
||
|
|
||
|
# if act_layer is None, the model default (passed to model init) will be used
|
||
|
act_layer = options['n'] if 'n' in options else None
|
||
|
exp_kernel_size = _parse_ksize(options['a']) if 'a' in options else 1
|
||
|
pw_kernel_size = _parse_ksize(options['p']) if 'p' in options else 1
|
||
4 years ago
|
force_in_chs = int(options['fc']) if 'fc' in options else 0 # FIXME hack to deal with in_chs issue in TPU def
|
||
5 years ago
|
num_repeat = int(options['r'])
|
||
3 years ago
|
|
||
5 years ago
|
# each type of block has different valid arguments, fill accordingly
|
||
3 years ago
|
block_args = dict(
|
||
|
block_type=block_type,
|
||
|
out_chs=int(options['c']),
|
||
|
stride=int(options['s']),
|
||
|
act_layer=act_layer,
|
||
|
)
|
||
5 years ago
|
if block_type == 'ir':
|
||
3 years ago
|
block_args.update(dict(
|
||
5 years ago
|
dw_kernel_size=_parse_ksize(options['k']),
|
||
|
exp_kernel_size=exp_kernel_size,
|
||
|
pw_kernel_size=pw_kernel_size,
|
||
|
exp_ratio=float(options['e']),
|
||
4 years ago
|
se_ratio=float(options['se']) if 'se' in options else 0.,
|
||
|
noskip=skip is False,
|
||
3 years ago
|
))
|
||
5 years ago
|
if 'cc' in options:
|
||
|
block_args['num_experts'] = int(options['cc'])
|
||
|
elif block_type == 'ds' or block_type == 'dsa':
|
||
3 years ago
|
block_args.update(dict(
|
||
5 years ago
|
dw_kernel_size=_parse_ksize(options['k']),
|
||
|
pw_kernel_size=pw_kernel_size,
|
||
4 years ago
|
se_ratio=float(options['se']) if 'se' in options else 0.,
|
||
5 years ago
|
pw_act=block_type == 'dsa',
|
||
4 years ago
|
noskip=block_type == 'dsa' or skip is False,
|
||
3 years ago
|
))
|
||
5 years ago
|
elif block_type == 'er':
|
||
3 years ago
|
block_args.update(dict(
|
||
5 years ago
|
exp_kernel_size=_parse_ksize(options['k']),
|
||
|
pw_kernel_size=pw_kernel_size,
|
||
|
exp_ratio=float(options['e']),
|
||
4 years ago
|
force_in_chs=force_in_chs,
|
||
|
se_ratio=float(options['se']) if 'se' in options else 0.,
|
||
|
noskip=skip is False,
|
||
3 years ago
|
))
|
||
5 years ago
|
elif block_type == 'cn':
|
||
3 years ago
|
block_args.update(dict(
|
||
5 years ago
|
kernel_size=int(options['k']),
|
||
4 years ago
|
skip=skip is True,
|
||
3 years ago
|
))
|
||
5 years ago
|
else:
|
||
|
assert False, 'Unknown block type (%s)' % block_type
|
||
3 years ago
|
if 'gs' in options:
|
||
|
block_args['group_size'] = options['gs']
|
||
5 years ago
|
|
||
|
return block_args, num_repeat
|
||
|
|
||
|
|
||
|
def _scale_stage_depth(stack_args, repeats, depth_multiplier=1.0, depth_trunc='ceil'):
|
||
|
""" Per-stage depth scaling
|
||
|
Scales the block repeats in each stage. This depth scaling impl maintains
|
||
|
compatibility with the EfficientNet scaling method, while allowing sensible
|
||
|
scaling for other models that may have multiple block arg definitions in each stage.
|
||
|
"""
|
||
|
|
||
|
# We scale the total repeat count for each stage, there may be multiple
|
||
|
# block arg defs per stage so we need to sum.
|
||
|
num_repeat = sum(repeats)
|
||
|
if depth_trunc == 'round':
|
||
|
# Truncating to int by rounding allows stages with few repeats to remain
|
||
|
# proportionally smaller for longer. This is a good choice when stage definitions
|
||
|
# include single repeat stages that we'd prefer to keep that way as long as possible
|
||
|
num_repeat_scaled = max(1, round(num_repeat * depth_multiplier))
|
||
|
else:
|
||
|
# The default for EfficientNet truncates repeats to int via 'ceil'.
|
||
|
# Any multiplier > 1.0 will result in an increased depth for every stage.
|
||
|
num_repeat_scaled = int(math.ceil(num_repeat * depth_multiplier))
|
||
|
|
||
|
# Proportionally distribute repeat count scaling to each block definition in the stage.
|
||
|
# Allocation is done in reverse as it results in the first block being less likely to be scaled.
|
||
|
# The first block makes less sense to repeat in most of the arch definitions.
|
||
|
repeats_scaled = []
|
||
|
for r in repeats[::-1]:
|
||
|
rs = max(1, round((r / num_repeat * num_repeat_scaled)))
|
||
|
repeats_scaled.append(rs)
|
||
|
num_repeat -= r
|
||
|
num_repeat_scaled -= rs
|
||
|
repeats_scaled = repeats_scaled[::-1]
|
||
|
|
||
|
# Apply the calculated scaling to each block arg in the stage
|
||
|
sa_scaled = []
|
||
|
for ba, rep in zip(stack_args, repeats_scaled):
|
||
|
sa_scaled.extend([deepcopy(ba) for _ in range(rep)])
|
||
|
return sa_scaled
|
||
|
|
||
|
|
||
3 years ago
|
def decode_arch_def(
|
||
|
arch_def,
|
||
|
depth_multiplier=1.0,
|
||
|
depth_trunc='ceil',
|
||
|
experts_multiplier=1,
|
||
|
fix_first_last=False,
|
||
|
group_size=None,
|
||
|
):
|
||
|
""" Decode block architecture definition strings -> block kwargs
|
||
|
|
||
|
Args:
|
||
|
arch_def: architecture definition strings, list of list of strings
|
||
|
depth_multiplier: network depth multiplier
|
||
|
depth_trunc: networ depth truncation mode when applying multiplier
|
||
|
experts_multiplier: CondConv experts multiplier
|
||
|
fix_first_last: fix first and last block depths when multiplier is applied
|
||
|
group_size: group size override for all blocks that weren't explicitly set in arch string
|
||
|
|
||
|
Returns:
|
||
|
list of list of block kwargs
|
||
|
"""
|
||
5 years ago
|
arch_args = []
|
||
4 years ago
|
if isinstance(depth_multiplier, tuple):
|
||
|
assert len(depth_multiplier) == len(arch_def)
|
||
|
else:
|
||
|
depth_multiplier = (depth_multiplier,) * len(arch_def)
|
||
|
for stack_idx, (block_strings, multiplier) in enumerate(zip(arch_def, depth_multiplier)):
|
||
5 years ago
|
assert isinstance(block_strings, list)
|
||
|
stack_args = []
|
||
|
repeats = []
|
||
|
for block_str in block_strings:
|
||
|
assert isinstance(block_str, str)
|
||
|
ba, rep = _decode_block_str(block_str)
|
||
|
if ba.get('num_experts', 0) > 0 and experts_multiplier > 1:
|
||
|
ba['num_experts'] *= experts_multiplier
|
||
3 years ago
|
if group_size is not None:
|
||
|
ba.setdefault('group_size', group_size)
|
||
5 years ago
|
stack_args.append(ba)
|
||
|
repeats.append(rep)
|
||
5 years ago
|
if fix_first_last and (stack_idx == 0 or stack_idx == len(arch_def) - 1):
|
||
|
arch_args.append(_scale_stage_depth(stack_args, repeats, 1.0, depth_trunc))
|
||
|
else:
|
||
4 years ago
|
arch_args.append(_scale_stage_depth(stack_args, repeats, multiplier, depth_trunc))
|
||
5 years ago
|
return arch_args
|
||
|
|
||
|
|
||
|
class EfficientNetBuilder:
|
||
|
""" Build Trunk Blocks
|
||
|
|
||
|
This ended up being somewhat of a cross between
|
||
|
https://github.com/tensorflow/tpu/blob/master/models/official/mnasnet/mnasnet_models.py
|
||
|
and
|
||
|
https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/maskrcnn_benchmark/modeling/backbone/fbnet_builder.py
|
||
|
|
||
|
"""
|
||
4 years ago
|
def __init__(self, output_stride=32, pad_type='', round_chs_fn=round_channels, se_from_exp=False,
|
||
4 years ago
|
act_layer=None, norm_layer=None, se_layer=None, drop_path_rate=0., feature_location=''):
|
||
5 years ago
|
self.output_stride = output_stride
|
||
|
self.pad_type = pad_type
|
||
4 years ago
|
self.round_chs_fn = round_chs_fn
|
||
4 years ago
|
self.se_from_exp = se_from_exp # calculate se channel reduction from expanded (mid) chs
|
||
5 years ago
|
self.act_layer = act_layer
|
||
|
self.norm_layer = norm_layer
|
||
4 years ago
|
self.se_layer = get_attn(se_layer)
|
||
|
try:
|
||
4 years ago
|
self.se_layer(8, rd_ratio=1.0) # test if attn layer accepts rd_ratio arg
|
||
4 years ago
|
self.se_has_ratio = True
|
||
4 years ago
|
except TypeError:
|
||
4 years ago
|
self.se_has_ratio = False
|
||
5 years ago
|
self.drop_path_rate = drop_path_rate
|
||
4 years ago
|
if feature_location == 'depthwise':
|
||
|
# old 'depthwise' mode renamed 'expansion' to match TF impl, old expansion mode didn't make sense
|
||
4 years ago
|
_logger.warning("feature_location=='depthwise' is deprecated, using 'expansion'")
|
||
4 years ago
|
feature_location = 'expansion'
|
||
5 years ago
|
self.feature_location = feature_location
|
||
4 years ago
|
assert feature_location in ('bottleneck', 'expansion', '')
|
||
4 years ago
|
self.verbose = _DEBUG_BUILDER
|
||
5 years ago
|
|
||
|
# state updated during build, consumed by model
|
||
|
self.in_chs = None
|
||
5 years ago
|
self.features = []
|
||
5 years ago
|
|
||
|
def _make_block(self, ba, block_idx, block_count):
|
||
5 years ago
|
drop_path_rate = self.drop_path_rate * block_idx / block_count
|
||
5 years ago
|
bt = ba.pop('block_type')
|
||
|
ba['in_chs'] = self.in_chs
|
||
4 years ago
|
ba['out_chs'] = self.round_chs_fn(ba['out_chs'])
|
||
|
if 'force_in_chs' in ba and ba['force_in_chs']:
|
||
|
# NOTE this is a hack to work around mismatch in TF EdgeEffNet impl
|
||
|
ba['force_in_chs'] = self.round_chs_fn(ba['force_in_chs'])
|
||
5 years ago
|
ba['pad_type'] = self.pad_type
|
||
|
# block act fn overrides the model default
|
||
|
ba['act_layer'] = ba['act_layer'] if ba['act_layer'] is not None else self.act_layer
|
||
|
assert ba['act_layer'] is not None
|
||
4 years ago
|
ba['norm_layer'] = self.norm_layer
|
||
4 years ago
|
ba['drop_path_rate'] = drop_path_rate
|
||
4 years ago
|
if bt != 'cn':
|
||
4 years ago
|
se_ratio = ba.pop('se_ratio')
|
||
|
if se_ratio and self.se_layer is not None:
|
||
|
if not self.se_from_exp:
|
||
|
# adjust se_ratio by expansion ratio if calculating se channels from block input
|
||
|
se_ratio /= ba.get('exp_ratio', 1.0)
|
||
|
if self.se_has_ratio:
|
||
|
ba['se_layer'] = partial(self.se_layer, rd_ratio=se_ratio)
|
||
|
else:
|
||
|
ba['se_layer'] = self.se_layer
|
||
4 years ago
|
|
||
|
if bt == 'ir':
|
||
4 years ago
|
_log_info_if(' InvertedResidual {}, Args: {}'.format(block_idx, str(ba)), self.verbose)
|
||
4 years ago
|
block = CondConvResidual(**ba) if ba.get('num_experts', 0) else InvertedResidual(**ba)
|
||
5 years ago
|
elif bt == 'ds' or bt == 'dsa':
|
||
4 years ago
|
_log_info_if(' DepthwiseSeparable {}, Args: {}'.format(block_idx, str(ba)), self.verbose)
|
||
5 years ago
|
block = DepthwiseSeparableConv(**ba)
|
||
|
elif bt == 'er':
|
||
4 years ago
|
_log_info_if(' EdgeResidual {}, Args: {}'.format(block_idx, str(ba)), self.verbose)
|
||
5 years ago
|
block = EdgeResidual(**ba)
|
||
|
elif bt == 'cn':
|
||
4 years ago
|
_log_info_if(' ConvBnAct {}, Args: {}'.format(block_idx, str(ba)), self.verbose)
|
||
5 years ago
|
block = ConvBnAct(**ba)
|
||
|
else:
|
||
|
assert False, 'Uknkown block type (%s) while building model.' % bt
|
||
|
|
||
4 years ago
|
self.in_chs = ba['out_chs'] # update in_chs for arg of next block
|
||
5 years ago
|
return block
|
||
|
|
||
|
def __call__(self, in_chs, model_block_args):
|
||
|
""" Build the blocks
|
||
|
Args:
|
||
|
in_chs: Number of input-channels passed to first block
|
||
|
model_block_args: A list of lists, outer list defines stages, inner
|
||
|
list contains strings defining block configuration(s)
|
||
|
Return:
|
||
|
List of block stacks (each stack wrapped in nn.Sequential)
|
||
|
"""
|
||
4 years ago
|
_log_info_if('Building model trunk with %d stages...' % len(model_block_args), self.verbose)
|
||
5 years ago
|
self.in_chs = in_chs
|
||
|
total_block_count = sum([len(x) for x in model_block_args])
|
||
|
total_block_idx = 0
|
||
|
current_stride = 2
|
||
|
current_dilation = 1
|
||
|
stages = []
|
||
4 years ago
|
if model_block_args[0][0]['stride'] > 1:
|
||
|
# if the first block starts with a stride, we need to extract first level feat from stem
|
||
|
feature_info = dict(
|
||
|
module='act1', num_chs=in_chs, stage=0, reduction=current_stride,
|
||
|
hook_type='forward' if self.feature_location != 'bottleneck' else '')
|
||
|
self.features.append(feature_info)
|
||
|
|
||
|
# outer list of block_args defines the stacks
|
||
|
for stack_idx, stack_args in enumerate(model_block_args):
|
||
|
last_stack = stack_idx + 1 == len(model_block_args)
|
||
|
_log_info_if('Stack: {}'.format(stack_idx), self.verbose)
|
||
|
assert isinstance(stack_args, list)
|
||
5 years ago
|
|
||
|
blocks = []
|
||
4 years ago
|
# each stack (stage of blocks) contains a list of block arguments
|
||
|
for block_idx, block_args in enumerate(stack_args):
|
||
|
last_block = block_idx + 1 == len(stack_args)
|
||
|
_log_info_if(' Block: {}'.format(block_idx), self.verbose)
|
||
|
|
||
5 years ago
|
assert block_args['stride'] in (1, 2)
|
||
4 years ago
|
if block_idx >= 1: # only the first block in any stack can have a stride > 1
|
||
5 years ago
|
block_args['stride'] = 1
|
||
|
|
||
4 years ago
|
extract_features = False
|
||
|
if last_block:
|
||
|
next_stack_idx = stack_idx + 1
|
||
|
extract_features = next_stack_idx >= len(model_block_args) or \
|
||
|
model_block_args[next_stack_idx][0]['stride'] > 1
|
||
5 years ago
|
|
||
|
next_dilation = current_dilation
|
||
|
if block_args['stride'] > 1:
|
||
|
next_output_stride = current_stride * block_args['stride']
|
||
|
if next_output_stride > self.output_stride:
|
||
|
next_dilation = current_dilation * block_args['stride']
|
||
|
block_args['stride'] = 1
|
||
4 years ago
|
_log_info_if(' Converting stride to dilation to maintain output_stride=={}'.format(
|
||
|
self.output_stride), self.verbose)
|
||
5 years ago
|
else:
|
||
|
current_stride = next_output_stride
|
||
|
block_args['dilation'] = current_dilation
|
||
|
if next_dilation != current_dilation:
|
||
|
current_dilation = next_dilation
|
||
|
|
||
|
# create the block
|
||
|
block = self._make_block(block_args, total_block_idx, total_block_count)
|
||
|
blocks.append(block)
|
||
|
|
||
|
# stash feature module name and channel info for model feature extraction
|
||
|
if extract_features:
|
||
4 years ago
|
feature_info = dict(
|
||
|
stage=stack_idx + 1, reduction=current_stride, **block.feature_info(self.feature_location))
|
||
|
module_name = f'blocks.{stack_idx}.{block_idx}'
|
||
|
leaf_name = feature_info.get('module', '')
|
||
|
feature_info['module'] = '.'.join([module_name, leaf_name]) if leaf_name else module_name
|
||
5 years ago
|
self.features.append(feature_info)
|
||
5 years ago
|
|
||
|
total_block_idx += 1 # incr global block idx (across all stacks)
|
||
|
stages.append(nn.Sequential(*blocks))
|
||
|
return stages
|
||
|
|
||
|
|
||
5 years ago
|
def _init_weight_goog(m, n='', fix_group_fanout=True):
|
||
5 years ago
|
""" Weight initialization as per Tensorflow official implementations.
|
||
|
|
||
5 years ago
|
Args:
|
||
|
m (nn.Module): module to init
|
||
|
n (str): module name
|
||
5 years ago
|
fix_group_fanout (bool): enable correct (matching Tensorflow TPU impl) fanout calculation w/ group convs
|
||
5 years ago
|
|
||
5 years ago
|
Handles layers in EfficientNet, EfficientNet-CondConv, MixNet, MnasNet, MobileNetV3, etc:
|
||
|
* https://github.com/tensorflow/tpu/blob/master/models/official/mnasnet/mnasnet_model.py
|
||
|
* https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/efficientnet_model.py
|
||
|
"""
|
||
5 years ago
|
if isinstance(m, CondConv2d):
|
||
|
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
||
5 years ago
|
if fix_group_fanout:
|
||
|
fan_out //= m.groups
|
||
5 years ago
|
init_weight_fn = get_condconv_initializer(
|
||
4 years ago
|
lambda w: nn.init.normal_(w, 0, math.sqrt(2.0 / fan_out)), m.num_experts, m.weight_shape)
|
||
5 years ago
|
init_weight_fn(m.weight)
|
||
|
if m.bias is not None:
|
||
4 years ago
|
nn.init.zeros_(m.bias)
|
||
5 years ago
|
elif isinstance(m, nn.Conv2d):
|
||
|
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
||
5 years ago
|
if fix_group_fanout:
|
||
|
fan_out //= m.groups
|
||
4 years ago
|
nn.init.normal_(m.weight, 0, math.sqrt(2.0 / fan_out))
|
||
5 years ago
|
if m.bias is not None:
|
||
4 years ago
|
nn.init.zeros_(m.bias)
|
||
5 years ago
|
elif isinstance(m, nn.BatchNorm2d):
|
||
4 years ago
|
nn.init.ones_(m.weight)
|
||
|
nn.init.zeros_(m.bias)
|
||
5 years ago
|
elif isinstance(m, nn.Linear):
|
||
|
fan_out = m.weight.size(0) # fan-out
|
||
|
fan_in = 0
|
||
|
if 'routing_fn' in n:
|
||
|
fan_in = m.weight.size(1)
|
||
|
init_range = 1.0 / math.sqrt(fan_in + fan_out)
|
||
4 years ago
|
nn.init.uniform_(m.weight, -init_range, init_range)
|
||
|
nn.init.zeros_(m.bias)
|
||
5 years ago
|
|
||
|
|
||
5 years ago
|
def efficientnet_init_weights(model: nn.Module, init_fn=None):
|
||
|
init_fn = init_fn or _init_weight_goog
|
||
|
for n, m in model.named_modules():
|
||
|
init_fn(m, n)
|
||
|
|