""" EfficientNet, MobileNetV3, etc Builder Assembles EfficieNet and related network feature blocks from string definitions. Handles stride, dilation calculations, and selects feature extraction points. Hacked together by / Copyright 2019, Ross Wightman """ from typing import Callable, Optional import logging import math import re from copy import deepcopy from functools import partial from typing import Any, Dict, List import torch.nn as nn from timm.layers import CondConv2d, get_condconv_initializer, get_act_layer, get_attn, make_divisible, LayerType from ._efficientnet_blocks import * from ._manipulate import named_modules __all__ = ["EfficientNetBuilder", "BlockArgs", "decode_arch_def", "efficientnet_init_weights", 'resolve_bn_args', 'resolve_act_layer', 'round_channels', 'BN_MOMENTUM_TF_DEFAULT', 'BN_EPS_TF_DEFAULT'] _logger = logging.getLogger(__name__) _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) BlockArgs = List[List[Dict[str, Any]]] def get_bn_args_tf(): return _BN_ARGS_TF.copy() def resolve_bn_args(kwargs): bn_args = {} 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'): return get_act_layer(kwargs.pop('act_layer', default)) 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) def _log_info_if(msg, condition): if condition: _logger.info(msg) 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 = {} skip = None for op in ops: # string options being checked on individual basis, combine if they grow if op == 'noskip': skip = False # force no skip connection elif op == 'skip': skip = True # force a skip connection elif op.startswith('n'): # activation fn key = op[0] v = op[1:] if v == 're': value = get_act_layer('relu') elif v == 'r6': value = get_act_layer('relu6') elif v == 'hs': value = get_act_layer('hard_swish') elif v == 'sw': value = get_act_layer('swish') # aka SiLU elif v == 'mi': value = get_act_layer('mish') 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 start_kernel_size = _parse_ksize(options['a']) if 'a' in options else 1 end_kernel_size = _parse_ksize(options['p']) if 'p' in options else 1 force_in_chs = int(options['fc']) if 'fc' in options else 0 # FIXME hack to deal with in_chs issue in TPU def num_repeat = int(options['r']) # each type of block has different valid arguments, fill accordingly block_args = dict( block_type=block_type, out_chs=int(options['c']), stride=int(options['s']), act_layer=act_layer, ) if block_type == 'ir': block_args.update(dict( dw_kernel_size=_parse_ksize(options['k']), exp_kernel_size=start_kernel_size, pw_kernel_size=end_kernel_size, exp_ratio=float(options['e']), se_ratio=float(options.get('se', 0.)), noskip=skip is False, s2d=int(options.get('d', 0)) > 0, )) if 'cc' in options: block_args['num_experts'] = int(options['cc']) elif block_type == 'ds' or block_type == 'dsa': block_args.update(dict( dw_kernel_size=_parse_ksize(options['k']), pw_kernel_size=end_kernel_size, se_ratio=float(options.get('se', 0.)), pw_act=block_type == 'dsa', noskip=block_type == 'dsa' or skip is False, s2d=int(options.get('d', 0)) > 0, )) elif block_type == 'er': block_args.update(dict( exp_kernel_size=_parse_ksize(options['k']), pw_kernel_size=end_kernel_size, exp_ratio=float(options['e']), force_in_chs=force_in_chs, se_ratio=float(options.get('se', 0.)), noskip=skip is False, )) elif block_type == 'cn': block_args.update(dict( kernel_size=int(options['k']), skip=skip is True, )) elif block_type == 'uir': # override exp / proj kernels for start/end in uir block start_kernel_size = _parse_ksize(options['a']) if 'a' in options else 0 end_kernel_size = _parse_ksize(options['p']) if 'p' in options else 0 block_args.update(dict( dw_kernel_size_start=start_kernel_size, # overload exp ks arg for dw start dw_kernel_size_mid=_parse_ksize(options['k']), dw_kernel_size_end=end_kernel_size, # overload pw ks arg for dw end exp_ratio=float(options['e']), se_ratio=float(options.get('se', 0.)), noskip=skip is False, )) elif block_type == 'mha': kv_dim = int(options['d']) block_args.update(dict( dw_kernel_size=_parse_ksize(options['k']), num_heads=int(options['h']), key_dim=kv_dim, value_dim=kv_dim, kv_stride=int(options.get('v', 1)), noskip=skip is False, )) elif block_type == 'mqa': kv_dim = int(options['d']) block_args.update(dict( dw_kernel_size=_parse_ksize(options['k']), num_heads=int(options['h']), key_dim=kv_dim, value_dim=kv_dim, kv_stride=int(options.get('v', 1)), noskip=skip is False, )) else: assert False, 'Unknown block type (%s)' % block_type if 'gs' in options: block_args['group_size'] = int(options['gs']) 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 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 """ arch_args = [] 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)): 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 if group_size is not None: ba.setdefault('group_size', group_size) stack_args.append(ba) repeats.append(rep) 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: arch_args.append(_scale_stage_depth(stack_args, repeats, multiplier, depth_trunc)) 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 """ def __init__( self, output_stride: int = 32, pad_type: str = '', round_chs_fn: Callable = round_channels, se_from_exp: bool = False, act_layer: Optional[LayerType] = None, norm_layer: Optional[LayerType] = None, aa_layer: Optional[LayerType] = None, se_layer: Optional[LayerType] = None, drop_path_rate: float = 0., layer_scale_init_value: Optional[float] = None, feature_location: str = '', ): self.output_stride = output_stride self.pad_type = pad_type self.round_chs_fn = round_chs_fn self.se_from_exp = se_from_exp # calculate se channel reduction from expanded (mid) chs self.act_layer = act_layer self.norm_layer = norm_layer self.aa_layer = aa_layer self.se_layer = get_attn(se_layer) try: self.se_layer(8, rd_ratio=1.0) # test if attn layer accepts rd_ratio arg self.se_has_ratio = True except TypeError: self.se_has_ratio = False self.drop_path_rate = drop_path_rate self.layer_scale_init_value = layer_scale_init_value if feature_location == 'depthwise': # old 'depthwise' mode renamed 'expansion' to match TF impl, old expansion mode didn't make sense _logger.warning("feature_location=='depthwise' is deprecated, using 'expansion'") feature_location = 'expansion' self.feature_location = feature_location assert feature_location in ('bottleneck', 'expansion', '') self.verbose = _DEBUG_BUILDER # state updated during build, consumed by model self.in_chs = None self.features = [] def _make_block(self, ba, block_idx, block_count): drop_path_rate = self.drop_path_rate * block_idx / block_count bt = ba.pop('block_type') ba['in_chs'] = self.in_chs ba['out_chs'] = self.round_chs_fn(ba['out_chs']) s2d = ba.get('s2d', 0) if s2d > 0: # adjust while space2depth active ba['out_chs'] *= 4 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']) 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 ba['norm_layer'] = self.norm_layer ba['drop_path_rate'] = drop_path_rate if self.aa_layer is not None: ba['aa_layer'] = self.aa_layer se_ratio = ba.pop('se_ratio', None) 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 s2d == 1: # adjust for start of space2depth se_ratio /= 4 if self.se_has_ratio: ba['se_layer'] = partial(self.se_layer, rd_ratio=se_ratio) else: ba['se_layer'] = self.se_layer if bt == 'ir': _log_info_if(' InvertedResidual {}, Args: {}'.format(block_idx, str(ba)), self.verbose) block = CondConvResidual(**ba) if ba.get('num_experts', 0) else InvertedResidual(**ba) elif bt == 'ds' or bt == 'dsa': _log_info_if(' DepthwiseSeparable {}, Args: {}'.format(block_idx, str(ba)), self.verbose) block = DepthwiseSeparableConv(**ba) elif bt == 'er': _log_info_if(' EdgeResidual {}, Args: {}'.format(block_idx, str(ba)), self.verbose) block = EdgeResidual(**ba) elif bt == 'cn': _log_info_if(' ConvBnAct {}, Args: {}'.format(block_idx, str(ba)), self.verbose) block = ConvBnAct(**ba) elif bt == 'uir': _log_info_if(' UniversalInvertedResidual {}, Args: {}'.format(block_idx, str(ba)), self.verbose) block = UniversalInvertedResidual(**ba, layer_scale_init_value=self.layer_scale_init_value) elif bt == 'mqa': _log_info_if(' MobileMultiQueryAttention {}, Args: {}'.format(block_idx, str(ba)), self.verbose) block = MobileAttention(**ba, use_multi_query=True, layer_scale_init_value=self.layer_scale_init_value) elif bt == 'mha': _log_info_if(' MobileMultiHeadAttention {}, Args: {}'.format(block_idx, str(ba)), self.verbose) block = MobileAttention(**ba, layer_scale_init_value=self.layer_scale_init_value) else: assert False, 'Unknown block type (%s) while building model.' % bt self.in_chs = ba['out_chs'] # update in_chs for arg of next block 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) """ _log_info_if('Building model trunk with %d stages...' % len(model_block_args), self.verbose) 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 = [] 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='bn1', num_chs=in_chs, stage=0, reduction=current_stride) self.features.append(feature_info) # outer list of block_args defines the stacks space2depth = 0 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) blocks = [] # 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) assert block_args['stride'] in (1, 2) if block_idx >= 1: # only the first block in any stack can have a stride > 1 block_args['stride'] = 1 if not space2depth and block_args.pop('s2d', False): assert block_args['stride'] == 1 space2depth = 1 if space2depth > 0: # FIXME s2d is a WIP if space2depth == 2 and block_args['stride'] == 2: block_args['stride'] = 1 # to end s2d region, need to correct expansion and se ratio relative to input block_args['exp_ratio'] /= 4 space2depth = 0 else: block_args['s2d'] = space2depth 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 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 _log_info_if(' Converting stride to dilation to maintain output_stride=={}'.format( self.output_stride), self.verbose) 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) if space2depth == 1: space2depth = 2 # stash feature module name and channel info for model feature extraction if extract_features: feature_info = dict( stage=stack_idx + 1, reduction=current_stride, **block.feature_info(self.feature_location), ) leaf_name = feature_info.get('module', '') if leaf_name: feature_info['module'] = '.'.join([f'blocks.{stack_idx}.{block_idx}', leaf_name]) else: assert last_block feature_info['module'] = f'blocks.{stack_idx}' self.features.append(feature_info) total_block_idx += 1 # incr global block idx (across all stacks) stages.append(nn.Sequential(*blocks)) return stages def _init_weight_goog(m, n='', fix_group_fanout=True): """ Weight initialization as per Tensorflow official implementations. Args: m (nn.Module): module to init n (str): module name fix_group_fanout (bool): enable correct (matching Tensorflow TPU impl) fanout calculation w/ group convs 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 """ if isinstance(m, CondConv2d): fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels if fix_group_fanout: fan_out //= m.groups init_weight_fn = get_condconv_initializer( lambda w: nn.init.normal_(w, 0, math.sqrt(2.0 / fan_out)), m.num_experts, m.weight_shape) init_weight_fn(m.weight) if m.bias is not None: nn.init.zeros_(m.bias) elif isinstance(m, nn.Conv2d): fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels if fix_group_fanout: fan_out //= m.groups nn.init.normal_(m.weight, 0, math.sqrt(2.0 / fan_out)) if m.bias is not None: nn.init.zeros_(m.bias) elif isinstance(m, nn.BatchNorm2d): nn.init.ones_(m.weight) nn.init.zeros_(m.bias) 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) nn.init.uniform_(m.weight, -init_range, init_range) nn.init.zeros_(m.bias) 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) # iterate and call any module.init_weights() fn, children first for n, m in named_modules(model): if hasattr(m, 'init_weights'): m.init_weights()