import paddle from paddle import ParamAttr import paddle.nn as nn import paddle.nn.functional as F from paddle.nn import Conv2d, BatchNorm, Linear, Dropout from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d import math import collections import re import copy __all__ = [ 'EfficientNet', 'EfficientNetB0_small', 'EfficientNetB0', 'EfficientNetB1', 'EfficientNetB2', 'EfficientNetB3', 'EfficientNetB4', 'EfficientNetB5', 'EfficientNetB6', 'EfficientNetB7' ] GlobalParams = collections.namedtuple('GlobalParams', [ 'batch_norm_momentum', 'batch_norm_epsilon', 'dropout_rate', 'num_classes', 'width_coefficient', 'depth_coefficient', 'depth_divisor', 'min_depth', 'drop_connect_rate', ]) BlockArgs = collections.namedtuple('BlockArgs', [ 'kernel_size', 'num_repeat', 'input_filters', 'output_filters', 'expand_ratio', 'id_skip', 'stride', 'se_ratio' ]) GlobalParams.__new__.__defaults__ = (None, ) * len(GlobalParams._fields) BlockArgs.__new__.__defaults__ = (None, ) * len(BlockArgs._fields) def efficientnet_params(model_name): """ Map EfficientNet model name to parameter coefficients. """ params_dict = { # Coefficients: width,depth,resolution,dropout 'efficientnet-b0': (1.0, 1.0, 224, 0.2), 'efficientnet-b1': (1.0, 1.1, 240, 0.2), 'efficientnet-b2': (1.1, 1.2, 260, 0.3), 'efficientnet-b3': (1.2, 1.4, 300, 0.3), 'efficientnet-b4': (1.4, 1.8, 380, 0.4), 'efficientnet-b5': (1.6, 2.2, 456, 0.4), 'efficientnet-b6': (1.8, 2.6, 528, 0.5), 'efficientnet-b7': (2.0, 3.1, 600, 0.5), } return params_dict[model_name] def efficientnet(width_coefficient=None, depth_coefficient=None, dropout_rate=0.2, drop_connect_rate=0.2): """ Get block arguments according to parameter and coefficients. """ blocks_args = [ 'r1_k3_s11_e1_i32_o16_se0.25', 'r2_k3_s22_e6_i16_o24_se0.25', 'r2_k5_s22_e6_i24_o40_se0.25', 'r3_k3_s22_e6_i40_o80_se0.25', 'r3_k5_s11_e6_i80_o112_se0.25', 'r4_k5_s22_e6_i112_o192_se0.25', 'r1_k3_s11_e6_i192_o320_se0.25', ] blocks_args = BlockDecoder.decode(blocks_args) global_params = GlobalParams( batch_norm_momentum=0.99, batch_norm_epsilon=1e-3, dropout_rate=dropout_rate, drop_connect_rate=drop_connect_rate, num_classes=1000, width_coefficient=width_coefficient, depth_coefficient=depth_coefficient, depth_divisor=8, min_depth=None) return blocks_args, global_params def get_model_params(model_name, override_params): """ Get the block args and global params for a given model """ if model_name.startswith('efficientnet'): w, d, _, p = efficientnet_params(model_name) blocks_args, global_params = efficientnet( width_coefficient=w, depth_coefficient=d, dropout_rate=p) else: raise NotImplementedError('model name is not pre-defined: %s' % model_name) if override_params: global_params = global_params._replace(**override_params) return blocks_args, global_params def round_filters(filters, global_params): """ Calculate and round number of filters based on depth multiplier. """ multiplier = global_params.width_coefficient if not multiplier: return filters divisor = global_params.depth_divisor min_depth = global_params.min_depth filters *= multiplier min_depth = min_depth or divisor new_filters = max(min_depth, int(filters + divisor / 2) // divisor * divisor) if new_filters < 0.9 * filters: # prevent rounding by more than 10% new_filters += divisor return int(new_filters) def round_repeats(repeats, global_params): """ Round number of filters based on depth multiplier. """ multiplier = global_params.depth_coefficient if not multiplier: return repeats return int(math.ceil(multiplier * repeats)) class BlockDecoder(object): """ Block Decoder, straight from the official TensorFlow repository. """ @staticmethod def _decode_block_string(block_string): """ Gets a block through a string notation of arguments. """ assert isinstance(block_string, str) ops = block_string.split('_') options = {} for op in ops: splits = re.split(r'(\d.*)', op) if len(splits) >= 2: key, value = splits[:2] options[key] = value # Check stride cond_1 = ('s' in options and len(options['s']) == 1) cond_2 = ((len(options['s']) == 2) and (options['s'][0] == options['s'][1])) assert (cond_1 or cond_2) return BlockArgs( kernel_size=int(options['k']), num_repeat=int(options['r']), input_filters=int(options['i']), output_filters=int(options['o']), expand_ratio=int(options['e']), id_skip=('noskip' not in block_string), se_ratio=float(options['se']) if 'se' in options else None, stride=[int(options['s'][0])]) @staticmethod def _encode_block_string(block): """Encodes a block to a string.""" args = [ 'r%d' % block.num_repeat, 'k%d' % block.kernel_size, 's%d%d' % (block.strides[0], block.strides[1]), 'e%s' % block.expand_ratio, 'i%d' % block.input_filters, 'o%d' % block.output_filters ] if 0 < block.se_ratio <= 1: args.append('se%s' % block.se_ratio) if block.id_skip is False: args.append('noskip') return '_'.join(args) @staticmethod def decode(string_list): """ Decode a list of string notations to specify blocks in the network. string_list: list of strings, each string is a notation of block return list of BlockArgs namedtuples of block args """ assert isinstance(string_list, list) blocks_args = [] for block_string in string_list: blocks_args.append(BlockDecoder._decode_block_string(block_string)) return blocks_args @staticmethod def encode(blocks_args): """ Encodes a list of BlockArgs to a list of strings. :param blocks_args: a list of BlockArgs namedtuples of block args :return: a list of strings, each string is a notation of block """ block_strings = [] for block in blocks_args: block_strings.append(BlockDecoder._encode_block_string(block)) return block_strings def initial_type(name, use_bias=False): param_attr = ParamAttr(name=name + "_weights") if use_bias: bias_attr = ParamAttr(name=name + "_offset") else: bias_attr = False return param_attr, bias_attr def init_batch_norm_layer(name="batch_norm"): param_attr = ParamAttr(name=name + "_scale") bias_attr = ParamAttr(name=name + "_offset") return param_attr, bias_attr def init_fc_layer(name="fc"): param_attr = ParamAttr(name=name + "_weights") bias_attr = ParamAttr(name=name + "_offset") return param_attr, bias_attr def cal_padding(img_size, stride, filter_size, dilation=1): """Calculate padding size.""" if img_size % stride == 0: out_size = max(filter_size - stride, 0) else: out_size = max(filter_size - (img_size % stride), 0) return out_size // 2, out_size - out_size // 2 inp_shape = { "b0_small": [224, 112, 112, 56, 28, 14, 14, 7], "b0": [224, 112, 112, 56, 28, 14, 14, 7], "b1": [240, 120, 120, 60, 30, 15, 15, 8], "b2": [260, 130, 130, 65, 33, 17, 17, 9], "b3": [300, 150, 150, 75, 38, 19, 19, 10], "b4": [380, 190, 190, 95, 48, 24, 24, 12], "b5": [456, 228, 228, 114, 57, 29, 29, 15], "b6": [528, 264, 264, 132, 66, 33, 33, 17], "b7": [600, 300, 300, 150, 75, 38, 38, 19] } def _drop_connect(inputs, prob, is_test): if is_test: return inputs keep_prob = 1.0 - prob inputs_shape = paddle.shape(inputs) random_tensor = keep_prob + paddle.rand(shape=[inputs_shape[0], 1, 1, 1]) binary_tensor = paddle.floor(random_tensor) output = inputs / keep_prob * binary_tensor return output class Conv2ds(nn.Layer): def __init__(self, input_channels, output_channels, filter_size, stride=1, padding=0, groups=None, name="conv2d", act=None, use_bias=False, padding_type=None, model_name=None, cur_stage=None): super(Conv2ds, self).__init__() assert act in [None, "swish", "sigmoid"] self.act = act param_attr, bias_attr = initial_type(name=name, use_bias=use_bias) def get_padding(filter_size, stride=1, dilation=1): padding = ((stride - 1) + dilation * (filter_size - 1)) // 2 return padding inps = 1 if model_name == None and cur_stage == None else inp_shape[ model_name][cur_stage] self.need_crop = False if padding_type == "SAME": top_padding, bottom_padding = cal_padding(inps, stride, filter_size) left_padding, right_padding = cal_padding(inps, stride, filter_size) height_padding = bottom_padding width_padding = right_padding if top_padding != bottom_padding or left_padding != right_padding: height_padding = top_padding + stride width_padding = left_padding + stride self.need_crop = True padding = [height_padding, width_padding] elif padding_type == "VALID": height_padding = 0 width_padding = 0 padding = [height_padding, width_padding] elif padding_type == "DYNAMIC": padding = get_padding(filter_size, stride) else: padding = padding_type groups = 1 if groups is None else groups self._conv = Conv2d( input_channels, output_channels, filter_size, groups=groups, stride=stride, # act=act, padding=padding, weight_attr=param_attr, bias_attr=bias_attr) def forward(self, inputs): x = self._conv(inputs) if self.act == "swish": x = F.swish(x) elif self.act == "sigmoid": x = F.sigmoid(x) if self.need_crop: x = x[:, :, 1:, 1:] return x class ConvBNLayer(nn.Layer): def __init__(self, input_channels, filter_size, output_channels, stride=1, num_groups=1, padding_type="SAME", conv_act=None, bn_act="swish", use_bn=True, use_bias=False, name=None, conv_name=None, bn_name=None, model_name=None, cur_stage=None): super(ConvBNLayer, self).__init__() self._conv = Conv2ds( input_channels=input_channels, output_channels=output_channels, filter_size=filter_size, stride=stride, groups=num_groups, act=conv_act, padding_type=padding_type, name=conv_name, use_bias=use_bias, model_name=model_name, cur_stage=cur_stage) self.use_bn = use_bn if use_bn is True: bn_name = name + bn_name param_attr, bias_attr = init_batch_norm_layer(bn_name) self._bn = BatchNorm( num_channels=output_channels, act=bn_act, momentum=0.99, epsilon=0.001, moving_mean_name=bn_name + "_mean", moving_variance_name=bn_name + "_variance", param_attr=param_attr, bias_attr=bias_attr) def forward(self, inputs): if self.use_bn: x = self._conv(inputs) x = self._bn(x) return x else: return self._conv(inputs) class ExpandConvNorm(nn.Layer): def __init__(self, input_channels, block_args, padding_type, name=None, model_name=None, cur_stage=None): super(ExpandConvNorm, self).__init__() self.oup = block_args.input_filters * block_args.expand_ratio self.expand_ratio = block_args.expand_ratio if self.expand_ratio != 1: self._conv = ConvBNLayer( input_channels, 1, self.oup, bn_act=None, padding_type=padding_type, name=name, conv_name=name + "_expand_conv", bn_name="_bn0", model_name=model_name, cur_stage=cur_stage) def forward(self, inputs): if self.expand_ratio != 1: return self._conv(inputs) else: return inputs class DepthwiseConvNorm(nn.Layer): def __init__(self, input_channels, block_args, padding_type, name=None, model_name=None, cur_stage=None): super(DepthwiseConvNorm, self).__init__() self.k = block_args.kernel_size self.s = block_args.stride if isinstance(self.s, list) or isinstance(self.s, tuple): self.s = self.s[0] oup = block_args.input_filters * block_args.expand_ratio self._conv = ConvBNLayer( input_channels, self.k, oup, self.s, num_groups=input_channels, bn_act=None, padding_type=padding_type, name=name, conv_name=name + "_depthwise_conv", bn_name="_bn1", model_name=model_name, cur_stage=cur_stage) def forward(self, inputs): return self._conv(inputs) class ProjectConvNorm(nn.Layer): def __init__(self, input_channels, block_args, padding_type, name=None, model_name=None, cur_stage=None): super(ProjectConvNorm, self).__init__() final_oup = block_args.output_filters self._conv = ConvBNLayer( input_channels, 1, final_oup, bn_act=None, padding_type=padding_type, name=name, conv_name=name + "_project_conv", bn_name="_bn2", model_name=model_name, cur_stage=cur_stage) def forward(self, inputs): return self._conv(inputs) class SEBlock(nn.Layer): def __init__(self, input_channels, num_squeezed_channels, oup, padding_type, name=None, model_name=None, cur_stage=None): super(SEBlock, self).__init__() self._pool = AdaptiveAvgPool2d(1) self._conv1 = Conv2ds( input_channels, num_squeezed_channels, 1, use_bias=True, padding_type=padding_type, act="swish", name=name + "_se_reduce") self._conv2 = Conv2ds( num_squeezed_channels, oup, 1, act="sigmoid", use_bias=True, padding_type=padding_type, name=name + "_se_expand") def forward(self, inputs): x = self._pool(inputs) x = self._conv1(x) x = self._conv2(x) return paddle.multiply(inputs, x) class MbConvBlock(nn.Layer): def __init__(self, input_channels, block_args, padding_type, use_se, name=None, drop_connect_rate=None, model_name=None, cur_stage=None): super(MbConvBlock, self).__init__() oup = block_args.input_filters * block_args.expand_ratio self.block_args = block_args self.has_se = use_se and (block_args.se_ratio is not None) and ( 0 < block_args.se_ratio <= 1) self.id_skip = block_args.id_skip self.expand_ratio = block_args.expand_ratio self.drop_connect_rate = drop_connect_rate if self.expand_ratio != 1: self._ecn = ExpandConvNorm( input_channels, block_args, padding_type=padding_type, name=name, model_name=model_name, cur_stage=cur_stage) self._dcn = DepthwiseConvNorm( input_channels * block_args.expand_ratio, block_args, padding_type=padding_type, name=name, model_name=model_name, cur_stage=cur_stage) if self.has_se: num_squeezed_channels = max( 1, int(block_args.input_filters * block_args.se_ratio)) self._se = SEBlock( input_channels * block_args.expand_ratio, num_squeezed_channels, oup, padding_type=padding_type, name=name, model_name=model_name, cur_stage=cur_stage) self._pcn = ProjectConvNorm( input_channels * block_args.expand_ratio, block_args, padding_type=padding_type, name=name, model_name=model_name, cur_stage=cur_stage) def forward(self, inputs): x = inputs if self.expand_ratio != 1: x = self._ecn(x) x = F.swish(x) x = self._dcn(x) x = F.swish(x) if self.has_se: x = self._se(x) x = self._pcn(x) if self.id_skip and \ self.block_args.stride == 1 and \ self.block_args.input_filters == self.block_args.output_filters: if self.drop_connect_rate: x = _drop_connect(x, self.drop_connect_rate, not self.training) x = paddle.elementwise_add(x, inputs) return x class ConvStemNorm(nn.Layer): def __init__(self, input_channels, padding_type, _global_params, name=None, model_name=None, cur_stage=None): super(ConvStemNorm, self).__init__() output_channels = round_filters(32, _global_params) self._conv = ConvBNLayer( input_channels, filter_size=3, output_channels=output_channels, stride=2, bn_act=None, padding_type=padding_type, name="", conv_name="_conv_stem", bn_name="_bn0", model_name=model_name, cur_stage=cur_stage) def forward(self, inputs): return self._conv(inputs) class ExtractFeatures(nn.Layer): def __init__(self, input_channels, _block_args, _global_params, padding_type, use_se, model_name=None): super(ExtractFeatures, self).__init__() self._global_params = _global_params self._conv_stem = ConvStemNorm( input_channels, padding_type=padding_type, _global_params=_global_params, model_name=model_name, cur_stage=0) self.block_args_copy = copy.deepcopy(_block_args) idx = 0 block_size = 0 for block_arg in self.block_args_copy: block_arg = block_arg._replace( input_filters=round_filters(block_arg.input_filters, _global_params), output_filters=round_filters(block_arg.output_filters, _global_params), num_repeat=round_repeats(block_arg.num_repeat, _global_params)) block_size += 1 for _ in range(block_arg.num_repeat - 1): block_size += 1 self.conv_seq = [] cur_stage = 1 for block_args in _block_args: block_args = block_args._replace( input_filters=round_filters(block_args.input_filters, _global_params), output_filters=round_filters(block_args.output_filters, _global_params), num_repeat=round_repeats(block_args.num_repeat, _global_params)) drop_connect_rate = self._global_params.drop_connect_rate if drop_connect_rate: drop_connect_rate *= float(idx) / block_size _mc_block = self.add_sublayer( "_blocks." + str(idx) + ".", MbConvBlock( block_args.input_filters, block_args=block_args, padding_type=padding_type, use_se=use_se, name="_blocks." + str(idx) + ".", drop_connect_rate=drop_connect_rate, model_name=model_name, cur_stage=cur_stage)) self.conv_seq.append(_mc_block) idx += 1 if block_args.num_repeat > 1: block_args = block_args._replace( input_filters=block_args.output_filters, stride=1) for _ in range(block_args.num_repeat - 1): drop_connect_rate = self._global_params.drop_connect_rate if drop_connect_rate: drop_connect_rate *= float(idx) / block_size _mc_block = self.add_sublayer( "block." + str(idx) + ".", MbConvBlock( block_args.input_filters, block_args, padding_type=padding_type, use_se=use_se, name="_blocks." + str(idx) + ".", drop_connect_rate=drop_connect_rate, model_name=model_name, cur_stage=cur_stage)) self.conv_seq.append(_mc_block) idx += 1 cur_stage += 1 def forward(self, inputs): x = self._conv_stem(inputs) x = F.swish(x) for _mc_block in self.conv_seq: x = _mc_block(x) return x class EfficientNet(nn.Layer): def __init__(self, name="b0", padding_type="SAME", override_params=None, use_se=True, class_dim=1000): super(EfficientNet, self).__init__() model_name = 'efficientnet-' + name self.name = name self._block_args, self._global_params = get_model_params( model_name, override_params) self.padding_type = padding_type self.use_se = use_se self._ef = ExtractFeatures( 3, self._block_args, self._global_params, self.padding_type, self.use_se, model_name=self.name) output_channels = round_filters(1280, self._global_params) if name == "b0_small" or name == "b0" or name == "b1": oup = 320 elif name == "b2": oup = 352 elif name == "b3": oup = 384 elif name == "b4": oup = 448 elif name == "b5": oup = 512 elif name == "b6": oup = 576 elif name == "b7": oup = 640 self._conv = ConvBNLayer( oup, 1, output_channels, bn_act="swish", padding_type=self.padding_type, name="", conv_name="_conv_head", bn_name="_bn1", model_name=self.name, cur_stage=7) self._pool = AdaptiveAvgPool2d(1) if self._global_params.dropout_rate: self._drop = Dropout( p=self._global_params.dropout_rate, mode="upscale_in_train") param_attr, bias_attr = init_fc_layer("_fc") self._fc = Linear( output_channels, class_dim, weight_attr=param_attr, bias_attr=bias_attr) def forward(self, inputs): x = self._ef(inputs) x = self._conv(x) x = self._pool(x) if self._global_params.dropout_rate: x = self._drop(x) x = paddle.squeeze(x, axis=[2, 3]) x = self._fc(x) return x def EfficientNetB0_small(padding_type='DYNAMIC', override_params=None, use_se=False, **args): model = EfficientNet( name='b0', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model def EfficientNetB0(padding_type='SAME', override_params=None, use_se=True, **args): model = EfficientNet( name='b0', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model def EfficientNetB1(padding_type='SAME', override_params=None, use_se=True, **args): model = EfficientNet( name='b1', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model def EfficientNetB2(padding_type='SAME', override_params=None, use_se=True, **args): model = EfficientNet( name='b2', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model def EfficientNetB3(padding_type='SAME', override_params=None, use_se=True, **args): model = EfficientNet( name='b3', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model def EfficientNetB4(padding_type='SAME', override_params=None, use_se=True, **args): model = EfficientNet( name='b4', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model def EfficientNetB5(padding_type='SAME', override_params=None, use_se=True, **args): model = EfficientNet( name='b5', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model def EfficientNetB6(padding_type='SAME', override_params=None, use_se=True, **args): model = EfficientNet( name='b6', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model def EfficientNetB7(padding_type='SAME', override_params=None, use_se=True, **args): model = EfficientNet( name='b7', padding_type=padding_type, override_params=override_params, use_se=use_se, **args) return model