mobilenetV3_small.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### mobilenetV3_small"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tensorflow.keras.layers import Conv2D, DepthwiseConv2D, Dense, GlobalAveragePooling2D,Input\n",
    "from tensorflow.keras.layers import Activation, BatchNormalization, Add, Multiply, Reshape\n",
    "from tensorflow.keras.models import Model\n",
    "from tensorflow.keras import backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "alpha = 1\n",
    "def relu6(x):\n",
    "    # relu函数\n",
    "    return K.relu(x, max_value=6.0)\n",
    "\n",
    "def hard_swish(x):\n",
    "    # 利用relu函数乘上x模拟sigmoid\n",
    "    return x * K.relu(x + 3.0, max_value=6.0) / 6.0\n",
    "\n",
    "def return_activation(x, nl):\n",
    "    # 用于判断使用哪个激活函数\n",
    "    if nl == 'HS':\n",
    "        x = Activation(hard_swish)(x)\n",
    "    if nl == 'RE':\n",
    "        x = Activation(relu6)(x)\n",
    "\n",
    "    return x\n",
    "\n",
    "def conv_block(inputs, filters, kernel, strides, nl):\n",
    "    # 一个卷积单元，也就是conv2d + batchnormalization + activation\n",
    "    channel_axis = 1 if K.image_data_format() == 'channels_first' else -1\n",
    "\n",
    "    x = Conv2D(filters, kernel, padding='same', strides=strides)(inputs)\n",
    "    x = BatchNormalization(axis=channel_axis)(x)\n",
    "\n",
    "    return return_activation(x, nl)\n",
    "\n",
    "def squeeze(inputs):\n",
    "    # 注意力机制单元\n",
    "    input_channels = int(inputs.shape[-1])\n",
    "\n",
    "    x = GlobalAveragePooling2D()(inputs)\n",
    "    x = Dense(int(input_channels/4))(x)\n",
    "    x = Activation(relu6)(x)\n",
    "    x = Dense(input_channels)(x)\n",
    "    x = Activation(hard_swish)(x)\n",
    "    x = Reshape((1, 1, input_channels))(x)\n",
    "    x = Multiply()([inputs, x])\n",
    "\n",
    "    return x\n",
    "\n",
    "def bottleneck(inputs, filters, kernel, up_dim, stride, sq, nl):\n",
    "    channel_axis = 1 if K.image_data_format() == 'channels_first' else -1\n",
    "\n",
    "    input_shape = K.int_shape(inputs)\n",
    "\n",
    "    tchannel = int(up_dim)\n",
    "    cchannel = int(alpha * filters)\n",
    "\n",
    "    r = stride == 1 and input_shape[3] == filters\n",
    "    # 1x1卷积调整通道数，通道数上升\n",
    "    x = conv_block(inputs, tchannel, (1, 1), (1, 1), nl)\n",
    "    # 进行3x3深度可分离卷积\n",
    "    x = DepthwiseConv2D(kernel, strides=(stride, stride), depth_multiplier=1, padding='same')(x)\n",
    "    x = BatchNormalization(axis=channel_axis)(x)\n",
    "    x = return_activation(x, nl)\n",
    "    # 引入注意力机制\n",
    "    if sq:\n",
    "        x = squeeze(x)\n",
    "    # 下降通道数\n",
    "    x = Conv2D(cchannel, (1, 1), strides=(1, 1), padding='same')(x)\n",
    "    x = BatchNormalization(axis=channel_axis)(x)\n",
    "\n",
    "\n",
    "    if r:\n",
    "        x = Add()([x, inputs])\n",
    "\n",
    "    return x\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def MobileNetv3_small(shape = (224,224,3),n_class = 1000):\n",
    "    inputs = Input(shape)\n",
    "    # 224,224,3 -> 112,112,16\n",
    "    x = conv_block(inputs, 16, (3, 3), strides=(2, 2), nl='HS')\n",
    "\n",
    "    # 112,112,16 -> 56,56,16\n",
    "    x = bottleneck(x, 16, (3, 3), up_dim=16, stride=2, sq=True, nl='RE')\n",
    "\n",
    "    # 56,56,16 -> 28,28,24\n",
    "    x = bottleneck(x, 24, (3, 3), up_dim=72, stride=2, sq=False, nl='RE')\n",
    "    x = bottleneck(x, 24, (3, 3), up_dim=88, stride=1, sq=False, nl='RE')\n",
    "    \n",
    "    # 28,28,24 -> 14,14,40\n",
    "    x = bottleneck(x, 40, (5, 5), up_dim=96, stride=2, sq=True, nl='HS')\n",
    "    x = bottleneck(x, 40, (5, 5), up_dim=240, stride=1, sq=True, nl='HS')\n",
    "    x = bottleneck(x, 40, (5, 5), up_dim=240, stride=1, sq=True, nl='HS')\n",
    "    # 14,14,40 -> 14,14,48\n",
    "    x = bottleneck(x, 48, (5, 5), up_dim=120, stride=1, sq=True, nl='HS')\n",
    "    x = bottleneck(x, 48, (5, 5), up_dim=144, stride=1, sq=True, nl='HS')\n",
    "\n",
    "    # 14,14,48 -> 7,7,96\n",
    "    x = bottleneck(x, 96, (5, 5), up_dim=288, stride=2, sq=True, nl='HS')\n",
    "    x = bottleneck(x, 96, (5, 5), up_dim=576, stride=1, sq=True, nl='HS')\n",
    "    x = bottleneck(x, 96, (5, 5), up_dim=576, stride=1, sq=True, nl='HS')\n",
    "\n",
    "    x = conv_block(x, 576, (1, 1), strides=(1, 1), nl='HS')\n",
    "    x = GlobalAveragePooling2D()(x)\n",
    "    x = Reshape((1, 1, 576))(x)\n",
    "\n",
    "    x = Conv2D(1024, (1, 1), padding='same')(x)\n",
    "    x = return_activation(x, 'HS')\n",
    "\n",
    "    x = Conv2D(n_class, (1, 1), padding='same', activation='softmax')(x)\n",
    "    x = Reshape((n_class,))(x)\n",
    "\n",
    "    model = Model(inputs, x)\n",
    "\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = MobileNetv3_small()\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}