save_and_load.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "pycharm": {
     "is_executing": false
    }
   },
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "pycharm": {
     "is_executing": false
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "text": [
      "2.1.0\n",
      "1.18.1\n"
     ],
     "output_type": "stream"
    }
   ],
   "source": [
    "for gpu in tf.config.experimental.list_physical_devices('GPU'):\n",
    "    tf.config.experimental.set_memory_growth(gpu, True)\n",
    "\n",
    "print(tf.__version__)\n",
    "print(np.__version__)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "outputs": [],
   "source": [
    "pretrained_model_names = ['ade20k', 'cityscapes', 'voc2012']\n",
    "pretrained_model_name = pretrained_model_names[2]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n",
     "is_executing": false
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "outputs": [],
   "source": [
    "from pspnet import PSPNet101, PSPNet50\n",
    "nb_classes_config = {\n",
    "    'ade20k': 150,\n",
    "    'cityscapes': 19,\n",
    "    'voc2012': 21\n",
    "}\n",
    "pspnet_layers_config = {\n",
    "    'ade20k': 'pspnet50',\n",
    "    'cityscapes': 'pspnet101',\n",
    "    'voc2012': 'pspnet101'\n",
    "}\n",
    "input_shape_config = {\n",
    "    'ade20k': (473, 473),\n",
    "    'cityscapes': (713, 713),\n",
    "    'voc2012': (473, 473)\n",
    "}\n",
    "pspnet_config = {\n",
    "    'pspnet50': PSPNet50,\n",
    "    'pspnet101': PSPNet101\n",
    "}\n",
    "nb_classes = nb_classes_config[pretrained_model_name]\n",
    "pspnet_layers = pspnet_layers_config[pretrained_model_name]\n",
    "input_shape = input_shape_config[pretrained_model_name]\n",
    "PSPNet = pspnet_config[pspnet_layers]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n",
     "is_executing": false
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "outputs": [
    {
     "name": "stdout",
     "text": [
      "No Keras model & weights found, import from npy weights.\n",
      "Building a PSPNet based on ResNet 101 expecting inputs of shape (473, 473) predicting 21 classes\n",
      "PSP module will interpolate to a final feature map size of (60, 60)\n",
      "Importing weights from weights\\npy\\pspnet101_voc2012.npy\n",
      "input_2\n",
      "conv1_1_3x3_s2\n",
      "conv1_1_3x3_s2_bn\n",
      "activation_58\n",
      "conv1_2_3x3\n",
      "conv1_2_3x3_bn\n",
      "activation_59\n",
      "conv1_3_3x3\n",
      "conv1_3_3x3_bn\n",
      "activation_60\n",
      "max_pooling2d_1\n",
      "activation_61\n",
      "conv2_1_1x1_reduce\n",
      "conv2_1_1x1_reduce_bn\n",
      "activation_62\n",
      "zero_padding2d_16\n",
      "conv2_1_3x3\n",
      "conv2_1_3x3_bn\n",
      "activation_63\n",
      "conv2_1_1x1_increase\n",
      "conv2_1_1x1_proj\n",
      "conv2_1_1x1_increase_bn\n",
      "conv2_1_1x1_proj_bn\n",
      "add_16\n",
      "activation_64\n",
      "conv2_2_1x1_reduce\n",
      "conv2_2_1x1_reduce_bn\n",
      "activation_65\n",
      "zero_padding2d_17\n",
      "conv2_2_3x3\n",
      "conv2_2_3x3_bn\n",
      "activation_66\n",
      "conv2_2_1x1_increase\n",
      "conv2_2_1x1_increase_bn\n",
      "add_17\n",
      "activation_67\n",
      "conv2_3_1x1_reduce\n",
      "conv2_3_1x1_reduce_bn\n",
      "activation_68\n",
      "zero_padding2d_18\n",
      "conv2_3_3x3\n",
      "conv2_3_3x3_bn\n",
      "activation_69\n",
      "conv2_3_1x1_increase\n",
      "conv2_3_1x1_increase_bn\n",
      "add_18\n",
      "activation_70\n",
      "conv3_1_1x1_reduce\n",
      "conv3_1_1x1_reduce_bn\n",
      "activation_71\n",
      "zero_padding2d_19\n",
      "conv3_1_3x3\n",
      "conv3_1_3x3_bn\n",
      "activation_72\n",
      "conv3_1_1x1_increase\n",
      "conv3_1_1x1_proj\n",
      "conv3_1_1x1_increase_bn\n",
      "conv3_1_1x1_proj_bn\n",
      "add_19\n",
      "activation_73\n",
      "conv3_2_1x1_reduce\n",
      "conv3_2_1x1_reduce_bn\n",
      "activation_74\n",
      "zero_padding2d_20\n",
      "conv3_2_3x3\n",
      "conv3_2_3x3_bn\n",
      "activation_75\n",
      "conv3_2_1x1_increase\n",
      "conv3_2_1x1_increase_bn\n",
      "add_20\n",
      "activation_76\n",
      "conv3_3_1x1_reduce\n",
      "conv3_3_1x1_reduce_bn\n",
      "activation_77\n",
      "zero_padding2d_21\n",
      "conv3_3_3x3\n",
      "conv3_3_3x3_bn\n",
      "activation_78\n",
      "conv3_3_1x1_increase\n",
      "conv3_3_1x1_increase_bn\n",
      "add_21\n",
      "activation_79\n",
      "conv3_4_1x1_reduce\n",
      "conv3_4_1x1_reduce_bn\n",
      "activation_80\n",
      "zero_padding2d_22\n",
      "conv3_4_3x3\n",
      "conv3_4_3x3_bn\n",
      "activation_81\n",
      "conv3_4_1x1_increase\n",
      "conv3_4_1x1_increase_bn\n",
      "add_22\n",
      "activation_82\n",
      "conv4_1_1x1_reduce\n",
      "conv4_1_1x1_reduce_bn\n",
      "activation_83\n",
      "zero_padding2d_23\n",
      "conv4_1_3x3\n",
      "conv4_1_3x3_bn\n",
      "activation_84\n",
      "conv4_1_1x1_increase\n",
      "conv4_1_1x1_proj\n",
      "conv4_1_1x1_increase_bn\n",
      "conv4_1_1x1_proj_bn\n",
      "add_23\n",
      "activation_85\n",
      "conv4_2_1x1_reduce\n",
      "conv4_2_1x1_reduce_bn\n",
      "activation_86\n",
      "zero_padding2d_24\n",
      "conv4_2_3x3\n",
      "conv4_2_3x3_bn\n",
      "activation_87\n",
      "conv4_2_1x1_increase\n",
      "conv4_2_1x1_increase_bn\n",
      "add_24\n",
      "activation_88\n",
      "conv4_3_1x1_reduce\n",
      "conv4_3_1x1_reduce_bn\n",
      "activation_89\n",
      "zero_padding2d_25\n",
      "conv4_3_3x3\n",
      "conv4_3_3x3_bn\n",
      "activation_90\n",
      "conv4_3_1x1_increase\n",
      "conv4_3_1x1_increase_bn\n",
      "add_25\n",
      "activation_91\n",
      "conv4_4_1x1_reduce\n",
      "conv4_4_1x1_reduce_bn\n",
      "activation_92\n",
      "zero_padding2d_26\n",
      "conv4_4_3x3\n",
      "conv4_4_3x3_bn\n",
      "activation_93\n",
      "conv4_4_1x1_increase\n",
      "conv4_4_1x1_increase_bn\n",
      "add_26\n",
      "activation_94\n",
      "conv4_5_1x1_reduce\n",
      "conv4_5_1x1_reduce_bn\n",
      "activation_95\n",
      "zero_padding2d_27\n",
      "conv4_5_3x3\n",
      "conv4_5_3x3_bn\n",
      "activation_96\n",
      "conv4_5_1x1_increase\n",
      "conv4_5_1x1_increase_bn\n",
      "add_27\n",
      "activation_97\n",
      "conv4_6_1x1_reduce\n",
      "conv4_6_1x1_reduce_bn\n",
      "activation_98\n",
      "zero_padding2d_28\n",
      "conv4_6_3x3\n",
      "conv4_6_3x3_bn\n",
      "activation_99\n",
      "conv4_6_1x1_increase\n",
      "conv4_6_1x1_increase_bn\n",
      "add_28\n",
      "activation_100\n",
      "conv4_7_1x1_reduce\n",
      "conv4_7_1x1_reduce_bn\n",
      "activation_101\n",
      "zero_padding2d_29\n",
      "conv4_7_3x3\n",
      "conv4_7_3x3_bn\n",
      "activation_102\n",
      "conv4_7_1x1_increase\n",
      "conv4_7_1x1_increase_bn\n",
      "add_29\n",
      "activation_103\n",
      "conv4_8_1x1_reduce\n",
      "conv4_8_1x1_reduce_bn\n",
      "activation_104\n",
      "zero_padding2d_30\n",
      "conv4_8_3x3\n",
      "conv4_8_3x3_bn\n",
      "activation_105\n",
      "conv4_8_1x1_increase\n",
      "conv4_8_1x1_increase_bn\n",
      "add_30\n",
      "activation_106\n",
      "conv4_9_1x1_reduce\n",
      "conv4_9_1x1_reduce_bn\n",
      "activation_107\n",
      "zero_padding2d_31\n",
      "conv4_9_3x3\n",
      "conv4_9_3x3_bn\n",
      "activation_108\n",
      "conv4_9_1x1_increase\n",
      "conv4_9_1x1_increase_bn\n",
      "add_31\n",
      "activation_109\n",
      "conv4_10_1x1_reduce\n",
      "conv4_10_1x1_reduce_bn\n",
      "activation_110\n",
      "zero_padding2d_32\n",
      "conv4_10_3x3\n",
      "conv4_10_3x3_bn\n",
      "activation_111\n",
      "conv4_10_1x1_increase\n",
      "conv4_10_1x1_increase_bn\n",
      "add_32\n",
      "activation_112\n",
      "conv4_11_1x1_reduce\n",
      "conv4_11_1x1_reduce_bn\n",
      "activation_113\n",
      "zero_padding2d_33\n",
      "conv4_11_3x3\n",
      "conv4_11_3x3_bn\n",
      "activation_114\n",
      "conv4_11_1x1_increase\n",
      "conv4_11_1x1_increase_bn\n",
      "add_33\n",
      "activation_115\n",
      "conv4_12_1x1_reduce\n",
      "conv4_12_1x1_reduce_bn\n",
      "activation_116\n",
      "zero_padding2d_34\n",
      "conv4_12_3x3\n",
      "conv4_12_3x3_bn\n",
      "activation_117\n",
      "conv4_12_1x1_increase\n",
      "conv4_12_1x1_increase_bn\n",
      "add_34\n",
      "activation_118\n",
      "conv4_13_1x1_reduce\n",
      "conv4_13_1x1_reduce_bn\n",
      "activation_119\n",
      "zero_padding2d_35\n",
      "conv4_13_3x3\n",
      "conv4_13_3x3_bn\n",
      "activation_120\n",
      "conv4_13_1x1_increase\n",
      "conv4_13_1x1_increase_bn\n",
      "add_35\n",
      "activation_121\n",
      "conv4_14_1x1_reduce\n",
      "conv4_14_1x1_reduce_bn\n",
      "activation_122\n",
      "zero_padding2d_36\n",
      "conv4_14_3x3\n",
      "conv4_14_3x3_bn\n",
      "activation_123\n",
      "conv4_14_1x1_increase\n",
      "conv4_14_1x1_increase_bn\n",
      "add_36\n",
      "activation_124\n",
      "conv4_15_1x1_reduce\n",
      "conv4_15_1x1_reduce_bn\n",
      "activation_125\n",
      "zero_padding2d_37\n",
      "conv4_15_3x3\n",
      "conv4_15_3x3_bn\n",
      "activation_126\n",
      "conv4_15_1x1_increase\n",
      "conv4_15_1x1_increase_bn\n",
      "add_37\n",
      "activation_127\n",
      "conv4_16_1x1_reduce\n",
      "conv4_16_1x1_reduce_bn\n",
      "activation_128\n",
      "zero_padding2d_38\n",
      "conv4_16_3x3\n",
      "conv4_16_3x3_bn\n",
      "activation_129\n",
      "conv4_16_1x1_increase\n",
      "conv4_16_1x1_increase_bn\n",
      "add_38\n",
      "activation_130\n",
      "conv4_17_1x1_reduce\n",
      "conv4_17_1x1_reduce_bn\n",
      "activation_131\n",
      "zero_padding2d_39\n",
      "conv4_17_3x3\n",
      "conv4_17_3x3_bn\n",
      "activation_132\n",
      "conv4_17_1x1_increase\n",
      "conv4_17_1x1_increase_bn\n",
      "add_39\n",
      "activation_133\n",
      "conv4_18_1x1_reduce\n",
      "conv4_18_1x1_reduce_bn\n",
      "activation_134\n",
      "zero_padding2d_40\n",
      "conv4_18_3x3\n",
      "conv4_18_3x3_bn\n",
      "activation_135\n",
      "conv4_18_1x1_increase\n",
      "conv4_18_1x1_increase_bn\n",
      "add_40\n",
      "activation_136\n",
      "conv4_19_1x1_reduce\n",
      "conv4_19_1x1_reduce_bn\n",
      "activation_137\n",
      "zero_padding2d_41\n",
      "conv4_19_3x3\n",
      "conv4_19_3x3_bn\n",
      "activation_138\n",
      "conv4_19_1x1_increase\n",
      "conv4_19_1x1_increase_bn\n",
      "add_41\n",
      "activation_139\n",
      "conv4_20_1x1_reduce\n",
      "conv4_20_1x1_reduce_bn\n",
      "activation_140\n",
      "zero_padding2d_42\n",
      "conv4_20_3x3\n",
      "conv4_20_3x3_bn\n",
      "activation_141\n",
      "conv4_20_1x1_increase\n",
      "conv4_20_1x1_increase_bn\n",
      "add_42\n",
      "activation_142\n",
      "conv4_21_1x1_reduce\n",
      "conv4_21_1x1_reduce_bn\n",
      "activation_143\n",
      "zero_padding2d_43\n",
      "conv4_21_3x3\n",
      "conv4_21_3x3_bn\n",
      "activation_144\n",
      "conv4_21_1x1_increase\n",
      "conv4_21_1x1_increase_bn\n",
      "add_43\n",
      "activation_145\n",
      "conv4_22_1x1_reduce\n",
      "conv4_22_1x1_reduce_bn\n",
      "activation_146\n",
      "zero_padding2d_44\n",
      "conv4_22_3x3\n",
      "conv4_22_3x3_bn\n",
      "activation_147\n",
      "conv4_22_1x1_increase\n",
      "conv4_22_1x1_increase_bn\n",
      "add_44\n",
      "activation_148\n",
      "conv4_23_1x1_reduce\n",
      "conv4_23_1x1_reduce_bn\n",
      "activation_149\n",
      "zero_padding2d_45\n",
      "conv4_23_3x3\n",
      "conv4_23_3x3_bn\n",
      "activation_150\n",
      "conv4_23_1x1_increase\n",
      "conv4_23_1x1_increase_bn\n",
      "add_45\n",
      "activation_151\n",
      "conv5_1_1x1_reduce\n",
      "conv5_1_1x1_reduce_bn\n",
      "activation_152\n",
      "zero_padding2d_46\n",
      "conv5_1_3x3\n",
      "conv5_1_3x3_bn\n",
      "activation_153\n",
      "conv5_1_1x1_increase\n",
      "conv5_1_1x1_proj\n",
      "conv5_1_1x1_increase_bn\n",
      "conv5_1_1x1_proj_bn\n",
      "add_46\n",
      "activation_154\n",
      "conv5_2_1x1_reduce\n",
      "conv5_2_1x1_reduce_bn\n",
      "activation_155\n",
      "zero_padding2d_47\n",
      "conv5_2_3x3\n",
      "conv5_2_3x3_bn\n",
      "activation_156\n",
      "conv5_2_1x1_increase\n",
      "conv5_2_1x1_increase_bn\n",
      "add_47\n",
      "activation_157\n",
      "conv5_3_1x1_reduce\n",
      "conv5_3_1x1_reduce_bn\n",
      "activation_158\n",
      "zero_padding2d_48\n",
      "conv5_3_3x3\n",
      "conv5_3_3x3_bn\n",
      "activation_159\n",
      "conv5_3_1x1_increase\n",
      "conv5_3_1x1_increase_bn\n",
      "add_48\n",
      "activation_160\n",
      "average_pooling2d_7\n",
      "average_pooling2d_6\n",
      "average_pooling2d_5\n",
      "average_pooling2d_4\n",
      "conv5_3_pool6_conv\n",
      "conv5_3_pool3_conv\n",
      "conv5_3_pool2_conv\n",
      "conv5_3_pool1_conv\n",
      "conv5_3_pool6_conv_bn\n",
      "conv5_3_pool3_conv_bn\n",
      "conv5_3_pool2_conv_bn\n",
      "conv5_3_pool1_conv_bn\n",
      "activation_164\n",
      "activation_163\n",
      "activation_162\n",
      "activation_161\n",
      "interp_8\n",
      "interp_7\n",
      "interp_6\n",
      "interp_5\n",
      "concatenate_1\n",
      "conv5_4\n",
      "conv5_4_bn\n",
      "activation_165\n",
      "dropout_1\n",
      "conv6\n",
      "interp_9\n",
      "activation_166\n",
      "Finished importing weights.\n",
      "Writing keras model & weights\n",
      "Finished writing Keras model & weights\n"
     ],
     "output_type": "stream"
    }
   ],
   "source": [
    "pspnet_model = PSPNet(nb_classes=nb_classes, input_shape=input_shape, weights=f'{pspnet_layers}_{pretrained_model_name}')\n",
    "# pspnet_model.model.summary()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n",
     "is_executing": false
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "outputs": [],
   "source": [
    "# They will get saved into `weights/keras` as well\n",
    "# These cells save in the root directory\n",
    "with open(f'{pspnet_layers}_{pretrained_model_name}.json', 'w+') as json_file:\n",
    "    print(pspnet_model.model.to_json(), file=json_file)\n",
    "    \n",
    "tf.keras.models.save_model(pspnet_model.model, f'{pspnet_layers}_{pretrained_model_name}.h5')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n",
     "is_executing": false
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "outputs": [],
   "source": [
    "from layers_builder import Interp\n",
    "with open(f'{pspnet_layers}_{pretrained_model_name}.json', 'r') as json_model_file:\n",
    "    model_json = json_model_file.read()\n",
    "    model = tf.keras.models.model_from_json(model_json, custom_objects={'Interp': Interp})"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n",
     "is_executing": false
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "pycharm": {
     "is_executing": false,
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "model.load_weights(f'{pspnet_layers}_{pretrained_model_name}.h5')\n"
   ]
  }
 ],
 "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": "3.7.5"
  },
  "pycharm": {
   "stem_cell": {
    "cell_type": "raw",
    "source": [],
    "metadata": {
     "collapsed": false
    }
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}