caffe2keras.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import caffe.proto.caffe_pb2 as caffe\n",
    "# import caffe_pb2 as caffe\n",
    "import numpy as np\n",
    "\n",
    "p = caffe.NetParameter()\n",
    "with open('/src/conv3d_deepnetA_sport1m_iter_1900000', 'rb') as f:\n",
    "    p.ParseFromString(\n",
    "        f.read()\n",
    "    )\n",
    "def rot90(W):\n",
    "    for i in range(W.shape[0]):\n",
    "        for j in range(W.shape[1]):\n",
    "            for k in range(W.shape[2]):\n",
    "                W[i, j, k] = np.rot90(W[i, j, k], 2)\n",
    "    return W"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "params = []\n",
    "conv_layers_indx = [1, 4, 7, 9, 12, 14, 17, 19]\n",
    "for i in conv_layers_indx:\n",
    "    layer = p.layers[i]\n",
    "    weights_b = np.array(layer.blobs[1].diff, dtype=np.float32)\n",
    "    weights_p = np.array(layer.blobs[0].diff, dtype=np.float32).reshape(\n",
    "        layer.blobs[0].num, layer.blobs[0].channels, -1,\n",
    "        layer.blobs[0].height, layer.blobs[0].width\n",
    "    )\n",
    "    weights_p = rot90(weights_p)\n",
    "    params.append([weights_p, weights_b])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "fc_layers_indx = [22, 25, 28]\n",
    "for i in fc_layers_indx:\n",
    "    layer = p.layers[i]\n",
    "    weights_b = np.array(layer.blobs[1].diff, dtype=np.float32)\n",
    "    weights_p = np.array(layer.blobs[0].diff, dtype=np.float32).reshape(\n",
    "        layer.blobs[0].num, layer.blobs[0].channels, \n",
    "        layer.blobs[0].height,\n",
    "        {22:8192, 25:4096, 28:4096}[i], layer.blobs[0].width)[0,0,0,:,:].T\n",
    "    params.append([weights_p, weights_b])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import pickle\n",
    "with open('/src/params.pkl', mode='wb') as f:\n",
    "    pickle.dump(params,f)\n",
    "# with open('/src/params.pickle', 'rb') as f:\n",
    "#     params = pickle.load(f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python2.7/dist-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n",
      "Using TensorFlow backend.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv1 (Conv3D)               (None, 16, 112, 112, 64)  5248      \n",
      "_________________________________________________________________\n",
      "pool1 (MaxPooling3D)         (None, 16, 56, 56, 64)    0         \n",
      "_________________________________________________________________\n",
      "conv2 (Conv3D)               (None, 16, 56, 56, 128)   221312    \n",
      "_________________________________________________________________\n",
      "pool2 (MaxPooling3D)         (None, 8, 28, 28, 128)    0         \n",
      "_________________________________________________________________\n",
      "conv3a (Conv3D)              (None, 8, 28, 28, 256)    884992    \n",
      "_________________________________________________________________\n",
      "conv3b (Conv3D)              (None, 8, 28, 28, 256)    1769728   \n",
      "_________________________________________________________________\n",
      "pool3 (MaxPooling3D)         (None, 4, 14, 14, 256)    0         \n",
      "_________________________________________________________________\n",
      "conv4a (Conv3D)              (None, 4, 14, 14, 512)    3539456   \n",
      "_________________________________________________________________\n",
      "conv4b (Conv3D)              (None, 4, 14, 14, 512)    7078400   \n",
      "_________________________________________________________________\n",
      "pool4 (MaxPooling3D)         (None, 2, 7, 7, 512)      0         \n",
      "_________________________________________________________________\n",
      "conv5a (Conv3D)              (None, 2, 7, 7, 512)      7078400   \n",
      "_________________________________________________________________\n",
      "conv5b (Conv3D)              (None, 2, 7, 7, 512)      7078400   \n",
      "_________________________________________________________________\n",
      "zeropad5 (ZeroPadding3D)     (None, 2, 8, 8, 512)      0         \n",
      "_________________________________________________________________\n",
      "pool5 (MaxPooling3D)         (None, 1, 4, 4, 512)      0         \n",
      "_________________________________________________________________\n",
      "flatten_1 (Flatten)          (None, 8192)              0         \n",
      "_________________________________________________________________\n",
      "fc6 (Dense)                  (None, 4096)              33558528  \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 4096)              0         \n",
      "_________________________________________________________________\n",
      "fc7 (Dense)                  (None, 4096)              16781312  \n",
      "_________________________________________________________________\n",
      "dropout_2 (Dropout)          (None, 4096)              0         \n",
      "_________________________________________________________________\n",
      "fc8 (Dense)                  (None, 487)               1995239   \n",
      "=================================================================\n",
      "Total params: 79,991,015\n",
      "Trainable params: 79,991,015\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "None\n"
     ]
    }
   ],
   "source": [
    "from keras.models import Sequential\n",
    "from keras.layers import Conv3D, MaxPool3D, ZeroPadding3D, Dense, Dropout, Flatten\n",
    "def get_model(summary=False, backend='tf'):\n",
    "    \"\"\" Return the Keras model of the network\n",
    "    \"\"\"\n",
    "    model = Sequential()\n",
    "    if backend == 'tf':\n",
    "        input_shape = (16, 112, 112, 3)  # l, h, w, c\n",
    "    else:\n",
    "        input_shape = (3, 16, 112, 112)  # c, l, h, w\n",
    "    model.add(Conv3D(64, (3, 3, 3), activation='relu',\n",
    "                     padding='same', name='conv1',\n",
    "                     input_shape=input_shape))\n",
    "    model.add(MaxPool3D(pool_size=(1, 2, 2), strides=(1, 2, 2),\n",
    "                           padding='valid', name='pool1'))\n",
    "    # 2nd layer group\n",
    "    model.add(Conv3D(128, (3, 3, 3), activation='relu',\n",
    "                     padding='same', name='conv2'))\n",
    "    model.add(MaxPool3D(pool_size=(2, 2, 2), strides=(2, 2, 2),\n",
    "                           padding='valid', name='pool2'))\n",
    "    # 3rd layer group\n",
    "    model.add(Conv3D(256, (3, 3, 3), activation='relu',\n",
    "                     padding='same', name='conv3a'))\n",
    "    model.add(Conv3D(256, (3, 3, 3), activation='relu',\n",
    "                     padding='same', name='conv3b'))\n",
    "    model.add(MaxPool3D(pool_size=(2, 2, 2), strides=(2, 2, 2),\n",
    "                           padding='valid', name='pool3'))\n",
    "    # 4th layer group\n",
    "    model.add(Conv3D(512, (3, 3, 3), activation='relu',\n",
    "                     padding='same', name='conv4a'))\n",
    "    model.add(Conv3D(512, (3, 3, 3), activation='relu',\n",
    "                     padding='same', name='conv4b'))\n",
    "    model.add(MaxPool3D(pool_size=(2, 2, 2), strides=(2, 2, 2),\n",
    "                           padding='valid', name='pool4'))\n",
    "    # 5th layer group\n",
    "    model.add(Conv3D(512, (3, 3, 3), activation='relu',\n",
    "                     padding='same', name='conv5a'))\n",
    "    model.add(Conv3D(512, (3, 3, 3), activation='relu',\n",
    "                     padding='same', name='conv5b'))\n",
    "    model.add(ZeroPadding3D(padding=((0, 0), (0, 1), (0, 1)),\n",
    "                            name='zeropad5'))\n",
    "    model.add(MaxPool3D(pool_size=(2, 2, 2), strides=(2, 2, 2),\n",
    "                           padding='valid', name='pool5'))\n",
    "    model.add(Flatten())\n",
    "    # FC layers group\n",
    "    model.add(Dense(4096, activation='relu', name='fc6'))\n",
    "    model.add(Dropout(.5))\n",
    "    model.add(Dense(4096, activation='relu', name='fc7'))\n",
    "    model.add(Dropout(.5))\n",
    "    model.add(Dense(487, activation='softmax', name='fc8'))\n",
    "    model.compile(optimizer='sgd', loss='mean_squared_error')\n",
    "\n",
    "    if summary:\n",
    "        print(model.summary())\n",
    "    return model\n",
    "model = get_model(summary=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(0, 0, 'conv1')\n",
      "((3, 3, 3, 3, 64), (64,), (3, 3, 3, 3, 64), (64,))\n",
      "(2, 1, 'conv2')\n",
      "((3, 3, 3, 64, 128), (128,), (3, 3, 3, 64, 128), (128,))\n",
      "(4, 2, 'conv3a')\n",
      "((3, 3, 3, 128, 256), (256,), (3, 3, 3, 128, 256), (256,))\n",
      "(5, 3, 'conv3b')\n",
      "((3, 3, 3, 256, 256), (256,), (3, 3, 3, 256, 256), (256,))\n",
      "(7, 4, 'conv4a')\n",
      "((3, 3, 3, 256, 512), (512,), (3, 3, 3, 256, 512), (512,))\n",
      "(8, 5, 'conv4b')\n",
      "((3, 3, 3, 512, 512), (512,), (3, 3, 3, 512, 512), (512,))\n",
      "(10, 6, 'conv5a')\n",
      "((3, 3, 3, 512, 512), (512,), (3, 3, 3, 512, 512), (512,))\n",
      "(11, 7, 'conv5b')\n",
      "((3, 3, 3, 512, 512), (512,), (3, 3, 3, 512, 512), (512,))\n",
      "(15, 8, 'fc6')\n",
      "((8192, 4096), (4096,), (8192, 4096), (4096,))\n",
      "(17, 9, 'fc7')\n",
      "((4096, 4096), (4096,), (4096, 4096), (4096,))\n",
      "(19, 10, 'fc8')\n",
      "((4096, 487), (487,), (4096, 487), (487,))\n"
     ]
    }
   ],
   "source": [
    "model_layers_indx = [0, 2, 4, 5, 7, 8, 10, 11] + [15, 17, 19]\n",
    "for j,i in enumerate(model_layers_indx):\n",
    "    print(i,j, model.layers[i].name)\n",
    "    wp,wb = model.layers[i].get_weights()\n",
    "    pp, pb = params[j]\n",
    "    print(wp.shape, wb.shape, pp.transpose().shape, pb.shape)\n",
    "    model.layers[i].set_weights([pp.transpose(), pb])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import h5py\n",
    "\n",
    "model.save_weights('/src/sports1M_weights.h5', overwrite=True)\n",
    "json_string = model.to_json()\n",
    "with open('/src/sports1M_model.json', 'w') as f:\n",
    "    f.write(json_string)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}