https://github.com/keras-team/keras/issues/12334

keras_subclass_bug.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "from keras.layers import Dense, Input\n",
    "from keras.models import Model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Working Example"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Model1(Model):\n",
    "    def __init__(self, **kwargs):\n",
    "        super(Model1, self).__init__(**kwargs)\n",
    "        \n",
    "        self.m1 = Dense(64)\n",
    "        self.m2 = Dense(64)\n",
    "        self.m3 = Dense(64)\n",
    "        self.l = Dense(32, activation=\"relu\")\n",
    "    \n",
    "    def call(self, i):\n",
    "        o = i\n",
    "        o = self.m1(o)\n",
    "        o = self.m2(o)\n",
    "        o = self.m3(o)\n",
    "        return self.l(o)\n",
    "    \n",
    "    def compute_output_shape(self, input_shape):\n",
    "        return (input_shape[0], 32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "i = Input((32,))\n",
    "net = Model1()(i)\n",
    "net = Dense(2, activation=\"softmax\")(net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = Model(i, net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_1 (InputLayer)         (None, 32)                0         \n",
      "_________________________________________________________________\n",
      "model1_1 (Model1)            (None, 32)                12512     \n",
      "_________________________________________________________________\n",
      "dense_5 (Dense)              (None, 2)                 66        \n",
      "=================================================================\n",
      "Total params: 12,578\n",
      "Trainable params: 12,578\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Buggy Example"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Model2(Model):\n",
    "    def __init__(self, **kwargs):\n",
    "        super(Model2, self).__init__(**kwargs)\n",
    "        \n",
    "        self.m = [Dense(64) for _ in range(3)]\n",
    "        self.l = Dense(32, activation=\"relu\")\n",
    "    \n",
    "    def call(self, i):\n",
    "        o = i\n",
    "        for l in self.m:\n",
    "            print(l)\n",
    "            o = l(o)\n",
    "        return self.l(o)\n",
    "    \n",
    "    def compute_output_shape(self, input_shape):\n",
    "        return (input_shape[0], 32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<keras.layers.core.Dense object at 0xb387f5358>\n",
      "<keras.layers.core.Dense object at 0xb387f5278>\n",
      "<keras.layers.core.Dense object at 0xb387f57f0>\n"
     ]
    }
   ],
   "source": [
    "i = Input((32,))\n",
    "net = Model2()(i)\n",
    "net = Dense(2, activation=\"softmax\")(net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = Model(i, net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_2 (InputLayer)         (None, 32)                0         \n",
      "_________________________________________________________________\n",
      "model2_1 (Model2)            (None, 32)                2080      \n",
      "_________________________________________________________________\n",
      "dense_10 (Dense)             (None, 2)                 66        \n",
      "=================================================================\n",
      "Total params: 2,146\n",
      "Trainable params: 2,146\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "NOTE: should say 12,578"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<keras.engine.input_layer.InputLayer at 0xb387f52b0>,\n",
       " <__main__.Model2 at 0xb387f51d0>,\n",
       " <keras.layers.core.Dense at 0xb387f5400>]"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.layers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<keras.layers.core.Dense at 0xb387f5898>]"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.layers[1].layers"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "NOTE: only 1 instead of 4"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### tensorflow.keras"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Model3(tf.keras.models.Model):\n",
    "    def __init__(self, **kwargs):\n",
    "        super(Model3, self).__init__(**kwargs)\n",
    "        \n",
    "        self.m = [tf.keras.layers.Dense(64) for _ in range(3)]\n",
    "        self.l = tf.keras.layers.Dense(32, activation=\"relu\")\n",
    "    \n",
    "    def call(self, i):\n",
    "        o = i\n",
    "        for l in self.m:\n",
    "            print(l)\n",
    "            o = l(o)\n",
    "        return self.l(o)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<tensorflow.python.keras.layers.core.Dense object at 0xb388c9be0>\n",
      "<tensorflow.python.keras.layers.core.Dense object at 0xb388c9e48>\n",
      "<tensorflow.python.keras.layers.core.Dense object at 0xb388d40b8>\n"
     ]
    }
   ],
   "source": [
    "i = tf.keras.layers.Input((32,))\n",
    "net = Model3()(i)\n",
    "net = tf.keras.layers.Dense(2, activation=\"softmax\")(net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = tf.keras.models.Model(i, net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_1 (InputLayer)         (None, 32)                0         \n",
      "_________________________________________________________________\n",
      "model3 (Model3)              (None, 32)                12512     \n",
      "_________________________________________________________________\n",
      "dense_4 (Dense)              (None, 2)                 66        \n",
      "=================================================================\n",
      "Total params: 12,578\n",
      "Trainable params: 12,578\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<tensorflow.python.keras.engine.input_layer.InputLayer at 0xb388c90b8>,\n",
       " <__main__.Model3 at 0xb388c9080>,\n",
       " <tensorflow.python.keras.layers.core.Dense at 0xb388c9a20>]"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.layers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<tensorflow.python.keras.layers.core.Dense at 0xb388c9be0>,\n",
       " <tensorflow.python.keras.layers.core.Dense at 0xb388c9e48>,\n",
       " <tensorflow.python.keras.layers.core.Dense at 0xb388d40b8>,\n",
       " <tensorflow.python.keras.layers.core.Dense at 0xb388d4240>]"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.layers[1].layers"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "NOTE: working as expected"
   ]
  }
 ],
 "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": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}