Corrected metalearning example

metalearning_qaoa/metalearning_qaoa.ipynb

@@ -72,7 +72,7 @@
         "\n",
         "Created : 2020-Feb-06\n",
         "\n",
-        "Last updated : 2020-Mar-02"
+        "Last updated : 2020-Apr-09"
       ]
     },
     {
@@ -175,11 +175,11 @@
       "source": [
         "The QAOA ansatz consists of repeated applications of a mixer Hamiltonian $\\hat{H}_M$ and the cost Hamiltonian $\\hat{H}_C$.  The total applied unitary is\n",
         "$$\\hat{U}(\\eta,\\gamma) = \\prod_{j=1}^{p}e^{-i\\eta_{j}\\hat{H}_M}e^{-i\\gamma_{j} \\hat{H}_C},$$\n",
-        "where $p$ is the number of timesthe mixer and cost are applied; the parameters $\\eta_j, \\gamma_j$ are to be optimized to produce a bitstring of minimal energy with respect to $\\hat{H}_C$.\n",
+        "where $p$ is the number of times the mixer and cost are applied; the parameters $\\eta_j, \\gamma_j$ are to be optimized to produce a bitstring of minimal energy with respect to $\\hat{H}_C$.\n",
         "\n",
         "One traditional family of Hamiltonians used in QAOA are the Ising models.  These are defined as\n",
         "$$\\hat{H}_\\mathrm{P}=\\sum_i h_i \\hat{Z}^{(i)}+\\sum_{i,j} J_{ij} \\hat{Z}^{(i)}\\hat{Z}^{(j)}.$$\n",
-        "There is a one-to-one mapping between weighted graphs and Ising models: $h_i$ can be thought of as the weight of a graph node $i$ and $J_{ij}$ can be thought of as the weight of a graph edge between nodes $i$ and $j$.  In applications such as [MaxCut](https://en.wikipedia.org/wiki/Maximum_cut), we have $h_i = 0$ and $J_{ij} = 1 \\forall i, j$.  We this define a function that takes a graph and outputs the corresponding Ising model:"
+        "There is a one-to-one mapping between weighted graphs and Ising models: $h_i$ can be thought of as the weight of a graph node $i$ and $J_{ij}$ can be thought of as the weight of a graph edge between nodes $i$ and $j$.  In applications such as [MaxCut](https://en.wikipedia.org/wiki/Maximum_cut), we have $h_i = 0$ and $J_{ij} = 1$ for all indices $i$ and $j$.  The importance of this graph correspondence motivates us to define the following function, which takes a graph and returns the corresponding Ising model:"
       ]
     },
     {
@@ -294,7 +294,7 @@
         "    ops = inputs[1]\n",
         "    state = inputs[2]\n",
         "    params = inputs[3]\n",
-        "    prev_output = inputs[3]\n",
+        "    prev_output = inputs[4]\n",
         "    joined = tf.keras.layers.concatenate([state, params, prev_output])\n",
         "    shared = self.shared(joined)\n",
         "    s_inp = self.state(shared)\n",
@@ -421,15 +421,15 @@
         "\n",
         "# Our model will output it's parameter guesses along with the loss value that is\n",
         "# computed over them. This way we can use the model to guess parameters later on\n",
-        "model = tf.keras.Model(inputs=[state_inp, params_inp, exp_inp, op_inp, circuit_inp],\n",
+        "model = tf.keras.Model(inputs=[circuit_inp, op_inp, state_inp, params_inp, exp_inp],\n",
         "    outputs=[\n",
         "        output_0[3], output_1[3], output_2[3], output_3[3], output_4[3],\n",
         "        full_loss\n",
         "    ])\n",
         "model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001),\n",
         "              loss=value_loss, loss_weights=[0, 0, 0, 0, 0, 1])\n",
         "\n",
-        "model.fit(x=[initial_state, initial_params, initial_exp, ops_tensor, circuit_tensor],\n",
+        "model.fit(x=[circuit_tensor, ops_tensor, initial_state, initial_params, initial_exp],\n",
         "          y=[\n",
         "              np.zeros((N_POINTS, 1)),\n",
         "              np.zeros((N_POINTS, 1)),\n",
@@ -473,7 +473,7 @@
         "initial_exp = np.zeros((N_POINTS // 2, 25)).astype(np.float32)\n",
         "\n",
         "out1, out2, out3, out4, out5, _ = model(\n",
-        "    [initial_state, initial_guesses, initial_exp, ops_tensor, circuit_tensor])\n",
+        "    [circuit_tensor, ops_tensor, initial_state, initial_guesses, initial_exp])\n",
         "\n",
         "one_vals = tf.reduce_mean(tfq.layers.Expectation()(\n",
         "    circuit_tensor,\n",
@@ -564,11 +564,11 @@
         "plt.imshow(output_vals)\n",
         "\n",
         "guess_0, guess_1, guess_2, guess_3, guess_4, _ = model([\n",
+        "    tfq.convert_to_tensor([test_graph_circuit]),\n",
+        "    tfq.convert_to_tensor([[test_graph_op]]),\n",
         "    np.zeros((1, 25)).astype(np.float32),\n",
         "    np.zeros((1, 2)).astype(np.float32),\n",
         "    np.zeros((1, 25)).astype(np.float32),\n",
-        "    tfq.convert_to_tensor([[test_graph_op]]),\n",
-        "    tfq.convert_to_tensor([test_graph_circuit]),    \n",
         "])\n",
         "all_guesses = [guess_0, guess_1, guess_2, guess_3, guess_4]\n",
         "all_guesses = [list(a.numpy()[0]) for a in all_guesses]\n",