first (to be fixed) layers test with keras

src/qiboml/models/layers/__init__.py

@@ -0,0 +1 @@
+from qiboml.models.layers.training import QuantumLayer

src/qiboml/models/layers/encoding.py

@@ -0,0 +1 @@
+"""Encoding layers with Qiboml."""

src/qiboml/models/layers/training.py

@@ -0,0 +1,32 @@
+import keras
+import qibo
+import tensorflow as tf
+
+
+class QuantumLayer(keras.layers.Layer):
+    def __init__(self, circuit: qibo.Circuit, **kwargs):
+        super().__init__(**kwargs)
+        self.circuit = circuit
+        nparams = len(circuit.get_parameters())
+        self.qparams = self.add_weight(
+            shape=(nparams,),
+            initializer="random_normal",
+            trainable=True,
+            dtype="float64",
+        )
+
+    def call(self, inputs):
+        # Check input shape
+        if inputs.shape[1] != 2**self.circuit.nqubits:
+            raise ValueError(
+                f"Input of a QuantumLayer has to be a vector of length 2**{self.circuit.nqubits}"
+            )
+        # Update circuit parameters with trainable weights
+        self.circuit.set_parameters(self.qparams.value)
+        # Execute circuit
+        state = self.circuit(initial_state=inputs).state()
+        # Ensure return type is compatible with TensorFlow
+        return tf.abs(state)
+
+    def compute_output_shape(self, input_shape):
+        return (input_shape[0], 2**self.circuit.nqubits)

test_keras.py

@@ -0,0 +1,49 @@
+import os
+
+import keras
+import matplotlib.pyplot as plt
+import numpy as np
+
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+
+import tensorflow as tf
+
+tf.get_logger().setLevel("ERROR")
+
+
+import qibo
+
+import qiboml
+import qiboml.models
+import qiboml.models.layers
+from qiboml.models.ansatze import reuploading_circuit
+
+qibo.set_backend("tensorflow")
+
+x = np.linspace(-1, 1, 500)
+y = np.sin(6 * x) ** 2 + np.random.normal(0, 0.1, 500)
+
+nqubits = 3
+nlayers = 1
+
+circuit = reuploading_circuit(nqubits, nlayers)
+
+model = keras.Sequential()
+model.add(keras.layers.Input(shape=(1,)))
+model.add(keras.layers.Dense(2**nqubits, activation="relu"))
+model.add(qiboml.models.layers.QuantumLayer(circuit=circuit))
+model.add(keras.layers.Dense(32, input_shape=(8,), activation="relu"))
+model.add(keras.layers.Dense(1, activation="linear"))
+model.compile(optimizer="adam", loss="mse")
+model.summary()
+
+model.fit(x, y, batch_size=32, epochs=100)
+
+test_x = np.linspace(-1, 1, 100)
+test_y = model.predict(test_x)
+
+plt.figure(figsize=(6, 6 * 6 / 8))
+plt.plot(x, y, marker=".", color="royalblue", label="Training")
+plt.plot(test_x, test_y, marker=".", color="red", label="Test")
+plt.legend()
+plt.savefig("data.png")