Migrating TensorFlowBackend to Qiboml

doc/source/code-examples/advancedexamples.rst

@@ -78,19 +78,8 @@ or programmatically, during runtime, as follows:
     # retrieve the current number of threads
     current_threads = qibo.get_threads()
 
-On the other hand, when using the ``tensorflow`` backend Qibo inherits
-Tensorflow's defaults for CPU thread configuration.
-Tensorflow allows restricting the number of threads as follows:
-
-.. code-block:: python
-
-    import tensorflow as tf
-    tf.config.threading.set_inter_op_parallelism_threads(1)
-    tf.config.threading.set_intra_op_parallelism_threads(1)
-    import qibo
-
-Note that this should be run during Tensorflow initialization in the beginning
-of the script and before creating the qibo backend.
+For similar wariness when using a machine learning backend (such as TensorFlow or Pytorch)
+please refer to the Qiboml documentation.
 
 Using multiple GPUs
 ^^^^^^^^^^^^^^^^^^^
@@ -707,13 +696,18 @@ circuit output matches a target state using the fidelity as the corresponding lo
 Note that, as in the following example, the rotation angles have to assume real values
 to ensure the rotational gates are representing unitary operators.
 
+Qibo doesn't provide Tensorflow and Pytorch as native backends; Qiboml has to be
+installed and used as provider of these quantum machine learning backends.
+
 .. code-block:: python
 
     import qibo
-    qibo.set_backend("tensorflow")
-    import tensorflow as tf
+    qibo.set_backend(backend="qiboml", platform="tensorflow")
     from qibo import gates, models
 
+    backend = qibo.get_backend()
+    tf = backend.tf
+
     # Optimization parameters
     nepochs = 1000
     optimizer = tf.keras.optimizers.Adam()
@@ -737,8 +731,9 @@ to ensure the rotational gates are representing unitary operators.
         optimizer.apply_gradients(zip([grads], [params]))
 
 
-Note that the ``"tensorflow"`` backend has to be used here because other custom
-backends do not support automatic differentiation.
+Note that the ``"tensorflow"`` backend has to be used here since it provides
+automatic differentiation tools. To be constructed, the Qiboml package has to be
+installed and used.
 
 The optimization procedure may also be compiled, however in this case it is not
 possible to use :meth:`qibo.circuit.Circuit.set_parameters` as the
@@ -748,10 +743,12 @@ For example:
 .. code-block:: python
 
     import qibo
-    qibo.set_backend("tensorflow")
-    import tensorflow as tf
+    qibo.set_backend(backend="qiboml", platform="tensorflow")
     from qibo import gates, models
 
+    backend = qibo.get_backend()
+    tf = backend.tf
+
     nepochs = 1000
     optimizer = tf.keras.optimizers.Adam()
     target_state = tf.ones(4, dtype=tf.complex128) / 2.0

examples/anomaly_detection/test.py

@@ -3,7 +3,6 @@
 
 import matplotlib.pyplot as plt
 import numpy as np
-import tensorflow as tf
 
 import qibo
 from qibo import Circuit, gates
@@ -22,7 +21,8 @@ def main(n_layers, train_size, filename, plot, save_loss):
         save_loss (bool): save losses for standard and anomalous data (default False).
     """
 
-    qibo.set_backend("tensorflow")
+    qibo.set_backend(backend="qiboml", platform="tensorflow")
+    tf = qibo.get_backend().tf
 
     # Circuit ansatz
     def make_encoder(n_qubits, n_layers, params, q_compression):

examples/anomaly_detection/train.py

@@ -3,7 +3,6 @@
 from pathlib import Path
 
 import numpy as np
-import tensorflow as tf
 
 import qibo
 from qibo import Circuit, gates
@@ -23,7 +22,8 @@ def main(n_layers, batch_size, nepochs, train_size, filename, lr_boundaries):
         lr_boundaries (list): epochs when learning rate is reduced, 6 monotone growing values from 0 to nepochs (default [3,6,9,12,15,18]).
     """
 
-    qibo.set_backend("tensorflow")
+    qibo.set_backend(backend="qiboml", platform="tensorflow")
+    tf = qibo.get_backend().tf
 
     # Circuit ansatz
     def make_encoder(n_qubits, n_layers, params, q_compression):

examples/benchmarks/main.py

@@ -47,7 +47,10 @@ def limit_gpu_memory(memory_limit=None):
     Args:
         memory_limit: Memory limit in MBs.
     """
-    import tensorflow as tf
+    import qibo
+
+    qibo.set_backend(backend="qiboml", platform="tensorflow")
+    tf = qibo.get_backend().tf
 
     if memory_limit is None:
         print("\nNo GPU memory limiter used.\n")

examples/qclustering/minimization.py

@@ -1,10 +1,10 @@
 import math
 
 import numpy as np
-import tensorflow as tf
 from grover import grover_qc
 from oracle import create_oracle_circ
 
+import qibo
 from qibo import gates
 from qibo.models import Circuit
 
@@ -23,6 +23,8 @@ def duerr_hoyer_algo(distances):
     int
         New cluster assigned for that point.
     """
+    qibo.set_backend(backend="qiboml", platform="tensorflow")
+    tf = qibo.get_backend().tf
 
     k = len(distances)
     n = int(math.floor(math.log2(k)) + 1)

examples/reuploading_classifier/qlassifier.py

@@ -5,6 +5,7 @@
 from matplotlib.cm import get_cmap
 from matplotlib.colors import Normalize
 
+import qibo
 from qibo import Circuit, gates
 
 
@@ -117,7 +118,8 @@ def minimize(self, method="BFGS", options=None, compile=True):
             parameters = r[1].result.xbest
 
         elif method == "sgd":
-            import tensorflow as tf
+            qibo.set_backend(backend="qiboml", platform="tensorflow")
+            tf = qibo.get_backend().tf
 
             circuit = self.circuit(self.training_set[0])