qTransformer is a quantum circuit neural network classifier based on chained quantum attention mechanism layers, residual layers, and feed-forward neural network layers. Attention is computed using a trainable, parameterized general two-body interaction between word embedding statevectors in a sentence system; residual connections are represented using CX-gates; feed-forward neural network layers are represented using custom variational ansatzes such as RealAmplitudes and EfficientSU2.
The construction for the circuit representing the model is as follows:
# ------------------------------------------------------------------------------
# Construct Circuit.
# Feature map encodes the word embedding vectors.
feature_map = PauliFeatureMap(n)
# Attention transforms the sentence system to contain information regarding how
# much each word qubit should attend to every other word qubit.
if include_attention:
attention_layers = [attention(n, l=i) for i in range(reps)]
# Residual adds information regarding the original word emebddings back to the
# post-attention sentence system.
if include_residual:
residual_layers = [residual(n) for _ in range(reps)]
# Ansatz serves as a feed-forward layer at the end of the transformer block.
feed_forward_layers = [
EfficientSU2(n, parameter_prefix=str(i)) for i in range(reps)
]
self.circuit = QuantumCircuit(2 * n if include_residual else n)
self.circuit.append(feature_map, range(n))
if include_residual:
self.circuit.append(feature_map, range(n, 2 * n))
for i in range(reps):
if include_attention:
self.circuit.append(attention_layers[i], range(n))
if include_residual:
self.circuit.append(residual_layers[i], range(2 * n))
self.circuit.append(feed_forward_layers[i], range(n))
# ------------------------------------------------------------------------------The wrapper class for the qTransformer lives in /qtransformer/__init__.py.