rdisipio

class MultiHeadAttentionQuantum(MultiHeadAttentionBase):
    def __init__(self, 
                 embed_dim, num_heads, 
                 n_qubits, n_qlayers=1, q_device='default.qubit'):
        super(MultiHeadAttentionQuantum, self).__init__(embed_dim, num_heads)
        # todo: add intermediate layer to "dress" quantum circuit
        assert n_qubits == embed_dim, f"Number of qubits ({n_qubits}) does not match embedding dim ({embed_dim})"
        self.dev = qml.device(q_device, wires=n_qubits)
        
        weight_shapes = {"weights": (n_qlayers, n_qubits)}

rdisipio

class MultiHeadAttentionClassical(MultiHeadAttentionBase):
    def __init__(self, embed_dim, num_heads):
        super(MultiHeadAttentionClassical, self).__init__(embed_dim, num_heads)
        self.wq = tf.keras.layers.Dense(embed_dim)
        self.wk = tf.keras.layers.Dense(embed_dim)
        self.wv = tf.keras.layers.Dense(embed_dim)
        self.dense = tf.keras.layers.Dense(embed_dim)
    
    def apply_dense_layers(self, v, k, q):
        q = self.wq(q)  # (batch_size, seq_len, embed_dim)

rdisipio

class LSTMTagger(nn.Module):
    def __init__(self, embedding_dim, hidden_dim, vocab_size, tagset_size, n_qubits=0):
        super(LSTMTagger, self).__init__()
        self.hidden_dim = hidden_dim

        self.word_embeddings = nn.Embedding(vocab_size, embedding_dim)

        # The LSTM takes word embeddings as inputs, and outputs hidden states
        # with dimensionality hidden_dim.
        if n_qubits > 0:

rdisipio

concat_size = inputs_dim + hidden_dim
clayer_in = torch.nn.Linear(concat_size, n_qubits)
VQC = [qml.qnn.TorchLayer(qlayer, weight_shapes) for _ in range(4)]
clayer_out = torch.nn.Linear(n_qubits, hidden_size)

hidden_seq = []
for t in range(seq_length):
    # get features from the t-th element in seq, for all entries in the batch
    x_t = x[:, t, :]
            

rdisipio

n_qubits = 4
dev = qml.device("default.qubit", wires=n_qubits)
def _circuit(inputs, weights):
    qml.templates.AngleEmbedding(inputs, wires=range(n_qubits))
    qml.templates.BasicEntanglerLayers(weights, wires=range(n_qubits))
    return [qml.expval(qml.PauliZ(wires=i)) for i in range(n_qubits)]
qlayer = qml.QNode(_circuit, dev, interface="torch")

rdisipio

def make_model_quantum(n_categories, n_qubits=4, n_layers=2, embedding_dim=512):

    text_in = keras.Input( shape=(embedding_dim,), dtype=tf.float64, name='text_in')

    x = layers.Dense(n_qubits, activation='tanh', dtype=tf.float64)(text_in)

    x = VariationalQuantumCircuit(
            n_categories=n_categories, 
            n_qubits=n_qubits, 
            n_layers=n_layers)(x)

rdisipio

def make_model_classical(n_categories, latent_dim=16, embedding_dim=512):
    
    text_in = keras.Input( shape=(embedding_dim,), dtype=tf.float64, name='text_in') # (None, 512)

    x = layers.Dense(latent_dim, activation='tanh', dtype=tf.float64)(text_in)

    x_out = layers.Dense(n_categories, activation='softmax')(x)

    return keras.Model(inputs=text_in, outputs=x_out, name="ClassicalPreprintClassifier")

rdisipio

#!/usr/bin/env python

import os, sys
import pickle

import numpy as np
import pandas as pd
import urllib.request
import re
import feedparser

rdisipio

def top_k_filtering( logits, top_k = 5):
    
    # a[...,1] equivalent to a[: ,: ,1 ]
    indices_to_remove = logits < tf.math.top_k(logits,top_k)[0][..., -1, None]
    # indices_to_remove is a tensor of bool values e.g. [ True, False, False, ..., True ]

    # 1d indices
    idx_remove = tf.where( indices_to_remove == True )[:,-1]
    idx_keep = tf.where( indices_to_remove == False )[:,-1]