neural_state.py

# Copyright 2019 PIQuIL - All Rights Reserved.

# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at

#     http://www.apache.org/licenses/LICENSE-2.0

# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


import abc
from itertools import chain
from math import ceil

import numpy as np
import torch
from tqdm import tqdm, tqdm_notebook

from qucumber.callbacks import CallbackList, Timer
from qucumber.utils import cplx
from qucumber.utils.data import extract_refbasis_samples
from qucumber.utils.gradients_utils import vector_to_grads


class NeuralStateBase(abc.ABC):
    """Abstract Base Class for Neural Network Quantum States."""

    _stop_training = False

    @property
    def stop_training(self):
        """If `True`, will not train.

        If this property is set to `True` during the training cycle, training
        will terminate once the current batch or epoch ends (depending on when
        `stop_training` was set).
        """
        return self._stop_training

    @stop_training.setter
    def stop_training(self, new_val):
        if isinstance(new_val, bool):
            self._stop_training = new_val
        else:
            raise ValueError("stop_training must be a boolean value!")

    @property
    def max_size(self):
        """Maximum size of the Hilbert space for full enumeration"""
        return 20

    @property
    @abc.abstractmethod
    def networks(self):
        """A list of the names of the internal RBMs."""

    @property
    @abc.abstractmethod
    def rbm_am(self):
        """The RBM to be used to learn the wavefunction amplitude."""

    @rbm_am.setter
    @abc.abstractmethod
    def rbm_am(self, new_val):
        raise NotImplementedError

    @property
    @abc.abstractmethod
    def device(self):
        """The device that the model is on."""

    @device.setter
    @abc.abstractmethod
    def device(self, new_val):
        raise NotImplementedError

    def reinitialize_parameters(self):
        """Randomize the parameters of the internal RBMs."""
        for net in self.networks:
            getattr(self, net).initialize_parameters()

    def __getattr__(self, attr):
        return getattr(self.rbm_am, attr)

    def probability(self, v, Z=1.0):
        """Evaluates the probability of the given vector(s) of visible
        states. Assumes the visible states were measured in the computational
        basis.

        :param v: The visible states.
        :type v: torch.Tensor
        :param Z: The partition function / normalization constant.
                  Defaults to 1, producing unnormalized probabilities.
        :type Z: float

        :returns: The probability of the given vector(s) of visible units.
        :rtype: torch.Tensor
        """
        v = v.to(device=self.device, dtype=torch.double)
        return (-self.rbm_am.effective_energy(v)).exp() / Z

    def sample(self, k, num_samples=1, initial_state=None, overwrite=False):
        r"""Performs k steps of Block Gibbs sampling. One step consists of sampling
        the hidden state :math:`\bm{h}` from the conditional distribution
        :math:`p_{\bm{\lambda}}(\bm{h}\:|\:\bm{v})`, and sampling the visible
        state :math:`\bm{v}` from the conditional distribution
        :math:`p_{\bm{\lambda}}(\bm{v}\:|\:\bm{h})`.

        :param k: Number of Block Gibbs steps.
        :type k: int
        :param num_samples: The number of samples to generate.
        :type num_samples: int
        :param initial_state: The initial state of the Markov Chains. If given,
                              `num_samples` will be ignored.
        :type initial_state: torch.Tensor
        :param overwrite: Whether to overwrite the initial_state tensor, if it is provided.
        :type overwrite: bool
        """
        if initial_state is None:
            dist = torch.distributions.Bernoulli(probs=0.5)
            sample_size = torch.Size((num_samples, self.num_visible))
            initial_state = dist.sample(sample_size).to(
                device=self.device, dtype=torch.double
            )

        return self.rbm_am.gibbs_steps(k, initial_state, overwrite=overwrite)

    def subspace_vector(self, num, size=None, device=None):
        r"""Generates a single vector from the Hilbert space of dimension
        :math:`2^{\text{size}}`.

        :param size: The size of each element of the Hilbert space.
        :type size: int
        :param num: The specific vector to return from the Hilbert space. Since
                    the Hilbert space can be represented by the set of binary strings
                    of length `size`, `num` is equivalent to the decimal representation
                    of the returned vector.
        :type num: int
        :param device: The device to create the vector on. Defaults to the
                       device this model is on.

        :returns: A state from the Hilbert space.
        :rtype: torch.Tensor
        """
        device = device if device is not None else self.device
        size = size if size else self.num_visible
        space = ((num & (1 << np.arange(size))) > 0)[::-1]
        space = space.astype(int)
        return torch.tensor(space, dtype=torch.double, device=device)

    def generate_hilbert_space(self, size=None, device=None):
        r"""Generates Hilbert space of dimension :math:`2^{\text{size}}`.

        :param size: The size of each element of the Hilbert space. Defaults to
                     the number of visible units.
        :type size: int
        :param device: The device to create the Hilbert space matrix on.
                       Defaults to the device this model is on.

        :returns: A tensor with all the basis states of the Hilbert space.
        :rtype: torch.Tensor
        """
        device = device if device is not None else self.device
        size = size if size else self.rbm_am.num_visible
        if size > self.max_size:
            raise ValueError("Size of the Hilbert space is too large!")
        else:
            dim = np.arange(2 ** size)
            space = ((dim[:, None] & (1 << np.arange(size))) > 0)[:, ::-1]
            space = space.astype(int)
            return torch.tensor(space, dtype=torch.double, device=device)

    def normalization(self, space):
        r"""Compute the normalization constant of the state.
        In the case of a pure state, this is the norm of the unnormalized wavefunction.
        In the case of a mixed state, this is the trace of the unnormalized density
        matrix.

        .. math::

            Z_{\bm{\lambda}}=
            \sum_{\bm{\sigma}} p_{\bm{\lambda}}(\bm{\sigma})

        :param space: A rank 2 tensor of the entire visible space.
        :type space: torch.Tensor

        """
        return self.rbm_am.partition(space)

    def compute_normalization(self, space):
        """Alias for :func:`normalization<qucumber.nn_states.NeuralStateBase.normalization>`"""
        return self.normalization(space)

    def save(self, location, metadata=None):
        """Saves the NeuralState parameters to the given location along with
        any given metadata.

        :param location: The location to save the data.
        :type location: str or file
        :param metadata: Any extra metadata to store alongside the NeuralState
                         parameters.
        :type metadata: dict
        """
        # add extra metadata to dictionary before saving it to disk
        metadata = metadata if metadata else {}

        if hasattr(self, "unitary_dict"):
            if "unitary_dict" in metadata.keys():
                raise ValueError(
                    "Invalid key in metadata; unitary_dict cannot be a key!"
                )
            metadata["unitary_dict"] = self.unitary_dict

        # validate metadata
        for net in self.networks:
            if net in metadata.keys():
                raise ValueError(f"Invalid key in metadata; '{net}' cannot be a key!")

        data = {net: getattr(self, net).state_dict() for net in self.networks}
        data.update(**metadata)
        torch.save(data, location)

    def load(self, location):
        """Loads the NeuralState parameters from the given location ignoring any
        metadata stored in the file. Overwrites the NeuralState's parameters.

        .. note::
            The NeuralState object on which this function is called must
            have the same parameter shapes as the one who's parameters are being
            loaded.

        :param location: The location to load the NeuralState parameters from.
        :type location: str or file
        """
        state_dict = torch.load(location, map_location=self.device)

        for net in self.networks:
            getattr(self, net).load_state_dict(state_dict[net])

        if hasattr(self, "unitary_dict") and "unitary_dict" in state_dict.keys():
            self.unitary_dict = state_dict["unitary_dict"]

    @staticmethod
    @abc.abstractmethod
    def autoload(location, gpu=False):
        """Initializes a NeuralState from the parameters in the given
        location.

        :param location: The location to load the model parameters from.
        :type location: str or file
        :param gpu: Whether the returned model should be on the GPU.
        :type gpu: bool

        :returns: A new NeuralState initialized from the given parameters.
                  The returned NeuralState will be of whichever type this function
                  was called on. An error may be thrown if the loaded parameters
                  correspond to a different type of NeuralState than the caller.
        """

    @abc.abstractmethod
    def importance_sampling_numerator(self, vp, v):
        r"""Compute the numerator of the weight of sample `vp`,
        with respect to the sample `v`.

        In the case of a mixed state, this quantity is :math:`\rho(\bm{\sigma'}, \bm{\sigma})`,
        while in the pure case it is :math:`\psi(\bm{\sigma'})`

        :param vp: A batch containing the samples :math:`\bm{\sigma'}`
        :type vp: torch.Tensor
        :param v: A batch containing the samples :math:`\bm{\sigma}`
        :type v: torch.Tensor

        :returns: A complex tensor containing the numerator of the weights of
                  :math:`\bm{\sigma'}` with respect to :math:`\bm{\sigma}`
        :rtype: torch.Tensor
        """

    @abc.abstractmethod
    def importance_sampling_denominator(self, v):
        r"""Compute the denominator of the weight of an arbitrary sample,
        with respect to the sample `v`.

        In the case of a mixed state, this quantity is :math:`\rho(\bm{\sigma}, \bm{\sigma})`,
        while in the pure case it is :math:`\psi(\bm{\sigma'})`

        :param v: A batch containing the samples :math:`\bm{\sigma}`
        :type v: torch.Tensor

        :returns: A complex tensor containing the denominator of the weights
                  with respect to :math:`\bm{\sigma}`
        :rtype: torch.Tensor
        """

    def importance_sampling_weight(self, vp, v):
        r"""Compute the weight of sample `vp`, with respect to the sample `v`.

        In the case of a mixed state, this ratio is:

        .. math::
            \frac{\rho(\bm{\sigma'}, \bm{\sigma})}{\rho(\bm{\sigma}, \bm{\sigma})}

        While in the pure case:

        .. math::
            \frac{\psi(\bm{\sigma'})}{\psi(\bm{\sigma})}

        :param vp: A batch containing the samples :math:`\bm{\sigma'}`
        :type vp: torch.Tensor
        :param v: A batch containing the samples :math:`\bm{\sigma}`
        :type v: torch.Tensor

        :returns: A complex tensor containing the weights of :math:`\bm{\sigma'}` with
                  respect to :math:`\bm{\sigma}`
        :rtype: torch.Tensor
        """
        return cplx.elementwise_division(
            self.importance_sampling_numerator(vp, v),
            self.importance_sampling_denominator(v),
        )

    def gradient(self, samples, bases=None):
        r"""Compute the gradient of a batch of sample, measured in given bases.

        :param sample: A batch of samples to compute the gradient of.
        :type sample: numpy.ndarray
        :param basis: A batch of bases.
        :type basis: numpy.ndarray or list[str] or None

        :returns: A list of 2 tensors containing the accumulated gradients
                  of each of the internal RBMs.
        :rtype: list[torch.Tensor]
        """
        grad = [
            torch.zeros(
                getattr(self, net).num_pars, dtype=torch.double, device=self.device
            )
            for net in self.networks
        ]
        if bases is None:
            grad[0] = self.rbm_am.effective_energy_gradient(samples)
        else:
            if samples.dim() < 2:
                samples = samples.unsqueeze(0)
                bases = np.array(list(bases)).reshape(1, -1)

            unique_bases, indices = np.unique(bases, axis=0, return_inverse=True)
            indices = torch.Tensor(indices).to(samples)

            for i in range(unique_bases.shape[0]):
                basis = unique_bases[i, :]
                rot_sites = np.where(basis != "Z")[0]

                if rot_sites.size != 0:
                    sample_grad = self.rotated_gradient(basis, samples[indices == i, :])
                else:
                    sample_grad = [
                        self.rbm_am.effective_energy_gradient(samples[indices == i, :]),
                        0.0,
                    ]

                grad[0] += sample_grad[0]  # Accumulate amplitude RBM gradient
                grad[1] += sample_grad[1]  # Accumulate phase RBM gradient

        return grad

    def positive_phase_gradients(self, samples_batch, bases_batch=None):
        r"""Computes the positive phase of the gradients of the parameters.

        :param samples_batch: The measurements
        :type samples_batch: torch.Tensor
        :param bases_batch: The bases in which the measurements are made
        :type bases_batch: numpy.ndarray

        :returns: A two-element list containing the amplitude and phase RBM gradients
        :rtype: list[torch.Tensor]
        """
        grad = self.gradient(samples_batch, bases=bases_batch)
        grad = [gr / float(samples_batch.shape[0]) for gr in grad]
        return grad

    def compute_exact_gradients(self, samples_batch, space, bases_batch=None):
        r"""Computes the gradients of the parameters, using exact sampling
        for the negative phase update instead of Gibbs sampling

        :param samples_batch: The measurements
        :type samples_batch: torch.Tensor
        :param space: A rank 2 tensor of the entire visible space.
        :type space: torch.Tensor
        :param bases_batch: The bases in which the measurements are made
        :type bases_batch: numpy.ndarray

        :returns: A two-element list containing the amplitude and phase RBM
                  gradients calculated with an exact negative phase update
        :rtype: list[torch.Tensor]
        """
        # Positive phase: learning signal driven by the data (and bases)
        grad = self.positive_phase_gradients(samples_batch, bases_batch=bases_batch)

        # Negative phase: learning signal driven by the amplitude RBM of
        # the NN state
        probs = self.probability(space, Z=1.0)  # unnormalized probs
        Z = probs.sum()
        probs /= Z

        all_grads = self.rbm_am.effective_energy_gradient(space, reduce=False)
        grad[0] -= torch.mv(
            all_grads.t(), probs
        )  # average the gradients, weighted by probs

        return grad

    def compute_batch_gradients(self, k, samples_batch, neg_batch, bases_batch=None):
        """Compute the gradients of a batch of the training data (`samples_batch`).

        If measurements are taken in bases other than the reference basis,
        a list of bases (`bases_batch`) must also be provided.

        :param k: Number of contrastive divergence steps in training.
        :type k: int
        :param samples_batch: Batch of the input samples.
        :type samples_batch: torch.Tensor
        :param neg_batch: Batch of the input samples for computing the
                          negative phase.
        :type neg_batch: torch.Tensor
        :param bases_batch: Batch of the input bases corresponding to the samples
                            in `samples_batch`.
        :type bases_batch: numpy.ndarray

        :returns: A two-element list containing the amplitude and phase RBM
                  gradients calculated with a Gibbs sampled negative phase update
        :rtype: list[torch.Tensor]
        """
        # Positive phase: learning signal driven by the data (and bases)
        grad = self.positive_phase_gradients(samples_batch, bases_batch=bases_batch)

        # Negative phase: learning signal driven by the amplitude RBM of
        # the NN state
        vk = self.rbm_am.gibbs_steps(k, neg_batch)
        grad_model = self.rbm_am.effective_energy_gradient(vk)
        grad[0] -= grad_model / float(neg_batch.shape[0])
        # No negative signal for the phase parameters
        return grad

    def _shuffle_data(
        self,
        pos_batch_size,
        neg_batch_size,
        num_batches,
        train_samples,
        input_bases,
        z_samples,
    ):
        pos_batch_perm = torch.randperm(train_samples.shape[0])

        shuffled_pos_samples = train_samples[pos_batch_perm]
        if input_bases is None:
            if neg_batch_size == pos_batch_size:
                neg_batch_perm = pos_batch_perm
            else:
                neg_batch_perm = torch.randint(
                    train_samples.shape[0],
                    size=(num_batches * neg_batch_size,),
                    dtype=torch.long,
                )
            shuffled_neg_samples = train_samples[neg_batch_perm]
        else:
            neg_batch_perm = torch.randint(
                z_samples.shape[0],
                size=(num_batches * neg_batch_size,),
                dtype=torch.long,
            )
            shuffled_neg_samples = z_samples[neg_batch_perm]

        # List of all the batches for positive phase.
        pos_batches = [
            shuffled_pos_samples[batch_start : (batch_start + pos_batch_size)]
            for batch_start in range(0, len(shuffled_pos_samples), pos_batch_size)
        ]

        neg_batches = [
            shuffled_neg_samples[batch_start : (batch_start + neg_batch_size)]
            for batch_start in range(0, len(shuffled_neg_samples), neg_batch_size)
        ]

        if input_bases is not None:
            shuffled_pos_bases = input_bases[pos_batch_perm]
            pos_batches_bases = [
                shuffled_pos_bases[batch_start : (batch_start + pos_batch_size)]
                for batch_start in range(0, len(train_samples), pos_batch_size)
            ]
            return zip(pos_batches, neg_batches, pos_batches_bases)
        else:
            return zip(pos_batches, neg_batches)

    def fit(
        self,
        data,
        epochs=100,
        pos_batch_size=100,
        neg_batch_size=None,
        k=1,
        lr=1e-3,
        input_bases=None,
        progbar=False,
        starting_epoch=1,
        time=False,
        callbacks=None,
        optimizer=torch.optim.SGD,
        optimizer_args=None,
        scheduler=None,
        scheduler_args=None,
        **kwargs,
    ):
        r"""Train the NeuralState.

        :param data: The training samples
        :type data: numpy.ndarray
        :param epochs: The number of full training passes through the dataset.
                       Technically, this specifies the index of the *last* training
                       epoch, which is relevant if `starting_epoch` is being set.
        :type epochs: int
        :param pos_batch_size: The size of batches for the positive phase
                               taken from the data.
        :type pos_batch_size: int
        :param neg_batch_size: The size of batches for the negative phase
                               taken from the data. Defaults to `pos_batch_size`.
        :type neg_batch_size: int
        :param k: The number of contrastive divergence steps.
        :type k: int
        :param lr: Learning rate
        :type lr: float
        :param input_bases: The measurement bases for each sample. Must be provided
                            if training a ComplexWaveFunction or DensityMatrix.
        :type input_bases: numpy.ndarray
        :param progbar: Whether or not to display a progress bar. If "notebook"
                        is passed, will use a Jupyter notebook compatible
                        progress bar.
        :type progbar: bool or str
        :param starting_epoch: The epoch to start from. Useful if continuing training
                               from a previous state.
        :type starting_epoch: int
        :param callbacks: Callbacks to run while training.
        :type callbacks: list[qucumber.callbacks.CallbackBase]
        :param optimizer: The constructor of a torch optimizer.
        :type optimizer: torch.optim.Optimizer
        :param scheduler: The constructor of a torch scheduler
        :param optimizer_args: Arguments to pass to the optimizer
        :type optimizer_args: dict
        :param scheduler_args: Arguments to pass to the scheduler
        :type scheduler_args: dict
        :param \**kwargs: Ignored; exists for backwards compatibility.
        """
        if self.stop_training:  # terminate immediately if stop_training is true
            return

        disable_progbar = progbar is False
        progress_bar = tqdm_notebook if progbar == "notebook" else tqdm

        callbacks = CallbackList(callbacks if callbacks else [])
        if time:
            callbacks.append(Timer())

        neg_batch_size = neg_batch_size if neg_batch_size else pos_batch_size

        if isinstance(data, torch.Tensor):
            train_samples = (
                data.clone().detach().to(device=self.device, dtype=torch.double)
            )
        else:
            train_samples = torch.tensor(data, device=self.device, dtype=torch.double)

        all_params = [getattr(self, net).parameters() for net in self.networks]
        all_params = list(chain(*all_params))

        optimizer_args = {} if optimizer_args is None else optimizer_args
        scheduler_args = {} if scheduler_args is None else scheduler_args

        optimizer = optimizer(all_params, lr=lr, **optimizer_args)

        if scheduler is not None:
            scheduler = scheduler(optimizer, **scheduler_args)

        if input_bases is not None:
            z_samples = extract_refbasis_samples(train_samples, input_bases).to(
                device=self.device
            )
        else:
            z_samples = None

        callbacks.on_train_start(self)

        num_batches = ceil(train_samples.shape[0] / pos_batch_size)
        for ep in progress_bar(
            range(starting_epoch, epochs + 1), desc="Epochs ", disable=disable_progbar
        ):
            data_iterator = self._shuffle_data(
                pos_batch_size,
                neg_batch_size,
                num_batches,
                train_samples,
                input_bases,
                z_samples,
            )
            callbacks.on_epoch_start(self, ep)

            for b, batch in enumerate(data_iterator):
                callbacks.on_batch_start(self, ep, b)

                all_grads = self.compute_batch_gradients(k, *batch)

                optimizer.zero_grad()  # clear any cached gradients

                # assign gradients to corresponding parameters
                for i, net in enumerate(self.networks):
                    rbm = getattr(self, net)
                    vector_to_grads(all_grads[i], rbm.parameters())

                optimizer.step()  # tell the optimizer to apply the gradients

                callbacks.on_batch_end(self, ep, b)
                if self.stop_training:  # check for stop_training signal
                    break

            if scheduler is not None:
                scheduler.step()

            callbacks.on_epoch_end(self, ep)
            if self.stop_training:  # check for stop_training signal
                break

        callbacks.on_train_end(self)


# make module path show up properly in sphinx docs
NeuralStateBase.__module__ = "qucumber.nn_states"