seqtoseq.py

"""Sequence to sequence translation models."""

from deepchem.models import KerasModel, layers
from heapq import heappush, heappushpop
import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import Input, Layer, Dense, Dropout, GRU, Lambda, Conv1D, Flatten, BatchNormalization


class VariationalRandomizer(Layer):
    """Add random noise to the embedding and include a corresponding loss."""

    def __init__(self, embedding_dimension, annealing_start_step,
                 annealing_final_step, **kwargs):
        super(VariationalRandomizer, self).__init__(**kwargs)
        self._embedding_dimension = embedding_dimension
        self._annealing_final_step = annealing_final_step
        self._annealing_start_step = annealing_start_step
        self.dense_mean = Dense(embedding_dimension)
        self.dense_stddev = Dense(embedding_dimension)
        self.combine = layers.CombineMeanStd(training_only=True)

    def call(self, inputs, training=True):
        input, global_step = inputs
        embedding_mean = self.dense_mean(input)
        embedding_stddev = self.dense_stddev(input)
        embedding = self.combine([embedding_mean, embedding_stddev],
                                 training=training)
        mean_sq = embedding_mean * embedding_mean
        stddev_sq = embedding_stddev * embedding_stddev
        kl = mean_sq + stddev_sq - tf.math.log(stddev_sq + 1e-20) - 1
        anneal_steps = self._annealing_final_step - self._annealing_start_step
        if anneal_steps > 0:
            current_step = tf.cast(global_step,
                                   tf.float32) - self._annealing_start_step
            anneal_frac = tf.maximum(0.0, current_step) / anneal_steps
            kl_scale = tf.minimum(1.0, anneal_frac * anneal_frac)
        else:
            kl_scale = 1.0
        self.add_loss(0.5 * kl_scale * tf.reduce_mean(kl))
        return embedding


class SeqToSeq(KerasModel):
    """Implements sequence to sequence translation models.

    The model is based on the description in Sutskever et al., "Sequence to
    Sequence Learning with Neural Networks" (https://arxiv.org/abs/1409.3215),
    although this implementation uses GRUs instead of LSTMs.  The goal is to
    take sequences of tokens as input, and translate each one into a different
    output sequence.  The input and output sequences can both be of variable
    length, and an output sequence need not have the same length as the input
    sequence it was generated from.  For example, these models were originally
    developed for use in natural language processing.  In that context, the
    input might be a sequence of English words, and the output might be a
    sequence of French words.  The goal would be to train the model to translate
    sentences from English to French.

    The model consists of two parts called the "encoder" and "decoder".  Each one
    consists of a stack of recurrent layers.  The job of the encoder is to
    transform the input sequence into a single, fixed length vector called the
    "embedding".  That vector contains all relevant information from the input
    sequence.  The decoder then transforms the embedding vector into the output
    sequence.

    These models can be used for various purposes.  First and most obviously,
    they can be used for sequence to sequence translation.  In any case where you
    have sequences of tokens, and you want to translate each one into a different
    sequence, a SeqToSeq model can be trained to perform the translation.

    Another possible use case is transforming variable length sequences into
    fixed length vectors.  Many types of models require their inputs to have a
    fixed shape, which makes it difficult to use them with variable sized inputs
    (for example, when the input is a molecule, and different molecules have
    different numbers of atoms).  In that case, you can train a SeqToSeq model as
    an autoencoder, so that it tries to make the output sequence identical to the
    input one.  That forces the embedding vector to contain all information from
    the original sequence.  You can then use the encoder for transforming
    sequences into fixed length embedding vectors, suitable to use as inputs to
    other types of models.

    Another use case is to train the decoder for use as a generative model.  Here
    again you begin by training the SeqToSeq model as an autoencoder.  Once
    training is complete, you can supply arbitrary embedding vectors, and
    transform each one into an output sequence.  When used in this way, you
    typically train it as a variational autoencoder.  This adds random noise to
    the encoder, and also adds a constraint term to the loss that forces the
    embedding vector to have a unit Gaussian distribution.  You can then pick
    random vectors from a Gaussian distribution, and the output sequences should
    follow the same distribution as the training data.

    When training as a variational autoencoder, it is best to use KL cost
    annealing, as described in https://arxiv.org/abs/1511.06349.  The constraint
    term in the loss is initially set to 0, so the optimizer just tries to
    minimize the reconstruction loss.  Once it has made reasonable progress
    toward that, the constraint term can be gradually turned back on.  The range
    of steps over which this happens is configurable.
    """

    sequence_end = object()

    def __init__(self,
                 input_tokens,
                 output_tokens,
                 max_output_length,
                 encoder_layers=4,
                 decoder_layers=4,
                 embedding_dimension=512,
                 dropout=0.0,
                 reverse_input=True,
                 variational=False,
                 annealing_start_step=5000,
                 annealing_final_step=10000,
                 **kwargs):
        """Construct a SeqToSeq model.

        In addition to the following arguments, this class also accepts all the keyword arguments
        from TensorGraph.

        Parameters
        ----------
        input_tokens: list
            a list of all tokens that may appear in input sequences
        output_tokens: list
            a list of all tokens that may appear in output sequences
        max_output_length: int
            the maximum length of output sequence that may be generated
        encoder_layers: int
            the number of recurrent layers in the encoder
        decoder_layers: int
            the number of recurrent layers in the decoder
        embedding_dimension: int
            the width of the embedding vector.  This also is the width of all
            recurrent layers.
        dropout: float
            the dropout probability to use during training
        reverse_input: bool
            if True, reverse the order of input sequences before sending them into
            the encoder.  This can improve performance when working with long sequences.
        variational: bool
            if True, train the model as a variational autoencoder.  This adds random
            noise to the encoder, and also constrains the embedding to follow a unit
            Gaussian distribution.
        annealing_start_step: int
            the step (that is, batch) at which to begin turning on the constraint term
            for KL cost annealing
        annealing_final_step: int
            the step (that is, batch) at which to finish turning on the constraint term
            for KL cost annealing
        """
        if SeqToSeq.sequence_end not in input_tokens:
            input_tokens = input_tokens + [SeqToSeq.sequence_end]
        if SeqToSeq.sequence_end not in output_tokens:
            output_tokens = output_tokens + [SeqToSeq.sequence_end]
        self._input_tokens = input_tokens
        self._output_tokens = output_tokens
        self._input_dict = dict((x, i) for i, x in enumerate(input_tokens))
        self._output_dict = dict((x, i) for i, x in enumerate(output_tokens))
        self._max_output_length = max_output_length
        self._embedding_dimension = embedding_dimension
        self._reverse_input = reverse_input
        self.encoder = self._create_encoder(encoder_layers, dropout)
        self.decoder = self._create_decoder(decoder_layers, dropout)
        features = self._create_features()
        gather_indices = Input(shape=(2,), dtype=tf.int32)
        global_step = Input(shape=tuple(), dtype=tf.int32)
        embedding = self.encoder([features, gather_indices])
        self._embedding = self.encoder([features, gather_indices],
                                       training=False)
        if variational:
            randomizer = VariationalRandomizer(self._embedding_dimension,
                                               annealing_start_step,
                                               annealing_final_step)
            embedding = randomizer([self._embedding, global_step])
            self._embedding = randomizer([self._embedding, global_step],
                                         training=False)
        output = self.decoder(embedding)
        model = tf.keras.Model(inputs=[features, gather_indices, global_step],
                               outputs=output)
        super(SeqToSeq, self).__init__(model, self._create_loss(), **kwargs)

    def _create_features(self):
        return Input(shape=(None, len(self._input_tokens)))

    def _create_encoder(self, n_layers, dropout):
        """Create the encoder as a tf.keras.Model."""
        input = self._create_features()
        gather_indices = Input(shape=(2,), dtype=tf.int32)
        prev_layer = input
        for i in range(n_layers):
            if dropout > 0.0:
                prev_layer = Dropout(rate=dropout)(prev_layer)
            prev_layer = GRU(self._embedding_dimension,
                             return_sequences=True)(prev_layer)
        prev_layer = Lambda(lambda x: tf.gather_nd(x[0], x[1]))(
            [prev_layer, gather_indices])
        return tf.keras.Model(inputs=[input, gather_indices],
                              outputs=prev_layer)

    def _create_decoder(self, n_layers, dropout):
        """Create the decoder as a tf.keras.Model."""
        input = Input(shape=(self._embedding_dimension,))
        prev_layer = layers.Stack()(self._max_output_length * [input])
        for i in range(n_layers):
            if dropout > 0.0:
                prev_layer = Dropout(dropout)(prev_layer)
            prev_layer = GRU(self._embedding_dimension,
                             return_sequences=True)(prev_layer)
        output = Dense(len(self._output_tokens),
                       activation=tf.nn.softmax)(prev_layer)
        return tf.keras.Model(inputs=input, outputs=output)

    def _create_loss(self):
        """Create the loss function."""

        def loss_fn(outputs, labels, weights):
            prob = tf.reduce_sum(outputs[0] * labels[0], axis=2)
            mask = tf.reduce_sum(labels[0], axis=2)
            log_prob = tf.math.log(prob + 1e-20) * mask
            loss = -tf.reduce_mean(tf.reduce_sum(log_prob, axis=1))
            return loss + sum(self.model.losses)

        return loss_fn

    def fit_sequences(self,
                      sequences,
                      max_checkpoints_to_keep=5,
                      checkpoint_interval=1000,
                      restore=False):
        """Train this model on a set of sequences

        Parameters
        ----------
        sequences: iterable
            the training samples to fit to.  Each sample should be
            represented as a tuple of the form (input_sequence, output_sequence).
        max_checkpoints_to_keep: int
            the maximum number of checkpoints to keep.  Older checkpoints are discarded.
        checkpoint_interval: int
            the frequency at which to write checkpoints, measured in training steps.
        restore: bool
            if True, restore the model from the most recent checkpoint and continue training
            from there.  If False, retrain the model from scratch.
        """
        self.fit_generator(self._generate_batches(sequences),
                           max_checkpoints_to_keep=max_checkpoints_to_keep,
                           checkpoint_interval=checkpoint_interval,
                           restore=restore)

    def predict_from_sequences(self, sequences, beam_width=5):
        """Given a set of input sequences, predict the output sequences.

        The prediction is done using a beam search with length normalization.

        Parameters
        ----------
        sequences: iterable
            the input sequences to generate a prediction for
        beam_width: int
            the beam width to use for searching.  Set to 1 to use a simple greedy search.
        """
        result = []
        for batch in self._batch_elements(sequences):
            features = self._create_input_array(batch)
            indices = np.array([(i, len(batch[i]) if i < len(batch) else 0)
                                for i in range(self.batch_size)])
            probs = self.predict_on_generator([[
                (features, indices, np.array(self.get_global_step())), None,
                None
            ]])
            for i in range(len(batch)):
                result.append(self._beam_search(probs[i], beam_width))
        return result

    def predict_from_embeddings(self, embeddings, beam_width=5):
        """Given a set of embedding vectors, predict the output sequences.

        The prediction is done using a beam search with length normalization.

        Parameters
        ----------
        embeddings: iterable
            the embedding vectors to generate predictions for
        beam_width: int
            the beam width to use for searching.  Set to 1 to use a simple greedy search.
        """
        result = []
        for batch in self._batch_elements(embeddings):
            embedding_array = np.zeros(
                (self.batch_size, self._embedding_dimension), dtype=np.float32)
            for i, e in enumerate(batch):
                embedding_array[i] = e
            probs = self.decoder(embedding_array, training=False)
            probs = probs.numpy()
            for i in range(len(batch)):
                result.append(self._beam_search(probs[i], beam_width))
        return result

    def predict_embeddings(self, sequences):
        """Given a set of input sequences, compute the embedding vectors.

        Parameters
        ----------
        sequences: iterable
            the input sequences to generate an embedding vector for
        """
        result = []
        for batch in self._batch_elements(sequences):
            features = self._create_input_array(batch)
            indices = np.array([(i, len(batch[i]) if i < len(batch) else 0)
                                for i in range(self.batch_size)])
            embeddings = self.predict_on_generator(
                [[(features, indices, np.array(self.get_global_step())), None,
                  None]],
                outputs=self._embedding)
            for i in range(len(batch)):
                result.append(embeddings[i])
        return np.array(result, dtype=np.float32)

    def _beam_search(self, probs, beam_width):
        """Perform a beam search for the most likely output sequence."""
        if beam_width == 1:
            # Do a simple greedy search.

            s = []
            for i in range(len(probs)):
                token = self._output_tokens[np.argmax(probs[i])]
                if token == SeqToSeq.sequence_end:
                    break
                s.append(token)
            return s

        # Do a beam search with length normalization.

        logprobs = np.log(probs)
        # Represent each candidate as (normalized prob, raw prob, sequence)
        candidates = [(0.0, 0.0, [])]
        for i in range(len(logprobs)):
            new_candidates = []
            for c in candidates:
                if len(c[2]) > 0 and c[2][-1] == SeqToSeq.sequence_end:
                    # This candidate sequence has already been terminated
                    if len(new_candidates) < beam_width:
                        heappush(new_candidates, c)
                    else:
                        heappushpop(new_candidates, c)
                else:
                    # Consider all possible tokens we could add to this candidate sequence.
                    for j, logprob in enumerate(logprobs[i]):
                        new_logprob = logprob + c[1]
                        newc = (new_logprob / (len(c[2]) + 1), new_logprob,
                                c[2] + [self._output_tokens[j]])
                        if len(new_candidates) < beam_width:
                            heappush(new_candidates, newc)
                        else:
                            heappushpop(new_candidates, newc)
            candidates = new_candidates
        return sorted(candidates)[-1][2][:-1]

    def _create_input_array(self, sequences):
        """Create the array describing the input sequences for a batch."""
        lengths = [len(x) for x in sequences]
        if self._reverse_input:
            sequences = [reversed(s) for s in sequences]
        features = np.zeros(
            (self.batch_size, max(lengths) + 1, len(self._input_tokens)),
            dtype=np.float32)
        for i, sequence in enumerate(sequences):
            for j, token in enumerate(sequence):
                features[i, j, self._input_dict[token]] = 1
        features[np.arange(len(sequences)), lengths,
                 self._input_dict[SeqToSeq.sequence_end]] = 1
        return features

    def _create_output_array(self, sequences):
        """Create the array describing the target sequences for a batch."""
        lengths = [len(x) for x in sequences]
        labels = np.zeros(
            (self.batch_size, self._max_output_length, len(
                self._output_tokens)),
            dtype=np.float32)
        end_marker_index = self._output_dict[SeqToSeq.sequence_end]
        for i, sequence in enumerate(sequences):
            for j, token in enumerate(sequence):
                labels[i, j, self._output_dict[token]] = 1
            for j in range(lengths[i], self._max_output_length):
                labels[i, j, end_marker_index] = 1
        return labels

    def _batch_elements(self, elements):
        """Combine elements into batches."""
        batch = []
        for s in elements:
            batch.append(s)
            if len(batch) == self.batch_size:
                yield batch
                batch = []
        if len(batch) > 0:
            yield batch

    def _generate_batches(self, sequences):
        """Create feed_dicts for fitting."""
        for batch in self._batch_elements(sequences):
            inputs = []
            outputs = []
            for input, output in batch:
                inputs.append(input)
                outputs.append(output)
            for i in range(len(inputs), self.batch_size):
                inputs.append([])
                outputs.append([])
            features = self._create_input_array(inputs)
            labels = self._create_output_array(outputs)
            gather_indices = np.array([(i, len(x)) for i, x in enumerate(inputs)
                                      ])
            yield ([features, gather_indices,
                    np.array(self.get_global_step())], [labels], [])


class AspuruGuzikAutoEncoder(SeqToSeq):
    """
    This is an implementation of Automatic Chemical Design Using a Continuous Representation of Molecules
    http://pubs.acs.org/doi/full/10.1021/acscentsci.7b00572

    Abstract
    --------
    We report a method to convert discrete representations of molecules to and
    from a multidimensional continuous representation. This model allows us to
    generate new molecules for efficient exploration and optimization through
    open-ended spaces of chemical compounds. A deep neural network was trained on
    hundreds of thousands of existing chemical structures to construct three
    coupled functions: an encoder, a decoder, and a predictor. The encoder
    converts the discrete representation of a molecule into a real-valued
    continuous vector, and the decoder converts these continuous vectors back to
    discrete molecular representations. The predictor estimates chemical
    properties from the latent continuous vector representation of the molecule.
    Continuous representations of molecules allow us to automatically generate
    novel chemical structures by performing simple operations in the latent space,
    such as decoding random vectors, perturbing known chemical structures, or
    interpolating between molecules. Continuous representations also allow the use
    of powerful gradient-based optimization to efficiently guide the search for
    optimized functional compounds. We demonstrate our method in the domain of
    drug-like molecules and also in a set of molecules with fewer that nine heavy
    atoms.

    Notes
    -------
    This is currently an imperfect reproduction of the paper.  One difference is
    that teacher forcing in the decoder is not implemented.  The paper also
    discusses co-learning molecular properties at the same time as training the
    encoder/decoder.  This is not done here.  The hyperparameters chosen are from
    ZINC dataset.

    This network also currently suffers from exploding gradients.  Care has to be taken when training.

    NOTE(LESWING): Will need to play around with annealing schedule to not have exploding gradients
    TODO(LESWING): Teacher Forcing
    TODO(LESWING): Sigmoid variational loss annealing schedule
    The output GRU layer had one
    additional input, corresponding to the character sampled from the softmax output of the
    previous time step and was trained using teacher forcing. 48 This increased the accuracy
    of generated SMILES strings, which resulted in higher fractions of valid SMILES strings
    for latent points outside the training data, but also made training more difficult, since the
    decoder showed a tendency to ignore the (variational) encoding and rely solely on the input
    sequence. The variational loss was annealed according to sigmoid schedule after 29 epochs,
    running for a total 120 epochs

    I also added a BatchNorm before the mean and std embedding layers.  This has empiracally
    made training more stable, and is discussed in Ladder Variational Autoencoders.
    https://arxiv.org/pdf/1602.02282.pdf
    Maybe if Teacher Forcing and Sigmoid variational loss annealing schedule are used the
    BatchNorm will no longer be neccessary.
    """

    def __init__(self,
                 num_tokens,
                 max_output_length,
                 embedding_dimension=196,
                 filter_sizes=[9, 9, 10],
                 kernel_sizes=[9, 9, 11],
                 decoder_dimension=488,
                 **kwargs):
        """
        Parameters
        ----------
        filter_sizes: list of int
            Number of filters for each 1D convolution in the encoder
        kernel_sizes: list of int
            Kernel size for each 1D convolution in the encoder
        decoder_dimension: int
            Number of channels for the GRU Decoder
        """
        if len(filter_sizes) != len(kernel_sizes):
            raise ValueError("Must have same number of layers and kernels")
        self._filter_sizes = filter_sizes
        self._kernel_sizes = kernel_sizes
        self._decoder_dimension = decoder_dimension
        super(AspuruGuzikAutoEncoder,
              self).__init__(input_tokens=num_tokens,
                             output_tokens=num_tokens,
                             max_output_length=max_output_length,
                             embedding_dimension=embedding_dimension,
                             variational=True,
                             reverse_input=False,
                             **kwargs)

    def _create_features(self):
        return Input(shape=(self._max_output_length, len(self._input_tokens)))

    def _create_encoder(self, n_layers, dropout):
        """Create the encoder as a tf.keras.Model."""
        input = self._create_features()
        gather_indices = Input(shape=(2,), dtype=tf.int32)
        prev_layer = input
        for i in range(len(self._filter_sizes)):
            filter_size = self._filter_sizes[i]
            kernel_size = self._kernel_sizes[i]
            if dropout > 0.0:
                prev_layer = Dropout(rate=dropout)(prev_layer)
            prev_layer = Conv1D(filters=filter_size,
                                kernel_size=kernel_size,
                                activation=tf.nn.relu)(prev_layer)
        prev_layer = Flatten()(prev_layer)
        prev_layer = Dense(self._decoder_dimension,
                           activation=tf.nn.relu)(prev_layer)
        prev_layer = BatchNormalization()(prev_layer)
        return tf.keras.Model(inputs=[input, gather_indices],
                              outputs=prev_layer)

    def _create_decoder(self, n_layers, dropout):
        """Create the decoder as a tf.keras.Model."""
        input = Input(shape=(self._embedding_dimension,))
        prev_layer = Dense(self._embedding_dimension,
                           activation=tf.nn.relu)(input)
        prev_layer = layers.Stack()(self._max_output_length * [prev_layer])
        for i in range(3):
            if dropout > 0.0:
                prev_layer = Dropout(dropout)(prev_layer)
            prev_layer = GRU(self._decoder_dimension,
                             return_sequences=True)(prev_layer)
        output = Dense(len(self._output_tokens),
                       activation=tf.nn.softmax)(prev_layer)
        return tf.keras.Model(inputs=input, outputs=output)

    def _create_input_array(self, sequences):
        return self._create_output_array(sequences)