ng.py

# Copyright 2018 Amazon.com, Inc. or its affiliates. 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.
# A copy of the License is located at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# or in the "license" file accompanying this file. This file 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.

from dataclasses import dataclass, field
from typing import Optional, List, Union

import numpy as np
from toolz import first, valmap

from gluonts import maybe
from gluonts.core.settings import Settings
from gluonts.itertools import (
    Cyclic,
    Map,
    batcher,
    IterableSlice,
    select,
)
from gluonts import zebras as zb


from gluonts.torch.modules.loss import DistributionLoss, NegativeLogLikelihood
from gluonts.torch.distributions import DistributionOutput, StudentTOutput
from gluonts.torch.model.simple_feedforward import (
    SimpleFeedForwardLightningModule,
)


@dataclass
class ObservedValuesIndicator:
    ref: str
    output: str

    def __call__(self, frame):
        return frame.set_like(
            self.ref,
            self.output,
            np.array(~np.isnan(frame[self.ref]), dtype=np.float32),
        )


@dataclass
class InstanceSampler:
    n: int

    def __call__(self, data, future_length):
        max_idx = len(data) - future_length

        if max_idx > 0:
            return np.random.randint(max_idx, size=self.n)

        return []


@dataclass
class InstanceSplitter:
    sampler: InstanceSampler

    past_length: int
    future_length: int

    def __call__(self, data):
        for element in data:
            for split_index in self.sampler(element, self.future_length):
                yield element.split(
                    split_index,
                    past_length=self.past_length,
                    future_length=self.future_length,
                )


class Env(Settings):
    training_sampler: InstanceSampler = InstanceSampler(1)
    cache_data: bool = False


env = Env()


class DeepLearningEstimator:
    def get_schema(self):
        raise NotImplementedError

    def create_training_instances(self, dataset):
        raise NotImplementedError

    def create_validation_instances(self, dataset):
        raise NotImplementedError

    def training_pipeline(self):
        return []

    def validation_pipeline(self):
        return self.training_pipeline()

    def predictor_pipeline(self):
        return self.training_pipeline()

    def into_batches(self, instances):
        batches = batcher(instances, self.batch_size)
        batches = Map(zb.batch(type=self.tensor_type), batches)
        batches = Map(lambda batch: batch.as_dict(), batches)
        return batches

    @env._inject("cache_data")
    def training_dataloader(self, training_data, cache_data: bool = False):
        schema = self.get_schema()
        training_data = Map(
            lambda entry: schema.load_timeframe(
                entry,
                start=entry["start"],
                freq="M",
            ),
            training_data,
        )

        for step in self.training_pipeline():
            training_data = Map(step, training_data)

        if cache_data:
            training_data = list(training_data)

        training_data = Cyclic(training_data).stream()

        instances = self.create_training_instances(training_data)
        batches = self.into_batches(instances)

        return IterableSlice(
            batches,
            getattr(self, "num_batches_per_epoch", None),
        )

    @env._inject("cache_data")
    def validation_dataloader(self, validation_data, cache_data: bool = False):
        schema = self.get_schema()
        validation_data = Map(schema.load_timeframe, validation_data)

        for step in self.validation_pipeline():
            validation_data = Map(step, validation_data)

        if cache_data:
            validation_data = list(validation_data)

        instances = self.create_validation_instances(validation_data)
        batches = self.into_batches(instances)

        return IterableSlice(batches, None)


class SplittingEstimator(DeepLearningEstimator):
    @env._inject("training_sampler")
    def create_training_instances(
        self, dataset, training_sampler=InstanceSampler(1)
    ):
        instance_splitter = InstanceSplitter(
            training_sampler,
            past_length=self.past_length,
            future_length=self.prediction_length,
        )
        return instance_splitter(dataset)

    def create_validation_instances(self, dataset):
        return Map(
            lambda frame: frame.split(
                -self.prediction_length,
                past_length=self.past_length,
                future_length=self.prediction_length,
            ),
            dataset,
        )

    def train(self, training_data, validation_data=None):
        return self.train_model(
            self.training_dataloader(training_data),
            maybe.map(validation_data, self.validation_dataloader),
        )

    def train_model(self, training_data, validation_data):
        raise NotImplementedError


import torch
import pytorch_lightning as pl


class TorchEstimator(DeepLearningEstimator):
    tensor_type = torch.tensor

    def train_model(self, training_data, validation_data):
        trainer = pl.Trainer(**self.trainer_kwargs)
        training_network = self.create_lightning_module()
        trainer.fit(
            model=training_network,
            train_dataloaders=training_data,
            val_dataloaders=validation_data,
        )

        return self.create_predictor(training_network)


from gluonts.util import lazy_property


@dataclass(eq=False)
class DistributionForecast:
    distr_output: DistributionOutput
    args: Union[zb.BatchTimeFrame, zb.TimeFrame]
    arg_names: list

    @lazy_property
    def dist(self):
        return self.distr_output.distribution(
            [self.args.columns[name] for name in self.arg_names],
            self.args.static["loc"],
            self.args.static["scale"],
        )

    @property
    def mean(self):
        return self.args.like(
            {"mean": self.dist.mean},
        )["mean"]

    def __iter__(self):
        if isinstance(self.args, zb.BatchTimeFrame):
            for args in self.args:
                yield DistributionForecast(
                    self.distr_output, args, self.arg_names
                )
        else:
            yield self


@dataclass
class DistributionPredictor:
    network: ...
    schema: ...
    pipeline: ...

    prediction_length: int
    past_length: int
    distr_output: DistributionOutput
    tensor_type: ...

    def predict_one(self, data):
        return first(self.predict_batch([data]))

    def predict_batch(self, data):
        data = map(
            lambda x: self.schema.load_splitframe(
                x,
                future_length=self.prediction_length,
                freq="M",
                start=x["start"],
            ),
            data,
        )

        data = map(lambda x: x.resize(past_length=self.past_length), data)

        for step in self.pipeline:
            data = map(step, data)

        inputs = zb.batch(list(data), type=torch.tensor)

        distr_args, loc, scale = self.network(
            **select(
                self.network.model.describe_inputs(),
                inputs.as_dict(),
            )
        )

        arg_names = list(range(len(distr_args)))

        return DistributionForecast(
            self.distr_output,
            inputs.future.like(
                valmap(torch.detach, dict(zip(arg_names, distr_args))),
                static={"scale": scale, "loc": loc},
            ),
            arg_names=arg_names,
        )


@dataclass
class SimpleFeedForwardEstimator(TorchEstimator, SplittingEstimator):
    prediction_length: int
    context_length: Optional[int] = None
    hidden_dimensions: List[int] = field(default_factory=lambda: [20, 20])
    lr: float = 1e-3
    weight_decay: float = 1e-8
    distr_output: DistributionOutput = StudentTOutput()
    loss: DistributionLoss = NegativeLogLikelihood()
    batch_norm: bool = False
    batch_size: int = 32
    num_batches_per_epoch: int = 50
    trainer_kwargs: dict = field(default_factory=dict)

    def __post_init__(self):
        self.trainer_kwargs = dict(
            {
                "max_epochs": 100,
                "gradient_clip_val": 10.0,
            },
            **self.trainer_kwargs,
        )
        self.context_length = maybe.unwrap_or(
            self.context_length, 10 * self.prediction_length
        )

    @property
    def past_length(self):
        return self.context_length

    def get_schema(self):
        return zb.Schema(
            {
                "target": zb.Field(ndims=1, tdim=-1, past_only=True),
            }
        )

    def training_pipeline(self):
        return [ObservedValuesIndicator("target", "observed_values")]

    def predictor_pipeline(self):
        return []

    def create_lightning_module(self):
        return SimpleFeedForwardLightningModule(
            loss=self.loss,
            lr=self.lr,
            weight_decay=self.weight_decay,
            model_kwargs={
                "prediction_length": self.prediction_length,
                "context_length": self.context_length,
                "hidden_dimensions": self.hidden_dimensions,
                "distr_output": self.distr_output,
                "batch_norm": self.batch_norm,
            },
        )

    def create_predictor(self, model):
        return DistributionPredictor(
            model,
            schema=self.get_schema(),
            pipeline=self.predictor_pipeline(),
            prediction_length=self.prediction_length,
            past_length=self.past_length,
            distr_output=self.distr_output,
            tensor_type=torch.tensor,
        )


from gluonts.dataset.repository.datasets import get_dataset

airpassengers = get_dataset("airpassengers")

my_estimator = SimpleFeedForwardEstimator(
    prediction_length=8, batch_size=32, trainer_kwargs={"max_epochs": 1}
)

with env._let(cache_data=True):
    predictor = my_estimator.train(
        airpassengers.train,
        airpassengers.test,
    )

test_data = list(airpassengers.test)

forecasts = predictor.predict_batch(test_data * 2)

print(forecasts.mean)

# for forecast in forecasts:
#     print(forecast.mean)

# print(
#     predictor.predict_one(test_data[0]).mean,
# )