decagon-pytorch/src/icosagon/trainloop.py

from .model import Model
import torch
from .batch import PredictionsBatch, \
    flatten_predictions, \
    gather_batch_indices
from typing import Callable
from types import FunctionType


class TrainLoop(object):
    def __init__(self, model: Model, lr: float = 0.001,
        loss: Callable[[torch.Tensor, torch.Tensor], torch.Tensor] = \
            torch.nn.functional.binary_cross_entropy_with_logits,
        batch_size: int = 100, generator: torch.Generator = None) -> None:

        if not isinstance(model, Model):
            raise TypeError('model must be an instance of Model')

        lr = float(lr)

        if not isinstance(loss, FunctionType):
            raise TypeError('loss must be a function')

        batch_size = int(batch_size)

        if generator is not None and not isinstance(generator, torch.Generator):
            raise TypeError('generator must be an instance of torch.Generator')

        self.model = model
        self.lr = lr
        self.loss = loss
        self.batch_size = batch_size
        self.generator = generator or torch.default_generator

        self.opt = None

        self.build()

    def build(self) -> None:
        opt = torch.optim.Adam(self.model.parameters(), lr=self.lr)
        self.opt = opt

    def run_epoch(self):
        batch = PredictionsBatch(self.model.prep_d, batch_size=self.batch_size,
            generator=self.generator)
        # pred = self.model(None)
        # n = len(list(iter(batch)))
        loss_sum = 0
        for i, indices in enumerate(batch):
            print('%.2f%% (%d/%d)' % (i * batch.batch_size * 100 / batch.total_edge_count, i * batch.batch_size, batch.total_edge_count))
            self.opt.zero_grad()
            pred = self.model(None)
            pred = flatten_predictions(pred)
            # batch = PredictionsBatch(pred, batch_size=self.batch_size, shuffle=True)
            # seed = torch.rand(1).item()
            # rng_state = torch.get_rng_state()
            # torch.manual_seed(seed)
            #it = iter(batch)
            #torch.set_rng_state(rng_state)
            #for k in range(i):
                #_ = next(it)
            #(input, target) = next(it)
            (input, target) = gather_batch_indices(pred, indices)
            loss = self.loss(input, target)
            loss.backward()
            self.opt.step()
            loss_sum += loss.detach().cpu().item()
        return loss_sum


    def train(self, max_epochs):
        best_loss = None
        best_epoch = None
        for i in range(max_epochs):
            loss = self.run_epoch()
            if best_loss is None or loss < best_loss:
                best_loss = loss
                best_epoch = i
        return loss, best_loss, best_epoch