6

我是张量板的新手。

我正在使用相当简单的代码运行实验,这是输出:

在此处输入图像描述

我不记得要求hp_metric图表,但它在这里。

它是什么,我该如何摆脱它?


使用 Pytorch Lightning 重现的完整代码(不是我认为任何人都应该重现这个来回答):

请注意,唯一取消引用 TensorBoard 的行是

self.logger.experiment.add_scalars("losses", {"train_loss": loss}, global_step=self.current_epoch)
import torch
from torch import nn
import torch.nn.functional as F
from typing import List, Optional
from pytorch_lightning.core.lightning import LightningModule
from Testing.Research.toy_datasets.ClustersDataset import ClustersDataset
from torch.utils.data import DataLoader
from Testing.Research.config.ConfigProvider import ConfigProvider
from pytorch_lightning import Trainer, seed_everything
from torch import optim
import os
from pytorch_lightning.loggers import TensorBoardLogger


class VAEFC(LightningModule):
    # see https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73
    # for possible upgrades, see https://arxiv.org/pdf/1602.02282.pdf
    # https://stats.stackexchange.com/questions/332179/how-to-weight-kld-loss-vs-reconstruction-loss-in-variational-auto-encoder
    def __init__(self, encoder_layer_sizes: List, decoder_layer_sizes: List, config):
        super(VAEFC, self).__init__()
        self._config = config
        self.logger: Optional[TensorBoardLogger] = None

        assert len(encoder_layer_sizes) >= 3, "must have at least 3 layers (2 hidden)"
        # encoder layers
        self._encoder_layers = nn.ModuleList()
        for i in range(1, len(encoder_layer_sizes) - 1):
            enc_layer = nn.Linear(encoder_layer_sizes[i - 1], encoder_layer_sizes[i])
            self._encoder_layers.append(enc_layer)

        # predict mean and covariance vectors
        self._mean_layer = nn.Linear(encoder_layer_sizes[
                                         len(encoder_layer_sizes) - 2],
                                     encoder_layer_sizes[len(encoder_layer_sizes) - 1])
        self._logvar_layer = nn.Linear(encoder_layer_sizes[
                                           len(encoder_layer_sizes) - 2],
                                       encoder_layer_sizes[len(encoder_layer_sizes) - 1])

        # decoder layers
        self._decoder_layers = nn.ModuleList()
        for i in range(1, len(decoder_layer_sizes)):
            dec_layer = nn.Linear(decoder_layer_sizes[i - 1], decoder_layer_sizes[i])
            self._decoder_layers.append(dec_layer)

        self._recon_function = nn.MSELoss(reduction='mean')

    def _encode(self, x):
        for i in range(len(self._encoder_layers)):
            layer = self._encoder_layers[i]
            x = F.relu(layer(x))

        mean_output = self._mean_layer(x)
        logvar_output = self._logvar_layer(x)
        return mean_output, logvar_output

    def _reparametrize(self, mu, logvar):
        if not self.training:
            return mu
        std = logvar.mul(0.5).exp_()
        if std.is_cuda:
            eps = torch.cuda.FloatTensor(std.size()).normal_()
        else:
            eps = torch.FloatTensor(std.size()).normal_()
        reparameterized = eps.mul(std).add_(mu)
        return reparameterized

    def _decode(self, z):
        for i in range(len(self._decoder_layers) - 1):
            layer = self._decoder_layers[i]
            z = F.relu((layer(z)))

        decoded = self._decoder_layers[len(self._decoder_layers) - 1](z)
        # decoded = F.sigmoid(self._decoder_layers[len(self._decoder_layers)-1](z))
        return decoded

    def _loss_function(self, recon_x, x, mu, logvar, reconstruction_function):
        """
        recon_x: generating images
        x: origin images
        mu: latent mean
        logvar: latent log variance
        """
        binary_cross_entropy = reconstruction_function(recon_x, x)  # mse loss TODO see if mse or cross entropy
        # loss = 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
        kld_element = mu.pow(2).add_(logvar.exp()).mul_(-1).add_(1).add_(logvar)
        kld = torch.sum(kld_element).mul_(-0.5)
        # KL divergence Kullback–Leibler divergence, regularization term for VAE
        # It is a measure of how different two probability distributions are different from each other.
        # We are trying to force the distributions closer while keeping the reconstruction loss low.
        # see https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73

        # read on weighting the regularization term here:
        # https://stats.stackexchange.com/questions/332179/how-to-weight-kld-loss-vs-reconstruction-loss-in-variational
        # -auto-encoder
        return binary_cross_entropy + kld * self._config.regularization_factor

    def training_step(self, batch, batch_index):
        orig_batch, noisy_batch, _ = batch
        noisy_batch = noisy_batch.view(noisy_batch.size(0), -1)

        recon_batch, mu, logvar = self.forward(noisy_batch)

        loss = self._loss_function(
            recon_batch,
            orig_batch, mu, logvar,
            reconstruction_function=self._recon_function
        )
        # self.logger.experiment.add_scalars("losses", {"train_loss": loss})
        self.logger.experiment.add_scalars("losses", {"train_loss": loss}, global_step=self.current_epoch)
        # self.logger.experiment.add_scalar("train_loss", loss, self.current_epoch)
        self.logger.experiment.flush()
        return loss

    def train_dataloader(self):
        default_dataset, train_dataset, test_dataset = ClustersDataset.clusters_dataset_by_config()
        train_dataloader = DataLoader(train_dataset, batch_size=self._config.batch_size, shuffle=True)
        return train_dataloader

    def test_dataloader(self):
        default_dataset, train_dataset, test_dataset = ClustersDataset.clusters_dataset_by_config()
        test_dataloader = DataLoader(test_dataset, batch_size=self._config.batch_size, shuffle=True)
        return test_dataloader

    def configure_optimizers(self):
        optimizer = optim.Adam(model.parameters(), lr=self._config.learning_rate)
        return optimizer

    def forward(self, x):
        mu, logvar = self._encode(x)
        z = self._reparametrize(mu, logvar)
        decoded = self._decode(z)
        return decoded, mu, logvar


if __name__ == "__main__":
    config = ConfigProvider.get_config()
    seed_everything(config.random_seed)
    latent_dim = config.latent_dim
    enc_layer_sizes = config.enc_layer_sizes + [latent_dim]
    dec_layer_sizes = [latent_dim] + config.dec_layer_sizes
    model = VAEFC(config=config, encoder_layer_sizes=enc_layer_sizes, decoder_layer_sizes=dec_layer_sizes)

    logger = TensorBoardLogger(save_dir='tb_logs', name='VAEFC')
    logger.hparams = config  # TODO only put here relevant stuff
    # trainer = Trainer(gpus=1)
    trainer = Trainer(deterministic=config.is_deterministic,
                      #auto_lr_find=config.auto_lr_find,
                      #log_gpu_memory='all',
                      # min_epochs=99999,
                      max_epochs=config.num_epochs,
                      default_root_dir=os.getcwd(),
                      logger=logger
                      )
    # trainer.tune(model)
    trainer.fit(model)
    print("done training vae with lightning")

集群数据集.py

from torch.utils.data import Dataset
import matplotlib.pyplot as plt
import torch
import numpy as np
from Testing.Research.config.ConfigProvider import ConfigProvider


class ClustersDataset(Dataset):
    __default_dataset = None
    __default_dataset_train = None
    __default_dataset_test = None

    def __init__(self, cluster_size: int, noise_factor: float = 0, transform=None, n_clusters=2, centers_radius=4.0):
        super(ClustersDataset, self).__init__()
        self._cluster_size = cluster_size
        self._noise_factor = noise_factor
        self._n_clusters = n_clusters
        self._centers_radius = centers_radius
        # self._transform = transform
        self._size = self._cluster_size * self._n_clusters

        self._create_data_clusters()
        self._combine_clusters_to_array()
        self._normalize_data()
        self._add_noise()

        # self._plot()
        pass

    @staticmethod
    def clusters_dataset_by_config():
        if ClustersDataset.__default_dataset is not None:
            return \
                ClustersDataset.__default_dataset, \
                ClustersDataset.__default_dataset_train, \
                ClustersDataset.__default_dataset_test
        config = ConfigProvider.get_config()
        default_dataset = ClustersDataset(
            cluster_size=config.cluster_size,
            noise_factor=config.noise_factor,
            transform=None,
            n_clusters=config.n_clusters,
            centers_radius=config.centers_radius
        )
        
        train_size = int(config.train_size * len(default_dataset))
        test_size = len(default_dataset) - train_size
        train_dataset, test_dataset = torch.utils.data.random_split(default_dataset, [train_size, test_size])

        ClustersDataset.__default_dataset = default_dataset
        ClustersDataset.__default_dataset_train = train_dataset
        ClustersDataset.__default_dataset_test = test_dataset

        return default_dataset, train_dataset, test_dataset

    def _create_data_clusters(self):
        self._clusters = [torch.zeros((self._cluster_size, 2)) for _ in range(self._n_clusters)]
        centers_radius = self._centers_radius
        for i, c in enumerate(self._clusters):
            r, x, y = 3.0, centers_radius * np.cos(i * np.pi * 2 / self._n_clusters), centers_radius * np.sin(
                i * np.pi * 2 / self._n_clusters)
            cluster_length = 1.1
            cluster_start = i * 2 * np.pi / self._n_clusters
            cluster_end = cluster_length * (i + 1) * 2 * np.pi / self._n_clusters
            cluster_inds = torch.linspace(start=cluster_start, end=cluster_end, steps=self._cluster_size,
                                          dtype=torch.float)
            c[:, 0] = r * torch.sin(cluster_inds) + y
            c[:, 1] = r * torch.cos(cluster_inds) + x

    def _plot(self):
        plt.figure()
        plt.scatter(self._noisy_values[:, 0], self._noisy_values[:, 1], s=1, color='b', label="noisy_values")
        plt.scatter(self._values[:, 0], self._values[:, 1], s=1, color='r', label="values")
        plt.legend(loc="upper left")
        plt.show()

    def _combine_clusters_to_array(self):
        size = self._size
        self._values = torch.zeros(size, 2)
        self._labels = torch.zeros(size, dtype=torch.long)
        for i, c in enumerate(self._clusters):
            self._values[i * self._cluster_size: (i + 1) * self._cluster_size, :] = self._clusters[i]
            self._labels[i * self._cluster_size: (i + 1) * self._cluster_size] = i

    def _add_noise(self):
        size = self._size

        mean = torch.zeros(size, 2)
        std = torch.ones(size, 2)
        noise = torch.normal(mean, std)
        self._noisy_values = torch.zeros(size, 2)
        self._noisy_values[:] = self._values
        self._noisy_values = self._noisy_values + noise * self._noise_factor

    def _normalize_data(self):
        values_min, values_max = torch.min(self._values), torch.max(self._values)
        self._values = (self._values - values_min) / (values_max - values_min)
        self._values = self._values * 2 - 1

    def __len__(self):
        return self._size  # number of samples in the dataset

    def __getitem__(self, index):
        item = self._values[index, :]
        noisy_item = self._noisy_values[index, :]
        # if self._transform is not None:
        #     noisy_item = self._transform(item)
        return item, noisy_item, self._labels[index]

    @property
    def values(self):
        return self._values

    @property
    def noisy_values(self):
        return self._noisy_values

配置值(ConfigProvider 只是将它们作为对象返回)

num_epochs: 15
batch_size: 128
learning_rate: 0.0001
auto_lr_find: False

noise_factor: 0.1
regularization_factor: 0.0

cluster_size: 5000
n_clusters: 5
centers_radius: 4.0
train_size: 0.8

latent_dim: 8

enc_layer_sizes: [2, 200, 200, 200]
dec_layer_sizes: [200, 200, 200, 2]

retrain_vae: False
random_seed: 11
is_deterministic: True
4

2 回答 2

4

这是pytorch Lightning中tensorboard的默认设置。您可以设置default_hp_metric为 false 以摆脱此指标。

TensorBoardLogger(save_dir='tb_logs', name='VAEFC', default_hp_metric=False)

hp_metric帮助您跟踪不同超参数的模型性能。你可以hparams在你的tensorboard中检查它。

于 2021-01-04T00:03:23.860 回答
1

hp_metric(超参数度量)是为了帮助你调整你的超参数。

您可以将此指标设置为您喜欢的任何值,如pytorch 官方文档中所述。

然后,您可以查看您的超参数,并根据您选择的任何指标查看哪个结果最好。

或者,如果您不想要它,您可以按照@joe32140 的回答中的建议禁用它:

您可以将 default_hp_metric 设置为 false 以摆脱此指标。

TensorBoardLogger(save_dir='tb_logs', name='VAEFC', default_hp_metric=False)
于 2021-06-14T15:55:31.450 回答