fig clf inserted and not resize on kld
This commit is contained in:
Steffen Illium
+107
-194
@@ -1,19 +1,22 @@
|
||||
from random import choices, seed
|
||||
import numpy as np
|
||||
|
||||
import torch
|
||||
from functools import reduce
|
||||
from operator import mul
|
||||
|
||||
from random import choices, choice
|
||||
|
||||
import torch
|
||||
|
||||
from torch import nn
|
||||
from torch.optim import Adam
|
||||
from torchvision.datasets import MNIST
|
||||
|
||||
from datasets.mnist import MyMNIST
|
||||
from datasets.trajectory_dataset import TrajData
|
||||
from lib.evaluation.classification import ROCEvaluation
|
||||
from lib.modules.blocks import ConvModule, ResidualModule, DeConvModule
|
||||
from lib.modules.utils import LightningBaseModule, Flatten
|
||||
|
||||
import matplotlib.pyplot as plt
|
||||
import lib.variables as V
|
||||
from lib.visualization.generator_eval import GeneratorVisualizer
|
||||
|
||||
|
||||
class CNNRouteGeneratorModel(LightningBaseModule):
|
||||
@@ -24,48 +27,71 @@ class CNNRouteGeneratorModel(LightningBaseModule):
|
||||
return Adam(self.parameters(), lr=self.hparams.train_param.lr)
|
||||
|
||||
def training_step(self, batch_xy, batch_nb, *args, **kwargs):
|
||||
batch_x, alternative = batch_xy
|
||||
batch_x, target = batch_xy
|
||||
generated_alternative, z, mu, logvar = self(batch_x)
|
||||
element_wise_loss = self.criterion(generated_alternative, alternative)
|
||||
# see Appendix B from VAE paper:
|
||||
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
|
||||
# https://arxiv.org/abs/1312.6114
|
||||
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
|
||||
target = batch_x if 'ae' in self.hparams.data_param.mode else target
|
||||
element_wise_loss = self.criterion(generated_alternative, target)
|
||||
|
||||
kld_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
|
||||
# Dimensional Resizing TODO: Does This make sense? Sanity Check it!
|
||||
# kld_loss /= reduce(mul, self.in_shape)
|
||||
# kld_loss *= self.hparams.data_param.dataset_length / self.hparams.train_param.batch_size * 100
|
||||
if 'vae' in self.hparams.data_param.mode:
|
||||
# see Appendix B from VAE paper:
|
||||
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
|
||||
# https://arxiv.org/abs/1312.6114
|
||||
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
|
||||
kld_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
|
||||
# Dimensional Resizing TODO: Does This make sense? Sanity Check it!
|
||||
# kld_loss /= reduce(mul, self.in_shape)
|
||||
# kld_loss *= self.hparams.data_param.dataset_length / self.hparams.train_param.batch_size
|
||||
|
||||
loss = kld_loss + element_wise_loss
|
||||
loss = kld_loss + element_wise_loss
|
||||
else:
|
||||
loss = element_wise_loss
|
||||
kld_loss = 0
|
||||
return dict(loss=loss, log=dict(element_wise_loss=element_wise_loss, loss=loss, kld_loss=kld_loss))
|
||||
|
||||
def _test_val_step(self, batch_xy, batch_nb, *args):
|
||||
batch_x, _ = batch_xy
|
||||
map_array = batch_x[:, 0].unsqueeze(1)
|
||||
trajectory = batch_x[:, 1].unsqueeze(1)
|
||||
labels = batch_x[:, 2].unsqueeze(1).max(dim=-1).values.max(-1).values
|
||||
if 'vae' in self.hparams.data_param.mode:
|
||||
z, mu, logvar = self.encode(batch_x)
|
||||
else:
|
||||
z = self.encode(batch_x)
|
||||
mu, logvar = z, z
|
||||
|
||||
_, mu, _ = self.encode(batch_x)
|
||||
generated_alternative = self.generate(mu)
|
||||
return dict(maps=map_array, trajectories=trajectory, batch_nb=batch_nb, labels=labels,
|
||||
generated_alternative=generated_alternative, pred_label=-1)
|
||||
return_dict = dict(input=batch_x, batch_nb=batch_nb, output=generated_alternative, z=z, mu=mu, logvar=logvar)
|
||||
|
||||
if 'hom' in self.hparams.data_param.mode:
|
||||
labels = torch.full((batch_x.shape[0], 1), V.HOMOTOPIC)
|
||||
elif 'alt' in self.hparams.data_param.mode:
|
||||
labels = torch.full((batch_x.shape[0], 1), V.ALTERNATIVE)
|
||||
elif 'vae' in self.hparams.data_param.mode:
|
||||
labels = torch.full((batch_x.shape[0], 1), V.ANY)
|
||||
elif 'ae' in self.hparams.data_param.mode:
|
||||
labels = torch.full((batch_x.shape[0], 1), V.ANY)
|
||||
else:
|
||||
labels = batch_x[:, 2].unsqueeze(1).max(dim=-1).values.max(-1).values
|
||||
|
||||
return_dict.update(labels=self._move_to_model_device(labels))
|
||||
return return_dict
|
||||
|
||||
def _test_val_epoch_end(self, outputs, test=False):
|
||||
val_restul_dict = self.generate_random()
|
||||
plt.close('all')
|
||||
|
||||
from lib.visualization.generator_eval import GeneratorVisualizer
|
||||
g = GeneratorVisualizer(**val_restul_dict)
|
||||
fig = g.draw()
|
||||
g = GeneratorVisualizer(choice(outputs))
|
||||
fig = g.draw_io_bundle()
|
||||
self.logger.log_image(f'{self.name}_Output', fig, step=self.global_step)
|
||||
plt.clf()
|
||||
|
||||
fig = g.draw_latent()
|
||||
self.logger.log_image(f'{self.name}_Latent', fig, step=self.global_step)
|
||||
plt.clf()
|
||||
|
||||
return dict(epoch=self.current_epoch)
|
||||
|
||||
def on_epoch_start(self):
|
||||
self.dataset.seed(self.logger.version)
|
||||
# self.dataset.seed(self.logger.version)
|
||||
# torch.random.manual_seed(self.logger.version)
|
||||
# np.random.seed(self.logger.version)
|
||||
pass
|
||||
|
||||
def validation_step(self, *args):
|
||||
return self._test_val_step(*args)
|
||||
@@ -82,19 +108,23 @@ class CNNRouteGeneratorModel(LightningBaseModule):
|
||||
def __init__(self, *params, issubclassed=False):
|
||||
super(CNNRouteGeneratorModel, self).__init__(*params)
|
||||
|
||||
if not issubclassed:
|
||||
if False:
|
||||
# Dataset
|
||||
self.dataset = TrajData(self.hparams.data_param.map_root, mode='generator_all_in_map',
|
||||
self.dataset = TrajData(self.hparams.data_param.map_root,
|
||||
mode=self.hparams.data_param.mode,
|
||||
preprocessed=self.hparams.data_param.use_preprocessed,
|
||||
length=self.hparams.data_param.dataset_length, normalized=True)
|
||||
self.criterion = nn.MSELoss()
|
||||
self.criterion = nn.MSELoss()
|
||||
|
||||
self.dataset = MyMNIST()
|
||||
|
||||
# Additional Attributes #
|
||||
#######################################################
|
||||
self.in_shape = self.dataset.map_shapes_max
|
||||
self.use_res_net = self.hparams.model_param.use_res_net
|
||||
self.lat_dim = self.hparams.model_param.lat_dim
|
||||
self.feature_dim = self.lat_dim * 10
|
||||
self.feature_dim = self.lat_dim
|
||||
self.out_channels = 1 if 'generator' in self.hparams.data_param.mode else self.in_shape[0]
|
||||
########################################################
|
||||
|
||||
# NN Nodes
|
||||
@@ -119,7 +149,7 @@ class CNNRouteGeneratorModel(LightningBaseModule):
|
||||
conv_filters=self.hparams.model_param.filters[1],
|
||||
use_norm=self.hparams.model_param.use_norm,
|
||||
use_bias=self.hparams.model_param.use_bias)
|
||||
self.enc_conv_1b = ConvModule(self.enc_conv_1a.shape, conv_kernel=3, conv_stride=2, conv_padding=0,
|
||||
self.enc_conv_1b = ConvModule(self.enc_conv_1a.shape, conv_kernel=3, conv_stride=1, conv_padding=0,
|
||||
conv_filters=self.hparams.model_param.filters[1],
|
||||
use_norm=self.hparams.model_param.use_norm,
|
||||
use_bias=self.hparams.model_param.use_bias)
|
||||
@@ -137,20 +167,8 @@ class CNNRouteGeneratorModel(LightningBaseModule):
|
||||
use_norm=self.hparams.model_param.use_norm,
|
||||
use_bias=self.hparams.model_param.use_bias)
|
||||
|
||||
self.enc_res_3 = ResidualModule(self.enc_conv_2b.shape, ConvModule, 2, conv_kernel=7, conv_stride=1,
|
||||
conv_padding=3, conv_filters=self.hparams.model_param.filters[2],
|
||||
use_norm=self.hparams.model_param.use_norm,
|
||||
use_bias=self.hparams.model_param.use_bias)
|
||||
self.enc_conv_3a = ConvModule(self.enc_res_3.shape, conv_kernel=7, conv_stride=1, conv_padding=0,
|
||||
conv_filters=self.hparams.model_param.filters[2],
|
||||
use_norm=self.hparams.model_param.use_norm,
|
||||
use_bias=self.hparams.model_param.use_bias)
|
||||
self.enc_conv_3b = ConvModule(self.enc_conv_3a.shape, conv_kernel=7, conv_stride=1, conv_padding=0,
|
||||
conv_filters=self.hparams.model_param.filters[2],
|
||||
use_norm=self.hparams.model_param.use_norm,
|
||||
use_bias=self.hparams.model_param.use_bias)
|
||||
|
||||
self.enc_flat = Flatten(self.enc_conv_3b.shape)
|
||||
last_conv_shape = self.enc_conv_2b.shape
|
||||
self.enc_flat = Flatten(last_conv_shape)
|
||||
self.enc_lin_1 = nn.Linear(self.enc_flat.shape, self.feature_dim)
|
||||
|
||||
#
|
||||
@@ -160,46 +178,43 @@ class CNNRouteGeneratorModel(LightningBaseModule):
|
||||
|
||||
#
|
||||
# Variational Bottleneck
|
||||
self.mu = nn.Linear(self.feature_dim, self.lat_dim)
|
||||
self.logvar = nn.Linear(self.feature_dim, self.lat_dim)
|
||||
if 'vae' in self.hparams.data_param.mode:
|
||||
self.mu = nn.Linear(self.feature_dim, self.lat_dim)
|
||||
self.logvar = nn.Linear(self.feature_dim, self.lat_dim)
|
||||
|
||||
#
|
||||
# Linear Bottleneck
|
||||
else:
|
||||
self.z = nn.Linear(self.feature_dim, self.lat_dim)
|
||||
|
||||
#
|
||||
# Alternative Generator
|
||||
self.gen_lin_1 = nn.Linear(self.hparams.model_param.lat_dim, self.feature_dim)
|
||||
self.gen_lin_1 = nn.Linear(self.lat_dim, self.enc_flat.shape)
|
||||
|
||||
self.gen_lin_2 = nn.Linear(self.feature_dim, self.enc_flat.shape)
|
||||
# self.gen_lin_2 = nn.Linear(self.feature_dim, self.enc_flat.shape)
|
||||
|
||||
self.reshape_to_last_conv = Flatten(self.enc_flat.shape, self.enc_conv_3b.shape)
|
||||
self.reshape_to_last_conv = Flatten(self.enc_flat.shape, last_conv_shape)
|
||||
|
||||
self.gen_deconv_1a = DeConvModule(self.enc_conv_3b.shape, self.hparams.model_param.filters[2],
|
||||
conv_padding=0, conv_kernel=11, conv_stride=1,
|
||||
use_norm=self.hparams.model_param.use_norm)
|
||||
self.gen_deconv_1b = DeConvModule(self.gen_deconv_1a.shape, self.hparams.model_param.filters[2],
|
||||
conv_padding=0, conv_kernel=9, conv_stride=2,
|
||||
self.gen_deconv_1a = DeConvModule(last_conv_shape, self.hparams.model_param.filters[2],
|
||||
conv_padding=1, conv_kernel=9, conv_stride=1,
|
||||
use_norm=self.hparams.model_param.use_norm)
|
||||
|
||||
self.gen_deconv_2a = DeConvModule(self.gen_deconv_1b.shape, self.hparams.model_param.filters[1],
|
||||
conv_padding=0, conv_kernel=7, conv_stride=1,
|
||||
use_norm=self.hparams.model_param.use_norm)
|
||||
self.gen_deconv_2b = DeConvModule(self.gen_deconv_2a.shape, self.hparams.model_param.filters[1],
|
||||
conv_padding=0, conv_kernel=7, conv_stride=1,
|
||||
self.gen_deconv_2a = DeConvModule(self.gen_deconv_1a.shape, self.hparams.model_param.filters[1],
|
||||
conv_padding=1, conv_kernel=7, conv_stride=1,
|
||||
use_norm=self.hparams.model_param.use_norm)
|
||||
|
||||
self.gen_deconv_3a = DeConvModule(self.gen_deconv_2b.shape, self.hparams.model_param.filters[0],
|
||||
conv_padding=1, conv_kernel=5, conv_stride=1,
|
||||
use_norm=self.hparams.model_param.use_norm)
|
||||
self.gen_deconv_3b = DeConvModule(self.gen_deconv_3a.shape, self.hparams.model_param.filters[0],
|
||||
conv_padding=1, conv_kernel=4, conv_stride=1,
|
||||
use_norm=self.hparams.model_param.use_norm)
|
||||
|
||||
self.gen_deconv_out = DeConvModule(self.gen_deconv_3b.shape, 1, activation=None,
|
||||
self.gen_deconv_out = DeConvModule(self.gen_deconv_2a.shape, self.out_channels, activation=None,
|
||||
conv_padding=0, conv_kernel=3, conv_stride=1,
|
||||
use_norm=self.hparams.model_param.use_norm)
|
||||
|
||||
def forward(self, batch_x):
|
||||
#
|
||||
# Encode
|
||||
z, mu, logvar = self.encode(batch_x)
|
||||
if 'vae' in self.hparams.data_param.mode:
|
||||
z, mu, logvar = self.encode(batch_x)
|
||||
else:
|
||||
z = self.encode(batch_x)
|
||||
mu, logvar = z, z
|
||||
|
||||
#
|
||||
# Generate
|
||||
@@ -220,148 +235,46 @@ class CNNRouteGeneratorModel(LightningBaseModule):
|
||||
combined_tensor = self.enc_res_2(combined_tensor) if self.use_res_net else combined_tensor
|
||||
combined_tensor = self.enc_conv_2a(combined_tensor)
|
||||
combined_tensor = self.enc_conv_2b(combined_tensor)
|
||||
combined_tensor = self.enc_res_3(combined_tensor) if self.use_res_net else combined_tensor
|
||||
combined_tensor = self.enc_conv_3a(combined_tensor)
|
||||
combined_tensor = self.enc_conv_3b(combined_tensor)
|
||||
# combined_tensor = self.enc_res_3(combined_tensor) if self.use_res_net else combined_tensor
|
||||
# combined_tensor = self.enc_conv_3a(combined_tensor)
|
||||
# combined_tensor = self.enc_conv_3b(combined_tensor)
|
||||
|
||||
combined_tensor = self.enc_flat(combined_tensor)
|
||||
combined_tensor = self.enc_lin_1(combined_tensor)
|
||||
combined_tensor = self.enc_lin_2(combined_tensor)
|
||||
|
||||
combined_tensor = self.enc_norm(combined_tensor)
|
||||
combined_tensor = self.activation(combined_tensor)
|
||||
|
||||
combined_tensor = self.enc_lin_2(combined_tensor)
|
||||
combined_tensor = self.enc_norm(combined_tensor)
|
||||
combined_tensor = self.activation(combined_tensor)
|
||||
|
||||
#
|
||||
# Variational
|
||||
# Parameter and Sampling
|
||||
mu = self.mu(combined_tensor)
|
||||
logvar = self.logvar(combined_tensor)
|
||||
z = self.reparameterize(mu, logvar)
|
||||
return z, mu, logvar
|
||||
if 'vae' in self.hparams.data_param.mode:
|
||||
mu = self.mu(combined_tensor)
|
||||
logvar = self.logvar(combined_tensor)
|
||||
z = self.reparameterize(mu, logvar)
|
||||
return z, mu, logvar
|
||||
else:
|
||||
#
|
||||
# Linear Bottleneck
|
||||
z = self.z(combined_tensor)
|
||||
return z
|
||||
|
||||
def generate(self, z):
|
||||
alt_tensor = self.gen_lin_1(z)
|
||||
alt_tensor = self.activation(alt_tensor)
|
||||
alt_tensor = self.gen_lin_2(alt_tensor)
|
||||
alt_tensor = self.activation(alt_tensor)
|
||||
# alt_tensor = self.gen_lin_2(alt_tensor)
|
||||
# alt_tensor = self.activation(alt_tensor)
|
||||
alt_tensor = self.reshape_to_last_conv(alt_tensor)
|
||||
alt_tensor = self.gen_deconv_1a(alt_tensor)
|
||||
alt_tensor = self.gen_deconv_1b(alt_tensor)
|
||||
|
||||
alt_tensor = self.gen_deconv_2a(alt_tensor)
|
||||
alt_tensor = self.gen_deconv_2b(alt_tensor)
|
||||
alt_tensor = self.gen_deconv_3a(alt_tensor)
|
||||
alt_tensor = self.gen_deconv_3b(alt_tensor)
|
||||
|
||||
# alt_tensor = self.gen_deconv_3a(alt_tensor)
|
||||
# alt_tensor = self.gen_deconv_3b(alt_tensor)
|
||||
alt_tensor = self.gen_deconv_out(alt_tensor)
|
||||
# alt_tensor = self.activation(alt_tensor)
|
||||
alt_tensor = self.sigmoid(alt_tensor)
|
||||
# alt_tensor = self.sigmoid(alt_tensor)
|
||||
return alt_tensor
|
||||
|
||||
def generate_random(self, n=12):
|
||||
|
||||
samples, alternatives = zip(*[self.dataset.test_dataset[choice]
|
||||
for choice in choices(range(self.dataset.length), k=n)])
|
||||
samples = self._move_to_model_device(torch.stack([torch.as_tensor(x) for x in samples]))
|
||||
alternatives = self._move_to_model_device(torch.stack([torch.as_tensor(x) for x in alternatives]))
|
||||
|
||||
return self._test_val_step((samples, alternatives), -9999)
|
||||
|
||||
|
||||
class CNNRouteGeneratorDiscriminated(CNNRouteGeneratorModel):
|
||||
|
||||
name = 'CNNRouteGeneratorDiscriminated'
|
||||
|
||||
def training_step(self, batch_xy, batch_nb, *args, **kwargs):
|
||||
batch_x, label = batch_xy
|
||||
|
||||
generated_alternative, z, mu, logvar = self(batch_x)
|
||||
map_array, trajectory = batch_x
|
||||
|
||||
map_stack = torch.cat((map_array, trajectory, generated_alternative), dim=1)
|
||||
pred_label = self.discriminator(map_stack)
|
||||
discriminated_bce_loss = self.criterion(pred_label, label.float().unsqueeze(-1))
|
||||
|
||||
# see Appendix B from VAE paper:
|
||||
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
|
||||
# https://arxiv.org/abs/1312.6114
|
||||
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
|
||||
kld_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
|
||||
# Dimensional Resizing
|
||||
kld_loss /= reduce(mul, self.in_shape)
|
||||
|
||||
loss = (kld_loss + discriminated_bce_loss) / 2
|
||||
return dict(loss=loss, log=dict(loss=loss,
|
||||
discriminated_bce_loss=discriminated_bce_loss,
|
||||
kld_loss=kld_loss)
|
||||
)
|
||||
|
||||
def _test_val_step(self, batch_xy, batch_nb, *args):
|
||||
batch_x, label = batch_xy
|
||||
|
||||
generated_alternative, z, mu, logvar = self(batch_x)
|
||||
map_array, trajectory = batch_x
|
||||
|
||||
map_stack = torch.cat((map_array, trajectory, generated_alternative), dim=1)
|
||||
pred_label = self.discriminator(map_stack)
|
||||
|
||||
discriminated_bce_loss = self.criterion(pred_label, label.float().unsqueeze(-1))
|
||||
return dict(discriminated_bce_loss=discriminated_bce_loss, batch_nb=batch_nb,
|
||||
pred_label=pred_label, label=label, generated_alternative=generated_alternative)
|
||||
|
||||
def validation_step(self, *args):
|
||||
return self._test_val_step(*args)
|
||||
|
||||
def validation_epoch_end(self, outputs: list):
|
||||
return self._test_val_epoch_end(outputs)
|
||||
|
||||
def _test_val_epoch_end(self, outputs, test=False):
|
||||
evaluation = ROCEvaluation(plot_roc=True)
|
||||
pred_label = torch.cat([x['pred_label'] for x in outputs])
|
||||
labels = torch.cat([x['label'] for x in outputs]).unsqueeze(1)
|
||||
mean_losses = torch.stack([x['discriminated_bce_loss'] for x in outputs]).mean()
|
||||
|
||||
# Sci-py call ROC eval call is eval(true_label, prediction)
|
||||
roc_auc, tpr, fpr = evaluation(labels.cpu().numpy(), pred_label.cpu().numpy(), )
|
||||
if test:
|
||||
# self.logger.log_metrics(score_dict)
|
||||
self.logger.log_image(f'{self.name}_ROC-Curve', plt.gcf(), step=self.global_step)
|
||||
plt.clf()
|
||||
|
||||
maps, trajectories, labels, val_restul_dict = self.generate_random()
|
||||
|
||||
from lib.visualization.generator_eval import GeneratorVisualizer
|
||||
g = GeneratorVisualizer(maps, trajectories, labels, val_restul_dict)
|
||||
fig = g.draw()
|
||||
self.logger.log_image(f'{self.name}_Output', fig, step=self.global_step)
|
||||
plt.clf()
|
||||
|
||||
return dict(mean_losses=mean_losses, roc_auc=roc_auc, epoch=self.current_epoch)
|
||||
|
||||
def test_step(self, *args):
|
||||
return self._test_val_step(*args)
|
||||
|
||||
def test_epoch_end(self, outputs):
|
||||
return self._test_val_epoch_end(outputs, test=True)
|
||||
|
||||
@property
|
||||
def discriminator(self):
|
||||
if self._disc is None:
|
||||
raise RuntimeError('Set the Discriminator first; "set_discriminator(disc_model)')
|
||||
return self._disc
|
||||
|
||||
def set_discriminator(self, disc_model):
|
||||
if self._disc is not None:
|
||||
raise RuntimeError('Discriminator has already been set... What are trying to do?')
|
||||
self._disc = disc_model
|
||||
|
||||
def __init__(self, *params):
|
||||
raise NotImplementedError
|
||||
super(CNNRouteGeneratorDiscriminated, self).__init__(*params, issubclassed=True)
|
||||
|
||||
self._disc = None
|
||||
|
||||
self.criterion = nn.BCELoss()
|
||||
|
||||
self.dataset = TrajData(self.hparams.data_param.map_root, mode='just_route', preprocessed=True,
|
||||
length=self.hparams.data_param.dataset_length, normalized=True)
|
||||
|
||||
@@ -0,0 +1,116 @@
|
||||
from random import choices, seed
|
||||
import numpy as np
|
||||
|
||||
import torch
|
||||
from functools import reduce
|
||||
from operator import mul
|
||||
|
||||
from torch import nn
|
||||
from torch.optim import Adam
|
||||
|
||||
from datasets.trajectory_dataset import TrajData
|
||||
from lib.evaluation.classification import ROCEvaluation
|
||||
from lib.models.generators.cnn import CNNRouteGeneratorModel
|
||||
from lib.modules.blocks import ConvModule, ResidualModule, DeConvModule
|
||||
from lib.modules.utils import LightningBaseModule, Flatten
|
||||
|
||||
import matplotlib.pyplot as plt
|
||||
|
||||
|
||||
class CNNRouteGeneratorDiscriminated(CNNRouteGeneratorModel):
|
||||
|
||||
name = 'CNNRouteGeneratorDiscriminated'
|
||||
|
||||
def training_step(self, batch_xy, batch_nb, *args, **kwargs):
|
||||
batch_x, label = batch_xy
|
||||
|
||||
generated_alternative, z, mu, logvar = self(batch_x)
|
||||
map_array, trajectory = batch_x
|
||||
|
||||
map_stack = torch.cat((map_array, trajectory, generated_alternative), dim=1)
|
||||
pred_label = self.discriminator(map_stack)
|
||||
discriminated_bce_loss = self.criterion(pred_label, label.float().unsqueeze(-1))
|
||||
|
||||
# see Appendix B from VAE paper:
|
||||
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
|
||||
# https://arxiv.org/abs/1312.6114
|
||||
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
|
||||
kld_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
|
||||
# Dimensional Resizing
|
||||
kld_loss /= reduce(mul, self.in_shape)
|
||||
|
||||
loss = (kld_loss + discriminated_bce_loss) / 2
|
||||
return dict(loss=loss, log=dict(loss=loss,
|
||||
discriminated_bce_loss=discriminated_bce_loss,
|
||||
kld_loss=kld_loss)
|
||||
)
|
||||
|
||||
def _test_val_step(self, batch_xy, batch_nb, *args):
|
||||
batch_x, label = batch_xy
|
||||
|
||||
generated_alternative, z, mu, logvar = self(batch_x)
|
||||
map_array, trajectory = batch_x
|
||||
|
||||
map_stack = torch.cat((map_array, trajectory, generated_alternative), dim=1)
|
||||
pred_label = self.discriminator(map_stack)
|
||||
|
||||
discriminated_bce_loss = self.criterion(pred_label, label.float().unsqueeze(-1))
|
||||
return dict(discriminated_bce_loss=discriminated_bce_loss, batch_nb=batch_nb,
|
||||
pred_label=pred_label, label=label, generated_alternative=generated_alternative)
|
||||
|
||||
def validation_step(self, *args):
|
||||
return self._test_val_step(*args)
|
||||
|
||||
def validation_epoch_end(self, outputs: list):
|
||||
return self._test_val_epoch_end(outputs)
|
||||
|
||||
def _test_val_epoch_end(self, outputs, test=False):
|
||||
evaluation = ROCEvaluation(plot_roc=True)
|
||||
pred_label = torch.cat([x['pred_label'] for x in outputs])
|
||||
labels = torch.cat([x['label'] for x in outputs]).unsqueeze(1)
|
||||
mean_losses = torch.stack([x['discriminated_bce_loss'] for x in outputs]).mean()
|
||||
|
||||
# Sci-py call ROC eval call is eval(true_label, prediction)
|
||||
roc_auc, tpr, fpr = evaluation(labels.cpu().numpy(), pred_label.cpu().numpy(), )
|
||||
if test:
|
||||
# self.logger.log_metrics(score_dict)
|
||||
self.logger.log_image(f'{self.name}_ROC-Curve', plt.gcf(), step=self.global_step)
|
||||
plt.clf()
|
||||
|
||||
maps, trajectories, labels, val_restul_dict = self.generate_random()
|
||||
|
||||
from lib.visualization.generator_eval import GeneratorVisualizer
|
||||
g = GeneratorVisualizer(maps, trajectories, labels, val_restul_dict)
|
||||
fig = g.draw()
|
||||
self.logger.log_image(f'{self.name}_Output', fig, step=self.global_step)
|
||||
plt.clf()
|
||||
|
||||
return dict(mean_losses=mean_losses, roc_auc=roc_auc, epoch=self.current_epoch)
|
||||
|
||||
def test_step(self, *args):
|
||||
return self._test_val_step(*args)
|
||||
|
||||
def test_epoch_end(self, outputs):
|
||||
return self._test_val_epoch_end(outputs, test=True)
|
||||
|
||||
@property
|
||||
def discriminator(self):
|
||||
if self._disc is None:
|
||||
raise RuntimeError('Set the Discriminator first; "set_discriminator(disc_model)')
|
||||
return self._disc
|
||||
|
||||
def set_discriminator(self, disc_model):
|
||||
if self._disc is not None:
|
||||
raise RuntimeError('Discriminator has already been set... What are trying to do?')
|
||||
self._disc = disc_model
|
||||
|
||||
def __init__(self, *params):
|
||||
raise NotImplementedError
|
||||
super(CNNRouteGeneratorDiscriminated, self).__init__(*params, issubclassed=True)
|
||||
|
||||
self._disc = None
|
||||
|
||||
self.criterion = nn.BCELoss()
|
||||
|
||||
self.dataset = TrajData(self.hparams.data_param.map_root, mode='just_route', preprocessed=True,
|
||||
length=self.hparams.data_param.dataset_length, normalized=True)
|
||||
Reference in New Issue
Block a user