def test_outputShapes(self):
network = DescriminatorNet(DEFAULT_CONFIG)
preds = network(self.X)
self.assertEqual(preds.logits.shape, (20, 1))
self.assertEqual(preds.logprob.shape, (20, 1))
self.assertEqual(preds.prob.shape, (20, 1))
self.assertEqual(preds.entropy.shape, (20, 1))
preds_with_y = network(self.X, False, self.y)
self.assertEqual(preds_with_y.logits.shape, (20, 1))
self.assertEqual(preds_with_y.logprob.shape, (20, 1))
self.assertEqual(preds_with_y.prob.shape, (20, 1))
self.assertEqual(preds_with_y.entropy.shape, (20, 1))
def test_no_nans_for_simple_inputs(self):
network = DescriminatorNet(DEFAULT_CONFIG)
preds = network(self.X)
# No Nans
self.assertEqual(np.count_nonzero(np.isnan(preds.logits)), 0)
self.assertEqual(np.count_nonzero(np.isnan(preds.logprob)), 0)
self.assertEqual(np.count_nonzero(np.isnan(preds.prob)), 0)
self.assertEqual(np.count_nonzero(np.isnan(preds.entropy)), 0)
preds_with_y = network(self.X, False, self.y)
self.assertEqual(np.count_nonzero(np.isnan(preds_with_y.logits)), 0)
self.assertEqual(np.count_nonzero(np.isnan(preds_with_y.logprob)), 0)
self.assertEqual(np.count_nonzero(np.isnan(preds_with_y.prob)), 0)
self.assertEqual(np.count_nonzero(np.isnan(preds_with_y.entropy)), 0)
def test_probs_between_0_and_1(self):
network = DescriminatorNet(DEFAULT_CONFIG)
preds = network(self.X)
# No Nans
self.assertEqual(tf.reduce_sum(tf.cast((preds.prob < 0.0), "float32")),
0)
self.assertEqual(tf.reduce_sum(tf.cast((preds.prob > 1.0), "float32")),
0)
self.assertEqual(
tf.reduce_sum(tf.cast((preds.logprob > 1.0), "float32")), 0)
preds_with_y = network(self.X, False, self.y)
self.assertEqual(
tf.reduce_sum(tf.cast((preds_with_y.prob < 0.0), "float32")), 0)
self.assertEqual(
tf.reduce_sum(tf.cast((preds_with_y.prob > 1.0), "float32")), 0)
self.assertEqual(
tf.reduce_sum(tf.cast((preds_with_y.logprob > 1.0), "float32")), 0)
def __init__(self,
config: GanNet.Config,
fake_getter=None,
real_getter=None,
**kwargs):
super().__init__(**kwargs)
self.descriminator = DescriminatorNet(config.descriminator)
self.generatornet = config.generator.get_generatornet_type()(
config.generator)
self.fake_getter = (fake_getter if fake_getter else
config.generator.get_fake_output_getter()())
self.real_getter = (real_getter if real_getter else
config.generator.get_real_output_getter()())
self.gan_generative = GanGenerativeNet(self.descriminator,
self.generatornet,
self.fake_getter)
self.gan_descriminative = GanDescriminativeNet(
self.descriminator,
self.generatornet,
self.fake_getter,
self.real_getter,
)
from deeper.models.adversarial_autoencoder.network import (
AdversarialAuoencoderNet, )
from deeper.models.adversarial_autoencoder.network_loss import (
AdverasrialAutoencoderLossNet, )
N_ROWS = 25
DIM_REG = 10
DIM_BOOL = 15
DIM_ORD = (5, )
DIM_CAT = (10, 8)
DIM_X = DIM_REG + DIM_BOOL + sum(DIM_ORD) + sum(DIM_CAT)
EMB_DIM = 10
LAT_DIM = 5
NCATS = 5
desciminatorConfig = DescriminatorNet.Config(embedding_dimensions=[EMB_DIM])
vaeConfig = VaeNet.Config(
input_dimensions=MultipleObjectiveDimensions(
regression=DIM_REG,
boolean=DIM_BOOL,
ordinal=DIM_ORD,
categorical=DIM_CAT,
),
output_dimensions=MultipleObjectiveDimensions(
regression=DIM_REG,
boolean=DIM_BOOL,
ordinal=DIM_ORD,
categorical=DIM_CAT,
),
encoder_embedding_dimensions=[EMB_DIM],
X_test = (X_test > 0.5).astype(float)
#%% Filter to a single label
if False:
LABEL = 9
X_train_og = X_train_og[y_train == LABEL]
X_test_og = X_test_og[y_test == LABEL]
X_train = X_train[y_train == LABEL]
X_test = X_test[y_test == LABEL]
#%% Instantiate the model
BATCH_SIZE = 12
desciminatorConfig = DescriminatorNet.Config(
embedding_dimensions=[24, 16, 12, 8],
activation=tf.keras.layers.Activation("relu"),
# embedding_dropout=0.25,
# bn_before=True,
)
vaeConfig = StackedGmvae.Config(
components=20,
cat_embedding_dimensions=[512, 512, 256, 128, 64],
input_dimensions=MultipleObjectiveDimensions(
regression=0,
boolean=X_train.shape[-1],
ordinal=(0, ),
categorical=(0, ),
),
output_dimensions=MultipleObjectiveDimensions(
regression=0,
boolean=X_train.shape[-1],
ordinal=(0, ),
import tensorflow as tf
import pytest
import unittest
import numpy as np
import os
from deeper.models.gan.descriminator import DescriminatorNet
DEFAULT_CONFIG = DescriminatorNet.Config(embedding_dimensions=[20, 15], )
class TestDescriminatorNet(unittest.TestCase):
def setUp(self):
self.X = np.random.random((20, 25))
self.y = (np.random.random((20, 1)) > 0.5).astype("float32")
def test_outputShapes(self):
network = DescriminatorNet(DEFAULT_CONFIG)
preds = network(self.X)
self.assertEqual(preds.logits.shape, (20, 1))
self.assertEqual(preds.logprob.shape, (20, 1))
self.assertEqual(preds.prob.shape, (20, 1))
self.assertEqual(preds.entropy.shape, (20, 1))
preds_with_y = network(self.X, False, self.y)
self.assertEqual(preds_with_y.logits.shape, (20, 1))
self.assertEqual(preds_with_y.logprob.shape, (20, 1))
self.assertEqual(preds_with_y.prob.shape, (20, 1))
self.assertEqual(preds_with_y.entropy.shape, (20, 1))