DCGAN on the MNIST Dataset¶

In [1]:

import tempfile
import os
import numpy as np
import tensorflow as tf
from tensorflow.python.keras import layers
from matplotlib import pyplot as plt
import fastestimator as fe
from fastestimator.backend import binary_crossentropy, feed_forward
from fastestimator.dataset.data import mnist
from fastestimator.op.numpyop import NumpyOp
from fastestimator.op.numpyop.univariate import ExpandDims, Normalize
from fastestimator.op.tensorop import TensorOp
from fastestimator.op.tensorop.model import ModelOp, UpdateOp
from fastestimator.trace.io import ModelSaver

import tempfile import os import numpy as np import tensorflow as tf from tensorflow.python.keras import layers from matplotlib import pyplot as plt import fastestimator as fe from fastestimator.backend import binary_crossentropy, feed_forward from fastestimator.dataset.data import mnist from fastestimator.op.numpyop import NumpyOp from fastestimator.op.numpyop.univariate import ExpandDims, Normalize from fastestimator.op.tensorop import TensorOp from fastestimator.op.tensorop.model import ModelOp, UpdateOp from fastestimator.trace.io import ModelSaver

In [2]:

parameters

batch_size = 256
epochs = 50
max_train_steps_per_epoch = None
save_dir = tempfile.mkdtemp()
model_name = 'model_epoch_50.h5'

batch_size = 256 epochs = 50 max_train_steps_per_epoch = None save_dir = tempfile.mkdtemp() model_name = 'model_epoch_50.h5'

Building components

Step 1: Prepare training and define a Pipeline¶

We are loading data from tf.keras.datasets.mnist and defining a series of operations to perform on the data before the training:

In [3]:

train_data, _ = mnist.load_data()
pipeline = fe.Pipeline(
    train_data=train_data,
    batch_size=batch_size,
    ops=[
        ExpandDims(inputs="x", outputs="x"),
        Normalize(inputs="x", outputs="x", mean=1.0, std=1.0, max_pixel_value=127.5),
        NumpyOp(inputs=lambda: np.random.normal(size=[100]).astype('float32'), outputs="z")
    ])

train_data, _ = mnist.load_data() pipeline = fe.Pipeline( train_data=train_data, batch_size=batch_size, ops=[ ExpandDims(inputs="x", outputs="x"), Normalize(inputs="x", outputs="x", mean=1.0, std=1.0, max_pixel_value=127.5), NumpyOp(inputs=lambda: np.random.normal(size=[100]).astype('float32'), outputs="z") ])

Step 2: Create a model and FastEstimator Network¶

First, we have to define the network architecture for both our Generator and Discriminator. After defining the architecture, users are expected to feed the architecture definition, along with associated model names and optimizers, to fe.build.

In [4]:

def generator():
    model = tf.keras.Sequential()
    model.add(layers.Dense(7 * 7 * 256, use_bias=False, input_shape=(100, )))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())
    model.add(layers.Reshape((7, 7, 256)))
    model.add(layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())
    model.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())
    model.add(layers.Conv2DTranspose(1, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))
    return model

def generator(): model = tf.keras.Sequential() model.add(layers.Dense(7 * 7 * 256, use_bias=False, input_shape=(100, ))) model.add(layers.BatchNormalization()) model.add(layers.LeakyReLU()) model.add(layers.Reshape((7, 7, 256))) model.add(layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False)) model.add(layers.BatchNormalization()) model.add(layers.LeakyReLU()) model.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False)) model.add(layers.BatchNormalization()) model.add(layers.LeakyReLU()) model.add(layers.Conv2DTranspose(1, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh')) return model

In [5]:

def discriminator():
    model = tf.keras.Sequential()
    model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same', input_shape=[28, 28, 1]))
    model.add(layers.LeakyReLU())
    model.add(layers.Dropout(0.3))
    model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same'))
    model.add(layers.LeakyReLU())
    model.add(layers.Dropout(0.3))
    model.add(layers.Flatten())
    model.add(layers.Dense(1))
    return model

def discriminator(): model = tf.keras.Sequential() model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same', input_shape=[28, 28, 1])) model.add(layers.LeakyReLU()) model.add(layers.Dropout(0.3)) model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same')) model.add(layers.LeakyReLU()) model.add(layers.Dropout(0.3)) model.add(layers.Flatten()) model.add(layers.Dense(1)) return model

In [6]:

gen_model = fe.build(model_fn=generator, optimizer_fn=lambda: tf.optimizers.Adam(1e-4))
disc_model = fe.build(model_fn=discriminator, optimizer_fn=lambda: tf.optimizers.Adam(1e-4))

gen_model = fe.build(model_fn=generator, optimizer_fn=lambda: tf.optimizers.Adam(1e-4)) disc_model = fe.build(model_fn=discriminator, optimizer_fn=lambda: tf.optimizers.Adam(1e-4))

We define the generator and discriminator losses. These can have multiple inputs and outputs.

In [7]:

class GLoss(TensorOp):
    """Compute generator loss."""
    def forward(self, data, state):
        return binary_crossentropy(y_pred=data, y_true=tf.ones_like(data), from_logits=True)

class GLoss(TensorOp): """Compute generator loss.""" def forward(self, data, state): return binary_crossentropy(y_pred=data, y_true=tf.ones_like(data), from_logits=True)

In [8]:

class DLoss(TensorOp):
    """Compute discriminator loss."""
    def forward(self, data, state):
        true_score, fake_score = data
        real_loss = binary_crossentropy(y_pred=true_score, y_true=tf.ones_like(true_score), from_logits=True)
        fake_loss = binary_crossentropy(y_pred=fake_score, y_true=tf.zeros_like(fake_score), from_logits=True)
        total_loss = real_loss + fake_loss
        return total_loss

class DLoss(TensorOp): """Compute discriminator loss.""" def forward(self, data, state): true_score, fake_score = data real_loss = binary_crossentropy(y_pred=true_score, y_true=tf.ones_like(true_score), from_logits=True) fake_loss = binary_crossentropy(y_pred=fake_score, y_true=tf.zeros_like(fake_score), from_logits=True) total_loss = real_loss + fake_loss return total_loss

fe.Network takes series of operators. Here we pass our models wrapped into ModelOps along with our loss functions and some update rules:

In [9]:

network = fe.Network(ops=[
        ModelOp(model=gen_model, inputs="z", outputs="x_fake"),
        ModelOp(model=disc_model, inputs="x_fake", outputs="fake_score"),
        GLoss(inputs="fake_score", outputs="gloss"),
        UpdateOp(model=gen_model, loss_name="gloss"),
        ModelOp(inputs="x", model=disc_model, outputs="true_score"),
        DLoss(inputs=("true_score", "fake_score"), outputs="dloss"),
        UpdateOp(model=disc_model, loss_name="dloss")
    ])

network = fe.Network(ops=[ ModelOp(model=gen_model, inputs="z", outputs="x_fake"), ModelOp(model=disc_model, inputs="x_fake", outputs="fake_score"), GLoss(inputs="fake_score", outputs="gloss"), UpdateOp(model=gen_model, loss_name="gloss"), ModelOp(inputs="x", model=disc_model, outputs="true_score"), DLoss(inputs=("true_score", "fake_score"), outputs="dloss"), UpdateOp(model=disc_model, loss_name="dloss") ])

Step 3: Prepare Estimator and configure the training loop¶

We will define an Estimator that has four notable arguments: network, pipeline, epochs and traces. Our Network and Pipeline objects are passed here as an argument along with the number of epochs and a Trace, in this case one designed to save our model every 5 epochs.

In [10]:

traces=ModelSaver(model=gen_model, save_dir=save_dir, frequency=5)

traces=ModelSaver(model=gen_model, save_dir=save_dir, frequency=5)

In [11]:

estimator = fe.Estimator(pipeline=pipeline,
                         network=network,
                         epochs=epochs,
                         traces=traces,
                         max_train_steps_per_epoch=max_train_steps_per_epoch)

estimator = fe.Estimator(pipeline=pipeline, network=network, epochs=epochs, traces=traces, max_train_steps_per_epoch=max_train_steps_per_epoch)

Training

In [12]:

estimator.fit()

estimator.fit()

    ______           __  ______     __  _                 __            
   / ____/___ ______/ /_/ ____/____/ /_(_)___ ___  ____ _/ /_____  _____
  / /_  / __ `/ ___/ __/ __/ / ___/ __/ / __ `__ \/ __ `/ __/ __ \/ ___/
 / __/ / /_/ (__  ) /_/ /___(__  ) /_/ / / / / / / /_/ / /_/ /_/ / /    
/_/    \__,_/____/\__/_____/____/\__/_/_/ /_/ /_/\__,_/\__/\____/_/     
                                                                        

FastEstimator-Start: step: 1; model_lr: 1e-04; model1_lr: 1e-04; 
FastEstimator-Train: step: 1; gloss: 0.7122225; dloss: 1.3922014; 
FastEstimator-Train: step: 100; gloss: 0.906471; dloss: 0.8187004; steps/sec: 10.09; 
FastEstimator-Train: step: 200; gloss: 0.59155834; dloss: 1.5896755; steps/sec: 9.93; 
FastEstimator-Train: step: 235; epoch: 1; epoch_time: 28.53 sec; 
FastEstimator-Train: step: 300; gloss: 0.7163421; dloss: 1.3333399; steps/sec: 8.9; 
FastEstimator-Train: step: 400; gloss: 0.6816584; dloss: 1.6007018; steps/sec: 9.88; 
FastEstimator-Train: step: 470; epoch: 2; epoch_time: 23.95 sec; 
FastEstimator-Train: step: 500; gloss: 0.7051203; dloss: 1.4395489; steps/sec: 9.69; 
FastEstimator-Train: step: 600; gloss: 0.75529504; dloss: 1.2358603; steps/sec: 9.86; 
FastEstimator-Train: step: 700; gloss: 0.8082159; dloss: 1.2728964; steps/sec: 9.84; 
FastEstimator-Train: step: 705; epoch: 3; epoch_time: 24.03 sec; 
FastEstimator-Train: step: 800; gloss: 0.8434949; dloss: 1.3006642; steps/sec: 9.65; 
FastEstimator-Train: step: 900; gloss: 0.84470236; dloss: 1.2344811; steps/sec: 9.79; 
FastEstimator-Train: step: 940; epoch: 4; epoch_time: 24.16 sec; 
FastEstimator-Train: step: 1000; gloss: 0.9431131; dloss: 1.0444374; steps/sec: 9.66; 
FastEstimator-Train: step: 1100; gloss: 0.6982814; dloss: 1.5213135; steps/sec: 9.77; 
Saved model to /tmp/tmpspul4xo8/model_epoch_5.h5
FastEstimator-Train: step: 1175; epoch: 5; epoch_time: 24.17 sec; 
FastEstimator-Train: step: 1200; gloss: 1.2540445; dloss: 0.8278078; steps/sec: 9.63; 
FastEstimator-Train: step: 1300; gloss: 0.70465124; dloss: 1.7595482; steps/sec: 9.76; 
FastEstimator-Train: step: 1400; gloss: 0.83103234; dloss: 1.168882; steps/sec: 9.77; 
FastEstimator-Train: step: 1410; epoch: 6; epoch_time: 24.22 sec; 
FastEstimator-Train: step: 1500; gloss: 0.86833733; dloss: 1.2078841; steps/sec: 9.57; 
FastEstimator-Train: step: 1600; gloss: 0.82795817; dloss: 1.2242851; steps/sec: 9.75; 
FastEstimator-Train: step: 1645; epoch: 7; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 1700; gloss: 0.9743507; dloss: 1.0731742; steps/sec: 9.59; 
FastEstimator-Train: step: 1800; gloss: 0.89325964; dloss: 1.1766281; steps/sec: 9.76; 
FastEstimator-Train: step: 1880; epoch: 8; epoch_time: 24.25 sec; 
FastEstimator-Train: step: 1900; gloss: 1.0287898; dloss: 0.9916363; steps/sec: 9.6; 
FastEstimator-Train: step: 2000; gloss: 0.8240694; dloss: 1.313368; steps/sec: 9.74; 
FastEstimator-Train: step: 2100; gloss: 0.9738071; dloss: 1.1259043; steps/sec: 9.73; 
FastEstimator-Train: step: 2115; epoch: 9; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 2200; gloss: 1.0899432; dloss: 1.0272337; steps/sec: 9.57; 
FastEstimator-Train: step: 2300; gloss: 0.868231; dloss: 1.2400149; steps/sec: 9.72; 
Saved model to /tmp/tmpspul4xo8/model_epoch_10.h5
FastEstimator-Train: step: 2350; epoch: 10; epoch_time: 24.35 sec; 
FastEstimator-Train: step: 2400; gloss: 0.9001913; dloss: 1.1931081; steps/sec: 9.58; 
FastEstimator-Train: step: 2500; gloss: 1.0865673; dloss: 0.8990781; steps/sec: 9.71; 
FastEstimator-Train: step: 2585; epoch: 11; epoch_time: 24.34 sec; 
FastEstimator-Train: step: 2600; gloss: 0.7485407; dloss: 1.672249; steps/sec: 9.58; 
FastEstimator-Train: step: 2700; gloss: 1.045316; dloss: 1.0383615; steps/sec: 9.73; 
FastEstimator-Train: step: 2800; gloss: 0.7666995; dloss: 1.4343789; steps/sec: 9.72; 
FastEstimator-Train: step: 2820; epoch: 12; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 2900; gloss: 1.1756387; dloss: 0.96622103; steps/sec: 9.6; 
FastEstimator-Train: step: 3000; gloss: 0.9090629; dloss: 1.1984154; steps/sec: 9.72; 
FastEstimator-Train: step: 3055; epoch: 13; epoch_time: 24.29 sec; 
FastEstimator-Train: step: 3100; gloss: 0.9301505; dloss: 1.113826; steps/sec: 9.59; 
FastEstimator-Train: step: 3200; gloss: 0.99965835; dloss: 1.0707076; steps/sec: 9.74; 
FastEstimator-Train: step: 3290; epoch: 14; epoch_time: 24.31 sec; 
FastEstimator-Train: step: 3300; gloss: 0.80838567; dloss: 1.6384692; steps/sec: 9.55; 
FastEstimator-Train: step: 3400; gloss: 0.8714433; dloss: 1.326818; steps/sec: 9.74; 
FastEstimator-Train: step: 3500; gloss: 0.9549879; dloss: 1.2086997; steps/sec: 9.73; 
Saved model to /tmp/tmpspul4xo8/model_epoch_15.h5
FastEstimator-Train: step: 3525; epoch: 15; epoch_time: 24.34 sec; 
FastEstimator-Train: step: 3600; gloss: 1.0164418; dloss: 1.0690243; steps/sec: 9.57; 
FastEstimator-Train: step: 3700; gloss: 1.0357686; dloss: 1.0537144; steps/sec: 9.72; 
FastEstimator-Train: step: 3760; epoch: 16; epoch_time: 24.32 sec; 
FastEstimator-Train: step: 3800; gloss: 0.7402923; dloss: 1.4840925; steps/sec: 9.6; 
FastEstimator-Train: step: 3900; gloss: 0.91192436; dloss: 1.3617609; steps/sec: 9.72; 
FastEstimator-Train: step: 3995; epoch: 17; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 4000; gloss: 1.2626994; dloss: 0.9568275; steps/sec: 9.59; 
FastEstimator-Train: step: 4100; gloss: 0.97824305; dloss: 1.300906; steps/sec: 9.74; 
FastEstimator-Train: step: 4200; gloss: 0.93075603; dloss: 1.387594; steps/sec: 9.73; 
FastEstimator-Train: step: 4230; epoch: 18; epoch_time: 24.31 sec; 
FastEstimator-Train: step: 4300; gloss: 1.0180345; dloss: 1.0898602; steps/sec: 9.59; 
FastEstimator-Train: step: 4400; gloss: 1.051662; dloss: 1.3392837; steps/sec: 9.71; 
FastEstimator-Train: step: 4465; epoch: 19; epoch_time: 24.33 sec; 
FastEstimator-Train: step: 4500; gloss: 1.0151768; dloss: 1.1482071; steps/sec: 9.56; 
FastEstimator-Train: step: 4600; gloss: 1.107022; dloss: 0.96815336; steps/sec: 9.71; 
FastEstimator-Train: step: 4700; gloss: 1.0924942; dloss: 1.1389236; steps/sec: 9.72; 
Saved model to /tmp/tmpspul4xo8/model_epoch_20.h5
FastEstimator-Train: step: 4700; epoch: 20; epoch_time: 24.39 sec; 
FastEstimator-Train: step: 4800; gloss: 1.1345683; dloss: 1.2068424; steps/sec: 9.54; 
FastEstimator-Train: step: 4900; gloss: 1.142304; dloss: 0.9673606; steps/sec: 9.74; 
FastEstimator-Train: step: 4935; epoch: 21; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 5000; gloss: 0.9886; dloss: 1.0960109; steps/sec: 9.57; 
FastEstimator-Train: step: 5100; gloss: 0.8936993; dloss: 1.2922779; steps/sec: 9.74; 
FastEstimator-Train: step: 5170; epoch: 22; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 5200; gloss: 1.1095095; dloss: 1.1243165; steps/sec: 9.57; 
FastEstimator-Train: step: 5300; gloss: 1.2485275; dloss: 0.89292765; steps/sec: 9.74; 
FastEstimator-Train: step: 5400; gloss: 1.0476826; dloss: 1.2994311; steps/sec: 9.75; 
FastEstimator-Train: step: 5405; epoch: 23; epoch_time: 24.29 sec; 
FastEstimator-Train: step: 5500; gloss: 1.3308836; dloss: 0.871735; steps/sec: 9.63; 
FastEstimator-Train: step: 5600; gloss: 1.115385; dloss: 1.2837725; steps/sec: 9.74; 
FastEstimator-Train: step: 5640; epoch: 24; epoch_time: 24.23 sec; 
FastEstimator-Train: step: 5700; gloss: 1.1920481; dloss: 1.0993654; steps/sec: 9.62; 
FastEstimator-Train: step: 5800; gloss: 1.3005233; dloss: 0.914361; steps/sec: 9.74; 
Saved model to /tmp/tmpspul4xo8/model_epoch_25.h5
FastEstimator-Train: step: 5875; epoch: 25; epoch_time: 24.27 sec; 
FastEstimator-Train: step: 5900; gloss: 1.3146336; dloss: 0.8816396; steps/sec: 9.6; 
FastEstimator-Train: step: 6000; gloss: 0.9764897; dloss: 1.289681; steps/sec: 9.75; 
FastEstimator-Train: step: 6100; gloss: 1.1467731; dloss: 1.1918977; steps/sec: 9.75; 
FastEstimator-Train: step: 6110; epoch: 26; epoch_time: 24.26 sec; 
FastEstimator-Train: step: 6200; gloss: 1.6301311; dloss: 0.9541445; steps/sec: 9.6; 
FastEstimator-Train: step: 6300; gloss: 1.2840165; dloss: 0.9587291; steps/sec: 9.73; 
FastEstimator-Train: step: 6345; epoch: 27; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 6400; gloss: 1.1097628; dloss: 1.0090048; steps/sec: 9.59; 
FastEstimator-Train: step: 6500; gloss: 1.2495477; dloss: 0.89897555; steps/sec: 9.73; 
FastEstimator-Train: step: 6580; epoch: 28; epoch_time: 24.32 sec; 
FastEstimator-Train: step: 6600; gloss: 1.1773547; dloss: 1.1330662; steps/sec: 9.58; 
FastEstimator-Train: step: 6700; gloss: 1.246088; dloss: 0.8964198; steps/sec: 9.73; 
FastEstimator-Train: step: 6800; gloss: 1.2250234; dloss: 1.0358574; steps/sec: 9.73; 
FastEstimator-Train: step: 6815; epoch: 29; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 6900; gloss: 1.1256618; dloss: 1.196687; steps/sec: 9.59; 
FastEstimator-Train: step: 7000; gloss: 1.1131527; dloss: 1.0596428; steps/sec: 9.73; 
Saved model to /tmp/tmpspul4xo8/model_epoch_30.h5
FastEstimator-Train: step: 7050; epoch: 30; epoch_time: 24.31 sec; 
FastEstimator-Train: step: 7100; gloss: 1.1662202; dloss: 1.0555116; steps/sec: 9.57; 
FastEstimator-Train: step: 7200; gloss: 1.0653521; dloss: 1.1444951; steps/sec: 9.71; 
FastEstimator-Train: step: 7285; epoch: 31; epoch_time: 24.36 sec; 
FastEstimator-Train: step: 7300; gloss: 1.1732882; dloss: 1.1456137; steps/sec: 9.56; 
FastEstimator-Train: step: 7400; gloss: 1.0872216; dloss: 1.128233; steps/sec: 9.74; 
FastEstimator-Train: step: 7500; gloss: 1.2431256; dloss: 1.1538315; steps/sec: 9.73; 
FastEstimator-Train: step: 7520; epoch: 32; epoch_time: 24.31 sec; 
FastEstimator-Train: step: 7600; gloss: 1.0806718; dloss: 1.2206206; steps/sec: 9.57; 
FastEstimator-Train: step: 7700; gloss: 1.1804712; dloss: 1.1420157; steps/sec: 9.71; 
FastEstimator-Train: step: 7755; epoch: 33; epoch_time: 24.36 sec; 
FastEstimator-Train: step: 7800; gloss: 1.1762993; dloss: 1.0413929; steps/sec: 9.54; 
FastEstimator-Train: step: 7900; gloss: 1.2267275; dloss: 0.98290396; steps/sec: 9.71; 
FastEstimator-Train: step: 7990; epoch: 34; epoch_time: 24.37 sec; 
FastEstimator-Train: step: 8000; gloss: 1.1847881; dloss: 1.0905983; steps/sec: 9.55; 
FastEstimator-Train: step: 8100; gloss: 1.1490288; dloss: 1.1739209; steps/sec: 9.72; 
FastEstimator-Train: step: 8200; gloss: 1.0283768; dloss: 1.059457; steps/sec: 9.73; 
Saved model to /tmp/tmpspul4xo8/model_epoch_35.h5
FastEstimator-Train: step: 8225; epoch: 35; epoch_time: 24.33 sec; 
FastEstimator-Train: step: 8300; gloss: 1.2351133; dloss: 1.1085691; steps/sec: 9.59; 
FastEstimator-Train: step: 8400; gloss: 1.1488228; dloss: 1.1410246; steps/sec: 9.74; 
FastEstimator-Train: step: 8460; epoch: 36; epoch_time: 24.31 sec; 
FastEstimator-Train: step: 8500; gloss: 1.152446; dloss: 1.1371456; steps/sec: 9.55; 
FastEstimator-Train: step: 8600; gloss: 1.2175394; dloss: 1.1543391; steps/sec: 9.74; 
FastEstimator-Train: step: 8695; epoch: 37; epoch_time: 24.31 sec; 
FastEstimator-Train: step: 8700; gloss: 1.1803217; dloss: 1.1517241; steps/sec: 9.59; 
FastEstimator-Train: step: 8800; gloss: 0.9561673; dloss: 1.2418871; steps/sec: 9.77; 
FastEstimator-Train: step: 8900; gloss: 1.0239995; dloss: 1.243228; steps/sec: 9.74; 
FastEstimator-Train: step: 8930; epoch: 38; epoch_time: 24.26 sec; 
FastEstimator-Train: step: 9000; gloss: 0.98074543; dloss: 1.2558163; steps/sec: 9.59; 
FastEstimator-Train: step: 9100; gloss: 1.0084043; dloss: 1.2773612; steps/sec: 9.74; 
FastEstimator-Train: step: 9165; epoch: 39; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 9200; gloss: 1.0313301; dloss: 1.2312038; steps/sec: 9.58; 
FastEstimator-Train: step: 9300; gloss: 1.0100834; dloss: 1.2482088; steps/sec: 9.73; 
FastEstimator-Train: step: 9400; gloss: 0.9327201; dloss: 1.3188391; steps/sec: 9.73; 
Saved model to /tmp/tmpspul4xo8/model_epoch_40.h5
FastEstimator-Train: step: 9400; epoch: 40; epoch_time: 24.34 sec; 
FastEstimator-Train: step: 9500; gloss: 1.1315899; dloss: 1.1447232; steps/sec: 9.55; 
FastEstimator-Train: step: 9600; gloss: 1.1352619; dloss: 1.0802212; steps/sec: 9.72; 
FastEstimator-Train: step: 9635; epoch: 41; epoch_time: 24.3 sec; 
FastEstimator-Train: step: 9700; gloss: 1.01453; dloss: 1.0975459; steps/sec: 9.59; 
FastEstimator-Train: step: 9800; gloss: 0.90930146; dloss: 1.2942967; steps/sec: 9.76; 
FastEstimator-Train: step: 9870; epoch: 42; epoch_time: 24.27 sec; 
FastEstimator-Train: step: 9900; gloss: 1.0540565; dloss: 1.170531; steps/sec: 9.58; 
FastEstimator-Train: step: 10000; gloss: 1.061863; dloss: 1.2722391; steps/sec: 9.76; 
FastEstimator-Train: step: 10100; gloss: 0.9647354; dloss: 1.1689386; steps/sec: 9.73; 
FastEstimator-Train: step: 10105; epoch: 43; epoch_time: 24.29 sec; 
FastEstimator-Train: step: 10200; gloss: 1.2080085; dloss: 1.075758; steps/sec: 9.57; 
FastEstimator-Train: step: 10300; gloss: 1.0741084; dloss: 1.1352613; steps/sec: 9.72; 
FastEstimator-Train: step: 10340; epoch: 44; epoch_time: 24.34 sec; 
FastEstimator-Train: step: 10400; gloss: 1.1394867; dloss: 0.9788251; steps/sec: 9.58; 
FastEstimator-Train: step: 10500; gloss: 1.1983887; dloss: 1.0792823; steps/sec: 9.73; 
Saved model to /tmp/tmpspul4xo8/model_epoch_45.h5
FastEstimator-Train: step: 10575; epoch: 45; epoch_time: 24.31 sec; 
FastEstimator-Train: step: 10600; gloss: 0.96989757; dloss: 1.2448618; steps/sec: 9.59; 
FastEstimator-Train: step: 10700; gloss: 0.9579427; dloss: 1.2773428; steps/sec: 9.7; 
FastEstimator-Train: step: 10800; gloss: 0.97453225; dloss: 1.2194138; steps/sec: 9.72; 
FastEstimator-Train: step: 10810; epoch: 46; epoch_time: 24.36 sec; 
FastEstimator-Train: step: 10900; gloss: 1.0218571; dloss: 1.1765122; steps/sec: 9.57; 
FastEstimator-Train: step: 11000; gloss: 1.1221988; dloss: 1.1675267; steps/sec: 9.71; 
FastEstimator-Train: step: 11045; epoch: 47; epoch_time: 24.34 sec; 
FastEstimator-Train: step: 11100; gloss: 0.900293; dloss: 1.1741953; steps/sec: 9.57; 
FastEstimator-Train: step: 11200; gloss: 1.1080045; dloss: 1.1090837; steps/sec: 9.73; 
FastEstimator-Train: step: 11280; epoch: 48; epoch_time: 24.31 sec; 
FastEstimator-Train: step: 11300; gloss: 1.1028197; dloss: 1.1044464; steps/sec: 9.59; 
FastEstimator-Train: step: 11400; gloss: 1.0530615; dloss: 1.2150866; steps/sec: 9.74; 
FastEstimator-Train: step: 11500; gloss: 0.8997061; dloss: 1.3101699; steps/sec: 9.75; 
FastEstimator-Train: step: 11515; epoch: 49; epoch_time: 24.29 sec; 
FastEstimator-Train: step: 11600; gloss: 1.0087067; dloss: 1.3297874; steps/sec: 9.58; 
FastEstimator-Train: step: 11700; gloss: 1.0053492; dloss: 1.2499433; steps/sec: 9.74; 
Saved model to /tmp/tmpspul4xo8/model_epoch_50.h5
FastEstimator-Train: step: 11750; epoch: 50; epoch_time: 24.3 sec; 
FastEstimator-Finish: step: 11750; total_time: 1219.47 sec; model_lr: 1e-04; model1_lr: 1e-04; 

Inferencing

For inferencing, first we have to load the trained model weights. We will load the trained generator weights using fe.build

In [19]:

model_path = os.path.join(save_dir, model_name)
trained_model = fe.build(model_fn=generator, weights_path=model_path, optimizer_fn=lambda: tf.optimizers.Adam(1e-4))

model_path = os.path.join(save_dir, model_name) trained_model = fe.build(model_fn=generator, weights_path=model_path, optimizer_fn=lambda: tf.optimizers.Adam(1e-4))

Loaded model weights from /tmp/tmpspul4xo8/model_epoch_50.h5

We will the generate some images from random noise:

In [20]:

images = feed_forward(trained_model, np.random.normal(size=(16, 100)), training=False)

images = feed_forward(trained_model, np.random.normal(size=(16, 100)), training=False)

WARNING:tensorflow:Layer dense_4 is casting an input tensor from dtype float64 to the layer's dtype of float32, which is new behavior in TensorFlow 2.  The layer has dtype float32 because it's dtype defaults to floatx.

If you intended to run this layer in float32, you can safely ignore this warning. If in doubt, this warning is likely only an issue if you are porting a TensorFlow 1.X model to TensorFlow 2.

To change all layers to have dtype float64 by default, call `tf.keras.backend.set_floatx('float64')`. To change just this layer, pass dtype='float64' to the layer constructor. If you are the author of this layer, you can disable autocasting by passing autocast=False to the base Layer constructor.

In [21]:

fig, axes = plt.subplots(4, 4)
axes = np.ravel(axes)
for i in range(images.shape[0]):
    axes[i].axis('off')
    axes[i].imshow(np.squeeze(images[i, ...] * 127.5 + 127.5), cmap='gray')

fig, axes = plt.subplots(4, 4) axes = np.ravel(axes) for i in range(images.shape[0]): axes[i].axis('off') axes[i].imshow(np.squeeze(images[i, ...] * 127.5 + 127.5), cmap='gray')