decoder-generative-models/cwgan.py at master · gvidigal-psr/decoder-generative-models · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
import os
import numpy as np
import tensorflow as tf
import tensorflow_probability as tfp

try:
  from misc import *
  from conditional_decoder_generator import cDecoderGenerator
  from wgan import Wasserstein_GAN
except ImportError:
  from .misc import *
  from .conditional_decoder_generator import cDecoderGenerator
  from .wgan import Wasserstein_GAN

class cWasserstein_GAN(cDecoderGenerator, Wasserstein_GAN):

  def __init__(self, data_sampler, **kw):
    super().__init__(data_sampler, **kw)

  @tf.function
  def transform(self, gen_batch, **call_kw):
    return self.generator( gen_batch, **call_kw )

  def generate(self, n_cond = 1, n_fakes = 10, ds = 'val'):
    # Sample conditions
    sampler = self.data_sampler
    all_generated = []
    all_conditions = []
    all_samples = []
    for _ in range(n_cond):
      raw_sample = self.sample_parser_fcn( sampler.sample( n_samples = 1, ds = ds ) )
      raw_sample = self._ensure_batch_size_dim(self.generator, raw_sample)
      samples = [tf.repeat( x, [n_fakes], axis = 0 ) for x in raw_sample ]
      generator_input = self.sample_generator_input(sampled_input = samples)
      generated_samples = self.transform( generator_input )
      all_generated.append(generated_samples)
      all_samples.append(raw_sample)
    return ( tf.stack( all_generated )  if n_cond > 1 else all_generated[0]
           , tf.stack( all_samples )    if n_cond > 1 else all_samples[0] )

  def extract_critic_input(self, data):
    """Only required when applying lipschitz smoothening"""
    return data[1]

  def extract_critic_conditioning(self, data):
    """Only required when applying lipschitz smoothening"""
    return data[0]

  @tf.function
  def _compute_u_hat(self, x, x_hat):
    x_shape = tf.concat([x.shape[:1],tf.ones_like(x.shape[1:],dtype=tf.int32)],axis=0)
    epsilon = tf.random.uniform(x_shape, 0.0, 1.0)
    u_hat = epsilon * x + (1 - epsilon) * x_hat
    return u_hat

  @tf.function
  def _lipschitz_penalty(self, x, x_hat, lipschitz_conditioning):
    u_hat = self._compute_u_hat(x, x_hat)
    with tf.GradientTape() as penalty_tape:
      penalty_tape.watch(u_hat)
      if not isinstance(lipschitz_conditioning, list):
        lipschitz_conditioning = [lipschitz_conditioning]
      critic_input = lipschitz_conditioning + [u_hat]
      func = self.critic(critic_input, **self._training_kw)
    grads = penalty_tape.gradient(func, u_hat)
    norm_grads = tf.sqrt(tf.reduce_sum(tf.square(grads), axis=tf.range(1,tf.size(tf.shape(x)))))
    lipschitz = tf.math.square( tf.reduce_mean((norm_grads - 1)))
    lipschitz = tf.multiply(self._grad_weight, lipschitz)
    return lipschitz

  @tf.function
  def _surrogate_loss( self, data_output, gen_output, critic_lipschitz ):
    wasserstein_loss, critic_data, critic_generator = self.wasserstein_loss(data_output, gen_output)
    critic_total = tf.add( wasserstein_loss, critic_lipschitz )
    return { "critic_total":     critic_total
           , "lipschitz":        critic_lipschitz
           , "critic_data":      critic_data
           , "critic_gen":       critic_generator
           , "wasserstein":      wasserstein_loss }

  @tf.function
  def _split_lipschitz( self, data_batch, gen_batch ):
    data_idxs = tf.random.shuffle( tf.range(start = 0, limit = tf.constant(self.data_sampler._batch_size)) )[:self.data_sampler._batch_size//2]
    gen_idxs = tf.random.shuffle( tf.range(start = 0, limit = tf.constant(self.data_sampler._batch_size)) )[:self.data_sampler._batch_size//2]
    lipschitz_conditioning = [tf.concat([tf.gather(x,data_idxs)
                                        ,tf.gather(y,gen_idxs)], axis = 0)
                              for x, y in zip(self.extract_critic_conditioning(data_batch)
                                             ,self.extract_critic_conditioning(gen_batch))]
    data_inputs = tf.concat([tf.gather(self.extract_critic_input(data_batch), data_idxs)
                            ,tf.gather(self.extract_critic_input(gen_batch), gen_idxs)], axis = 0)
    sampled_input = tf.gather( self.extract_generator_input_from_standard_batch_fcn(data_batch), data_idxs)
    if not isinstance(sampled_input, list):
      sampled_input = [sampled_input]
    generator_inpput = sampled_input + [self.sample_latent( self.data_sampler._batch_size//2 ) ]
    gen_inputs = tf.concat([self.transform( generator_input, **self._training_kw )
                            ,tf.gather(self.extract_critic_input(gen_batch), gen_idxs)], axis = 0)
    return data_inputs, gen_inputs, lipschitz_conditioning

  @tf.function
  def _train_step(self, data_batch, gen_batch, critic_only = False):
    with tf.GradientTape() as gen_tape, tf.GradientTape() as critic_tape:
      gen_batch  = self.transform(gen_batch, **self._training_kw)
      data_output = self.critic(data_batch, **self._training_kw)
      gen_output = self.critic(gen_batch, **self._training_kw)
      if self._use_lipschitz_penalty:
        if not(self.use_same_real_fake_conditioning):
          data_inputs, gen_inputs, lipschitz_conditioning = self._split_lipschitz( data_batch, gen_batch )
        else:
          data_inputs            = self.extract_critic_input(data_batch)
          gen_inputs             = self.extract_critic_input(gen_batch)
          lipschitz_conditioning = self.extract_critic_conditioning(data_batch)
        lipschitz = self._lipschitz_penalty(data_inputs, gen_inputs, lipschitz_conditioning)
      else:
        lipschitz = tf.constant(0.)
      critic_loss_dict = self._surrogate_loss( data_output, gen_output, lipschitz )
    # gen_tape, critic_tape
    self._apply_critic_update( critic_tape, critic_loss_dict["critic_total"] )
    if not critic_only:
      self._apply_gen_update( gen_tape, critic_loss_dict["critic_gen"] )
    return critic_loss_dict