Why is my Generator loss so negative and my discriminator is not learning?

Research & Models
#1

ganite.py

# Necessary packages
import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np
from utils import xavier_init, batch_generator


def ganite (train_x, train_t, train_y, test_x, parameters):
  """GANITE module.
  
  Args:
    - train_x: features in training data
    - train_t: treatments in training data
    - train_y: observed outcomes in training data
    - test_x: features in testing data
    - parameters: GANITE network parameters
      - h_dim: hidden dimensions
      - batch_size: the number of samples in each batch
      - iterations: the number of iterations for training
      - alpha: hyper-parameter to adjust the loss importance
      
  Returns:
    - test_y_hat: estimated potential outcome for testing set
  """
  # Parameters 
  h_dim = parameters['h_dim']
  batch_size = parameters['batch_size']
  iterations = parameters['iteration']
  alpha = parameters['alpha']
  
  no = train_x.shape[0]
  dim = 1
  g_loss_array=[]
  d_loss_array=[]

  # Reset graph
  tf.compat.v1.reset_default_graph()

  ## 1. Placeholder
  # 1.1. Feature (X)
  X = tf.placeholder(tf.float32, shape = [None, dim])
  # 1.2. Treatment (T)
  T = tf.placeholder(tf.float32, shape = [None, 1])
  # 1.3. Outcome (Y)
  Y = tf.placeholder(tf.float32, shape = [None, 1])

  ## 2. Variables
  # 2.1 Generator
  G_W1 = tf.Variable(xavier_init([(dim+2), h_dim])) # Inputs: X + Treatment + Factual outcome
  G_b1 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  G_W2 = tf.Variable(xavier_init([h_dim, h_dim]))
  G_b2 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  # Multi-task outputs for increasing the flexibility of the generator 
  G_W31 = tf.Variable(xavier_init([h_dim, h_dim]))
  G_b31 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  G_W32 = tf.Variable(xavier_init([h_dim, 1]))
  G_b32 = tf.Variable(tf.zeros(shape = [1])) # Output: Estimated outcome when t = 0
  
  G_W41 = tf.Variable(xavier_init([h_dim, h_dim]))
  G_b41 = tf.Variable(tf.zeros(shape = [h_dim])) 
  
  G_W42 = tf.Variable(xavier_init([h_dim, 1]))
  G_b42 = tf.Variable(tf.zeros(shape = [1])) # Output: Estimated outcome when t = 1
  
  # Generator variables
  theta_G = [G_W1, G_W2, G_W31, G_W32, G_W41, G_W42, G_b1, G_b2, G_b31, G_b32, G_b41, G_b42]
  
  # 2.2 Discriminator
  D_W1 = tf.Variable(xavier_init([(dim+2), h_dim])) # Inputs: X + Factual outcomes + Estimated counterfactual outcomes
  D_b1 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  D_W2 = tf.Variable(xavier_init([h_dim, h_dim]))
  D_b2 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  D_W3 = tf.Variable(xavier_init([h_dim, 1]))
  D_b3 = tf.Variable(tf.zeros(shape = [1]))
  
  # Discriminator variables
  theta_D = [D_W1, D_W2, D_W3, D_b1, D_b2, D_b3]
  
  # 2.3 Inference network
  I_W1 = tf.Variable(xavier_init([(dim), h_dim])) # Inputs: X
  I_b1 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  I_W2 = tf.Variable(xavier_init([h_dim, h_dim]))
  I_b2 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  # Multi-task outputs for increasing the flexibility of the inference network
  I_W31 = tf.Variable(xavier_init([h_dim, h_dim]))
  I_b31 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  I_W32 = tf.Variable(xavier_init([h_dim, 1]))
  I_b32 = tf.Variable(tf.zeros(shape = [1])) # Output: Estimated outcome when t = 0
  
  I_W41 = tf.Variable(xavier_init([h_dim, h_dim]))
  I_b41 = tf.Variable(tf.zeros(shape = [h_dim]))
  
  I_W42 = tf.Variable(xavier_init([h_dim, 1]))
  I_b42 = tf.Variable(tf.zeros(shape = [1])) # Output: Estimated outcome when t = 1
  
  # Inference network variables
  theta_I = [I_W1, I_W2, I_W31, I_W32, I_W41, I_W42, I_b1, I_b2, I_b31, I_b32, I_b41, I_b42]

  ## 3. Definitions of generator, discriminator and inference networks
  # 3.1 Generator
  def generator(x, t, y):
    """Generator function.
    
    Args:
      - x: features
      - t: treatments
      - y: observed labels
      
    Returns:
      - G_logit: estimated potential outcomes
    """
    # Concatenate feature, treatments, and observed labels as input
    inputs = tf.concat(axis = 1, values = [x,t,y])
    G_h1 = tf.nn.relu(tf.matmul(inputs, G_W1) + G_b1)
    G_h2 = tf.nn.relu(tf.matmul(G_h1, G_W2) + G_b2)
      
    # Estimated outcome if t = 0
    G_h31 = tf.nn.relu(tf.matmul(G_h2, G_W31) + G_b31)
    G_logit1 = tf.matmul(G_h31, G_W32) + G_b32
      
    # Estimated outcome if t = 1
    G_h41 = tf.nn.relu(tf.matmul(G_h2, G_W41) + G_b41)
    G_logit2 = tf.matmul(G_h41, G_W42) + G_b42
      
    G_logit = tf.concat(axis = 1, values = [G_logit1, G_logit2])
    return G_logit
      
  # 3.2. Discriminator
  def discriminator(x, t, y, hat_y):
    """Discriminator function.
    
    Args:
      - x: features
      - t: treatments
      - y: observed labels #outcome
      - hat_y: estimated counterfactuals
      
    Returns:
      - D_logit: estimated potential outcomes
    """
    # Concatenate factual & counterfactual outcomes
    input0 = (1.-t) * y + t * tf.reshape(hat_y[:,0], [-1,1]) # if t = 0
    input1 = t * y + (1.-t) * tf.reshape(hat_y[:,1], [-1,1]) # if t = 1
      
    inputs = tf.concat(axis = 1, values = [x, input0,input1])
      
    D_h1 = tf.nn.relu(tf.matmul(inputs, D_W1) + D_b1)
    D_h2 = tf.nn.relu(tf.matmul(D_h1, D_W2) + D_b2)
    D_logit = tf.matmul(D_h2, D_W3) + D_b3
    return D_logit
  
  # 3.3. Inference Nets
  def inference(x):
    """Inference function.
    
    Args:
      - x: features
      
    Returns:
      - I_logit: estimated potential outcomes
    """
    I_h1 = tf.nn.relu(tf.matmul(x, I_W1) + I_b1)
    I_h2 = tf.nn.relu(tf.matmul(I_h1, I_W2) + I_b2)
      
    # Estimated outcome if t = 0
    I_h31 = tf.nn.relu(tf.matmul(I_h2, I_W31) + I_b31)
    I_logit1 = tf.matmul(I_h31, I_W32) + I_b32
      
    # Estimated outcome if t = 1
    I_h41 = tf.nn.relu(tf.matmul(I_h2, I_W41) + I_b41)
    I_logit2 = tf.matmul(I_h41, I_W42) + I_b42
      
    I_logit = tf.concat(axis = 1, values = [I_logit1, I_logit2])
    return I_logit

  ## Structure
  # 1. Generator
  Y_tilde_logit = generator(X, T, Y)
  Y_tilde = tf.nn.sigmoid(Y_tilde_logit)
  
  # 2. Discriminator
  D_logit = discriminator(X,T,Y,Y_tilde)

  # 3. Inference network
  Y_hat_logit = inference(X)
  Y_hat = tf.nn.sigmoid(Y_hat_logit)

  ## Loss functions
  # 1. Discriminator loss
  D_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels = T, logits = D_logit )) 
  
  # 2. Generator loss
  G_loss_GAN = -D_loss 
  G_loss_Factual = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
      labels = Y, logits = (T * tf.reshape(Y_tilde_logit[:,1],[-1,1]) + \
                            (1. - T) * tf.reshape(Y_tilde_logit[:,0],[-1,1]) ))) 
  
  G_loss = G_loss_Factual + alpha * G_loss_GAN
  
  # 3. Inference loss
  I_loss1 = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
      labels = (T) * Y + (1-T) * tf.reshape(Y_tilde[:,1],[-1,1]), logits = tf.reshape(Y_hat_logit[:,1],[-1,1]) )) 
  I_loss2 = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
      labels = (1-T) * Y +  (T) * tf.reshape(Y_tilde[:,0],[-1,1]), logits = tf.reshape(Y_hat_logit[:,0],[-1,1]) )) 
  
  I_loss = I_loss1 + I_loss2

  ## Solver
  G_solver = tf.train.AdamOptimizer().minimize(G_loss, var_list=theta_G)
  D_solver = tf.train.AdamOptimizer().minimize(D_loss, var_list=theta_D)
  I_solver = tf.train.AdamOptimizer().minimize(I_loss, var_list=theta_I)

  ## GANITE training
  sess = tf.Session()
  sess.run(tf.global_variables_initializer())
      
  print('Start training Generator and Discriminator')
  # 1. Train Generator and Discriminator
  for it in range(iterations):
    
    for _ in range(2):
      # Discriminator training
      train_x=train_x.reshape((-1,1))
      X_mb, T_mb, Y_mb = batch_generator(train_x, train_t, train_y, batch_size)      
      _, D_loss_curr = sess.run([D_solver, D_loss], feed_dict = {X: X_mb, T: T_mb, Y: Y_mb})
      d_loss_array.append(D_loss_curr)

    # Generator training
    X_mb, T_mb, Y_mb = batch_generator(train_x, train_t, train_y, batch_size)        
    _, G_loss_curr = sess.run([G_solver, G_loss], feed_dict = {X: X_mb, T: T_mb, Y: Y_mb})
    g_loss_array.append(G_loss_curr)

    # Check point
    if it % 1000 == 0:
      print('Iteration: ' + str(it) + '/' + str(iterations) + ', D loss: ' + \
            str(np.round(D_loss_curr, 4)) + ', G loss: ' + str(np.round(G_loss_curr, 4)))
      plt.plot(g_loss_array,label='Generator loss')
      plt.plot( d_loss_array,label='Discriminator loss')
      plt.suptitle('GANS: Generator & Discriminator loss for')
      plt.xlabel('Iterations')
      plt.ylabel('Loss')
      plt.show()
      
  print('Start training Inference network')
  # 2. Train Inference network
  for it in range(iterations):
    
    X_mb, T_mb, Y_mb = batch_generator(train_x, train_t, train_y, batch_size) 
    _, I_loss_curr = sess.run([I_solver, I_loss], feed_dict = {X: X_mb, T: T_mb, Y: Y_mb})    
  
    # Check point
    if it % 1000 == 0:      
      print('Iteration: ' + str(it) + '/' + str(iterations) + 
            ', I loss: ' + str(np.round(I_loss_curr, 4)))
            
  ## Generate the potential outcomes
  test_x=test_x.reshape((-1,1))
  test_y_hat = sess.run(Y_hat, feed_dict = {X: test_x})
  
  return test_y_hat

data_loading.py

# Necessary packages
import numpy as np
from scipy.special import expit


def data_loading_twin(train_rate = 0.8):
  """Load data.
  
  Args:
    - train_rate: the ratio of training data
    
  Returns:
    - train_x: features in training data
    - train_t: treatments in training data
    - train_y: observed outcomes in training data
    - train_potential_y: potential outcomes in training data
    - test_x: features in testing data
    - test_potential_y: potential outcomes in testing data      
  """
  
  # Load original data (11400 patients, 30 features, 2 dimensional potential outcomes)
  ori_data = np.loadtxt("toydata100_2.csv", delimiter=",",skiprows=1)

  # Define features
  x = ori_data[:,2]
  no= x.shape[0]
                
  ## Assign treatment    
  t = ori_data[:,3]

  ## Define observable outcomes
  y = ori_data[:,4]

  ## Train/test division
  idx = np.random.uniform(no)
  train_idx = idx[:int(train_rate * no),:]
  test_idx = idx[int(train_rate * no):,:]
    
  train_x = x[train_idx,]
  train_t = t[train_idx]
  train_y = y[train_idx]

  test_x = x[test_idx,]
  test_y = y[test_idx,]
        
  return train_x, train_t, train_y, test_x, test_y

main_ganite.py

## Necessary packages
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import numpy as np
import warnings
warnings.filterwarnings("ignore")

# 1. GANITE model
from ganite import ganite
# 2. Data loading
from data_loading import data_loading_twin
# 3. Metrics
from metrics import PEHE, ATE


def main (args):
  """Main function for GANITE experiments.
  
  Args:
    - data_name: twin
    - train_rate: ratio of training data
    - Network parameters (should be optimized for different datasets)
      - h_dim: hidden dimensions
      - iteration: number of training iterations
      - batch_size: the number of samples in each batch
      - alpha: hyper-parameter to adjust the loss importance
  
  Returns:
    - test_y_hat: estimated potential outcomes
    - metric_results: performance on testing data
  """
  ## Data loading
  train_x, train_t, train_y, test_x, test_potential_y = \
  data_loading_twin(args.train_rate)
    
  print(args.data_name + ' dataset is ready.')
    
  ## Potential outcome estimations by GANITE
  # Set newtork parameters
  parameters = dict()    
  parameters['h_dim'] = args.h_dim
  parameters['iteration'] = args.iteration
  parameters['batch_size'] = args.batch_size
  parameters['alpha'] = args.alpha
      
  test_y_hat = ganite(train_x, train_t, train_y, test_x, parameters)   
  print('Finish GANITE training and potential outcome estimations')
  
  ## Performance metrics   
  # Output initialization
  metric_results = dict()
  
  # 1. PEHE
  test_PEHE = PEHE(test_potential_y, test_y_hat)
  metric_results['PEHE'] = np.round(test_PEHE, 4)
    
  # 2. ATE
  test_ATE = ATE(test_potential_y, test_y_hat)  
  metric_results['ATE'] = np.round(test_ATE, 4)
  
  ## Print performance metrics on testing data
  print(metric_results)

  return test_y_hat, metric_results


if __name__ == '__main__':  
  
  # Inputs for the main function
  parser = argparse.ArgumentParser()
  parser.add_argument(
      '--data_name',
      choices=['twin'],
      default='twin',
      type=str)
  parser.add_argument(
      '--train_rate',
      help='the ratio of training data',
      default=0.8,
      type=float)
  parser.add_argument(
      '--h_dim',
      help='hidden state dimensions (should be optimized)',
      default=30,
      type=int)
  parser.add_argument(
      '--iteration',
      help='Training iterations (should be optimized)',
      default=10000,
      type=int)
  parser.add_argument(
      '--batch_size',
      help='the number of samples in mini-batch (should be optimized)',
      default=256,
      type=int)
  parser.add_argument(
      '--alpha',
      help='hyper-parameter to adjust the loss importance (should be optimized)',
      default=1,
      type=int)
  
  args = parser.parse_args() 
  
  # Calls main function  
  test_y_hat, metrics = main(args)

metrics.py

Note: Metric functions for GANITE.
Reference: Jennifer L Hill, “Bayesian nonparametric modeling for causal inference”, Journal of Computational and Graphical Statistics, 2011.

(1) PEHE: Precision in Estimation of Heterogeneous Effect
(2) ATE: Average Treatment Effect

# Necessary packages
import numpy as np


def PEHE(y, y_hat):
  """Compute Precision in Estimation of Heterogeneous Effect.
  
  Args:
    - y: potential outcomes
    - y_hat: estimated potential outcomes
    
  Returns:
    - PEHE_val: computed PEHE
  """
  PEHE_val = np.mean( np.abs( (y[0]) - (y_hat[0]) ))
  return PEHE_val

def ATE(y, y_hat):
  """Compute Average Treatment Effect.
  
  Args:
    - y: potential outcomes
    - y_hat: estimated potential outcomes
    
  Returns:
    - ATE_val: computed ATE
  """
  ATE_val = np.abs(np.mean(y[0]) - np.mean(y_hat[0]))
  return ATE_val
#3

Hi @Shrija_Sheth ,

In general, a negative generator loss and a discriminator that is not learning can indicate that there is a problem with the training process.

Maybe you can try to follow the steps below and check if it is improving.

1.Incorrect loss function: If you accidentally swapped the loss functions, you may see a negative generator loss and a discriminator that is not learning.
2.Poor hyperparameters: The choice of hyperparameters, such as the learning rate, batch size, and number of epochs, can have a significant impact on the training process.
3.Insufficient training data: A deep learning model needs a significant amount of training data to learn patterns in the data effectively.
4.Model architecture issues: If the architecture of the generator or discriminator is not suitable for the problem you are trying to solve.
5.Initialization of the model parameters: If the model’s weights are initialized randomly, it may cause issues during training.
6.Learning rate mismatch: Another possible reason for these issues is a mismatch in the learning rates of the generator and discriminator. If the generator is learning too quickly, it may produce unrealistic samples and result in a negative loss.

It may be helpful to try adjusting the loss function, network architecture, hyperparameters, dataset, etc.

Thanks

#5

Thanks a lot @Laxma_Reddy_Patlolla for your reply. Although how to know the perfect learning rate for the model?

#6

Hi @Shrija_Sheth ,

It’s very difficult to say which learning rate is best for your model. But you can try following the steps below and finding the best learning rate for your model.

  1. Start with a small learning rate and increase it gradually.
  2. Use a validation set to evaluate the model’s performance as the learning rate increases.
  3. Experiment with different learning rates and see what works best for your model.

Here are some common methods for finding the perfect learning rate:

1.Learning Rate Range Test
2.Grid Search
3.Cyclical Learning Rates
4.Using learning rate schedules
5. Cross-validation
6. Bayesian optimization

Please let me know if it helps you.

Thanks.

#8

Well nothing is working out so far. I tried experimenting with different learning rates, changing the hyper-parameters, rechecking the loss functions but I am not getting the exact problem here… I don’t know what to do.