Coverage for hypergan/trainers/proportional_control_trainer.py : 19%

import tensorflow as tf 

import numpy as np 

import hyperchamber as hc 

import inspect 

from hypergan.trainers.base_trainer import BaseTrainer 

class ProportionalControlTrainer(BaseTrainer): 

def _create(self): 

d_loss = gan.graph.d_loss 

g_loss = gan.graph.g_loss 

g_lr = np.float32(config.g_learn_rate) 

d_lr = np.float32(config.d_learn_rate) 

gan.graph.d_vars = d_vars 

g_defk = {k[2:]: v for k, v in config.items() if k[2:] in inspect.getargspec(config.g_trainer).args and k.startswith("d_")} 

d_defk = {k[2:]: v for k, v in config.items() if k[2:] in inspect.getargspec(config.d_trainer).args and k.startswith("g_")} 

g_optimizer = config.g_trainer(g_lr, **g_defk) 

d_optimizer = config.d_trainer(d_lr, **d_defk) 

if(config.clipped_gradients): 

g_optimizer = capped_optimizer(g_optimizer, config.clipped_gradients, g_loss, g_vars) 

d_optimizer = capped_optimizer(d_optimizer, config.clipped_gradients, d_loss, d_vars) 

else: 

g_optimizer = g_optimizer.minimize(g_loss, var_list=g_vars) 

d_optimizer = d_optimizer.minimize(d_loss, var_list=d_vars) 

gan.graph.clip = [tf.assign(d,tf.clip_by_value(d, -config.d_clipped_weights, config.d_clipped_weights)) for d in d_vars] 

return g_optimizer, d_optimizer 

def _step(gan, feed_dict): 

sess = gan.sess 

config = gan.config 

x_t = gan.graph.x 

g_t = gan.graph.g 

d_log_t = gan.graph.d_log 

g_loss = gan.graph.g_loss 

d_loss = gan.graph.d_loss 

g_optimizer = gan.graph.g_optimizer 

d_optimizer = gan.graph.d_optimizer 

d_class_loss = gan.graph.d_class_loss 

d_vars = gan.graph.d_vars 

# in WGAN paper, values are clipped. This might not work, and is slow. 

if(config.d_clipped_weights): 

sess.run(gan.graph.clip) 

_, d_cost, d_log = sess.run([d_optimizer, d_loss, d_log_t], feed_dict) 

_, g_cost, g_k, measure= sess.run([g_optimizer, g_loss, gan.graph.update_k, gan.graph.measure], feed_dict) 

if self.step % 100 == 0: 

print("%2d: g cost %.2f d_loss %.2f k %.2f m %.2f gamma %.2f" % (self.step, g_cost , d_cost, g_k, measure, gan.graph.gamma)) 

self.step+=1 

return d_cost, g_cost