tuning hyperparameters

tuning.py

best_model = None
best_val_acc = 0
best_lr = None

num_trials=10
for trial in range(num_trials):
  # best so far is 2.74799e-4, aka 10 ** -3.5609848520174077
  #lr_exp = np.random.uniform(-5, 2)
  lr_exp = np.random.uniform(-3.65, -3.2)
  lr = 10 ** lr_exp
    
  reg_exp = np.random.uniform(-5.02, -4.4) # (-5.6, -5.5)
  reg = 10 ** reg_exp
    
  weight_scale_exp = np.random.uniform(-1.8, -1.31)
  weight_scale = 10 ** weight_scale_exp  # orig: 5e-2

  model = GeorgeNet(num_convnets=2, num_affine=2,
                    hidden_dim=100, # default: 500
                    weight_scale=weight_scale,
                    use_batchnorm=True,
                    reg=reg)

  num_epochs = 1 # orig: 5

  solver = Solver(model, data,
                  num_epochs=num_epochs, batch_size=100,
                  update_rule='adam',
                  optim_config={
                    'learning_rate': lr
                  },
                  verbose=False)
  solver.train()
  print "trial: %d, lr_exp: %g, reg_exp: %g, weight_scale_exp: %g, val_acc: %g" % (trial, lr_exp, reg_exp, weight_scale_exp, solver.best_val_acc)

  if solver.best_val_acc > best_val_acc:
      best_model = model        
      best_val_acc = solver.best_val_acc
      best_lr = lr


print "best_val_acc: ", best_val_acc
print "best_lr: ", best_lr