Computer Vision News - January 2019

26 Computer Vision News Focus on 30. model.add(Dense({{choice([ 256 , 512 , 1024 ])}})) 31. model.add(Activation({{choice([ 'relu' , 'sigmoid' ])}})) 32. model.add(Dropout({{uniform( 0 , 1 )}})) 33. 34. # If we choose 'four', add an additional fourth layer 35. if {{choice([ 'three' , 'four' ])}} == 'four' : 36. model.add(Dense( 100 )) 37. model.add({{choice([Dropout( 0.5 ), Activation( 'linear' )])}}) 38. model.add(Activation( 'relu' )) 39. 40. model.add(Dense( 10 )) 41. model.add(Activation( 'softmax' )) 42. 43. model.compile( loss = 'categorical_crossentropy' , metrics =[ 'accuracy' ], 44. optimizer ={{choice([ 'rmsprop' , 'adam' , 'sgd' ])}}) 45. 46. with experiment.train(): 47. result = model.fit(x_train, y_train, 48. batch_size ={{choice([ 64 , 128 ])}}, 49. epochs ={{choice([ 50 , 100 ])}}, verbose = 2 , 50. validation_split = 0.1 ) 51. print (model.summary()) 52. params = { 'batch_size' : result.params[ 'batch_size' ], 53. 'epochs' : result.params[ 'epochs' ], 54. 'layer1_type' : 'Dense' , 55. 'layer1_num_nodes' : model.get_layer( "dense_1" ) .get_config()[ 'units' ], 56. 'layer1_activation' : model.get_layer( "activation_5" ) 57. .get_config()[ 'activation' ], 58. 'optimizer' : type (model.optimizer) 59. } 60. #get the highest validation accuracy of the training epochs 61. validation_acc = np.amax(result.history[ 'val_acc' ]) 62. print ( 'Best validation acc of epoch:' , validation_acc) 63. return { 'loss' : -validation_acc, 'status' : STATUS_OK, 64. 'model' : model} 65. 66. if __name__ == '__main__' : 67. best_run, best_model = optim.minimize( model =create_model, 68. data =data, 69. algo =tpe.suggest, 70. max_evals = 1 , 71. trials =Trials()) 72. X_train, Y_train, X_test, Y_test = data() 73. print ( "Evalutation of best performing model:" ) 74. 75. experiment = Experiment( api_key = "" , 76. project_name = "" , workspace = "" ) 77. 78. with experiment.test(): 79. print (best_model.evaluate(X_test, Y_test)) 80. print ( "Best performing model chosen hyper-parameters:" ) 81. print (best_run) Tip - Train Your Network