Scikeras Tutorial - 5: Wrapping the MIMO Estimator and giving a CV run

MIMO4sklearn_CV.py

# Helper Method to get data into shape to pass to GridSearchClassifier
def get_sciki_xy(X,y):
  X_sciki = np.column_stack([X.reshape((y.shape[0], np.prod(X.shape[1:]))), y])
  y_sciki = np.column_stack([y,X.reshape((y.shape[0], np.prod(X.shape[1:])))])

  return X_sciki,y_sciki

def do_cross_val():
  (x_train_, y_train_), (x_test_, y_test_) = load_mnist() #load the dataset
  
  # Optional - trim the data size for faster epochs
  x_train, y_train = get_sciki_xy(x_train_[:1000], y_train_[:1000]) 
  x_test, y_test = get_sciki_xy(x_test_[:100],y_test_[:100])

  # Create a MIMOEstimator with get_model function
  # Parameters that need to be passed to get_model function are prefixed with model__
  clf = MIMOEstimator(model = get_model,
                        model__input_shape=x_train_.shape[1:],
                        model__n_class=len(np.unique(np.argmax(y_train_, 1))),
                        model__routings=args.routings,
                        model__batch_size = args.batch_size,
                        model__n_filters_c1=256, 
                        # epochs=args.epochs, 
                        # callbacks=[log, checkpoint, lr_decay],
                        model__model_type = 'train')
  
  # Print the shapes of X and Y
  print("X input shape = ", x_train.shape)
  print("Y input shape = ", y_train.shape)
  
  # Define the parameter grid to perform Grid-Search
  params = {'model__n_filters_c1': [128,256],
            'model__routings': [4,5]}
  
  # no. of examples/cv should be completely divisible by batch_size
  gs = GridSearchCV(estimator=clf, param_grid=params, cv=5 verbose=True)
  gs_res = gs.fit(X=x_train, 
                  y=y_train)
  print("Grid Search Results: ")
  print(gs_res)
  
  best_est = gs_res.best_estimator_
  best_score = gs_res.best_score_
  best_params = gs_res.best_params_
  
  print('Best score obtained after GridSearchCV: ', best_score)
  
  return best_est, best_params
  
est,params = do_cross_val() # call the function to begin training