def sample(self, T, g, g0=None): if g0==None: g0 = g v, h = self.v, self.h VH, HH, b_init = self V = zeros((T, v)) H = zeros((T, h)) B = zeros((T, h)) VH_t = 1*VH VH_t[2] = VH[2] + b_init V[[0]], H_t_stoch = rbm.sample(VH_t, g0, 1, self.vis_gauss) H[[0]] = sigmoid(VH_t * V[[0]]) if self.vis_gauss: V[[0]] = VH_t.T() * H_t_stoch else: V[[0]] = sigmoid(VH_t.T() * H_t_stoch) for t in range(1, T): B[[t]] = HH*H[[t-1]] VH_t[2] = VH[2] + B[t] V[[t]], H_t_stoch = rbm.sample(VH_t, g, 1, self.vis_gauss) H[[t]] = sigmoid(VH_t * V[[t]]) if self.vis_gauss: V[[t]] = VH_t.T() * H_t_stoch else: V[[t]] = sigmoid(VH_t.T() * H_t_stoch) return V
def sample(self, T, g, g0=None): if g0==None: g0 = g v, h = self.v, self.h VH, HH, b_init = self V = zeros((T, v)) H = zeros((T, h)) B = zeros((T, h)) VH_t = 1*VH VH_t[2] = VH[2] + b_init V[[0]], H[[0]] = rbm.sample(VH_t, g0, 1, self.vis_gauss) ## mean-fieldize the output: if self.vis_gauss: V[[0]] = VH_t.T() * H[[0]] else: V[[0]] = sigmoid(VH_t.T() * H[[0]]) for t in range(1, T): B[[t]] = HH*H[[t-1]] VH_t[2] = VH[2] + B[t] V[[t]], H[[t]] = rbm.sample(VH_t, g, 1, self.vis_gauss) ## mean-field-ize the output. if self.vis_gauss: V[[t]] = VH_t.T() * H[[t]] else: V[[t]] = sigmoid(VH_t.T() * H[[t]]) return V
def sample(self, T, g, g0=None): if g0 == None: g0 = g v, h = self.v, self.h VH, HH, b_init = self V = zeros((T, v)) H = zeros((T, h)) B = zeros((T, h)) VH_t = 1 * VH VH_t[2] = VH[2] + b_init V[[0]], H_t_stoch = rbm.sample(VH_t, g0, 1, self.vis_gauss) H[[0]] = sigmoid(VH_t * V[[0]]) if self.vis_gauss: V[[0]] = VH_t.T() * H_t_stoch else: V[[0]] = sigmoid(VH_t.T() * H_t_stoch) for t in range(1, T): B[[t]] = HH * H[[t - 1]] VH_t[2] = VH[2] + B[t] V[[t]], H_t_stoch = rbm.sample(VH_t, g, 1, self.vis_gauss) H[[t]] = sigmoid(VH_t * V[[t]]) if self.vis_gauss: V[[t]] = VH_t.T() * H_t_stoch else: V[[t]] = sigmoid(VH_t.T() * H_t_stoch) return V
0], :, :] ### LEARNING ### # propagate visible input to hidden units posHiddenProb = rbm.visibleToHiddenVec( v, weightsForUser) # get positive gradient # note that we only update the movies that this user has seen! posprods[ ratingsForUser[:, 0], :, :] = rbm.probProduct( v, posHiddenProb) ### UNLEARNING ### # sample from hidden distribution sampledHidden = rbm.sample(posHiddenProb) # propagate back to get "negative data" negData = rbm.hiddenToVisible( sampledHidden, weightsForUser) # propagate negative data to hidden units negHiddenProb = rbm.visibleToHiddenVec( negData, weightsForUser) # get negative gradient # note that we only update the movies that this user has seen! negprods[ ratingsForUser[:, 0], :, :] = rbm.probProduct( negData, negHiddenProb) # we average over the number of users in the batch (if we use mini-batch)
def run_RBM(F, epochs, epsilon, B, weightcost, momentum, f, mode): # Initialise all our arrays W = rbm.getInitialWeights(trStats["n_movies"], F, K) bestW = np.zeros(W.shape) bestRMSE = 100 posprods = np.zeros(W.shape) negprods = np.zeros(W.shape) grad = np.zeros(W.shape) for epoch in range(1, epochs): # in each epoch, we'll visit all users in a random order visitingOrder = np.array(trStats["u_users"]) np.random.shuffle(visitingOrder) visitingOrder = np.array_split(visitingOrder, visitingOrder.shape[0] / B) # print(visitingOrder) for batch in visitingOrder: # print(batch) temp = np.zeros(W.shape) for user in batch: # get the ratings of that user ratingsForUser = lib.getRatingsForUser(user, training) # build the visible input v = rbm.getV(ratingsForUser) # get the weights associated to movies the user has seen ratingsWithBias = np.append(ratingsForUser[:,0],W.shape[0]-1) weightsForUser = W[ratingsWithBias, :, :] ### LEARNING ### # propagate visible input to hidden units posHiddenProb = rbm.visibleToHiddenVec(v, weightsForUser) # get positive gradient # note that we only update the movies that this user has seen! posprods[ratingsWithBias, :, :] += rbm.probProduct(v, posHiddenProb) ### UNLEARNING ### # sample from hidden distribution sampledHidden = rbm.sample(posHiddenProb) # propagate back to get "negative data" negData = rbm.hiddenToVisible(sampledHidden, weightsForUser) # propagate negative data to hidden units negData[-1,:] = np.array([1,1,1,1,1]) negHiddenProb = rbm.visibleToHiddenVec(negData, weightsForUser) # get negative gradient # note that we only update the movies that this user has seen! negprods[ratingsWithBias, :, :] += rbm.probProduct(negData, negHiddenProb) # we average over the number of users delta = epsilon / epochs gradientLearningRate = epsilon - epoch * delta # print(gradientLearningRate) grad = momentum * grad + (1 - momentum) * gradientLearningRate * ( (posprods - negprods) / trStats['n_users'] - weightcost * np.linalg.norm(W)) temp += grad # hiddenBiases += hiddenBiasGrad # visibleBiases += visibleBiasGrad W += temp # Print the current RMSE for training and validation sets # this allows you to control for overfitting e.g # We predict over the training set tr_r_hat = rbm.predict(trStats["movies"], trStats["users"], W, training) trRMSE = lib.rmse(trStats["ratings"], tr_r_hat) # We predict over the validation set vl_r_hat = rbm.predict(vlStats["movies"], vlStats["users"], W, training) vlRMSE = lib.rmse(vlStats["ratings"], vl_r_hat) if vlRMSE < bestRMSE: bestRMSE = vlRMSE bestW = W # if mode == 'single': print("### EPOCH %d ###" % epoch) print("Learning rate = %f" % gradientLearningRate) print("Training loss = %f" % trRMSE) print("Validation loss = %f" % vlRMSE) content_to_be_written = "" + str(F) + ", " + str(epoch) + ", " + str(epsilon) + ", " + str(B) + ", " + str( weightcost) + ", " content_to_be_written += str(momentum) + ", " + str(trRMSE) + ", " + str(vlRMSE) + '\n' f.write(content_to_be_written) # np.save('best_weight.npy', bestW) return (bestW,bestRMSE)
def main_rbm(training=training, validation=validation, trStats=trStats, vlStats=vlStats, K=5, F=5, epochs=30, gradientLearningRate=0.0001,gradientLearningRate_v = 0.001, gradientLearningRate_h = 0.001, minibatch_size=10, alpha=0.9, stopping=False, momentum=False, learning_rate_type='time', learning_rate_k=0.1, learning_rate_drop=0.5, learning_rate_epochs_drop=10.0, _lambda = 0.3): # Print current hyperparams # frame = inspect.currentframe() # args, _, _, values = inspect.getargvalues(frame) # print ('Training and Predicting with the following hyperparameters:') # for i in args[4:]: # print (" %s = %s" % (i, values[i])) # Initialise all our arrays num_movies=trStats["n_movies"] num_users=trStats["n_users"] W = rbm.getInitialWeights(trStats["n_movies"], F, K) posprods = np.zeros(W.shape) negprods = np.zeros(W.shape) grad_w = np.zeros(W.shape) m_w=np.zeros((W.shape[0],F,5)) train_loss = [] validation_loss = [] vis_bias=np.zeros((num_movies,5)) m_v=np.zeros((num_movies,5)) hid_bias=np.zeros((F,)) m_h=np.zeros((F,)) best_train_loss = 100 best_validation_loss = 100 for epoch in range(1, epochs+1): # mini_batch_grads = [] # in each epoch, we'll visit all users in a random order visitingOrder = np.array(trStats["u_users"]) np.random.shuffle(visitingOrder) # for i in range(0, visitingOrder.shape[0], minibatch_size): # # Get pair of (X, y) of the current minibatch/chunk # visitingOrderMini = visitingOrder[i:i + minibatch_size] # y_train_mini = y_train[i:i + minibatch_size] # for i in range(0, visitingOrder.shape[0], minibatch_size): batches = batch_get(visitingOrder, minibatch_size) for batch in batches: prev_grad = grad_w grad_w = np.zeros(W.shape) for user in batch: # get the ratings of that user ratingsForUser = lib.getRatingsForUser(user, training) # build the visible input v = rbm.getV(ratingsForUser) # get the weights associated to movies the user has seen weightsForUser = W[ratingsForUser[:, 0], :, :] ### LEARNING ### # propagate visible input to hidden units posHiddenProb = rbm.visibleToHiddenVec(v, weightsForUser) #, hid_bias) # get positive gradient # note that we only update the movies that this user has seen! posprods[ratingsForUser[:, 0], :, :] += probProduct(v, posHiddenProb) ### UNLEARNING ### # sample from hidden distribution sampledHidden = rbm.sample(posHiddenProb) # propagate back to get "negative data" negData = rbm.hiddenToVisible(sampledHidden, weightsForUser)#, vis_bias[ratingsForUser[:,0]]) # propagate negative data to hidden units negHiddenProb = rbm.visibleToHiddenVec(negData, weightsForUser) #, hid_bias) # get negative gradient # note that we only update the movies that this user has seen! negprods[ratingsForUser[:, 0], :, :] += rbm.probProduct(negData, negHiddenProb) poshidact = sum(posHiddenProb) posvisact = sum(v) neghidact = sum(negHiddenProb) negvisact = sum(negData) # we average over the number of users in the batch (if we use mini-batch) # grad = (gradientLearningRate/epoch)*(posprods-negprods) ''' Regularization - ''' grad_w += adaptiveLearn(learning_rate_type=learning_rate_type, k=learning_rate_k, drop=learning_rate_drop, epochs_drop=learning_rate_epochs_drop, epoch=epoch)*((posprods-negprods)/trStats["n_users"]-_lambda*W) # mini_batch_grads.append(grad) # m = alpha*m+grad ''' Ask about the implementation of biases (should we create matrix of biases for hidden and visible layers?) ''' m_w = alpha*m_w + grad_w # m_v = alpha*m_v+(gradientLearningRate_v) * (posvisact - negvisact) # m_h = alpha*m_h+(gradientLearningRate_h) * (poshidact - neghidact) if momentum == False: W += grad_w else: W += m_w vis_bias += m_v hid_bias += m_h # Print the current RMSE for training and validation sets # this allows you to control for overfitting e.g # We predict over the training set tr_r_hat = rbm.predict(trStats["movies"], trStats["users"], W, training) #, vis_bias, hid_bias, predictType='exp') # print (tr_r_hat) trRMSE = lib.rmse(trStats["ratings"], tr_r_hat) # print (trRMSE) if trRMSE < best_train_loss: best_train_loss = trRMSE best_training_weights = W best_train_predictions = tr_r_hat # We predict over the validation set vl_r_hat = rbm.predict(vlStats["movies"], vlStats["users"], W, training) #, vis_bias, hid_bias, predictType='exp') # vl_r_hat vlRMSE = lib.rmse(vlStats["ratings"], vl_r_hat) if vlRMSE < best_validation_loss: best_validation_loss = vlRMSE best_validation_weights = W best_validation_predictions = vl_r_hat # best_momentum = momentum # best_reg = regularization # best_epoch = epoch # best_alpha = alpha # best_B = B # best_F = F # min_rmse = vlRMSE # print('Best RMSE:', min_rmse) train_loss.append(trRMSE) validation_loss.append(vlRMSE) print ("### EPOCH %d ###" % epoch) print ("Training loss = %f" % trRMSE) print ("Validation loss = %f" % vlRMSE) ### END ### # This part you can write on your own # you could plot the evolution of the training and validation RMSEs for example # predictedRatings = np.array([predictForUser(user, W, training) for user in trStats["u_users"]]) # np.savetxt("predictedRatings.txt", predictedRatings) # fig1 = plt.figure() # ax1 = fig1.add_subplot(121) # ax1.plot(train_loss) # ax2 = fig1.add_subplot(122) # ax2.plot(validation_loss) plt.plot(train_loss) plt.plot(validation_loss) plt.show() if stopping==True: print('Best training loss = %f' % best_train_loss) print('Best validation loss = %f' % best_validation_loss) else: print('Final training loss = %f' % trRMSE) print('Final validation loss = %f' % vlRMSE) return [best_validation_loss, best_validation_predictions]
def run_RBM(F, epochs, epsilon, B, weightcost, momentum, f, mode): # Initialise all our arrays W = rbm.getInitialWeights(trStats["n_movies"], F, K) bestW = np.zeros(W.shape) bestRMSE = 100 posprods = np.zeros(W.shape) negprods = np.zeros(W.shape) grad = np.zeros(W.shape) for epoch in range(1, epochs): # in each epoch, we'll visit all users in a random order visitingOrder = np.array(trStats["u_users"]) np.random.shuffle(visitingOrder) visitingOrder = np.array_split(visitingOrder, visitingOrder.shape[0] / B) # print(visitingOrder) for batch in visitingOrder: # print(batch) temp = np.zeros(W.shape) for user in batch: # get the ratings of that user ratingsForUser = lib.getRatingsForUser(user, training) # build the visible input v = rbm.getV(ratingsForUser) # get the weights associated to movies the user has seen ratingsWithBias = np.append(ratingsForUser[:, 0], W.shape[0] - 1) weightsForUser = W[ratingsWithBias, :, :] ### LEARNING ### # propagate visible input to hidden units posHiddenProb = rbm.visibleToHiddenVec(v, weightsForUser) # get positive gradient # note that we only update the movies that this user has seen! posprods[ratingsWithBias, :, :] += rbm.probProduct( v, posHiddenProb) ### UNLEARNING ### # sample from hidden distribution sampledHidden = rbm.sample(posHiddenProb) # propagate back to get "negative data" negData = rbm.hiddenToVisible(sampledHidden, weightsForUser) # propagate negative data to hidden units negData[-1, :] = np.array([1, 1, 1, 1, 1]) negHiddenProb = rbm.visibleToHiddenVec(negData, weightsForUser) # get negative gradient # note that we only update the movies that this user has seen! negprods[ratingsWithBias, :, :] += rbm.probProduct( negData, negHiddenProb) # we average over the number of users delta = epsilon / epochs gradientLearningRate = epsilon - epoch * delta # print(gradientLearningRate) grad = momentum * grad + ( 1 - momentum) * gradientLearningRate * ( (posprods - negprods) / trStats['n_users'] - weightcost * np.linalg.norm(W)) temp += grad # hiddenBiases += hiddenBiasGrad # visibleBiases += visibleBiasGrad W += temp # Print the current RMSE for training and validation sets # this allows you to control for overfitting e.g # We predict over the training set tr_r_hat = rbm.predict(trStats["movies"], trStats["users"], W, training) trRMSE = lib.rmse(trStats["ratings"], tr_r_hat) # We predict over the validation set vl_r_hat = rbm.predict(vlStats["movies"], vlStats["users"], W, training) vlRMSE = lib.rmse(vlStats["ratings"], vl_r_hat) if vlRMSE < bestRMSE: bestRMSE = vlRMSE bestW = W # if mode == 'single': print("### EPOCH %d ###" % epoch) print("Learning rate = %f" % gradientLearningRate) print("Training loss = %f" % trRMSE) print("Validation loss = %f" % vlRMSE) content_to_be_written = "" + str(F) + ", " + str(epoch) + ", " + str( epsilon) + ", " + str(B) + ", " + str(weightcost) + ", " content_to_be_written += str(momentum) + ", " + str( trRMSE) + ", " + str(vlRMSE) + '\n' f.write(content_to_be_written) # np.save('best_weight.npy', bestW) return (bestW, bestRMSE)