def ksne_obj(params,X,K,P,pijs,count_pot_dict,T, fast=True): N = X.shape[0] Z = params.reshape((N, P)) tdijs = squareform(pdist(Z, 'sqeuclidean')) # tilde dijs from numpy import log tdijs = -log(1+tdijs) term1 = np.sum(tdijs * pijs) # diagonal is 0 if fast: from chain.chain_alg_wrapper import ChainAlg chain = ChainAlg(N, Kmin=K, Kmax=K, minibatch_size = N) mask = np.ones((N,N)) - np.diag(np.ones(N)) log_mask = -np.diag(np.ones(N))*1e100 pot = tdijs + log_mask #exp_pot = mask * np.exp(tdijs) marginals, samples, logZs = chain.infer(pot) sum_log_Z = logZs.sum() else: sum_log_Z = 0 mask = np.ones(N) for i in xrange(N): if i > 0: mask[i-1] = 1 mask[i] = 0 node_margs, count_margs, log_Z = \ sc.conv_tree(mask * np.exp(tdijs[i,:]), count_pot_dict, T, use_fft=False) sum_log_Z += log_Z return term1 - sum_log_Z
def ksne_grad_correct(params,X,K,P,pijs,count_pot_dict,T): Z = params.reshape((N, P)) tdijs = -squareform(pdist(Z, 'sqeuclidean')) # tilde dijs grad = pijs.copy() mask = np.ones(N) for i in xrange(N): if i > 0: mask[i-1] = 1 mask[i] = 0 node_margs, count_margs, log_Z = \ sc.conv_tree(mask * np.exp(tdijs[i,:]), count_pot_dict, T, use_fft=False) grad[i,:] -= node_margs g2 = 2*(grad+grad.T) dZ = g2.dot(Z)-g2.sum(1)[:,np.newaxis]*Z return dZ
def train_sgd(X,Y,dataset_name,P,K,Z=None,num_iters=1,perplexity=5,batch_size=10,eta=0.1,mo=0,L2=0,class_to_keep=[],seed=1,distance='sqeuclidean',text_labels=None): """ X, Y: points and their respective labels dataset_name: the name given to this dataset P: the dimensionality of the embedding space K: the number of neighbors Z: coordinates of the points in embedded space num_iters: number of passes over the set of points perplexity: perplexity of the distribution over datapoints. Check: 'Visualizing Data using t-SNE by van der Maaten and Hinton in JMLR (09) 2008', for details. batch_size: size of mini batches eta: learning rate mo: momentum L2: size of L2 regularizer class_to_keep: subset of labels that have been kept (for naming results and figure files) seed: the value of the seed (for naming results and figure files) distance: distance metric {sqeuclidean, emd} test_labels: alternative labels used in plots """ N = X.shape[0] # num points D = X.shape[1] # input dimension ################################################################################## if Z is None: Z = np.random.randn(N,P)*0.1 # points in embedded space if P == 2: # plot if embedding is in 2D plot_embedding(Z,Y,K,eta,mo,0,target_entropy=perplexity,iters=0,tot_iters=num_iters,class_to_keep=class_to_keep,seed=seed,dataset_name=dataset_name,text_labels=text_labels) if distance == 'sqeuclidean': dijs = -squareform(pdist(X, 'sqeuclidean')) elif distance == 'emd': # earth mover's distance from emd.emddist import emddist dijs = -emddist(X) dijs = dist_normalize(dijs,Y,perplexity=perplexity,dataset_name=dataset_name) dijs -= HUGE_VAL * np.eye(N) pijs = np.zeros((N,N)) T = sc.make_balanced_binary_tree(N) root_idx = N + T.shape[0]-1 print "Root idx", root_idx global_count_potential = np.zeros(N+1) global_count_potential[K] = 1 only_global_constraint_dict = {} only_global_constraint_dict[root_idx] = global_count_potential # Precompute E_i[y_j] (or pijs) print 'precomute E_i[y_j]' from chain.chain_alg_wrapper import ChainAlg chain = ChainAlg(N, Kmin=K, Kmax=K, minibatch_size = N) pot = dijs #exp_pot = np.exp(dijs) marginals, samples, logZs = chain.infer(pot) pijs = marginals debug = False if debug: pijs_debug = pijs.copy() for i in xrange(N): thetas = dijs[i,:] #print np.exp(thetas) node_margs, count_margs, log_Z = \ sc.conv_tree(np.exp(thetas), only_global_constraint_dict, T, use_fft=False) pijs_debug[i,:] = node_margs diff = abs(pijs - pijs_debug).max() assert diff < 1e-8 print 'precomputation of marginals is correct!' ################################################################################## num_batches = np.ceil(np.double(N)/batch_size) randIndices = np.random.permutation(X.shape[0]) print 'num_batches = %s' % num_batches dZ = np.zeros(Z.shape) V = dZ*0 for i in range(num_iters): f_tot = 0 for batch in range(int(num_batches)): print "iteration " + str(i+1) + " batch " + str(batch+1) + " of " + str(int(num_batches)) ind = randIndices[np.mod(range(batch*batch_size,(batch+1)*batch_size),X.shape[0])] f_tot += ksne_obj(Z.flatten(),X,K,P,pijs,only_global_constraint_dict,T) g = ksne_grad((Z+V*mo).flatten(),X,K,P,pijs,only_global_constraint_dict,T).reshape(Z.shape) #dZ = mo*dZ - eta*g V = V*mo + eta*(g-L2*(Z+V*mo)) #Z -= dZ Z += V print 'objective: %s, |g|=%s' % (str(f_tot/num_batches), abs(g).mean()) if P == 2: # plot if embedding is in 2D plot_embedding(Z,Y,K,eta,mo,f_tot/num_batches,target_entropy=perplexity,iters=i+1,tot_iters=num_iters,class_to_keep=class_to_keep,seed=seed,dataset_name=dataset_name,text_labels=text_labels) return Z