def runSP(): ds = StockPrice() # create hidden layer with 2 nodes, init weights in range -0.1 to 0.1 and add # a bias with value 1 hidden_layer = mlp.Sigmoid(layer_name='hidden', dim=10000, irange=.1, init_bias=1.) # create Softmax output layer output_layer = mlp.Linear(layer_name='output', dim=1, irange=.1, init_bias=1.) # create Stochastic Gradient Descent trainer that runs for 400 epochs trainer = sgd.SGD(learning_rate=.005, batch_size=500, termination_criterion=EpochCounter(10)) layers = [hidden_layer, output_layer] # create neural net that takes two inputs ann = mlp.MLP(layers, nvis=1000) trainer.setup(ann, ds) # train neural net until the termination criterion is true while True: trainer.train(dataset=ds) ann.monitor.report_epoch() ann.monitor() if not trainer.continue_learning(ann): break #accuracy = Accuracy() acc = Accuracy() for i, predict in enumerate(ann.fprop(theano.shared(ds.valid[0], name='inputs')).eval()): print predict, ds.valid[1][i] acc.evaluatePN(predict[0], ds.valid[1][i][0]) acc.printResult()
def runAutoencoder(): ds = StockPrice() #print ds.train[0][0] data = np.random.randn(10, 5).astype(config.floatX) #print data print BinomialCorruptor(.2) ae = DenoisingAutoencoder(BinomialCorruptor(corruption_level=.2), 1000, 100, act_enc='sigmoid', act_dec='linear', tied_weights=False) trainer = sgd.SGD(learning_rate=.005, batch_size=5, termination_criterion=EpochCounter(3), cost=cost_ae.MeanSquaredReconstructionError(), monitoring_batches=5, monitoring_dataset=ds) trainer.setup(ae, ds) while True: trainer.train(dataset=ds) ae.monitor() ae.monitor.report_epoch() if not trainer.continue_learning(ae): break #print ds.train[0][0] #print ae.reconstruct(ds.train[0][0]) w = ae.weights.get_value() #ae.hidbias.set_value(np.random.randn(1000).astype(config.floatX)) hb = ae.hidbias.get_value() #ae.visbias.set_value(np.random.randn(100).astype(config.floatX)) vb = ae.visbias.get_value() d = tensor.matrix() result = np.dot(1. / (1 + np.exp(-hb - np.dot(ds.train[0][0], w))), w.T) + vb
def _create_trainer(self, dataset): sgd.log.setLevel(logging.WARNING) # Aggregate all the dropout parameters into shared dictionaries. probs, scales = {}, {} for l in [l for l in self.layers if l.dropout is not None]: incl = 1.0 - l.dropout probs[l.name] = incl scales[l.name] = 1.0 / incl if self.cost == "Dropout" or len(probs) > 0: # Use the globally specified dropout rate when there are no layer-specific ones. incl = 1.0 - self.dropout default_prob, default_scale = incl, 1.0 / incl # Pass all the parameters to pylearn2 as a custom cost function. self.cost = Dropout(default_input_include_prob=default_prob, default_input_scale=default_scale, input_include_probs=probs, input_scales=scales) logging.getLogger('pylearn2.monitor').setLevel(logging.WARNING) if dataset is not None: termination_criterion = MonitorBased(channel_name='objective', N=self.n_stable, prop_decrease=self.f_stable) else: termination_criterion = None return sgd.SGD(cost=self.cost, batch_size=self.batch_size, learning_rule=self._learning_rule, learning_rate=self.learning_rate, termination_criterion=termination_criterion, monitoring_dataset=dataset)
def cnn_run_dropout_maxout(data_path, num_rows, num_cols, num_channels, input_path, pred_path): t = time.time() sub_window = gen_center_sub_window(76, num_cols) trn = SarDataset(ds[0][0], ds[0][1], sub_window) vld = SarDataset(ds[1][0], ds[1][1], sub_window) tst = SarDataset(ds[2][0], ds[2][1], sub_window) print 'Take {}s to read data'.format(time.time() - t) t = time.time() batch_size = 100 h1 = maxout.Maxout(layer_name='h2', num_units=1, num_pieces=100, irange=.1) hidden_layer = mlp.ConvRectifiedLinear(layer_name='h2', output_channels=8, irange=0.05, kernel_shape=[5, 5], pool_shape=[2, 2], pool_stride=[2, 2], max_kernel_norm=1.9365) hidden_layer2 = mlp.ConvRectifiedLinear(layer_name='h3', output_channels=8, irange=0.05, kernel_shape=[5, 5], pool_shape=[2, 2], pool_stride=[2, 2], max_kernel_norm=1.9365) #output_layer = mlp.Softplus(dim=1,layer_name='output',irange=0.1) output_layer = mlp.Linear(dim=1, layer_name='output', irange=0.05) trainer = sgd.SGD(learning_rate=0.001, batch_size=100, termination_criterion=EpochCounter(2000), cost=dropout.Dropout(), train_iteration_mode='even_shuffled_sequential', monitor_iteration_mode='even_shuffled_sequential', monitoring_dataset={ 'test': tst, 'valid': vld, 'train': trn }) layers = [hidden_layer, hidden_layer2, output_layer] input_space = space.Conv2DSpace(shape=[num_rows, num_cols], num_channels=num_channels) ann = mlp.MLP(layers, input_space=input_space, batch_size=batch_size) watcher = best_params.MonitorBasedSaveBest(channel_name='valid_objective', save_path='sar_cnn_mlp.pkl') experiment = Train(dataset=trn, model=ann, algorithm=trainer, extensions=[watcher]) print 'Take {}s to compile code'.format(time.time() - t) t = time.time() experiment.main_loop() print 'Training time: {}s'.format(time.time() - t) serial.save('cnn_hhv_{0}_{1}.pkl'.format(num_rows, num_cols), ann, on_overwrite='backup') #read hh and hv into a 3D numpy image = read_hhv(input_path) return ann, sar_predict(ann, image, pred_path)
def __init__(self, layers, dropout=False, input_scaler=None, output_scaler=None, learning_rate=0.01, verbose=0): """ :param layers: List of tuples of types of layers alongside the number of neurons :param learning_rate: The learning rate for all layers :param verbose: Verbosity level :return: """ self.layers = layers self.ds = None self.f = None self.verbose = verbose cost = None if (dropout): cost = Dropout() self.trainer = sgd.SGD(learning_rate=learning_rate, cost=cost, batch_size=100) self.input_normaliser = input_scaler self.output_normaliser = output_scaler
def get_ae_pretrainer(layer, data, batch_size): init_lr = 0.1 dec_fac = 1.0001 train_algo = sgd.SGD( batch_size=batch_size, learning_rate=init_lr, init_momentum=0.5, monitoring_batches=100 / batch_size, monitoring_dataset={'train': data}, #cost = MeanSquaredReconstructionError(), cost=CAE_cost(), termination_criterion=EpochCounter(20), update_callbacks=sgd.ExponentialDecay(decay_factor=dec_fac, min_lr=0.02)) return Train(model=layer, algorithm=train_algo, dataset=data, \ extensions=[sgd.MomentumAdjustor(final_momentum=0.9, start=0, saturate=15), ])
def get_finetuner(model, cost, trainset, validset=None, batch_size=100, iters=100): train_algo = sgd.SGD( batch_size=batch_size, init_momentum=0.5, learning_rate=0.5, #monitoring_batches = 100/batch_size, #monitoring_dataset = {'train': trainset, 'valid': validset}, cost=cost, termination_criterion=EpochCounter(iters), update_callbacks=sgd.ExponentialDecay(decay_factor=1.005, min_lr=0.05)) return Train(model=model, algorithm=train_algo, dataset=trainset, save_freq=0, \ extensions=[sgd.MomentumAdjustor(final_momentum=0.9, start=0, saturate=int(0.8*iters)), ])
def __init__(self, data): self.N = 5 * 5 self.predictionLength = 2 # create hidden layer with 2 nodes, init weights in range -0.1 to 0.1 and add # a bias with value 1 hidden_layer = mlp.Sigmoid(layer_name='hidden', dim=25, irange=.1, init_bias=1.) # create Linear output layer output_layer = mlp.Linear(1, 'output', irange=.1, init_bias=1.) # create Stochastic Gradient Descent trainer that runs for 400 epochs trainer = sgd.SGD(learning_rate=.005, batch_size=10, termination_criterion=EpochCounter(100)) layers = [hidden_layer, output_layer] # create neural net that takes two inputs nn = mlp.MLP(layers, nvis=self.N) NeuralNetwork.__init__(self, data, nn, trainer)
def runXOR(): ds = XOR() hidden_layer = mlp.Sigmoid(layer_name='hidden', dim=10, irange=.1, init_bias=1.) output_layer = mlp.Linear(layer_name='output', dim=1, irange=.1, init_bias=1.) trainer = sgd.SGD(learning_rate=.05, batch_size=1, termination_criterion=EpochCounter(1000)) layers = [hidden_layer, output_layer] # create neural net that takes two inputs ann = mlp.MLP(layers, nvis=4) trainer.setup(ann, ds) # train neural net until the termination criterion is true while True: trainer.train(dataset=ds) #ann.monitor.report_epoch() #ann.monitor() if not trainer.continue_learning(ann): break inputs= np.array([[0, 0, 0, 1]]) print ann.fprop(theano.shared(inputs, name='inputs')).eval() inputs = np.array([[0, 1, 0, 1]]) print ann.fprop(theano.shared(inputs, name='inputs')).eval() inputs = np.array([[1, 1, 1, 1]]) print ann.fprop(theano.shared(inputs, name='inputs')).eval() inputs = np.array([[1, 1, 0, 0]]) print ann.fprop(theano.shared(inputs, name='inputs')).eval()
def main(): base_name = sys.argv[ 1] # 获取第一个参数 sys.argv[ ]记录(获取)命令行参数 sys(system) argv(argument variable)参数变量,该变量为list列表 n_epoch = int(sys.argv[2]) #获取第二个参数 n_hidden = int(sys.argv[3]) #获取第三个参数作为隐层神经元个数 include_rate = float(sys.argv[4]) in_size = 1001 #输入层神经元个数(标记基因个数) out_size = 1 #输出层神经元个数 b_size = 200 #偏差值 l_rate = 5e-4 #学习速率 l_rate_min = 1e-5 #学习速率最小值 decay_factor = 0.9 #衰减因数 lr_scale = 3.0 momentum = 0.5 init_vals = np.sqrt(6.0 / (np.array([in_size, n_hidden]) + np.array([n_hidden, out_size]))) #初始值,返回平方根 print 'loading data...' #显示载入数据 X_tr = np.load( 'geno_X_tr_float64.npy') # tr(traing)以numpy专用二进制类型保存训练数据集的数据 Y_tr = np.load('pheno_Y_tr_0-4760_float64.npy') Y_tr_pheno = np.array(Y_tr) X_va = np.load( 'geno_X_va_float64.npy') #验证集(模型选择,在学习到不同复杂度的模型中,选择对验证集有最小预测误差的模型) Y_va = np.load('pheno_Y_va_0-4760_float64.npy') Y_va_target = np.array(Y_va) X_te = np.load('geno_te_float64.npy') #测试集(对学习方法的评估) Y_te = np.load('pheno_Y_te_0-4760_float64.npy') Y_te_target = np.array(Y_te) random.seed(0) #设置生成随机数用的整数起始值。调用任何其他random模块函数之前调用这个函数 monitor_idx_tr = random.sample(range(88807), 5000) #监测训练 #将训练数据集类型设为32位浮点型,The DenseDesignMatrix class and related code Functionality for representing data that can be described as a dense matrix (rather than a sparse matrix) with each row containing an example and each column corresponding to a different feature. data_tr = p2_dt_dd.DenseDesignMatrix(X=X_tr.astype('float32'), y=Y_tr.astype('float32')) X_tr_monitor, Y_tr_monitor_target = X_tr[monitor_idx_tr, :], Y_tr_target[ monitor_idx_tr, :] #一个隐层,用Tanh()作激活函数; 输出层用线性函数作激活函数 h1_layer = p2_md_mlp.Tanh(layer_name='h1', dim=n_hidden, irange=init_vals[0], W_lr_scale=1.0, b_lr_scale=1.0) o_layer = p2_md_mlp.Linear(layer_name='y', dim=out_size, irange=0.0001, W_lr_scale=lr_scale, b_lr_scale=1.0) #Multilayer Perceptron;nvis(Number of “visible units” input units) layers(a list of layer objects,最后1层指定MLP的输出空间) model = p2_md_mlp.MLP(nvis=in_size, layers=[h1_layer, o_layer], seed=1) dropout_cost = p2_ct_mlp_dropout.Dropout(input_include_probs={ 'h1': 1.0, 'y': include_rate }, input_scales={ 'h1': 1.0, 'y': np.float32(1.0 / include_rate) }) #随机梯度下降法 algorithm = p2_alg_sgd.SGD( batch_size=b_size, learning_rate=l_rate, learning_rule=p2_alg_lr.Momentum(momentum), termination_criterion=p2_termcri.EpochCounter(max_epochs=1000), cost=dropout_cost) #训练 根据前面的定义 :dataset为一个密集型矩阵,model为MLP多层神经网络,algorithm为SGD train = pylearn2.train.Train(dataset=data_tr, model=model, algorithm=algorithm) train.setup() x = T.matrix() #定义为一个二维数组 #fprop(state_below) does the forward prop transformation y = model.fprop(x) f = theano.function([x], y) #定义一个function函数,输入为x,输出为y MAE_va_old = 10.0 #平均绝对误差 MAE_va_best = 10.0 MAE_tr_old = 10.0 #训练误差 MAE_te_old = 10.0 MAE_1000G_old = 10.0 MAE_1000G_best = 10.0 MAE_GTEx_old = 10.0 #base_name = sys.argv[1] # 获取第一个参数 sys.argv[ ]记录(获取)命令行参数 outlog = open(base_name + '.log', 'w') log_str = '\t'.join( map(str, [ 'epoch', 'MAE_va', 'MAE_va_change', 'MAE_te', 'MAE_te_change', 'MAE_tr', 'MAE_tr_change', 'learing_rate', 'time(sec)' ])) print log_str #输出运行日志 outlog.write(log_str + '\n') #Python的标准输出缓冲(这意味着它收集“写入”标准出来之前,将其写入到终端的数据)。调用sys.stdout.flush()强制其“缓冲 sys.stdout.flush() for epoch in range(0, n_epoch): t_old = time.time() train.algorithm.train(train.dataset) Y_va_hat = f(X_va.astype('float32')).astype('float64') Y_te_hat = f(X_te.astype('float32')).astype('float64') Y_tr_hat_monitor = f(X_tr_monitor.astype('float32')).astype('float64') #计算平均绝对误差 MAE_va = np.abs(Y_va_target - Y_va_hat).mean() MAE_te = np.abs(Y_te_target - Y_te_hat).mean() MAE_tr = np.abs(Y_tr_monitor_target - Y_tr_hat_monitor).mean() #误差变换率 MAE_va_change = (MAE_va - MAE_va_old) / MAE_va_old MAE_te_change = (MAE_te - MAE_te_old) / MAE_te_old MAE_tr_change = (MAE_tr - MAE_tr_old) / MAE_tr_old #将old误差值更新为当前误差值 MAE_va_old = MAE_va MAE_te_old = MAE_te MAE_tr_old = MAE_tr #返回当前的时间戳(1970纪元后经过的浮点秒数) t_new = time.time() l_rate = train.algorithm.learning_rate.get_value() log_str = '\t'.join( map(str, [ epoch + 1, '%.6f' % MAE_va, '%.6f' % MAE_va_change, '%.6f' % MAE_te, '%.6f' % MAE_te_change, '%.6f' % MAE_tr, '%.6f' % MAE_tr_change, '%.5f' % l_rate, int(t_new - t_old) ])) print log_str outlog.write(log_str + '\n') sys.stdout.flush() if MAE_tr_change > 0: #训练误差变换率大于0时,学习速率乘上一个衰减因子 l_rate = l_rate * decay_factor if l_rate < l_rate_min: #学习速率小于最小速率时,更新为最小速率 l_rate = l_rate_min train.algorithm.learning_rate.set_value(np.float32(l_rate)) if MAE_va < MAE_va_best: MAE_va_best = MAE_va outmodel = open(base_name + '_bestva_model.pkl', 'wb') pkl.dump(model, outmodel) outmodel.close() np.save(base_name + '_bestva_Y_te_hat.npy', Y_te_hat) np.save(base_name + '_bestva_Y_va_hat.npy', Y_va_hat) print 'MAE_va_best : %.6f' % (MAE_va_best) outlog.write('MAE_va_best : %.6f' % (MAE_va_best) + '\n') outlog.close()
def main(): base_name = sys.argv[1] n_epoch = int(sys.argv[2]) n_hidden = int(sys.argv[3]) include_rate = float(sys.argv[4]) in_size = 943 out_size = 4760 b_size = 200 l_rate = 3e-4 l_rate_min = 1e-5 decay_factor = 0.9 lr_scale = 3.0 momentum = 0.5 init_vals = np.sqrt(6.0/(np.array([in_size, n_hidden, n_hidden, n_hidden])+np.array([n_hidden, n_hidden, n_hidden, out_size]))) print 'loading data...' X_tr = np.load('bgedv2_X_tr_float64.npy') Y_tr = np.load('bgedv2_Y_tr_4760-9520_float64.npy') Y_tr_target = np.array(Y_tr) X_va = np.load('bgedv2_X_va_float64.npy') Y_va = np.load('bgedv2_Y_va_4760-9520_float64.npy') Y_va_target = np.array(Y_va) X_te = np.load('bgedv2_X_te_float64.npy') Y_te = np.load('bgedv2_Y_te_4760-9520_float64.npy') Y_te_target = np.array(Y_te) X_1000G = np.load('1000G_X_float64.npy') Y_1000G = np.load('1000G_Y_4760-9520_float64.npy') Y_1000G_target = np.array(Y_1000G) X_GTEx = np.load('GTEx_X_float64.npy') Y_GTEx = np.load('GTEx_Y_4760-9520_float64.npy') Y_GTEx_target = np.array(Y_GTEx) random.seed(0) monitor_idx_tr = random.sample(range(88807), 5000) data_tr = p2_dt_dd.DenseDesignMatrix(X=X_tr.astype('float32'), y=Y_tr.astype('float32')) X_tr_monitor, Y_tr_monitor_target = X_tr[monitor_idx_tr, :], Y_tr_target[monitor_idx_tr, :] h1_layer = p2_md_mlp.Tanh(layer_name='h1', dim=n_hidden, irange=init_vals[0], W_lr_scale=1.0, b_lr_scale=1.0) h2_layer = p2_md_mlp.Tanh(layer_name='h2', dim=n_hidden, irange=init_vals[1], W_lr_scale=lr_scale, b_lr_scale=1.0) h3_layer = p2_md_mlp.Tanh(layer_name='h3', dim=n_hidden, irange=init_vals[2], W_lr_scale=lr_scale, b_lr_scale=1.0) o_layer = p2_md_mlp.Linear(layer_name='y', dim=out_size, irange=0.0001, W_lr_scale=lr_scale, b_lr_scale=1.0) model = p2_md_mlp.MLP(nvis=in_size, layers=[h1_layer, h2_layer, h3_layer, o_layer], seed=1) dropout_cost = p2_ct_mlp_dropout.Dropout(input_include_probs={'h1':1.0, 'h2':include_rate, 'h3':include_rate, 'y':include_rate}, input_scales={'h1':1.0, 'h2':np.float32(1.0/include_rate), 'h3':np.float32(1.0/include_rate), 'y':np.float32(1.0/include_rate)}) algorithm = p2_alg_sgd.SGD(batch_size=b_size, learning_rate=l_rate, learning_rule = p2_alg_lr.Momentum(momentum), termination_criterion=p2_termcri.EpochCounter(max_epochs=1000), cost=dropout_cost) train = pylearn2.train.Train(dataset=data_tr, model=model, algorithm=algorithm) train.setup() x = T.matrix() y = model.fprop(x) f = theano.function([x], y) MAE_va_old = 10.0 MAE_va_best = 10.0 MAE_tr_old = 10.0 MAE_te_old = 10.0 MAE_1000G_old = 10.0 MAE_1000G_best = 10.0 MAE_GTEx_old = 10.0 outlog = open(base_name + '.log', 'w') log_str = '\t'.join(map(str, ['epoch', 'MAE_va', 'MAE_va_change', 'MAE_te', 'MAE_te_change', 'MAE_1000G', 'MAE_1000G_change', 'MAE_GTEx', 'MAE_GTEx_change', 'MAE_tr', 'MAE_tr_change', 'learing_rate', 'time(sec)'])) print log_str outlog.write(log_str + '\n') sys.stdout.flush() for epoch in range(0, n_epoch): t_old = time.time() train.algorithm.train(train.dataset) Y_va_hat = f(X_va.astype('float32')).astype('float64') Y_te_hat = f(X_te.astype('float32')).astype('float64') Y_tr_hat_monitor = f(X_tr_monitor.astype('float32')).astype('float64') Y_1000G_hat = f(X_1000G.astype('float32')).astype('float64') Y_GTEx_hat = f(X_GTEx.astype('float32')).astype('float64') MAE_va = np.abs(Y_va_target - Y_va_hat).mean() MAE_te = np.abs(Y_te_target - Y_te_hat).mean() MAE_tr = np.abs(Y_tr_monitor_target - Y_tr_hat_monitor).mean() MAE_1000G = np.abs(Y_1000G_target - Y_1000G_hat).mean() MAE_GTEx = np.abs(Y_GTEx_target - Y_GTEx_hat).mean() MAE_va_change = (MAE_va - MAE_va_old)/MAE_va_old MAE_te_change = (MAE_te - MAE_te_old)/MAE_te_old MAE_tr_change = (MAE_tr - MAE_tr_old)/MAE_tr_old MAE_1000G_change = (MAE_1000G - MAE_1000G_old)/MAE_1000G_old MAE_GTEx_change = (MAE_GTEx - MAE_GTEx_old)/MAE_GTEx_old MAE_va_old = MAE_va MAE_te_old = MAE_te MAE_tr_old = MAE_tr MAE_1000G_old = MAE_1000G MAE_GTEx_old = MAE_GTEx t_new = time.time() l_rate = train.algorithm.learning_rate.get_value() log_str = '\t'.join(map(str, [epoch+1, '%.6f'%MAE_va, '%.6f'%MAE_va_change, '%.6f'%MAE_te, '%.6f'%MAE_te_change, '%.6f'%MAE_1000G, '%.6f'%MAE_1000G_change, '%.6f'%MAE_GTEx, '%.6f'%MAE_GTEx_change, '%.6f'%MAE_tr, '%.6f'%MAE_tr_change, '%.5f'%l_rate, int(t_new-t_old)])) print log_str outlog.write(log_str + '\n') sys.stdout.flush() if MAE_tr_change > 0: l_rate = l_rate*decay_factor if l_rate < l_rate_min: l_rate = l_rate_min train.algorithm.learning_rate.set_value(np.float32(l_rate)) if MAE_va < MAE_va_best: MAE_va_best = MAE_va outmodel = open(base_name + '_bestva_model.pkl', 'wb') pkl.dump(model, outmodel) outmodel.close() np.save(base_name + '_bestva_Y_te_hat.npy', Y_te_hat) np.save(base_name + '_bestva_Y_va_hat.npy', Y_va_hat) if MAE_1000G < MAE_1000G_best: MAE_1000G_best = MAE_1000G outmodel = open(base_name + '_best1000G_model.pkl', 'wb') pkl.dump(model, outmodel) outmodel.close() np.save(base_name + '_best1000G_Y_1000G_hat.npy', Y_1000G_hat) np.save(base_name + '_best1000G_Y_GTEx_hat.npy', Y_GTEx_hat) print 'MAE_va_best : %.6f' % (MAE_va_best) print 'MAE_1000G_best : %.6f' % (MAE_1000G_best) outlog.write('MAE_va_best : %.6f' % (MAE_va_best) + '\n') outlog.write('MAE_1000G_best : %.6f' % (MAE_1000G_best) + '\n') outlog.close()
def main(): training_data, validation_data, test_data, std_scale = load_training_data() kaggle_test_features = load_test_data(std_scale) ############### # pylearn2 ML hl1 = mlp.Sigmoid(layer_name='hl1', dim=200, irange=.1, init_bias=1.) hl2 = mlp.Sigmoid(layer_name='hl2', dim=100, irange=.1, init_bias=1.) # create Softmax output layer output_layer = mlp.Softmax(9, 'output', irange=.1) # create Stochastic Gradient Descent trainer that runs for 400 epochs trainer = sgd.SGD(learning_rate=.05, batch_size=300, learning_rule=learning_rule.Momentum(.5), termination_criterion=MonitorBased( channel_name='valid_objective', prop_decrease=0., N=10), monitoring_dataset={ 'valid': validation_data, 'train': training_data }) layers = [hl1, hl2, output_layer] # create neural net model = mlp.MLP(layers, nvis=93) watcher = best_params.MonitorBasedSaveBest( channel_name='valid_objective', save_path='pylearn2_results/pylearn2_test.pkl') velocity = learning_rule.MomentumAdjustor(final_momentum=.6, start=1, saturate=250) decay = sgd.LinearDecayOverEpoch(start=1, saturate=250, decay_factor=.01) ###################### experiment = Train(dataset=training_data, model=model, algorithm=trainer, extensions=[watcher, velocity, decay]) experiment.main_loop() #load best model and test ################ model = serial.load('pylearn2_results/pylearn2_test.pkl') # get an prediction of the accuracy from the test_data test_results = model.fprop(theano.shared(test_data[0], name='test_data')).eval() print test_results.shape loss = multiclass_log_loss(test_data[1], test_results) print 'Test multiclass log loss:', loss out_file = 'pylearn2_results/' + str(loss) + 'ann' #exp.save(out_file + '.pkl') #save the kaggle results results = model.fprop( theano.shared(kaggle_test_features, name='kaggle_test_data')).eval() save_results(out_file + '.csv', kaggle_test_features, results)
#output = mlp.HingeLoss(layer_name='y',n_classes=2,irange=.05) #layers = [l5, l6, output] layers = [l1, l2, l3, l4, l5, output] ann = mlp.MLP(layers, nvis=X[0].reshape(-1).shape[0]) lr = 0.1 epochs = 400 trainer = sgd.SGD( learning_rate=lr, batch_size=100, learning_rule=learning_rule.Momentum(.05), # Remember, default dropout is .5 #cost=Dropout(input_include_probs={'l1': .5}, # input_scales={'l1': 1.}), termination_criterion=EpochCounter(epochs), monitoring_dataset={ 'train': ds, 'valid': ds_test }) watcher = best_params.MonitorBasedSaveBest(channel_name='valid_roc_auc', save_path='saved_clf.pkl') velocity = learning_rule.MomentumAdjustor(final_momentum=.9, start=1, saturate=250) decay = sgd.LinearDecayOverEpoch(start=1, saturate=250, decay_factor=lr * .05) rocauc = roc_auc.RocAucChannel()
images_train = images[train_index] y_train = y[train_index] images_train, y_train = shuffle(images_train, y_train, random_state=7) X_train = DenseDesignMatrix(X=images_train, y=y_train,view_converter=view_converter) images_test = images[test_index] y_test = y[test_index] X_test = DenseDesignMatrix(X=images_test, y=y_test,view_converter=view_converter) if retrain: print "training on", X_train.X.shape, 'testing on', X_test.X.shape trainer = sgd.SGD(learning_rate=learn_rate, batch_size=batch_size, learning_rule=learning_rule.Momentum(momentum_start), cost=Dropout( input_include_probs={'l1':1., 'l2':1., 'l3':1., 'l4':1., 'l5':1., 'l6':1.}, input_scales={'l1':1., 'l2':1., 'l3':1., 'l4':1., 'l5':1., 'l6':1.} ), termination_criterion=EpochCounter(max_epochs=max_epochs), monitoring_dataset={'train':X_train, 'valid':X_test}, ) input_space = Conv2DSpace(shape=(central_window_shape, central_window_shape), axes = axes, num_channels = 1) ann = mlp.MLP(layers, input_space=input_space) velocity = learning_rule.MomentumAdjustor(final_momentum=momentum_end, start=1, saturate=momentum_saturate)
layerh3 = mlp.ConvRectifiedLinear(layer_name='h3', output_channels=64, irange=.05, kernel_shape=[5, 5], pool_shape=[4, 4], pool_stride=[2, 2], max_kernel_norm=1.9365) ''' Note: changed the number of classes ''' layery = mlp.Softmax(max_col_norm=1.9365, layer_name='y', n_classes=121, istdev=.05) print 'Setting up trainers' trainer = sgd.SGD(learning_rate=0.5, batch_size=50, termination_criterion=EpochCounter(200), learning_rule=Momentum(init_momentum=0.5)) layers = [layerh2, layerh3, layery] ann = mlp.MLP(layers, input_space=Conv2DSpace(shape=[28, 28], num_channels=1)) trainer.setup(ann, ds) print 'Start Training' while True: trainer.train(dataset=ds) ann.monitor.report_epoch() ann.monitor() if not trainer.continue_learning(ann): break # 3. Predict XReport, Y_info = plankton.loadReportData() probMatrix = ann.fprop(theano.shared(XReport, name='XReport')).eval()
irange=ir, dim=dim, max_col_norm=1.) l3 = RectifiedLinear(layer_name='l3', irange=ir, dim=dim, max_col_norm=1.) output = Softmax(layer_name='y', n_classes=9, irange=ir, max_col_norm=mcn_out) mdl = MLP([l1, l2, l3, output], nvis=X2.shape[1]) trainer = sgd.SGD(learning_rate=lr, batch_size=bs, learning_rule=learning_rule.Momentum(mm), cost=Dropout(default_input_include_prob=ip, default_input_scale=1 / ip), termination_criterion=EpochCounter(epochs), seed=seed) decay = sgd.LinearDecayOverEpoch(start=2, saturate=20, decay_factor=.1) experiment = Train(dataset=training, model=mdl, algorithm=trainer, extensions=[decay]) experiment.main_loop() epochs_current = epochs for s in range(n_add): trainer = sgd.SGD(learning_rate=lr * .1, batch_size=bs, learning_rule=learning_rule.Momentum(mm), cost=Dropout(default_input_include_prob=ip,
from csv_data import CSVData import numpy as np class MLPData(DenseDesignMatrix): def __init__(self, X, y): super(MLPData, self).__init__(X=X, y=y.astype(int), y_labels=2) threshold = 0.95 hidden_layer = mlp.Sigmoid(layer_name='h0', dim=10, sparse_init=10) output_layer = mlp.Softmax(layer_name='y', n_classes=2, irange=0.05) layers = [hidden_layer, output_layer] neural_net = mlp.MLP(layers, nvis=10) trainer = sgd.SGD(batch_size=5, learning_rate=.1, termination_criterion=EpochCounter(100)) first = True learning = True correct = 0 incorrect = 0 total = 0 data = CSVData("results2.csv") while True: X, y = data.get_data() if (X == None): break if learning: ds = MLPData(X, np.array([[0]]))
def cnn_train( train_path, test_path, valid_path, save_path, predict_path, image_path, num_rows=28, num_cols=28, num_channels=2, batch_size=128, output_channels=[64, 64], kernel_shape=[[12, 12], [5, 5]], pool_shape=[[4, 4], [2, 2]], pool_stride=[[2, 2], [2, 2]], irange=[0.05, 0.05, 0.05], max_kernel_norm=[1.9365, 1.9365], learning_rate=0.001, init_momentum=0.9, weight_decay=[0.0002, 0.0002, 0.0002], n_epoch=1000, ): #load data #t = time.time() ds = load_data(valid_path, num_rows, num_cols, num_channels) vld = SarDataset(np.array(ds[0]), ds[1]) ds = load_data(train_path, num_rows, num_cols, num_channels) trn = SarDataset(np.array(ds[0]), ds[1]) ds = load_data(test_path, num_rows, num_cols, num_channels) tst = SarDataset(np.array(ds[0]), ds[1]) #load balanced data #ds = load_data_balance_under_sample(train_path, num_rows,num_cols, num_channels) #trn = SarDataset(np.array(ds[0]),ds[1]) #ds = load_data_balance(valid_path, num_rows,num_cols, num_channels) #vld = SarDataset(np.array(ds[0]),ds[1]) #ds = load_data_balance(test_path, num_rows,num_cols, num_channels) #tst = SarDataset(np.array(ds[0]),ds[1]) #print 'Take {}s to read data'.format( time.time()-t) #use gaussian convlution on the origional image to see if it can concentrate in the center #trn,tst,vld = load_data_lidar() #mytransformer = transformer.TransformationPipeline(input_space=space.Conv2DSpace(shape=[num_rows,num_cols],num_channels=num_channels),transformations=[transformer.Rotation(),transformer.Flipping()]) #trn = contestTransformerDataset.TransformerDataset(trn,mytransformer,space_preserving=True) #tst = contestTransformerDataset.TransformerDataset(tst,mytransformer,space_preserving=True) #vld = contestTransformerDataset.TransformerDataset(vld,mytransformer,space_preserving=True) #trn = transformer_dataset.TransformerDataset(trn,mytransformer,space_preserving=True) #tst = transformer_dataset.TransformerDataset(tst,mytransformer,space_preserving=True) #vld = transformer_dataset.TransformerDataset(vld,mytransformer,space_preserving=True) #setup the network t = time.time() layers = [] for i in range(len(output_channels)): layer_name = 'h{}'.format(i + 1) convlayer = mlp.ConvRectifiedLinear(layer_name=layer_name, output_channels=output_channels[i], irange=irange[i], kernel_shape=kernel_shape[i], pool_shape=pool_shape[i], pool_stride=pool_stride[i], max_kernel_norm=max_kernel_norm[i]) layers.append(convlayer) output_mlp = mlp.Linear(dim=1, layer_name='output', irange=irange[-1], use_abs_loss=True) #output_mlp = mlp.linear_mlp_ace(dim=1,layer_name='output',irange=irange[-1]) layers.append(output_mlp) #ann = cPickle.load(open('../output/train_with_2010_2l_40_64/original_500/f/f0.pkl')) #layers = [] #for layer in ann.layers: # layer.set_mlp_force(None) # layers.append(layer) trainer = sgd.SGD( learning_rate=learning_rate, batch_size=batch_size, termination_criterion=EpochCounter(n_epoch), #termination_criterion = termination_criteria.And([termination_criteria.MonitorBased(channel_name = 'train_objective', prop_decrease=0.01,N=10),EpochCounter(n_epoch)]), #cost = dropout.Dropout(), cost=cost.SumOfCosts( [cost.MethodCost('cost_from_X'), WeightDecay(weight_decay)]), init_momentum=init_momentum, train_iteration_mode='even_shuffled_sequential', monitor_iteration_mode='even_shuffled_sequential', monitoring_dataset={ 'test': tst, 'valid': vld, 'train': trn }) input_space = space.Conv2DSpace(shape=[num_rows, num_cols], num_channels=num_channels) #ann = mlp.MLP(layers,input_space=input_space,batch_size=batch_size) ann = serial.load( '../output/train_with_2010_2l_40_64/original_500/f/f0.pkl') ann = monitor.push_monitor(ann, 'stage_0') watcher = best_params.MonitorBasedSaveBest(channel_name='valid_objective', save_path=predict_path + save_path) flip = window_flip.WindowAndFlip((num_rows, num_cols), randomize=[tst, vld, trn]) experiment = Train(dataset=trn, model=ann, algorithm=trainer, extensions=[watcher, flip]) print 'Take {}s to compile code'.format(time.time() - t) #train the network t = time.time() experiment.main_loop() print 'Training time: {}h'.format((time.time() - t) / 3600) utils.sms_notice('Training time:{}'.format((time.time() - t) / 3600)) return ann
def cnn_train_tranformer(train_path, test_path, valid_path, save_path, predict_path, num_rows=28, num_cols=28, num_channels=2, batch_size=128, output_channels=[64, 64], kernel_shape=[[12, 12], [5, 5]], pool_shape=[[4, 4], [2, 2]], pool_stride=[[2, 2], [2, 2]], irange=[0.05, 0.05, 0.05], max_kernel_norm=[1.9365, 1.9365], learning_rate=0.001, init_momentum=0.9, weight_decay=[0.0002, 0.0002, 0.0002], n_epoch=1000, image_path=''): ds = load_data_transformed(train_path, num_cols, batch_size) ds = (np.transpose(ds[0], axes=[0, 3, 1, 2]), ds[1]) trn = SarDataset(np.array(ds[0]), ds[1]) ds = load_data_transformed(valid_path, num_cols, batch_size) ds = (np.transpose(ds[0], axes=[0, 3, 1, 2]), ds[1]) vld = SarDataset(np.array(ds[0]), ds[1]) ds = load_data_transformed(test_path, num_cols, batch_size) ds = (np.transpose(ds[0], axes=[0, 3, 1, 2]), ds[1]) tst = SarDataset(np.array(ds[0]), ds[1]) #setup the network #X = np.random.random([400000,2,41,41]) #y = np.random.random([400000,1]) #trn = SarDataset(X,y) #X = np.random.random([60000,2,41,41]) #y = np.random.random([60000,1]) #tst = SarDataset(X,y) #X = np.random.random([60000,2,41,41]) #y = np.random.random([60000,1]) #vld = SarDataset(X,y) t = time.time() layers = [] for i in range(len(output_channels)): layer_name = 'h{}'.format(i + 1) convlayer = mlp.ConvRectifiedLinear(layer_name=layer_name, output_channels=output_channels[i], irange=irange[i], kernel_shape=kernel_shape[i], pool_shape=pool_shape[i], pool_stride=pool_stride[i], max_kernel_norm=max_kernel_norm[i]) layers.append(convlayer) output_mlp = mlp.Linear(dim=1, layer_name='output', irange=irange[-1]) #output_mlp = mlp.linear_mlp_bayesian_cost(dim=1,layer_name='output',irange=irange[-1]) layers.append(output_mlp) trainer = sgd.SGD( learning_rate=learning_rate, batch_size=batch_size, termination_criterion=EpochCounter(n_epoch), #termination_criterion = termination_criteria.And([termination_criteria.MonitorBased(channel_name = 'train_objective', prop_decrease=0.01,N=10),EpochCounter(n_epoch)]), #cost = dropout.Dropout(), cost=cost.SumOfCosts( [cost.MethodCost('cost_from_X'), WeightDecay(weight_decay)]), init_momentum=init_momentum, train_iteration_mode='even_shuffled_sequential', monitor_iteration_mode='even_shuffled_sequential', monitoring_dataset={ 'test': tst, 'valid': vld, 'train': trn }) input_space = space.Conv2DSpace(shape=[num_rows, num_cols], num_channels=num_channels) #ann = mlp.MLP(layers,input_space=input_space,batch_size=batch_size) watcher = best_params.MonitorBasedSaveBest(channel_name='valid_objective', save_path=predict_path + save_path) #flip = window_flip.WindowAndFlip((num_rows,num_cols),randomize=[tst,vld,trn]) experiment = Train(dataset=trn, model=ann, algorithm=trainer, extensions=[watcher]) print 'Take {}s to compile code'.format(time.time() - t) #train the network t = time.time() experiment.main_loop() print 'Training time: {}h'.format((time.time() - t) / 3600) utils.sms_notice('Training time:{}'.format((time.time() - t) / 3600)) return ann
irange=ir, dim=dim, max_col_norm=1.) l3 = RectifiedLinear(layer_name='l3', irange=ir, dim=dim, max_col_norm=1.) output = Softmax(layer_name='y', n_classes=9, irange=ir, max_col_norm=mcn_out) mdl = MLP([l1, l2, l3, output], nvis=X2.shape[1]) trainer = sgd.SGD(learning_rate=lr, batch_size=bs, learning_rule=learning_rule.Momentum(mm), cost=Dropout(default_input_include_prob=ip, default_input_scale=1 / ip), termination_criterion=EpochCounter(epochs), seed=seed) decay = sgd.LinearDecayOverEpoch(start=2, saturate=20, decay_factor=.1) #fname = path + 'model/TRI_' + 'kmax_'+ str(k_max) + '_seed_' + str(seed) + '.pkl' experiment = Train(dataset=training, model=mdl, algorithm=trainer, extensions=[decay]) # save_path = fname, save_freq = epochs) experiment.main_loop() pred_train = predict(mdl, X2[:num_train].astype(np.float32)) pred_test = predict(mdl, X2[num_train:].astype(np.float32)) predAll_train += pred_train predAll_test += pred_test
def train(d): print 'Creating dataset' # load mnist here # X = d.train_X # y = d.train_Y # test_X = d.test_X # test_Y = d.test_Y # nb_classes = len(np.unique(y)) # train_y = convert_one_hot(y) # train_set = DenseDesignMatrix(X=X, y=y) train = DenseDesignMatrix(X=d.train_X, y=convert_one_hot(d.train_Y)) valid = DenseDesignMatrix(X=d.valid_X, y=convert_one_hot(d.valid_Y)) test = DenseDesignMatrix(X=d.test_X, y=convert_one_hot(d.test_Y)) print 'Setting up' batch_size = 1000 conv = mlp.ConvRectifiedLinear( layer_name='c0', output_channels=20, irange=.05, kernel_shape=[5, 5], pool_shape=[4, 4], pool_stride=[2, 2], # W_lr_scale=0.25, max_kernel_norm=1.9365) mout = MaxoutConvC01B(layer_name='m0', num_pieces=4, num_channels=96, irange=.05, kernel_shape=[5, 5], pool_shape=[4, 4], pool_stride=[2, 2], W_lr_scale=0.25, max_kernel_norm=1.9365) mout2 = MaxoutConvC01B(layer_name='m1', num_pieces=4, num_channels=96, irange=.05, kernel_shape=[5, 5], pool_shape=[4, 4], pool_stride=[2, 2], W_lr_scale=0.25, max_kernel_norm=1.9365) sigmoid = mlp.Sigmoid( layer_name='Sigmoid', dim=500, sparse_init=15, ) smax = mlp.Softmax(layer_name='y', n_classes=10, irange=0.) in_space = Conv2DSpace(shape=[28, 28], num_channels=1, axes=['c', 0, 1, 'b']) net = mlp.MLP( layers=[mout, mout2, smax], input_space=in_space, # nvis=784, ) trainer = bgd.BGD(batch_size=batch_size, line_search_mode='exhaustive', conjugate=1, updates_per_batch=10, monitoring_dataset={ 'train': train, 'valid': valid, 'test': test }, termination_criterion=termination_criteria.MonitorBased( channel_name='valid_y_misclass')) trainer = sgd.SGD(learning_rate=0.15, cost=dropout.Dropout(), batch_size=batch_size, monitoring_dataset={ 'train': train, 'valid': valid, 'test': test }, termination_criterion=termination_criteria.MonitorBased( channel_name='valid_y_misclass')) trainer.setup(net, train) epoch = 0 while True: print 'Training...', epoch trainer.train(dataset=train) net.monitor() epoch += 1
def supervisedLayerwisePRL(trainset, testset): ''' The supervised layerwise training as used in the PRL Paper. Input ------ trainset : A path to an hdf5 file created through h5py. testset : A path to an hdf5 file created through h5py. ''' batch_size = 100 # Both train and test h5py files are expected to have a 'topo_view' and 'y' # datasets side them corresponding to the 'b01c' data format as used in pylearn2 # and 'y' equivalent to the one hot encoded labels trn = HDF5Dataset(filename=trainset, topo_view='topo_view', y='y', load_all=False) tst = HDF5Dataset(filename=testset, topo_view='topo_view', y='y', load_all=False) ''' The 1st Convolution and Pooling Layers are added below. ''' h1 = mlp.ConvRectifiedLinear(layer_name='h1', output_channels=64, irange=0.05, kernel_shape=[4, 4], pool_shape=[4, 4], pool_stride=[2, 2], max_kernel_norm=1.9365) fc = mlp.RectifiedLinear(layer_name='fc', dim=1500, irange=0.05) output = mlp.Softmax(layer_name='y', n_classes=171, irange=.005, max_col_norm=1.9365) layers = [h1, fc, output] mdl = mlp.MLP(layers, input_space=Conv2DSpace(shape=(70, 70), num_channels=1)) trainer = sgd.SGD( learning_rate=0.002, batch_size=batch_size, learning_rule=learning_rule.RMSProp(), cost=SumOfCosts( costs=[Default(), WeightDecay(coeffs=[0.0005, 0.0005, 0.0005])]), train_iteration_mode='shuffled_sequential', monitor_iteration_mode='sequential', termination_criterion=EpochCounter(max_epochs=15), monitoring_dataset={ 'test': tst, 'valid': vld }) watcher = best_params.MonitorBasedSaveBest( channel_name='valid_y_misclass', save_path='./Saved Models/conv_supervised_layerwise_best1.pkl') decay = sgd.LinearDecayOverEpoch(start=8, saturate=15, decay_factor=0.1) experiment = Train( dataset=trn, model=mdl, algorithm=trainer, extensions=[watcher, decay], ) experiment.main_loop() del mdl mdl = serial.load('./Saved Models/conv_supervised_layerwise_best1.pkl') mdl = push_monitor(mdl, 'k') ''' The 2nd Convolution and Pooling Layers are added below. ''' h2 = mlp.ConvRectifiedLinear(layer_name='h2', output_channels=64, irange=0.05, kernel_shape=[4, 4], pool_shape=[4, 4], pool_stride=[2, 2], max_kernel_norm=1.9365) fc = mlp.RectifiedLinear(layer_name='fc', dim=1500, irange=0.05) output = mlp.Softmax(layer_name='y', n_classes=171, irange=.005, max_col_norm=1.9365) del mdl.layers[-1] mdl.layer_names.remove('y') del mdl.layers[-1] mdl.layer_names.remove('fc') mdl.add_layers([h2, fc, output]) trainer = sgd.SGD(learning_rate=0.002, batch_size=batch_size, learning_rule=learning_rule.RMSProp(), cost=SumOfCosts(costs=[ Default(), WeightDecay(coeffs=[0.0005, 0.0005, 0.0005, 0.0005]) ]), train_iteration_mode='shuffled_sequential', monitor_iteration_mode='sequential', termination_criterion=EpochCounter(max_epochs=15), monitoring_dataset={ 'test': tst, 'valid': vld }) watcher = best_params.MonitorBasedSaveBest( channel_name='valid_y_misclass', save_path='./Saved Models/conv_supervised_layerwise_best2.pkl') decay = sgd.LinearDecayOverEpoch(start=8, saturate=15, decay_factor=0.1) experiment = Train( dataset=trn, model=mdl, algorithm=trainer, extensions=[watcher, decay], ) experiment.main_loop() del mdl mdl = serial.load('./Saved Models/conv_supervised_layerwise_best2.pkl') mdl = push_monitor(mdl, 'l') ''' The 3rd Convolution and Pooling Layers are added below. ''' h3 = mlp.ConvRectifiedLinear(layer_name='h2', output_channels=64, irange=0.05, kernel_shape=[4, 4], pool_shape=[4, 4], pool_stride=[2, 2], max_kernel_norm=1.9365) fc = mlp.RectifiedLinear(layer_name='h3', dim=1500, irange=0.05) output = mlp.Softmax(layer_name='y', n_classes=10, irange=.005, max_col_norm=1.9365) del mdl.layers[-1] mdl.layer_names.remove('y') del mdl.layers[-1] mdl.layer_names.remove('fc') mdl.add_layers([h3, output]) trainer = sgd.SGD( learning_rate=.002, batch_size=batch_size, learning_rule=learning_rule.RMSProp(), cost=SumOfCosts(costs=[ Default(), WeightDecay(coeffs=[0.0005, 0.0005, 0.0005, 0.0005, 0.0005]) ]), train_iteration_mode='shuffled_sequential', monitor_iteration_mode='sequential', termination_criterion=EpochCounter(max_epochs=15), monitoring_dataset={ 'test': tst, 'valid': vld }) watcher = best_params.MonitorBasedSaveBest( channel_name='valid_y_misclass', save_path='./Saved Models/conv_supervised_layerwise_best3.pkl') decay = sgd.LinearDecayOverEpoch(start=8, saturate=15, decay_factor=0.1) experiment = Train( dataset=trn, model=mdl, algorithm=trainer, extensions=[watcher, decay], ) experiment.main_loop()
def main(): base_name = sys.argv[1] #文件名前缀 n_epoch = int(sys.argv[2]) # epoch次数 n_hidden = int(sys.argv[3]) # 隐含层节点数 include_rate = float(sys.argv[4]) # 包含率(1-dropout) in_size = 943 # 输入层节点数目 out_size = 4760 #输出层节点数 b_size = 200 #batch的大小 l_rate = 5e-4 #学习速率 l_rate_min = 1e-5 #学习速率最小值 decay_factor = 0.9 # lr_scale = 3.0 # momentum = 0.5 #摄动因子 init_vals = np.sqrt(6.0/(np.array([in_size, n_hidden])+np.array([n_hidden, out_size]))) print 'loading data...' #读取数据Train,Validation,Test X_tr = np.load('bgedv2_X_tr_float64.npy') Y_tr = np.load('bgedv2_Y_tr_0-4760_float64.npy') Y_tr_target = np.array(Y_tr) X_va = np.load('bgedv2_X_va_float64.npy') Y_va = np.load('bgedv2_Y_va_0-4760_float64.npy') Y_va_target = np.array(Y_va) X_te = np.load('bgedv2_X_te_float64.npy') Y_te = np.load('bgedv2_Y_te_0-4760_float64.npy') Y_te_target = np.array(Y_te) X_1000G = np.load('1000G_X_float64.npy') Y_1000G = np.load('1000G_Y_0-4760_float64.npy') Y_1000G_target = np.array(Y_1000G) X_GTEx = np.load('GTEx_X_float64.npy') Y_GTEx = np.load('GTEx_Y_0-4760_float64.npy') Y_GTEx_target = np.array(Y_GTEx) #随机化 random.seed(0) #随机抽取5000样本进行训练 monitor_idx_tr = random.sample(range(88807), 5000) #将数据X,Y整合成DensenMatrix类型 data_tr = p2_dt_dd.DenseDesignMatrix(X=X_tr.astype('float32'), y=Y_tr.astype('float32')) #取出X中对应5000样本进行训练 X_tr_monitor, Y_tr_monitor_target = X_tr[monitor_idx_tr, :], Y_tr_target[monitor_idx_tr, :] #设置多层感知机的隐含层计算方式 h1_layer = p2_md_mlp.Tanh(layer_name='h1', dim=n_hidden, irange=init_vals[0], W_lr_scale=1.0, b_lr_scale=1.0) #设置多层感知机的输出层计算方式 o_layer = p2_md_mlp.Linear(layer_name='y', dim=out_size, irange=0.0001, W_lr_scale=lr_scale, b_lr_scale=1.0) #设置好模型 model = p2_md_mlp.MLP(nvis=in_size, layers=[h1_layer, o_layer], seed=1) #设置dropout比例 dropout_cost = p2_ct_mlp_dropout.Dropout(input_include_probs={'h1':1.0, 'y':include_rate}, input_scales={'h1':1.0, 'y':np.float32(1.0/include_rate)}) #设置训练算法(batch大小,学习速率,学习规则,终止条件,dropout比例) algorithm = p2_alg_sgd.SGD(batch_size=b_size, learning_rate=l_rate, learning_rule = p2_alg_lr.Momentum(momentum), termination_criterion=p2_termcri.EpochCounter(max_epochs=1000), cost=dropout_cost) #设置训练类(数据集,训练模型,训练算法) train = pylearn2.train.Train(dataset=data_tr, model=model, algorithm=algorithm) train.setup() x = T.matrix() y = model.fprop(x) #训练好的模型对X的预测值 f = theano.function([x], y) MAE_va_old = 10.0 MAE_va_best = 10.0 MAE_tr_old = 10.0 MAE_te_old = 10.0 MAE_1000G_old = 10.0 MAE_1000G_best = 10.0 MAE_GTEx_old = 10.0 outlog = open(base_name + '.log', 'w') log_str = '\t'.join(map(str, ['epoch', 'MAE_va', 'MAE_va_change', 'MAE_te', 'MAE_te_change', 'MAE_1000G', 'MAE_1000G_change', 'MAE_GTEx', 'MAE_GTEx_change', 'MAE_tr', 'MAE_tr_change', 'learing_rate', 'time(sec)'])) print log_str outlog.write(log_str + '\n') sys.stdout.flush() #刷新缓冲区 for epoch in range(0, n_epoch): t_old = time.time() #开始时间 train.algorithm.train(train.dataset)#训练 #计算不同数据集预测值 Y_va_hat = f(X_va.astype('float32')).astype('float64') Y_te_hat = f(X_te.astype('float32')).astype('float64') Y_tr_hat_monitor = f(X_tr_monitor.astype('float32')).astype('float64') Y_1000G_hat = f(X_1000G.astype('float32')).astype('float64') Y_GTEx_hat = f(X_GTEx.astype('float32')).astype('float64') #计算预测值与真实值的MAE MAE_va = np.abs(Y_va_target - Y_va_hat).mean() MAE_te = np.abs(Y_te_target - Y_te_hat).mean() MAE_tr = np.abs(Y_tr_monitor_target - Y_tr_hat_monitor).mean() MAE_1000G = np.abs(Y_1000G_target - Y_1000G_hat).mean() MAE_GTEx = np.abs(Y_GTEx_target - Y_GTEx_hat).mean() #计算迭代误差 MAE_va_change = (MAE_va - MAE_va_old)/MAE_va_old MAE_te_change = (MAE_te - MAE_te_old)/MAE_te_old MAE_tr_change = (MAE_tr - MAE_tr_old)/MAE_tr_old MAE_1000G_change = (MAE_1000G - MAE_1000G_old)/MAE_1000G_old MAE_GTEx_change = (MAE_GTEx - MAE_GTEx_old)/MAE_GTEx_old #更新MAE MAE_va_old = MAE_va MAE_te_old = MAE_te MAE_tr_old = MAE_tr MAE_1000G_old = MAE_1000G MAE_GTEx_old = MAE_GTEx t_new = time.time() #终止时间 l_rate = train.algorithm.learning_rate.get_value() log_str = '\t'.join(map(str, [epoch+1, '%.6f'%MAE_va, '%.6f'%MAE_va_change, '%.6f'%MAE_te, '%.6f'%MAE_te_change, '%.6f'%MAE_1000G, '%.6f'%MAE_1000G_change, '%.6f'%MAE_GTEx, '%.6f'%MAE_GTEx_change, '%.6f'%MAE_tr, '%.6f'%MAE_tr_change, '%.5f'%l_rate, int(t_new-t_old)])) print log_str outlog.write(log_str + '\n') sys.stdout.flush() if MAE_tr_change > 0: #如果误差增大,减小学习速率 l_rate = l_rate*decay_factor if l_rate < l_rate_min: #学习速率最小为l_rate_min l_rate = l_rate_min train.algorithm.learning_rate.set_value(np.float32(l_rate)) #更改训练类的学习速率参数 #更新Validation误差值 if MAE_va < MAE_va_best: MAE_va_best = MAE_va outmodel = open(base_name + '_bestva_model.pkl', 'wb') pkl.dump(model, outmodel) outmodel.close() np.save(base_name + '_bestva_Y_te_hat.npy', Y_te_hat) np.save(base_name + '_bestva_Y_va_hat.npy', Y_va_hat) #更新1000G误差值 if MAE_1000G < MAE_1000G_best: MAE_1000G_best = MAE_1000G outmodel = open(base_name + '_best1000G_model.pkl', 'wb') pkl.dump(model, outmodel) outmodel.close() np.save(base_name + '_best1000G_Y_1000G_hat.npy', Y_1000G_hat) np.save(base_name + '_best1000G_Y_GTEx_hat.npy', Y_GTEx_hat) print 'MAE_va_best : %.6f' % (MAE_va_best) print 'MAE_1000G_best : %.6f' % (MAE_1000G_best) outlog.write('MAE_va_best : %.6f' % (MAE_va_best) + '\n') outlog.write('MAE_1000G_best : %.6f' % (MAE_1000G_best) + '\n') outlog.close()
random.seed(0) #设置生成随机数用的整数起始值。调用任何其他random模块函数之前调用这个函数 monitor_idx_tr = random.sample(range(88807), 5000) #监测训练 #将训练数据集类型设为32位浮点型,The DenseDesignMatrix class and related code Functionality for representing data that can be described as a dense matrix (rather than a sparse matrix) with each row containing an example and each column corresponding to a different feature. data_tr = p2_dt_dd.DenseDesignMatrix(X=X_tr.astype('float32'), y=Y_tr.astype('float32')) X_tr_monitor, Y_tr_monitor_target = X_tr[monitor_idx_tr, :], Y_tr_target[monitor_idx_tr, :] #一个隐层,用Tanh()作激活函数; 输出层用线性函数作激活函数 h1_layer = p2_md_mlp.Tanh(layer_name='h1', dim=n_hidden, irange=init_vals[0], W_lr_scale=1.0, b_lr_scale=1.0) o_layer = p2_md_mlp.Linear(layer_name='y', dim=out_size, irange=0.0001, W_lr_scale=lr_scale, b_lr_scale=1.0) #Multilayer Perceptron;nvis(Number of “visible units” input units) layers(a list of layer objects,最后1层指定MLP的输出空间) model = p2_md_mlp.MLP(nvis=in_size, layers=[h1_layer, o_layer], seed=1) dropout_cost = p2_ct_mlp_dropout.Dropout(input_include_probs={'h1':1.0, 'y':include_rate}, input_scales={'h1':1.0, 'y':np.float32(1.0/include_rate)}) #随机梯度下降法 algorithm = p2_alg_sgd.SGD(batch_size=b_size, learning_rate=l_rate, learning_rule = p2_alg_lr.Momentum(momentum), termination_criterion=p2_termcri.EpochCounter(max_epochs=1000), cost=dropout_cost) #训练 根据前面的定义 :dataset为一个密集型矩阵,model为MLP多层神经网络,algorithm为SGD train = pylearn2.train.Train(dataset=data_tr, model=model, algorithm=algorithm) train.setup() x = T.matrix() #定义为一个二维数组 #fprop(state_below) does the forward prop transformation y = model.fprop(x) f = theano.function([x], y) #定义一个function函数,输入为x,输出为y MAE_va_old = 10.0 #平均绝对误差 MAE_va_best = 10.0 MAE_tr_old = 10.0 #训练误差 MAE_te_old = 10.0
y.append([1, 0]) X = np.array(X) y = np.array(y) super(XOR, self).__init__(X=X, y=y) # create XOR dataset ds = XOR() # create hidden layer with 2 nodes, init weights in range -0.1 to 0.1 and add # a bias with value 1 hidden_layer = mlp.Sigmoid(layer_name='hidden', dim=2, irange=.1, init_bias=1.) # create Softmax output layer output_layer = mlp.Softmax(2, 'output', irange=.1) # create Stochastic Gradient Descent trainer that runs for 400 epochs trainer = sgd.SGD(learning_rate=.05, batch_size=10, termination_criterion=EpochCounter(400)) layers = [hidden_layer, output_layer] # create neural net that takes two inputs ann = mlp.MLP(layers, nvis=2) trainer.setup(ann, ds) # train neural net until the termination criterion is true while True: trainer.train(dataset=ds) ann.monitor.report_epoch() ann.monitor() if not trainer.continue_learning(ann): break inputs = np.array([[0, 0]]) print(ann.fprop(theanoShared(inputs, name='inputs')).eval())
def train(d=None): train_X = np.array(d.train_X) train_y = np.array(d.train_Y) valid_X = np.array(d.valid_X) valid_y = np.array(d.valid_Y) test_X = np.array(d.test_X) test_y = np.array(d.test_Y) nb_classes = len(np.unique(train_y)) train_y = convert_one_hot(train_y) valid_y = convert_one_hot(valid_y) # train_set = RotationalDDM(X=train_X, y=train_y) train_set = DenseDesignMatrix(X=train_X, y=train_y) valid_set = DenseDesignMatrix(X=valid_X, y=valid_y) print 'Setting up' batch_size = 100 c0 = mlp.ConvRectifiedLinear( layer_name='c0', output_channels=64, irange=.05, kernel_shape=[5, 5], pool_shape=[4, 4], pool_stride=[2, 2], # W_lr_scale=0.25, max_kernel_norm=1.9365) c1 = mlp.ConvRectifiedLinear( layer_name='c1', output_channels=64, irange=.05, kernel_shape=[5, 5], pool_shape=[4, 4], pool_stride=[2, 2], # W_lr_scale=0.25, max_kernel_norm=1.9365) c2 = mlp.ConvRectifiedLinear( layer_name='c2', output_channels=64, irange=.05, kernel_shape=[5, 5], pool_shape=[4, 4], pool_stride=[5, 4], W_lr_scale=0.25, # max_kernel_norm=1.9365 ) sp0 = mlp.SoftmaxPool( detector_layer_dim=16, layer_name='sp0', pool_size=4, sparse_init=512, ) sp1 = mlp.SoftmaxPool( detector_layer_dim=16, layer_name='sp1', pool_size=4, sparse_init=512, ) r0 = mlp.RectifiedLinear( layer_name='r0', dim=512, sparse_init=512, ) r1 = mlp.RectifiedLinear( layer_name='r1', dim=512, sparse_init=512, ) s0 = mlp.Sigmoid( layer_name='s0', dim=500, # max_col_norm=1.9365, sparse_init=15, ) out = mlp.Softmax( n_classes=nb_classes, layer_name='output', irange=.0, # max_col_norm=1.9365, # sparse_init=nb_classes, ) epochs = EpochCounter(100) layers = [s0, out] decay_coeffs = [.00005, .00005, .00005] in_space = Conv2DSpace( shape=[d.size, d.size], num_channels=1, ) vec_space = VectorSpace(d.size**2) nn = mlp.MLP( layers=layers, # input_space=in_space, nvis=d.size**2, # batch_size=batch_size, ) trainer = sgd.SGD( learning_rate=0.01, # cost=SumOfCosts(costs=[ # dropout.Dropout(), # MethodCost(method='cost_from_X'), # WeightDecay(decay_coeffs), # ]), # cost=MethodCost(method='cost_from_X'), batch_size=batch_size, # train_iteration_mode='even_shuffled_sequential', termination_criterion=epochs, # learning_rule=learning_rule.Momentum(init_momentum=0.5), ) trainer = bgd.BGD( batch_size=10000, line_search_mode='exhaustive', conjugate=1, updates_per_batch=10, termination_criterion=epochs, ) lr_adjustor = LinearDecayOverEpoch( start=1, saturate=10, decay_factor=.1, ) momentum_adjustor = learning_rule.MomentumAdjustor( final_momentum=.99, start=1, saturate=10, ) trainer.setup(nn, train_set) print 'Learning' test_X = vec_space.np_format_as(test_X, nn.get_input_space()) train_X = vec_space.np_format_as(train_X, nn.get_input_space()) i = 0 X = nn.get_input_space().make_theano_batch() Y = nn.fprop(X) predict = theano.function([X], Y) best = -40 best_iter = -1 while trainer.continue_learning(nn): print '--------------' print 'Training Epoch ' + str(i) trainer.train(dataset=train_set) nn.monitor() print 'Evaluating...' predictions = convert_categorical(predict(train_X[:2000])) score = accuracy_score(convert_categorical(train_y[:2000]), predictions) print 'Score on train: ' + str(score) predictions = convert_categorical(predict(test_X)) score = accuracy_score(test_y, predictions) print 'Score on test: ' + str(score) best, best_iter = (best, best_iter) if best > score else (score, i) print 'Current best: ' + str(best) + ' at iter ' + str(best_iter) print classification_report(test_y, predictions) print 'Adjusting parameters...' # momentum_adjustor.on_monitor(nn, valid_set, trainer) # lr_adjustor.on_monitor(nn, valid_set, trainer) i += 1 print ' '
irange=0.01, init_bias=0) hidden_layer3 = mlp.RectifiedLinear(layer_name='hidden3', dim=128, irange=0.01, init_bias=0) # create Softmax output layer output_layer = mlp.Softmax(3, 'output', irange=.1) # create Stochastic Gradient Descent trainer that runs for 400 epochs cost = NegativeLogLikelihoodCost() rule = Momentum(0.9) # rule = Momentum(0.9, True) # update_callbacks=ExponentialDecay(1 + 1e-5, 0.001) trainer = sgd.SGD(learning_rate=0.01, cost=cost, batch_size=128, termination_criterion=EpochCounter(1000), monitoring_dataset=vds, learning_rule=rule) layers = [hidden_layer, hidden_layer2, output_layer] # create neural net that takes two inputs ann = mlp.MLP(layers, nvis=ds.feat_cnt) trainer.setup(ann, ds) print trainer.cost # train neural net until the termination criterion is true iteration = 0 while True: trainer.train(dataset=ds) ann.monitor.report_epoch()
output = mlp.Softmax(layer_name='y', n_classes=10, irange=.005, max_col_norm=1.9365) layers = [l1, l2, l3, l4, output] mdl = mlp.MLP(layers, input_space=in_space) trainer = sgd.SGD(learning_rate=.17, batch_size=128, learning_rule=learning_rule.Momentum(.5), # Remember, default dropout is .5 cost=Dropout(input_include_probs={'l1': .8}, input_scales={'l1': 1.}), termination_criterion=EpochCounter(max_epochs=475), monitoring_dataset={'valid': tst, 'train': trn}) preprocessor = Pipeline([GlobalContrastNormalization(scale=55.), ZCA()]) trn.apply_preprocessor(preprocessor=preprocessor, can_fit=True) tst.apply_preprocessor(preprocessor=preprocessor, can_fit=False) serial.save('kaggle_cifar10_preprocessor.pkl', preprocessor) watcher = best_params.MonitorBasedSaveBest( channel_name='valid_y_misclass', save_path='kaggle_cifar10_maxout_zca.pkl') velocity = learning_rule.MomentumAdjustor(final_momentum=.65,
momentum_rule = learning_rule.Momentum(initial_momentum) # learning rate start = .1 saturate = 20 decay_factor = .00001 learning_rate_adjustor = sgd.LinearDecayOverEpoch(start, saturate, decay_factor) # termination criterion that stops after 50 epochs without # any increase in misclassification on the validation set termination_criterion = MonitorBased(channel_name='objective', N=20, prop_decrease=0.0) # create Stochastic Gradient Descent trainer trainer = sgd.SGD(learning_rate=.001, batch_size=10, monitoring_dataset=ds_valid, termination_criterion=termination_criterion, cost=L1_cost) #learning_rule=momentum_rule, trainer.setup(ann, ds_train) # add monitor for saving the model with best score monitor_save_best = best_params.MonitorBasedSaveBest('objective','./tmp/best.pkl') ##################################### #Train model #################################### # train neural net until the termination criterion is true while True:
from pylearn2.models import mlp from pylearn2.training_algorithms import sgd from pylearn2.termination_criteria import EpochCounter raw_ds = CLICK4DAY(which_set='train', which_day=21) transformer = Transformer(raw=raw_ds, nfeatures=1024, rng=None) ds = TransformerDataset(raw=raw_ds, transformer=transformer, cpu_only=False, \ space_preserving=False) hidden_layer = mlp.Sigmoid(layer_name='hidden', dim=256, irange=.1, init_bias=1.) output_layer = mlp.Softmax(2, 'output', irange=.1) trainer = sgd.SGD(learning_rate=.05, batch_size=1024, \ train_iteration_mode='even_sequential',termination_criterion=EpochCounter(400)) layers = [hidden_layer, output_layer] ann = mlp.MLP(layers, nvis=1024) trainer.setup(ann, ds) # train neural net until the termination criterion is true while True: trainer.train(dataset=ds) ann.monitor.report_epoch() ann.monitor() if not trainer.continue_learning(ann): break