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)
#serial.save(DATA_DIR+'cae6_005_pretrained.pkl', stack)
# construct DBN
dbn = construct_dbn_from_stack(stack)
# train DBN
if submission:
traindata = sup_data[2]
validdata = sup_data[2]
else:
traindata = sup_data[0]
validdata = sup_data[1]
cost = dropout.Dropout(input_include_probs={'h0': 0.5},
input_scales={'h0': 1. / 0.5},
default_input_include_prob=0.5,
default_input_scale=1. / 0.5)
# finetune softmax layer a bit
finetuner = get_finetuner(dbn,
cost,
traindata,
validdata,
batch_size,
iters=100)
finetuner.main_loop()
# now finetune layer-by-layer
lrs = [5., 2., 1., 0.5, 0.25]
for ii, lr in zip(range(len(structure) - 1), lrs):
# set lr to boosted value for current layer
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()
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 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()
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