Python Mlp.powerfit方法代码示例

# 需要导入模块: from breze.learn.mlp import Mlp [as 别名]
# 或者: from breze.learn.mlp.Mlp import powerfit [as 别名]
def run_mlp(n_job, pars):

    f = h5.File('../../../datasets/eigdata.hdf5', 'r')
    X = f['matrices'][...]
    Z = f['eigvals'][...]

    f = open('mlp_training_%d' %n_job, 'w')

    max_passes = 100
    batch_size = 2000
    max_iter = max_passes * X.shape[0] / batch_size
    n_report = X.shape[0] / batch_size

    stop = climin.stops.AfterNIterations(max_iter)
    pause = climin.stops.ModuloNIterations(n_report)

    m = Mlp(20000, pars['n_hidden'], 1, hidden_transfers=[pars['hidden_transfer']]*len(pars['n_hidden']), out_transfer='identity', loss='squared',
            optimizer=pars['optimizer'], batch_size=batch_size)
    climin.initialize.randomize_normal(m.parameters.data, 0, pars['par_std'])

    losses = []
    f.write('max iter: %d \n' %max_iter)

    weight_decay = ((m.parameters.in_to_hidden**2).sum()
                    + (m.parameters.hidden_to_out**2).sum())
    weight_decay /= m.exprs['inpt'].shape[0]
    m.exprs['true_loss'] = m.exprs['loss']
    c_wd = 0.001
    m.exprs['loss'] = m.exprs['loss'] + c_wd * weight_decay

    start = time.time()
    # Set up a nice printout.
    keys = '#', 'seconds', 'loss', 'val_loss'
    max_len = max(len(i) for i in keys)
    header = '\t'.join(i for i in keys)
    f.write(header + '\n')
    f.write(('-' * len(header)) + '\n')

    for i, info in enumerate(m.powerfit((X, Z), (X, Z), stop, pause)):
        if info['n_iter'] % n_report != 0:
            continue
        passed = time.time() - start
        losses.append((info['loss'], info['val_loss']))

        info.update({
            'time': passed})
        row = '%(n_iter)i\t%(time)g\t%(loss)g\t%(val_loss)g' % info
        f.write(row)

    f.write('best val_loss: %f \n' %info['best_loss'])
    f.close()

    cp.dump(info['best_pars'], open('best_pars_%d.pkl' %n_job, 'w'))

# 需要导入模块: from breze.learn.mlp import Mlp [as 别名]
# 或者: from breze.learn.mlp.Mlp import powerfit [as 别名]
weight_decay /= m.exprs['inpt'].shape[0]
m.exprs['true_loss'] = m.exprs['loss']
c_wd = 0.001
m.exprs['loss'] = m.exprs['loss'] + c_wd * weight_decay

n_wrong = 1 - T.eq(T.argmax(m.exprs['output'], axis=1), T.argmax(m.exprs['target'], axis=1)).mean()
f_n_wrong = m.function(['inpt', 'target'], n_wrong)

start = time.time()
# Set up a nice printout.
keys = '#', 'seconds', 'loss', 'val loss', 'train emp', 'val emp'
max_len = max(len(i) for i in keys)
header = '\t'.join(i for i in keys)
print header
print '-' * len(header)

for i, info in enumerate(m.powerfit((X, Z), (VX, VZ), stop, pause)):
    if info['n_iter'] % n_report != 0:
        continue
    passed = time.time() - start
    losses.append((info['loss'], info['val_loss']))

    #img = tile_raster_images(fe.parameters['in_to_hidden'].T, image_dims, feature_dims, (1, 1))
    #save_and_display(img, 'filters-%i.png' % i)
    info.update({
        'time': passed,
        'train_emp': f_n_wrong(X, Z),
        'val_emp': f_n_wrong(VX, VZ),
    })
    row = '%(n_iter)i\t%(time)g\t%(loss)g\t%(val_loss)g\t%(train_emp)g\t%(val_emp)g' % info
    print row

# 需要导入模块: from breze.learn.mlp import Mlp [as 别名]
# 或者: from breze.learn.mlp.Mlp import powerfit [as 别名]
def run_mlp(arch, func, step, batch, init, X, Z, VX, VZ, wd):

    max_passes = 200
    batch_size = batch
    max_iter = max_passes * X.shape[0] / batch_size
    n_report = X.shape[0] / batch_size

    input_size = len(X[0])

    stop = climin.stops.after_n_iterations(max_iter)
    pause = climin.stops.modulo_n_iterations(n_report)

    #optimizer = 'rmsprop', {'steprate': 0.0001, 'momentum': 0.95, 'decay': 0.8}
    optimizer = 'gd', {'steprate': step}

    m = Mlp(input_size, arch, 2, hidden_transfers=func, out_transfer='softmax', loss='cat_ce',
            optimizer=optimizer, batch_size=batch_size)
    climin.initialize.randomize_normal(m.parameters.data, 0, init)

    losses = []
    print 'max iter', max_iter

    weight_decay = ((m.parameters.in_to_hidden**2).sum()
                    + (m.parameters.hidden_to_out**2).sum()
                    + (m.parameters.hidden_to_hidden_0**2).sum())
    weight_decay /= m.exprs['inpt'].shape[0]
    m.exprs['true_loss'] = m.exprs['loss']
    c_wd = wd
    m.exprs['loss'] = m.exprs['loss'] + c_wd * weight_decay

    n_wrong = 1 - T.eq(T.argmax(m.exprs['output'], axis=1), T.argmax(m.exprs['target'], axis=1)).mean()
    f_n_wrong = m.function(['inpt', 'target'], n_wrong)

    start = time.time()
    # Set up a nice printout.
    keys = '#', 'seconds', 'loss', 'val loss', 'train emp', 'val emp'
    max_len = max(len(i) for i in keys)
    header = '\t'.join(i for i in keys)
    print header
    print '-' * len(header)
    results = open('results.txt','a')
    results.write(header + '\n')
    results.write('-' * len(header) + '\n')
    results.close()

    for i, info in enumerate(m.powerfit((X, Z), (VX, VZ), stop, pause)):
        if info['n_iter'] % n_report != 0:
            continue
        passed = time.time() - start
        losses.append((info['loss'], info['val_loss']))

        info.update({
            'time': passed,
            'train_emp': f_n_wrong(X, Z),
            'val_emp': f_n_wrong(VX, VZ),
        })

        row = '%(n_iter)i\t%(time)g\t%(loss)g\t%(val_loss)g\t%(train_emp)g\t%(val_emp)g' % info
        results = open('results.txt','a')
        print row
        results.write(row + '\n')
        results.close()

    m.parameters.data[...] = info['best_pars']
    cp.dump(info['best_pars'],open('best_%s_%s_%s_%s_%s.pkl' %(arch,func,step,batch,init),'w'))

# 需要导入模块: from breze.learn.mlp import Mlp [as 别名]
# 或者: from breze.learn.mlp.Mlp import powerfit [as 别名]
def do_one_eval(X, Z, TX, TZ, test_labels, train_labels, step_rate, momentum, decay, c_wd, counter, opt):
    seed = 3453
    np.random.seed(seed)
    max_passes = 200
    batch_size = 25
    max_iter = 5000000
    n_report = X.shape[0] / batch_size
    weights = []
    optimizer = 'gd', {'step_rate': step_rate, 'momentum': momentum, 'decay': decay}


    stop = climin.stops.AfterNIterations(max_iter)
    pause = climin.stops.ModuloNIterations(n_report)
    # This defines our NN. Since BayOpt does not support categorical data, we just
    # use a fixed hidden layer length and transfer functions.
    m = Mlp(2100, [400, 100], 1, X, Z, hidden_transfers=['tanh', 'tanh'], out_transfer='identity', loss='squared',
            optimizer=optimizer, batch_size=batch_size, max_iter=max_iter)

    #climin.initialize.randomize_normal(m.parameters.data, 0, 1e-3)

    # Transform the test data
    #TX = m.transformedData(TX)
    TX = np.array([m.transformedData(TX) for _ in range(10)]).mean(axis=0)
    losses = []
    print 'max iter', max_iter

    m.init_weights()

    for layer in m.mlp.layers:
        weights.append(m.parameters[layer.weights])


    weight_decay = ((weights[0]**2).sum()
                        + (weights[1]**2).sum()
                        + (weights[2]**2).sum())

    weight_decay /= m.exprs['inpt'].shape[0]
    m.exprs['true_loss'] = m.exprs['loss']
    c_wd = c_wd
    m.exprs['loss'] = m.exprs['loss'] + c_wd * weight_decay

    mae = T.abs_((m.exprs['output'] * np.std(train_labels) + np.mean(train_labels))- m.exprs['target']).mean()
    f_mae = m.function(['inpt', 'target'], mae)

    rmse = T.sqrt(T.square((m.exprs['output'] * np.std(train_labels) + np.mean(train_labels))- m.exprs['target']).mean())
    f_rmse = m.function(['inpt', 'target'], rmse)

    start = time.time()
    # Set up a nice printout.
    keys = '#', 'seconds', 'loss', 'val loss', 'mae_train', 'rmse_train', 'mae_test', 'rmse_test'
    max_len = max(len(i) for i in keys)
    header = '\t'.join(i for i in keys)
    print header
    print '-' * len(header)
    results = open('result.txt', 'a')
    results.write(header + '\n')
    results.write('-' * len(header) + '\n')
    results.write("%f %f %f %f %s" %(step_rate, momentum, decay, c_wd, opt))
    results.write('\n')
    results.close()

    EXP_DIR = os.getcwd()
    base_path = os.path.join(EXP_DIR, "pars_hp_"+opt+str(counter)+".pkl")
    n_iter = 0

    if os.path.isfile(base_path):
        with open("pars_hp_"+opt+str(counter)+".pkl", 'rb') as tp:
            n_iter, best_pars = dill.load(tp)
            m.parameters.data[...] = best_pars

    for i, info in enumerate(m.powerfit((X, Z), (TX, TZ), stop, pause)):
        if info['n_iter'] % n_report != 0:
            continue
        passed = time.time() - start
        if math.isnan(info['loss']) == True:
            info.update({'mae_test': f_mae(TX, test_labels)})
            n_iter = info['n_iter']
            break
        losses.append((info['loss'], info['val_loss']))
        info.update({
            'time': passed,
            'mae_train': f_mae(m.transformedData(X), train_labels),
            'rmse_train': f_rmse(m.transformedData(X), train_labels),
            'mae_test': f_mae(TX, test_labels),
            'rmse_test': f_rmse(TX, test_labels)

        })
        info['n_iter'] += n_iter
        row = '%(n_iter)i\t%(time)g\t%(loss)f\t%(val_loss)f\t%(mae_train)g\t%(rmse_train)g\t%(mae_test)g\t%(rmse_test)g' % info
        results = open('result.txt','a')
        print row
        results.write(row + '\n')
        results.close()
        with open("pars_hp_"+opt+str(counter)+".pkl", 'wb') as fp:
            dill.dump((info['n_iter'], info['best_pars']), fp)
        with open("apsis_pars_"+opt+str(counter)+".pkl", 'rb') as fp:
            LAss, opt, step_rate, momentum, decay, c_wd, counter, n_iter1, result1 = dill.load(fp)
        n_iter1 = info['n_iter']
        result1 = info['mae_test']
        with open("apsis_pars_"+opt+str(counter)+".pkl", 'wb') as fp:
#.........这里部分代码省略.........

# 需要导入模块: from breze.learn.mlp import Mlp [as 别名]
# 或者: from breze.learn.mlp.Mlp import powerfit [as 别名]
def run_mlp(arch, func, step, batch, X, Z, TX, TZ, wd, opt):
    batch_size = batch
    #max_iter = max_passes * X.shape[ 0] / batch_size
    max_iter = 100000
    n_report = X.shape[0] / batch_size
    weights = []
    input_size = len(X[0])
    train_labels = Z
    test_labels = TZ

    stop = climin.stops.AfterNIterations(max_iter)
    pause = climin.stops.ModuloNIterations(n_report)


    optimizer = opt, {'step_rate': step}

    typ = 'plain'
    if typ == 'plain':
        m = Mlp(input_size, arch, 1, X, Z, hidden_transfers=func, out_transfer='identity', loss='squared', optimizer=optimizer, batch_size=batch_size, max_iter=max_iter)

    elif typ == 'fd':
        m = FastDropoutNetwork(2099, [400, 100], 1, X, Z, TX, TZ,
                hidden_transfers=['tanh', 'tanh'], out_transfer='identity', loss='squared',
                p_dropout_inpt=.1,
                p_dropout_hiddens=.2,
                optimizer=optimizer, batch_size=batch_size, max_iter=max_iter)


    climin.initialize.randomize_normal(m.parameters.data, 0, 1 / np.sqrt(m.n_inpt))


    # Transform the test data
    #TX = m.transformedData(TX)
    TX = np.array([m.transformedData(TX) for _ in range(10)]).mean(axis=0)

    losses = []
    print 'max iter', max_iter

    m.init_weights()

    X, Z, TX, TZ = [breze.learn.base.cast_array_to_local_type(i) for i in (X, Z, TX, TZ)]

    for layer in m.mlp.layers:
        weights.append(m.parameters[layer.weights])


    weight_decay = ((weights[0]**2).sum()
                        + (weights[1]**2).sum()
                        + (weights[2]**2).sum()
			+ (weights[3]**2).sum()
			)


    weight_decay /= m.exprs['inpt'].shape[0]
    m.exprs['true_loss'] = m.exprs['loss']
    c_wd = wd
    m.exprs['loss'] = m.exprs['loss'] + c_wd * weight_decay


    '''
    weight_decay = ((m.parameters.in_to_hidden**2).sum()
                        + (m.parameters.hidden_to_out**2).sum()
                        + (m.parameters.hidden_to_hidden_0**2).sum())
    weight_decay /= m.exprs['inpt'].shape[0]
    m.exprs['true_loss'] = m.exprs['loss']
    c_wd = 0.1
    m.exprs['loss'] = m.exprs['loss'] + c_wd * weight_decay
    '''

    mae = T.abs_((m.exprs['output'] * np.std(train_labels) + np.mean(train_labels))- m.exprs['target']).mean()
    f_mae = m.function(['inpt', 'target'], mae)

    rmse = T.sqrt(T.square((m.exprs['output'] * np.std(train_labels) + np.mean(train_labels))- m.exprs['target']).mean())
    f_rmse = m.function(['inpt', 'target'], rmse)



    start = time.time()
    # Set up a nice printout.
    keys = '#', 'seconds', 'loss', 'val loss', 'mae_train', 'rmse_train', 'mae_test', 'rmse_test'
    max_len = max(len(i) for i in keys)
    header = '\t'.join(i for i in keys)
    print header
    print '-' * len(header)
    results = open('result.txt', 'a')
    results.write(header + '\n')
    results.write('-' * len(header) + '\n')
    results.close()



    for i, info in enumerate(m.powerfit((X, Z), (TX, TZ), stop, pause)):
        if info['n_iter'] % n_report != 0:
            continue
        passed = time.time() - start
        losses.append((info['loss'], info['val_loss']))
        info.update({
            'time': passed,
            'mae_train': f_mae(m.transformedData(X), train_labels),
            'rmse_train': f_rmse(m.transformedData(X), train_labels),
#.........这里部分代码省略.........

# 需要导入模块: from breze.learn.mlp import Mlp [as 别名]
# 或者: from breze.learn.mlp.Mlp import powerfit [as 别名]
def do_one_eval(X, Z, VX, VZ, step_rate, momentum, decay, c_wd):
    """
    Does one evaluation of a neural network with the above parameters.

    Parameters
    ----------
    X, Z : matrix
        Feature and Target matrices of the training set, one-hot encoded.
    VX, VZ : matrix
        Feature and Target matrices of the validation set, one-hot encoded.
    step_rate : float
        The step-rate/learning rate of the rmsprop-algorithm
    momentum : float
        The momentum of the rmsprop.
    decay : float
        The step-rate decay
    c_wd : float
        Penalty term for the weight

    Returns
    -------
    val_emp : float
        The percentage of wrongly classified samples.
    """

    max_passes = 100
    batch_size = 250
    max_iter = max_passes * X.shape[0] / batch_size
    n_report = X.shape[0] / batch_size
    optimizer = 'rmsprop', {'step_rate': step_rate, 'momentum': momentum, 'decay': decay}

    # This defines our NN. Since BayOpt does not support categorical data, we just
    # use a fixed hidden layer length and transfer functions.
    m = Mlp(784, [800], 10, hidden_transfers=['sigmoid'], out_transfer='softmax', loss='cat_ce',
            optimizer=optimizer, batch_size=batch_size)
    climin.initialize.randomize_normal(m.parameters.data, 0, 1e-1)
    losses = []
    weight_decay = ((m.parameters.in_to_hidden**2).sum()
                + (m.parameters.hidden_to_out**2).sum())
    weight_decay /= m.exprs['inpt'].shape[0]
    m.exprs['true_loss'] = m.exprs['loss']
    c_wd = c_wd
    m.exprs['loss'] = m.exprs['loss'] + c_wd * weight_decay
    n_wrong = 1 - T.eq(T.argmax(m.exprs['output'], axis=1), T.argmax(m.exprs['target'], axis=1)).mean()
    f_n_wrong = m.function(['inpt', 'target'], n_wrong)
    stop = climin.stops.AfterNIterations(max_iter)
    pause = climin.stops.ModuloNIterations(n_report)

    start = time.time()
    # Set up a nice printout.
    keys = '#', 'seconds', 'loss', 'val loss', 'train emp', 'val emp'
    max_len = max(len(i) for i in keys)
    header = '\t'.join(i for i in keys)
    #print header
    #print '-' * len(header)

    for i, info in enumerate(m.powerfit((X, Z), (VX, VZ), stop, pause)):
        passed = time.time() - start
        losses.append((info['loss'], info['val_loss']))

        #img = tile_raster_images(fe.parameters['in_to_hidden'].T, image_dims, feature_dims, (1, 1))
        #save_and_display(img, 'filters-%i.png' % i)
        info.update({
            'time': passed,
            'train_emp': f_n_wrong(X, Z),
            'val_emp': f_n_wrong(VX, VZ),
        })
        row = '%(n_iter)i\t%(time)g\t%(loss)g\t%(val_loss)g\t%(train_emp)g\t%(val_emp)g' % info

        # Comment in this row if you want updates during the computation.
        #print row
    return info["val_emp"]