Пример #1
0
    def test_one_node_updown(self):
        """Test flow updown phase with just one node."""
        flow = dbn_binodes.get_DBN_flow(1, hidden_dims=[2])
        N = 10000
        x = mdp.numx.zeros((N, 4))
        for i in range(N):
            r = mdp.numx.rand()
            if r>0.666:
                x[i,:] = [0.,1.,0.,1.]
            elif r>0.333:
                x[i,:] = [1.,0.,1.,0.]

        def data_gen():
            for _ in range(100):
                # (data, learning rate, decay, momentum)
                yield (x, 0.5, 0., 0.5)

        flow.train([data_gen()])
        print flow[0].w_gen

        flow.updown_phase([x], epsilon=0.5,
                          decay=0., momentum=0.5, n_updates=3,
                          max_iter=300, min_error=1e-2)
        print flow[0].w_gen
        
        error = flow.updown_phase([x], epsilon=0.1,
                                  decay=0., momentum=0.9, n_updates=3,
                                  max_iter=1000, min_error=0.05)

        assert error < 0.05
Пример #2
0
    def test_one_node_updown(self):
        """Test flow updown phase with just one node."""
        flow = dbn_binodes.get_DBN_flow(1, hidden_dims=[2])
        N = 10000
        x = mdp.numx.zeros((N, 4))
        for i in range(N):
            r = mdp.numx.rand()
            if r > 0.666:
                x[i, :] = [0., 1., 0., 1.]
            elif r > 0.333:
                x[i, :] = [1., 0., 1., 0.]

        def data_gen():
            for _ in range(100):
                # (data, learning rate, decay, momentum)
                yield (x, 0.5, 0., 0.5)

        flow.train([data_gen()])
        print flow[0].w_gen

        flow.updown_phase([x],
                          epsilon=0.5,
                          decay=0.,
                          momentum=0.5,
                          n_updates=3,
                          max_iter=300,
                          min_error=1e-2)
        print flow[0].w_gen

        error = flow.updown_phase([x],
                                  epsilon=0.1,
                                  decay=0.,
                                  momentum=0.9,
                                  n_updates=3,
                                  max_iter=1000,
                                  min_error=0.05)

        assert error < 0.05
Пример #3
0
# -*- coding: utf-8 -*-
# Generated by codesnippet sphinx extension on 2012-10-04

import mdp
import numpy as np
np.random.seed(0)
import bimdp
import dbn_binodes

n_layers = 2
flow = dbn_binodes.get_DBN_flow(2, hidden_dims=[2, 2])

n_samples = 10000  # number of data points
n_greedy_reps = 100  # repetitions in greedy phase
x = mdp.numx.zeros((n_samples, 4))
for i in range(n_samples):
    r = mdp.numx.rand()
    if r > 0.666:
        x[i, :] = [0., 1., 0., 1.]
    elif r > 0.333:
        x[i, :] = [1., 0., 1., 0.]

data_iterables = [None] + [[x] * n_greedy_reps] * n_layers + [[x]]
msg_iterables = ([None] + [[{
    "epsilon": 0.1,
    "decay": 0.0,
    "momentum": 0.0
}] * n_greedy_reps] * n_layers + [[{
    "top_updates": 3,
    "epsilon": 0.1,
    "decay": 0.0,
Пример #4
0
# -*- coding: utf-8 -*-
# Generated by codesnippet sphinx extension on 2012-10-04

import mdp
import numpy as np
np.random.seed(0)
import bimdp
import dbn_binodes

n_layers = 2
flow = dbn_binodes.get_DBN_flow(2, hidden_dims=[2,2])

n_samples = 10000  # number of data points
n_greedy_reps = 100  # repetitions in greedy phase
x = mdp.numx.zeros((n_samples, 4))
for i in range(n_samples):
    r = mdp.numx.rand()
    if r>0.666:
        x[i,:] = [0.,1.,0.,1.]
    elif r>0.333:
        x[i,:] = [1.,0.,1.,0.]

data_iterables = [None] + [[x] * n_greedy_reps] * n_layers + [[x]]
msg_iterables = ([None] +
                 [[{"epsilon": 0.1, "decay": 0.0,
                    "momentum": 0.0}] * n_greedy_reps] * n_layers +
                 [[{"top_updates": 3, "epsilon": 0.1, "decay": 0.0,
                    "momentum": 0.0,
                    "max_iter": 2, "min_error": -1.0}]])

bimdp.show_training(flow, data_iterables, msg_iterables, debug=True)