def test_train():
numpy.random.seed(1)
training_data = synthetic_data(ROWS-30)
validation_data = synthetic_data(30)
testing_data = synthetic_data(10)
data = (training_data, validation_data, testing_data)
data = [(a, a.sum(axis=1)[:, numpy.newaxis]) for a in data]
(training_data, validation_data, testing_data) = data
(x, x_sums) = training_data
smh = SMH(
numpy_rng = numpy.random.RandomState(123),
mean_doc_size = x.sum(axis=1).mean(),
first_layer_type = 'bernoulli',
n_ins = x.shape[1],
mid_layer_sizes = [2],
inner_code_length = 1,
)
smh.train(training_data, validation_data, testing_data,
finetune_lr = 0.3, pretraining_epochs = 100, pretrain_lr = 0.01, training_epochs = 100,
method = 'cd', k = 1, noise_std_dev = 0, cost_method = 'squared_diff',
batch_size = 2, skip_trace_images=True, weights_file=None)
layers = smh.export_model()
for ((W,b),expectedW, expectedB) in zip(layers, previously_recorded_W, previously_recorded_b):
assert numpy.allclose(W.get_value(), expectedW)
assert numpy.allclose(b.get_value(), expectedB)
def load_model(cost_method, n_ins=784, mid_layer_sizes = [200],
inner_code_length = 10, weights_file='data/last_smh_model_params.pkl.gz'):
numpy_rng = numpy.random.RandomState(212)
smh = SMH(numpy_rng = numpy_rng, mid_layer_sizes = mid_layer_sizes, inner_code_length = inner_code_length, n_ins = n_ins)
smh.unroll_layers(cost_method,0) #need to unroll before loading params so that we have right number of layers
save_file=open(weights_file,'rb')
smh_params = cPickle.load(save_file)
save_file.close()
smh.load_model(smh_params)
return smh