def test_MLBL_implementation():
""" Test Your Implementation of Forward and Backward """
import coco_proc
import trainer
import cPickle as pickle
z, zd, zt = coco_proc.process(context=5)
d = {}
d['name'] = 'testrun'
d['loc'] = 'models/mlbl_model.pkl'
d['context'] = 5
d['learning_rate'] = 0.43
d['momentum'] = 0.23
d['batch_size'] = 40
d['maxepoch'] = 10
d['hidden_size'] = 441
d['word_decay'] = 3e-7
d['context_decay'] = 1e-8
d['factors'] = 50
# Load the word embeddings
embed_map = trainer.load_embeddings()
# Unpack some stuff from the data
train_ngrams = z['ngrams']
train_labels = z['labels']
train_instances = z['instances']
word_dict = z['word_dict']
index_dict = z['index_dict']
context = z['context']
vocabsize = len(z['word_dict'])
trainIM = z['IM']
train_index = z['index']
net = MLBL(name=d['name'],
loc=d['loc'],
seed=1234,
V=vocabsize,
K=100,
D=trainIM.shape[1],
h=d['hidden_size'],
context=d['context'],
batchsize=1,
maxepoch=d['maxepoch'],
eta_t=d['learning_rate'],
gamma_r=d['word_decay'],
gamma_c=d['context_decay'],
f=0.99,
p_i=d['momentum'],
p_f=d['momentum'],
T=20.0,
verbose=1)
# Train the network
X = train_instances
indX = train_index
Y = train_labels
net.init_params(embed_map, index_dict)
context_size = d['context']
batchX = X[0:context_size]
batchY = Y[0:context_size].toarray()
batchindX = indX[0:context_size].astype(int).flatten()
batchindX = np.floor(batchindX / 5).astype(int)
batchIm = trainIM[batchindX]
# check forward implementation
ft = net.forward(net.params, batchX, batchIm, test=True)
# load gt feature
ft_gt = pickle.load(open("data/val_implementation.p", "rb"))
# it should be less than 1.0e-5
print 'Difference (L2 norm) between implemented and ground truth feature = {}'.format(
np.linalg.norm(ft_gt - ft))
import coco_proc, trainer
if __name__ == '__main__':
z, zd, zt = coco_proc.process(context=5)
trainer.trainer(z, zd)