def test():
learner = OMTL(epoch=640, matrix_interaction_lr=1e-10, divergence="logdet")
kpercepts = KPerceptrons(learn_rate=10000)
knb = KNaiveBayes()
# Will spit out 20 batches of size 8000
for task, inp, outp in get_minibatches(batch_size=8000, num_epochs=1, add_bias=True):
knb.fit(np.asarray(task).reshape((inp.shape[0],1)), inp, np.asarray(outp).reshape((inp.shape[0],1)))
# Will spit out 20*3000 batches of size 1
for task, inp, outp in get_minibatches(batch_size=1, num_epochs=3000, add_bias=True):
learner.fit(task[0].astype(int), inp, outp)
kpercepts.fit(task[0].astype(int), inp, outp)
# Test_tasks.shape = (28143, 1)
test_tasks = np.load("data/test_tasks.npy")
# We need to call item on test_inputs because it is an array that contains a sparse matrix
# We then gotta convert it from sparse scipymatrix to nparray and add a bias
test_inputs = np.load("data/test_inputs.npy").item().toarray()
test_inputs = np.hstack((np.ones((test_inputs.shape[0],1)), test_inputs))
# Test_outputs.shape = (28143,)
# So we need to reshape it to a (28143, 1) array
test_outputs = np.load("data/test_outputs.npy").reshape((test_tasks.shape[0],1))
# See KNB.score to see how I manipulate these three arrays so that they fit with normal
# SKlearn learners
print "KNB: %f" %knb.score(test_tasks, test_inputs, test_outputs)
print "OMTL: %f" % learner.score(test_tasks, test_inputs, test_outputs)
print "kpercepts: %f" % kpercepts.score(test_tasks, test_inputs, test_outputs)
print json.dumps(learner.A.tolist())
with open("task_relate.json","w") as f:
json.dump(learner.A.tolist(), f)
def grid_search():
out = open("omtl_log", "w")
val_tasks, val_x, val_y = get_validation_set()
omtl_lrs = np.logspace(-28,-14,8)
perceptron_lrs = np.logspace(-10,4,8)
epochs = [20, 40, 80, 160, 320, 640, 1280]
start = time.time()
kpercept_results = []
omtl_results = []
for cnt, (omtl_lr, perceptron_lr) in enumerate(zip(omtl_lrs, perceptron_lrs)):
omtls = [OMTL(epoch=epoch, matrix_interaction_lr=omtl_lr, divergence="logdet") for epoch in epochs]
kpercepts = KPerceptrons(learn_rate=perceptron_lr)
for batch, (task, inp, outp) in enumerate(get_minibatches(
batch_size=1, num_epochs=5000, add_bias=True)):
[learner.fit(task[0].astype(int), inp, outp) for learner in omtls]
kpercepts.fit(task[0].astype(int), inp, outp)
if batch%500 ==0:
print "Batch %d" % batch
kpercept_results.append((perceptron_lr, kpercepts.score(val_tasks, val_x, val_y)))
for omtl in omtls:
if omtl.has_diverged:
omtl_results.append((omtl_lr, omtl.epoch, 0.0))
else:
omtl_results.append((omtl_lr, omtl.epoch, omtl.score(val_tasks, val_x, val_y)))
print "Done with iteration %d of %d" %(cnt, len(omtl_lrs))
json.dump({"omtl":omtl_results, "kpercepts":kpercept_results}, out)
out.close()
def get_validation_set(batch_size=1500): # Returns 3 arrays (tasks, X, y) with batch_size*20 examples in each val_tasks, val_x, val_y = np.ndarray(0), None, np.ndarray(0) for task, inp, outp in get_minibatches(batch_size=batch_size, num_epochs=1, add_bias=True): val_tasks = np.hstack((val_tasks, task.ravel())) val_y = np.hstack((val_y, outp.ravel())) if val_x is None: val_x = inp else: val_x = np.vstack((val_x, inp)) return val_tasks, val_x, val_y
def timeline():
# Generates the data file to be used to create the examples vs accuracy graph
val_tasks, val_x, val_y = get_validation_set()
learner = OMTL(epoch=40, matrix_interaction_lr=1e-20, divergence="logdet")
kpercepts = KPerceptrons(learn_rate=10000)
knb = KNaiveBayes()
results = []
tasks, x, y = [], None, []
for batch, (task, inp, outp) in enumerate(get_minibatches(
batch_size=1, num_epochs=8000, add_bias=True)):
learner.fit(task[0].astype(int), inp, outp)
kpercepts.fit(task[0].astype(int), inp, outp)
# Used so we only call KNB.fit once per 100 examples
tasks.append(int(task[0][0]))
y.append(outp[0].astype(int))
if x is None:
x = inp
else:
x = np.vstack((x,inp))
if batch%100 ==0 and batch>50:
knb.fit(np.asarray(tasks).reshape((x.shape[0],1)), x, np.asarray(y).reshape((x.shape[0],1)))
omtl_score = learner.score(val_tasks, val_x, val_y)
kpercept_score = kpercepts.score(val_tasks, val_x, val_y)
knb_score = knb.score(val_tasks, val_x, val_y)
print "Batch %d" % batch
print "OMTL score: %f" % omtl_score
print "KPercept score %f" % kpercept_score
print "KNB score %f" % knb_score
results.append((batch, omtl_score, kpercept_score, knb_score))
tasks, x, y = [], None, []
with open("timeline.json", "w") as f:
json.dump(results, f)
