def getUncertaintyRankings(model, data, forward=True):
embeds = []
for i in range(10):
bn = buildNetwork(seed=i,dropout=np.random.random(1)[0])
m = bn.build(data[0], data[-1])
m = bn.compiler(m,q_net=False,actor=False)
embeds.append(forwardPass(m, data[0], len(model.layers)-1))
stacked = np.stack(embeds, axis=-1)
variance = []
for i in range(len(stacked)):
variance.append(np.var(stacked[:,0:1][i][0]))
variance = pd.DataFrame(variance)
variance = variance.sort_values(0, ascending=forward)
return variance
def vanillaCurriculum(model, batches, outs, data, recursion, plot=True):
for curric in ('m', 'c', 'w'):
bn = buildNetwork()
model = bn.build(data[0], pd.DataFrame(uniques))
model = bn.compiler(model, q_net=False, actor=False)
training_loss = []
validation_loss = []
training_acc = []
validation_acc = []
embs = dict()
embs[0] = getEmbeddeds(model, batches)
for r in range(repetitions):
model, l, a, va_l, va_a = train(model,
batches,
outs,
data[2],
data[3],
curriculum=curric,
recurs=1000)
embs[r + 1] = getEmbeddeds(model, batches)
training_loss.append(l)
validation_loss.append(va_l)
training_acc.append(a)
validation_acc.append(va_a)
validation_acc = flattenList(validation_acc)
training_acc = flattenList(training_acc)
training_loss = flattenList(training_loss)
validation_loss = flattenList(validation_loss)
if plot:
plt.plot(validation_acc)
plt.ylabel("validation accuracy")
plt.xlabel("Iteration")
plt.legend(['m', 'c', 'w'])
plt.show()
plt.plot(validation_loss)
plt.ylabel("validation loss")
plt.xlabel("Iteration")
plt.legend(['m', 'c', 'w'])
plt.show()
curric = curric.sort_values(0, ascending=False)
else:
curric = curric.sort_values(0)
for batch in range(numbatches):
if cumulative == True:
ba[batch] = datas[0].iloc[0:(batch+1)*k]
else:
ba[batch] = datas[0].iloc[batch*k:(batch+1)*k]
ou = datas[1].iloc[ba[batch].index]
batches[i] = ba
outs[i] = ou
return batches, outs
data = loadData()
uniques = data[-1]
datas = dict()
for z in range(len(data)):
datas[z] = toDataframe(data[z])
ma, co, wa = getRankings(datas[0])
batches, outs = createBatches(datas, 10, False)
bn = buildNetwork()
model = bn.build(data[0],pd.DataFrame(uniques))
model = bn.compiler(model, q_net=False, actor=False)