def test_max_delta2():
maxlayer = MaxPoolingLayer("maxpool1", 2, 2)
inputlayer = InputLayer("input", 0)
data = np.array(range(1, 26)).reshape((-1, 5, 5))
data[0, 1, 2] = 1000
inputlayer.feed(data)
input_delta = np.zeros(data.shape)
delta = np.array(range(1, 10)).reshape((-1, 3, 3))
fakeOutput = InputLayer("output", 0)
maxlayer.set_input_layer(inputlayer)
maxlayer.set_next_layer(fakeOutput)
maxlayer.init()
maxlayer.active()
maxlayer.delta = delta
maxlayer.calc_input_delta(input_delta)
print(maxlayer)
print("\ninput data:")
print(data)
print("\nmax pooling result:")
print(maxlayer.get_output())
print("\norigin pooling result:")
print(maxlayer.output)
print("\nindex-keeper:")
print(maxlayer.input_keeper)
print("\ninput_delta")
print(input_delta)
return
def test_max_active2():
maxlayer = MaxPoolingLayer("maxpool1", 2, 2)
inputlayer = InputLayer("input", 0)
data = np.array(range(1, 37)).reshape((-1, 6, 6))
data[0, 1, 2] = 1000
inputlayer.feed(data)
fakeOutput = InputLayer("output", 0)
maxlayer.set_input_layer(inputlayer)
maxlayer.set_next_layer(fakeOutput)
maxlayer.init()
maxlayer.active()
print(maxlayer)
print("\ninput data:")
print(data)
print("\nmax pooling result:")
print(maxlayer.get_output())
print("\norigin pooling result:")
print(maxlayer.output)
print("\nindex-keeper:")
print(maxlayer.input_keeper)
return
def test1():
m = 3
n = 5
y = get_random_vectors(m, n)
nn = NNetwork()
myin = InputLayer("input", 1)
emb = EmbeddingLayer("emb1", m, n)
myout = FakeOutputLayer("output", n)
nn.set_input(myin)
nn.set_output(myout)
nn.add_hidden_layer(emb)
nn.connect_layers()
nn.set_log_interval(1000)
print(emb.weights)
print("*"*40)
data = get_train_data(m)
for i in range(500):
train_it(nn, y, data)
if i % 10 == 0:
evaluate_it(nn, y, data, "epoch-%s" % i)
print(y)
print("*"*40)
print(emb.weights)
return
def construct_nn(l2=0.0):
seq_input = InputLayer("word sequence input", -1)
seq_output = SoftmaxOutputLayer("word predict", VOC_SIZE)
# 1. set input and output layers
nn = NNetwork()
nn.set_input(seq_input)
nn.set_output(seq_output)
#2. set embedding layer
emb = EmbeddingLayer("embedding", VOC_SIZE, WORD_DIM)
emb.set_lambda2(l2)
nn.add_hidden_layer(emb)
#3. set RNN layer
rnn = RNNLayer("rnn1", RNN_HIDDEN_DIM, MAX_BPTT_STEPS)
rnn.set_lambda2(l2)
nn.add_hidden_layer(rnn)
#4. add another RNN layer
#rnn2 = RNNLayer("rnn2", RNN_HIDDEN_DIM2, MAX_BPTT_STEPS)
#rnn2.set_lambda2(l2)
#nn.add_hidden_layer(rnn2)
#5. complete the nerual network
nn.connect_layers()
logging.info("NN information:\n" + nn.get_detail())
return nn
def test_normal_active():
img_input = InputLayer("fake_input", 784)
h1 = HiddenLayer("h1", 784, activation.tanhFunc)
h1.set_input_layer(img_input)
h1.init()
input_data = np.random.uniform(-1, 1, 784)
img_input.feed(input_data)
begin = datetime.now()
print("[%s] begin to active." % (str(begin)))
for i in range(1000):
h1.active()
end = datetime.now()
delta = end - begin
print("[%s] end of activation, delta=%s." % (str(end), str(delta)))
return
def test_active():
img_input = InputLayer("fake_input", 784)
conv = ConvLayer("conv1")
conv.set_kernels(get_kernels())
conv.set_input_layer(img_input)
input_data = np.random.uniform(-1, 1, 784)
input_data = input_data.reshape((1, 28, 28))
img_input.feed(input_data)
conv.init()
begin = datetime.now()
print("[%s] begin to active." % (str(begin)))
for i in range(1000):
conv.active()
end = datetime.now()
delta = end - begin
print("[%s] end of activation, delta=%s." % (str(end), str(delta)))
return