class AttentionWeight:
def __init__(self):
self.params, self.grads = [], []
self.softmax = Softmax()
self.cache = None
def forward(self, hs, h):
N, T, H = hs.shape
hr = h.reshape(N, 1, H) # .repeat(T, axis=1)
t = hs * hr
s = np.sum(t, axis=2)
a = self.softmax.forward(s)
self.cache = (hs, hr)
return a
def backward(self, da):
hs, hr = self.cache
N, T, H = hs.shape
ds = self.softmax.backward(da)
dt = ds.reshape(N, T, 1).repeat(H, axis=2)
dhs = dt * hr
dhr = dt * hs
dh = np.sum(dhr, axis=1)
return dhs, dh
def infererence(args):
groups = 8
print('Loading image')
image = Image.open(args.image)
print('Preprocessing')
transformer = get_transformer()
input_data = preprocess(image, transformer)
print('input_data ', input_data.shape)
#conv layer
w = np.load('./data/' + 'module.conv1.weight.npy')
b = np.load('./data/' + 'module.conv1.bias.npy')
conv_layer = Convolution(w, b, stride=2, pad=1)
out = conv_layer.forward(input_data)
#savetxt('./dump/' + 'conv1_out.txt', out)
#max pooling
maxpool_layer = Pooling(3, 3, 2, 1)
out = maxpool_layer.forward(out)
#savetxt('./dump/' + 'maxpool_out.txt', out)
out = stage_shuffle(out, stage2_str, 3, groups)
#savetxt('./dump/' + 'stage2.txt', out)
out = stage_shuffle(out, stage3_str, 7, groups)
#savetxt('./dump/' + 'stage3.txt', out)
out = stage_shuffle(out, stage4_str, 3, groups)
#savetxt('./dump/' + 'stage4.txt', out)
h, w = out.shape[-2:]
avgpool_layer = AVGPooling(h, w, 1, 0)
out = avgpool_layer.forward(out).reshape(1, -1)
w = np.load('./data/' + 'module.fc.weight.npy')
b = np.load('./data/' + 'module.fc.bias.npy')
w = w.transpose(1, 0)
fc_layer = Affine(w, b)
out = fc_layer.forward(out)
softmax_layer = Softmax()
out = softmax_layer.forward(out).reshape(-1)
result = []
with open(args.idx_to_class) as json_file:
json_data = json.load(json_file)
'''
for key in json_data:
print(key, json_data[key])
'''
for i in range(0, 1000):
item = (out[i], json_data[str(i)])
result.append(item)
result = sorted(result, key=lambda item: item[0], reverse=True)
for i in range(0, 10):
print(result[i])
class AttentionWeight:
def __init__(self):
self.params, self.grads = [], []
self.softmax = Softmax()
self.cache = None
def forward(self, hs, h):
"""順伝搬
Args:
hs (ndarray): Encorderで取得した各単語ベクトルを連結した隠れベクトル
h (ndarray): hsの最終単語に対応するベクトル
Returns:
a (ndarray): Attention用重みベクトル
"""
N, T, H = hs.shape
hr = h.reshape(N, 1, H).repeat(T, axis=1)
t = hs * hr
s = np.sum(t, axis=2) # スコア(重み付き和)
a = self.softmax.forward(s)
self.cache = (hs, hr)
return a
def backward(self, da):
"""逆伝搬
Args:
da (ndarray): Attention用重みベクトルの誤差微分
Returns:
dhs: Encorderで取得した各単語ベクトルを連結した隠れベクトルの誤差微分
dh: hsの最終単語に対応するベクトルの誤差微分
"""
hs, hr = self.cache
N, T, H = hs.shape
ds = self.softmax.backward(da)
dt = ds.reshape(N, T, 1).repeat(H, axis=2) # Sumの逆伝搬はRepeat
dhs = dt * hr
dhr = dt * hs
dh = np.sum(dhr, axis=1) # Repeatの逆伝搬はSum
return dhs, dh
def __init__(self):
self.params, self.grads = [], []
self.softmax = Softmax()
self.cache = None
import sys
sys.path.append('..')
from common.layers import Softmax
import numpy as np
N, T, H = 10, 5, 4
hs = np.random.randn(N, T, H)
h = np.random.randn(N, H)
hr = h.reshape(N, 1, H).repeat(T, axis=1)
t = hs * hr
print(t.shape)
# (10, 5, 4)
s = np.sum(t, axis=2)
print(s.shape)
# (10, 5)
softmax = Softmax()
a = softmax.forward(s)
print(a.shape)
# (10, 5)
# 각 단어의 중요도를 나타내는 가중치 a
# 이 가중치 a의 합(가중합)을 이용해 맥박 벡터를 얻을 수 있다
# Decoder의 LSTM 계층의 은닉 상태 벡터를 h라고 했을 때,
# 지금 목표는 h가 hs의 각 단어 벡터와 얼마나 비슷한가를 수치로 나타내는 것
# 여기서는 가장 단순한 방법인 벡터의 내적을 이용한다 - 두 벡터가 얼마나 같은 방향을 향하고 있는가, 두 벡터의 유사도
# 유사도를 산출 후 소프트맥스 함수를 적용하여 정규화
class AttentionWeight:
'''
Encoderの全系列の隠れ状態hs(N, T, H)と
Decoderの現系列の隠れ状態h(N, H)とのドット積をとり、
softmax関数にかけることで系列ごとのアライメントa(N, T)を
出力するレイヤ
'''
def __init__(self):
self.params, self.grads = [], []
self.softmax = Softmax()
self.cache = None
def forward(self, hs, h):
'''
Decoderの隠れ状態h(N, H)をnp.repeatで(N, T, H)に拡張し、
hsとのアダマール積を取ってHについて総和を取り、
Softmax関数で正規化してアライメントa(N, T)を得る
Parameters
----------
hs : np.ndarray(N, T, H)
Encoderの全系列の隠れ状態
h : np.ndarray(N, H)
Decoderの現系列の隠れ状態
Returns
-------
np.ndarray(N, T)
hsに対し、系列ごとの重みを示すアライメント
'''
N, T, H = hs.shape
hr = h.reshape(N, 1, H)#.repeat(T, axis=1)
t = hs * hr # (N, T, H)
s = t.sum(axis=2)
a = self.softmax.forward(s) # (N, T)
self.cache = (hs, hr)
return a
def backward(self, da):
'''
sumの逆伝播はrepeat
repeatの逆伝播はsum
Parameters
----------
da : np.ndarray(N, T)
アライメントの勾配
Returns
-------
dhs, dh : np.ndarray(N, T, H), np.ndarray(N, H)
全系列の隠れ状態hsの勾配と系列の隠れ状態hの勾配
'''
hs, hr = self.cache
N, T, H = hs.shape
ds = self.softmax.backward(da)
dt = ds.reshape(N, T, 1).repeat(H, axis=2)
dhs = dt * hr # (N, T, H)
dhr = dt * hs # (N, T, H)
dh = dhr.sum(axis=1) # (N, H)
return dhs, dh
def __init__(self, name, nout, numpy_rng, theano_rng, batchsize=128):
# CALL PARENT CONSTRUCTOR TO SETUP CONVENIENCE FUNCTIONS
# (SAVE/LOAD, ...)
super(HumanConvNet, self).__init__(name=name)
self.numpy_rng = numpy_rng
self.batchsize = batchsize
self.theano_rng = theano_rng
self.mode = theano.shared(np.int8(0), name='mode')
self.nout = nout
self.inputs = T.ftensor4('inputs')
self.inputs.tag.test_value = numpy_rng.randn(self.batchsize, 1, 28,
28).astype(np.float32)
self.targets = T.ivector('targets')
self.targets.tag.test_value = numpy_rng.randint(
nout, size=self.batchsize).astype(np.int32)
self.layers = OrderedDict()
self.layers['conv0'] = ConvLayer(rng=self.numpy_rng,
inputs=self.inputs,
filter_shape=(128, 1, 5, 5),
image_shape=(None, 1, 28, 28),
name='conv0',
pad=2)
self.layers['maxpool0'] = MaxPoolLayer(inputs=self.layers['conv0'],
pool_size=(2, 2),
stride=(2, 2),
name='maxpool0')
self.layers['bias0'] = ConvBiasLayer(inputs=self.layers['maxpool0'],
name='bias0')
self.layers['relu0'] = Relu(inputs=self.layers['bias0'], name='relu0')
self.layers['conv1'] = ConvLayer(rng=self.numpy_rng,
inputs=self.layers['relu0'],
filter_shape=(64, 128, 3, 3),
name='conv1',
pad=1)
self.layers['maxpool1'] = MaxPoolLayer(inputs=self.layers['conv1'],
pool_size=(2, 2),
stride=(2, 2),
name='maxpool1')
self.layers['bias1'] = ConvBiasLayer(inputs=self.layers['maxpool1'],
name='bias1')
self.layers['relu1'] = Relu(inputs=self.layers['bias1'], name='relu1')
self.layers['reshape1'] = Reshape(
inputs=self.layers['relu1'],
shape=(self.layers['relu1'].outputs_shape[0],
np.prod(self.layers['relu1'].outputs_shape[1:])),
name='reshape1')
self.layers['fc2'] = AffineLayer(rng=self.numpy_rng,
inputs=self.layers['reshape1'],
nouts=256,
name='fc2')
self.layers['relu2'] = Relu(inputs=self.layers['fc2'], name='relu2')
self.layers['fc3'] = AffineLayer(rng=self.numpy_rng,
inputs=self.layers['relu2'],
nouts=self.nout,
name='fc3')
self.layers['softmax3'] = Softmax(inputs=self.layers['fc3'],
name='softmax3')
self.layers['clip3'] = Clip(inputs=self.layers['softmax3'],
name='clip3',
min_val=1e-6,
max_val=1 - 1e-6)
self.probabilities = self.forward(self.inputs)
self._cost = T.nnet.categorical_crossentropy(self.probabilities,
self.targets).mean()
self.classification = T.argmax(self.probabilities, axis=1)
self.params = []
for l in self.layers.values():
self.params.extend(l.params)
self._grads = T.grad(self._cost, self.params)
self.classify = theano.function(
[self.inputs],
self.classification,
)