def __call__(self, x, train=True):
h1 = F.relu(self.conv1_1(x))
h2 = F.max_pooling_2d(h1, 2, stride=2)
h3 = F.relu(self.conv2_1(h2))
h4 = F.max_pooling_2d(h3, 2, stride=2)
h5 = F.relu(self.conv3_1(h4))
h6 = F.relu(self.conv3_2(h5))
h7 = F.max_pooling_2d(h6, 2, stride=2)
h8 = F.relu(self.conv4_1(h7))
h9 = F.relu(self.conv4_2(h8))
h10 = F.max_pooling_2d(h9, 2, stride=2)
h11 = F.relu(self.conv5_1(h10))
h12 = F.relu(self.conv5_2(h11))
h13 = F.max_pooling_2d(h12, 2, stride=2)
h14 = F.dropout(F.relu(self.fc6(h13)), train=train, ratio=0.5)
h15 = F.dropout(F.relu(self.fc7(h14)), train=train, ratio=0.5)
h16 = self.fc8(h15)
y = F.sigmoid(h16)
return y
def __call__(self, x):
"""Compute actions for given observations."""
h = F.relu(self.l0(x))
h = F.relu(self.l1(h))
return squash(self.l2(h),
self.xp.asarray(self.action_low),
self.xp.asarray(self.action_high))
def __call__(self, x):
h1 = F.relu(self.bn1(self.conv1(x)))
h1 = F.relu(self.bn2(self.conv2(h1)))
h1 = self.bn3(self.conv3(h1))
h2 = self.bn4(self.conv4(x))
return F.relu(h1 + h2)
def call_inception(self, x, name):
out1 = self[name + '/1x1'](x)
out3 = self[name + '/3x3'](F.relu(self[name + '/3x3_reduce'](x)))
out5 = self[name + '/5x5'](F.relu(self[name + '/5x5_reduce'](x)))
pool = self[name + '/pool_proj'](F.max_pooling_2d(x, 3, stride=1, pad=1))
y = F.relu(F.concat((out1, out3, out5, pool), axis=1))
return y
def __call__(self, z, test=False):
h = F.reshape(F.relu(self.bn0l(self.l0z(z), test=test)), (z.data.shape[0], 512, 6, 6))
h = F.relu(self.bn1(self.dc1(h), test=test))
h = F.relu(self.bn2(self.dc2(h), test=test))
h = F.relu(self.bn3(self.dc3(h), test=test))
x = self.dc4(h)
return x
def __call__(self, x, train=True):
self.train = train
h = F.max_pooling_2d( F.relu(self.bn1(self.conv1(x), test=not self.train)), 3, stride=2 )
h = F.max_pooling_2d( F.relu(self.bn2(self.conv2(h), test=not self.train)), 3, stride=2 )
h = F.max_pooling_2d( F.relu(self.bn3(self.conv3(h), test=not self.train)), 3, stride=2 )
h = self.fc1(h)
return h
def __call__(self, x):
x.data = x.data / np.linalg.norm(x.data)
h1 = F.relu(self.l1(x))
h2 = F.relu(self.l2(h1))
h3 = F.relu(self.l3(h2))
return h3;
def forward(x_data, y_data, train=True):
""" 順伝搬の処理を定義 """
# train ... 学習フラグ（Falseにすると学習しない）
# 訓練時は，dropoutを実行
# テスト時は，dropoutを無効
# 入力と教師データ
# データ配列は，Chainer.Variable型にしないといけない
x, t = chainer.Variable(x_data), chainer.Variable(y_data)
# 隠れ層1の出力
h1 = F.dropout(F.relu(model.l1(x)), train=train)
# 隠れ層2の出力
h2 = F.dropout(F.relu(model.l2(h1)), train=train)
# 出力層の出力
y = model.l3(h2)
# 訓練時とテスト時で返す値を変える
# y ... ネットワークの出力（仮説）
# t ... 教師データ
if train: # 訓練
# 誤差を返す
# 多クラス分類なので，誤差関数としてソフトマックス関数の
# クロスエントロピー関数を使う
loss = F.softmax_cross_entropy(y, t)
return loss
else: # テスト
# 精度を返す
acc = F.accuracy(y, t)
return acc
def __call__(self, x, im_info):
h, n = self.trunk(x), x.data.shape[0]
rpn_cls_score = self.rpn_cls_score(h)
c, hh, ww = rpn_cls_score.data.shape[1:]
rpn_bbox_pred = self.rpn_bbox_pred(h)
rpn_cls_score = F.reshape(rpn_cls_score, (n, 2, -1))
# RoI Proposal
rpn_cls_prob = F.softmax(rpn_cls_score)
rpn_cls_prob_reshape = F.reshape(rpn_cls_prob, (n, c, hh, ww))
rois = self.proposal_layer(
rpn_cls_prob_reshape, rpn_bbox_pred, im_info, self.train)
boxes = rois[:, 1:5] / im_info[0][2]
rois = chainer.Variable(rois, volatile=not self.train)
# RCNN
pool5 = F.roi_pooling_2d(self.trunk.relu5_3_out, rois, 7, 7, 0.0625)
fc6 = F.relu(self.fc6(pool5))
fc7 = F.relu(self.fc7(fc6))
self.scores = F.softmax(self.cls_score(fc7))
box_deltas = self.bbox_pred(fc7).data
pred_boxes = bbox_transform_inv(boxes, box_deltas)
self.pred_boxes = clip_boxes(pred_boxes, im_info[0][:2])
if self.train:
# loss_cls = F.softmax_cross_entropy(cls_score, labels)
# huber loss with delta=1 means SmoothL1Loss
return None
else:
return self.scores, self.pred_boxes
def extract_features(self, x, train=True):
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
self.h1 = h
h = F.max_pooling_2d(h,2)
h = F.relu(self.conv3(h))
h = F.relu(self.conv4(h))
self.h2 = h
h = F.max_pooling_2d(h,2)
h = F.relu(self.conv5(h))
h = F.relu(self.conv6(h))
self.h3 = h
h = F.max_pooling_2d(h,2)
h = F.relu(self.conv7(h))
h = F.relu(self.conv8(h))
self.h4 = h
h = F.max_pooling_2d(h,2)
h = F.relu(self.conv9(h))
h = F.relu(self.conv10(h))
self.h5 = h
h = F.max_pooling_2d(h,2)
return h
def __call__(self, x, train):
h1 = F.relu(self.bn1(self.conv1(x), test=not train))
h1 = F.relu(self.bn2(self.conv2(h1), test=not train))
h1 = self.bn3(self.conv3(h1), test=not train)
h2 = self.bn4(self.conv4(x), test=not train)
return F.relu(h1 + h2)
def forward_super(self, x, train=True):
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
y = self.conv3(h)
return y
def forward(x, is_train=True):
h1 = F.max_pooling_2d(F.relu(model.conv1(x)), 3)
h2 = F.max_pooling_2d(F.relu(model.conv2(h1)), 3)
h3 = F.dropout(F.relu(model.l1(h2)), train=is_train)
h4 = F.dropout(F.relu(model.l2(h3)), train=is_train)
p = model.l3(h4)
return p
def __call__(self, x, train=True):
h1 = F.relu(self.conv1(x))
h2 = F.relu(self.conv2(h1))
h3 = F.relu(self.conv3(h2))
h4 = self.lstm(h3)
q = self.q(h4)
return q
def forward(x_data, y_data, train=True):
# Neural net architecture
x, t = chainer.Variable(x_data), chainer.Variable(y_data)
h = F.max_pooling_2d(F.dropout(F.relu(model.bn2(model.cv1(x))), train=train),2)
h = F.dropout(F.relu(model.ln3(h)), train=train)
y = model.ln4(h)
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def forward(self, x_data, y_data, train=True):
x = Variable(x_data, volatile=not train)
t = Variable(y_data, volatile=not train)
h = F.relu(self.bn1_1(self.conv1_1(x)))
h = F.relu(self.bn1_2(self.conv1_2(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = F.relu(self.bn2_1(self.conv2_1(h)))
h = F.relu(self.bn2_2(self.conv2_2(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = F.relu(self.bn3_1(self.conv3_1(h)))
h = F.relu(self.bn3_2(self.conv3_2(h)))
h = F.relu(self.bn3_3(self.conv3_3(h)))
h = F.relu(self.bn3_4(self.conv3_4(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = F.dropout(F.relu(self.fc4(h)), train=train, ratio=0.5)
h = F.dropout(F.relu(self.fc5(h)), train=train, ratio=0.5)
h = self.fc6(h)
if train:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t)
else:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t), h
def forward(x_data, y_data): x = Variable(x_data) t = Variable(y_data) h1 = F.relu(model.l1(x)) h2 = F.relu(model.l2(h1)) y = model.l3(h2) return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def predict(self, test_x: np.ndarray):
test_x = Variable(test_x)
self.h1 = F.dropout(F.relu(self.l1(test_x)))
self.h2 = F.dropout(F.relu(self.l2(self.h1)))
y = self.l3(self.h2)
predict_list = list(map(np.argmax, F.softmax(y).data))
return predict_list
def forward(self, x_data, y_data, train=True, gpu=-1): x, t = Variable(x_data), Variable(y_data) h = F.max_pooling_2d(F.relu(self.conv1(x)), ksize=2, stride=2) h = F.max_pooling_2d(F.relu(self.conv2(h)), ksize=3, stride=3) h = F.dropout(F.relu(self.l3(h)), train=train) y = self.l4(h) return F.softmax_cross_entropy(y, t), F.accuracy(y,t)
def forward(x_data, y_data, print_conf_matrix=False):
'''
Neural net architecture
:param x_data:
:param y_data:
:param train:
:return:
'''
x, t = Variable(x_data), Variable(y_data)
h1 = F.relu(model.l1(x))
h1 = F.max_pooling_2d(h1,max_pool_window_1,stride=max_pool_stride_1)
h2 = F.dropout(F.relu(model.l2(h1)))
h2 = F.average_pooling_2d(h2, avg_pool_window_2, stride=avg_pool_stride_2)
h2 = F.max_pooling_2d(h2,max_pool_window_2,stride=max_pool_stride_2)
y = model.l3(h2)
# display confusion matrix
if print_conf_matrix:
pdb.set_trace()
print confusion_matrix(cuda.to_cpu(t.data), cuda.to_cpu(y.data).argmax(axis=1))
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def Q_func_target(self, state):
h1 = F.relu(self.CNN_model_target.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.CNN_model_target.l2(h1))
h3 = F.relu(self.CNN_model_target.l3(h2))
h4 = F.relu(self.model.l4(h3))
Q = self.model_target.q_value(h4)
return Q
def __call__(self, x, rois, roi_indices):
"""Forward the chain.
We assume that there are :math:`N` batches.
Args:
x (~chainer.Variable): 4D image variable.
rois (array): A bounding box array containing coordinates of
proposal boxes. This is a concatenation of bounding box
arrays from multiple images in the batch.
Its shape is :math:`(R', 4)`. Given :math:`R_i` proposed
RoIs from the :math:`i` th image,
:math:`R' = \\sum _{i=1} ^ N R_i`.
roi_indices (array): An array containing indices of images to
which bounding boxes correspond to. Its shape is :math:`(R',)`.
"""
roi_indices = roi_indices.astype(np.float32)
indices_and_rois = self.xp.concatenate(
(roi_indices[:, None], rois), axis=1)
pool = _roi_pooling_2d_yx(
x, indices_and_rois, self.roi_size, self.roi_size,
self.spatial_scale)
fc6 = F.relu(self.fc6(pool))
fc7 = F.relu(self.fc7(fc6))
roi_cls_locs = self.cls_loc(fc7)
roi_scores = self.score(fc7)
return roi_cls_locs, roi_scores
def compute_features(self, obs):
obs = F.cast(obs, np.float32)
obs = F.transpose(obs, (0, 3, 1, 2))
h1 = F.relu(self.conv1(obs))
h2 = F.relu(self.conv2(h1))
h3 = F.relu(self.fc(h2))
return h3
def forward(self, x_data, y_data, train=True, models=None):
VGG_mini = models["VGG_mini"]
VGG_mini2 = models["VGG_mini2"]
VGG_mini3 = models["VGG_mini3"]
x = Variable(x_data, volatile=not train)
t = Variable(y_data, volatile=not train)
h = F.relu(self.conv1_1(x))
h = F.relu(self.conv1_2(h))
h = F.relu(self.conv1_3(h))
h = F.relu(self.conv1_4(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = F.relu(self.conv1_5(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = self.fc(h)
if train:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t)
else:
# return F.softmax_cross_entropy(h, t), F.accuracy(h, t), h
return F.softmax_cross_entropy(h, t), F.accuracy(h, t)
def __call__(self, x):
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
h = F.relu(self.conv3(h))
h = F.relu(self.fc4(h))
h = self.fc5(h)
return h
def __call__(self, state):
h1 = F.relu(self.conv1(state))
h2 = F.relu(self.l1(h1))
h3 = F.relu(self.l2(h2))
h4 = F.relu(self.l3(h3))
h5 = self.l4(h4)
return h5
def predict(self,x):
h1 = F.dropout(F.relu(self.l1(x)),ratio=self.dropout1,train=False)
h2 = F.dropout(F.relu(self.l2(h1)),ratio=self.dropout2,train=False)
h3 = F.dropout(F.relu(self.l3(h2)),ratio=self.dropout3,train=False)
h4 = self.l4(h3)
return h4
def __call__(self, x, t):
h = F.relu(self.bn1_1(self.conv1_1(x), test=not self.train))
h = F.relu(self.bn1_2(self.conv1_2(h), test=not self.train))
h = F.max_pooling_2d(h, 2, 2)
h = F.dropout(h, ratio=0.25, train=self.train)
h = F.relu(self.bn2_1(self.conv2_1(h), test=not self.train))
h = F.relu(self.bn2_2(self.conv2_2(h), test=not self.train))
h = F.max_pooling_2d(h, 2, 2)
h = F.dropout(h, ratio=0.25, train=self.train)
h = F.relu(self.bn3_1(self.conv3_1(h), test=not self.train))
h = F.relu(self.bn3_2(self.conv3_2(h), test=not self.train))
h = F.relu(self.bn3_3(self.conv3_3(h), test=not self.train))
h = F.relu(self.bn3_4(self.conv3_4(h), test=not self.train))
h = F.max_pooling_2d(h, 2, 2)
h = F.dropout(h, ratio=0.25, train=self.train)
h = F.dropout(F.relu(self.fc4(h)), ratio=0.5, train=self.train)
h = F.dropout(F.relu(self.fc5(h)), ratio=0.5, train=self.train)
h = self.fc6(h)
self.pred = F.softmax(h)
self.loss = F.softmax_cross_entropy(h, t)
self.accuracy = F.accuracy(self.pred, t)
if self.train:
return self.loss
else:
return self.pred
def __call__(self,x,y,state,train=True,target=True):
h = Variable(x.reshape(len(x), 1, 4, 3), volatile=not train)
t = Variable(y.flatten(), volatile=not train)
h0 = F.max_pooling_2d(F.relu(self.conv1(h)), 2)
h0 = F.max_pooling_2d(F.relu(self.conv2(h0)), 2)
if target == False:
data = h0.data
self.data_first.append(data)
h1= F.dropout(F.relu(self.l1(h0)),ratio=0.5,train=train)
if target == False:
data = h1.data
self.data_hidden.append(data)
y = self.l2(h1)
if target ==False:
data = y.data
self.data_output.append(data)
self.loss = F.softmax_cross_entropy(y,t)
return state, self.loss
def forward(x_data, y_data, train=True):
# Neural net architecture
x, t = chainer.Variable(x_data), chainer.Variable(y_data)
h1 = F.dropout(F.relu(model.l1(x)), train=train)
h2 = F.dropout(F.relu(model.l2(h1)), train=train)
y = model.l3(h2)
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def forward(self, x):
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
h = F.max_pooling_2d(h, (1, 2), stride=(1, 2))
h = F.reshape(h, (h.data.shape[0], h.data.shape[2], h.data.shape[1],
h.data.shape[3]))
h = F.relu(self.conv3(h))
h = F.relu(self.conv4(h))
h = F.max_pooling_2d(h, (5, 3))
h = F.relu(self.conv5(h))
h = F.relu(self.conv6(h))
h = F.max_pooling_2d(h, (1, 2))
h = F.relu(self.conv7(h))
h = F.relu(self.conv8(h))
h = F.max_pooling_2d(h, (1, 2))
h = F.relu(self.conv9(h))
h = F.relu(self.conv10(h))
h = F.max_pooling_2d(h, (1, 2))
#Get power spectrum by FFT. data_shape = (6,)
# frequency by 1.25Hz
# sampling_rate=1000Hz
# freq_names = {'delta','theta','lalpha','halpha','beta','lgamma'};
# freq_bands = [1 4; 4 8; 8 10; 10 13; 13 30; 30 50];
# delta:1.25-3.75:
# theta: 5-7.5:
# lalpha:8.75-10:
# halpha:11.25-12.5:
# beta:13.75-30:
# lgamma:31.25-50:
tmp = cupy.abs(cupy.fft.fft(x))
delta = cupy.average(tmp[:, :, :, 1:4], axis=3)
theta = cupy.average(tmp[:, :, :, 4:7], axis=3)
lalpha = cupy.average(tmp[:, :, :, 7:9], axis=3)
halpha = cupy.average(tmp[:, :, :, 9:11], axis=3)
beta = cupy.average(tmp[:, :, :, 11:25], axis=3)
lgamma = cupy.average(tmp[:, :, :, 25:41], axis=3)
Sum = delta + theta + lalpha + halpha + beta + lgamma
power_spectral = cupy.zeros((x.shape[0], x.shape[1], x.shape[2], 6))
power_spectral[:, :, :, 0] = cupy.divide(delta, Sum)
power_spectral[:, :, :, 1] = cupy.divide(theta, Sum)
power_spectral[:, :, :, 2] = cupy.divide(lalpha, Sum)
power_spectral[:, :, :, 3] = cupy.divide(halpha, Sum)
power_spectral[:, :, :, 4] = cupy.divide(beta, Sum)
power_spectral[:, :, :, 5] = cupy.divide(lgamma, Sum)
power_spectral = chainer.Variable(power_spectral)
power_spectral = F.cast(power_spectral, cupy.float32)
h = F.reshape(h, (h.shape[0], h.shape[1] * h.shape[2] * h.shape[3]))
power_spectral = F.reshape(
power_spectral,
(power_spectral.shape[0], power_spectral.shape[1] *
power_spectral.shape[2] * power_spectral.shape[3]))
h = F.relu(self.norm1(self.fc11(h)))
h = F.dropout(h)
h = F.relu(self.norm2(self.fc12(h)))
h = F.dropout(h)
#Concatenate the features extracted by deep neural network and relative power spectrum
h = F.concat((h, power_spectral), axis=1)
h = self.fc13(h)
if chainer.config.train:
return h
return F.softmax(h)
def __call__(self, x, test=False):
h = F.relu(self.b1(self.c1(x), test=test))
h = F.relu(self.b2(self.c2(h), test=test))
h = F.relu(self.b3(self.c3(h), test=test))
return h
def predict_depth(self, rgb, mask_score, depth_viz):
# concatenate rgb and depth_viz
concat_input = F.concat((rgb, depth_viz), axis=1)
# conv_depth_1
h = F.relu(self.conv_depth_1_1(concat_input))
h = F.relu(self.conv_depth_1_2(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
depth_pool1 = h # 1/2
# conv_depth_2
h = F.relu(self.conv_depth_2_1(depth_pool1))
h = F.relu(self.conv_depth_2_2(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
depth_pool2 = h # 1/4
# conv_depth_3
h = F.relu(self.conv_depth_3_1(depth_pool2))
h = F.relu(self.conv_depth_3_2(h))
h = F.relu(self.conv_depth_3_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
depth_pool3 = h # 1/8
# conv_depth_4
h = F.relu(self.conv_depth_4_1(depth_pool3))
h = F.relu(self.conv_depth_4_2(h))
h = F.relu(self.conv_depth_4_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
depth_pool4 = h # 1/16
# conv_depth_5
h = F.relu(self.conv_depth_5_1(depth_pool4))
h = F.relu(self.conv_depth_5_2(h))
h = F.relu(self.conv_depth_5_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
depth_pool5 = h # 1/32
if self.masking is True:
# (N, C, H, W) -> (N, H, W)
mask_pred_tmp = F.argmax(self.score_label, axis=1)
# (N, H, W) -> (N, 1, H, W), float required for resizing
mask_pred_tmp = mask_pred_tmp[:, None, :, :].data.astype(
self.xp.float32) # 1/1
resized_mask_pred = F.resize_images(
mask_pred_tmp,
(depth_pool5.shape[2], depth_pool5.shape[3])) # 1/32
depth_pool5_cp = depth_pool5
masked_depth_pool5 = depth_pool5_cp * \
(resized_mask_pred.data == 0.0).astype(self.xp.float32)
else:
masked_depth_pool5 = depth_pool5
# depth_fc6
h = F.relu(self.depth_fc6(masked_depth_pool5))
h = F.dropout(h, ratio=.5)
depth_fc6 = h # 1/32
# depth_fc7
h = F.relu(self.depth_fc7(depth_fc6))
h = F.dropout(h, ratio=.5)
depth_fc7 = h # 1/32
# depth_score_fr
h = self.depth_score_fr(depth_fc7)
depth_score_fr = h # 1/32
# depth_score_pool3
scale_depth_pool3 = 0.0001 * depth_pool3
h = self.depth_score_pool3(scale_depth_pool3)
depth_score_pool3 = h # 1/8
# depth_score_pool4
scale_depth_pool4 = 0.01 * depth_pool4
h = self.depth_score_pool4(scale_depth_pool4)
depth_score_pool4 = h # 1/16
# depth upscore2
h = self.depth_upscore2(depth_score_fr)
depth_upscore2 = h # 1/16
# depth_score_pool4c
h = depth_score_pool4[:, :, 5:5 + depth_upscore2.data.shape[2],
5:5 + depth_upscore2.data.shape[3]]
depth_score_pool4c = h # 1/16
# depth_fuse_pool4
h = depth_upscore2 + depth_score_pool4c
depth_fuse_pool4 = h # 1/16
# depth_upscore_pool4
h = self.depth_upscore_pool4(depth_fuse_pool4)
depth_upscore_pool4 = h # 1/8
# depth_score_pool3c
h = depth_score_pool3[:, :, 9:9 + depth_upscore_pool4.data.shape[2],
9:9 + depth_upscore_pool4.data.shape[3]]
depth_score_pool3c = h # 1/8
# depth_fuse_pool3
h = depth_upscore_pool4 + depth_score_pool3c
depth_fuse_pool3 = h # 1/8
# depth_upscore8
h = self.depth_upscore8(depth_fuse_pool3)
depth_upscore8 = h # 1/1
# depth_score
h = depth_upscore8[:, :, 31:31 + rgb.shape[2], 31:31 + rgb.shape[3]]
depth_score = h # 1/1
# (-inf, inf) -> (0, 1) -> (min_depth, max_depth)
h = F.sigmoid(depth_score)
h = h * (self.max_depth - self.min_depth) + self.min_depth
depth_pred = h
return depth_pred
def predict_mask(self, rgb, return_pool5=False):
# conv_rgb_1
h = F.relu(self.conv_rgb_1_1(rgb))
h = F.relu(self.conv_rgb_1_2(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
rgb_pool1 = h # 1/2
# conv_rgb_2
h = F.relu(self.conv_rgb_2_1(rgb_pool1))
h = F.relu(self.conv_rgb_2_2(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
rgb_pool2 = h # 1/4
# conv_rgb_3
h = F.relu(self.conv_rgb_3_1(rgb_pool2))
h = F.relu(self.conv_rgb_3_2(h))
h = F.relu(self.conv_rgb_3_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
rgb_pool3 = h # 1/8
# conv_rgb_4
h = F.relu(self.conv_rgb_4_1(rgb_pool3))
h = F.relu(self.conv_rgb_4_2(h))
h = F.relu(self.conv_rgb_4_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
rgb_pool4 = h # 1/16
# conv_rgb_5
h = F.relu(self.conv_rgb_5_1(rgb_pool4))
h = F.relu(self.conv_rgb_5_2(h))
h = F.relu(self.conv_rgb_5_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
rgb_pool5 = h # 1/32
# rgb_fc6
h = F.relu(self.rgb_fc6(rgb_pool5))
h = F.dropout(h, ratio=.5)
rgb_fc6 = h # 1/32
# rgb_fc7
h = F.relu(self.rgb_fc7(rgb_fc6))
h = F.dropout(h, ratio=.5)
rgb_fc7 = h # 1/32
# mask_score_fr
h = self.mask_score_fr(rgb_fc7)
mask_score_fr = h # 1/32
# mask_score_pool3
scale_rgb_pool3 = 0.0001 * rgb_pool3
h = self.mask_score_pool3(scale_rgb_pool3)
mask_score_pool3 = h # 1/8
# mask_score_pool4
scale_rgb_pool4 = 0.01 * rgb_pool4
h = self.mask_score_pool4(scale_rgb_pool4)
mask_score_pool4 = h # 1/16
# mask upscore2
h = self.mask_upscore2(mask_score_fr)
mask_upscore2 = h # 1/16
# mask_score_pool4c
h = mask_score_pool4[:, :, 5:5 + mask_upscore2.data.shape[2],
5:5 + mask_upscore2.data.shape[3]]
mask_score_pool4c = h # 1/16
# mask_fuse_pool4
h = mask_upscore2 + mask_score_pool4c
mask_fuse_pool4 = h # 1/16
# mask_upscore_pool4
h = self.mask_upscore_pool4(mask_fuse_pool4)
mask_upscore_pool4 = h # 1/8
# mask_score_pool3c
h = mask_score_pool3[:, :, 9:9 + mask_upscore_pool4.data.shape[2],
9:9 + mask_upscore_pool4.data.shape[3]]
mask_score_pool3c = h # 1/8
# mask_fuse_pool3
h = mask_upscore_pool4 + mask_score_pool3c
mask_fuse_pool3 = h # 1/8
# mask_upscore8
h = self.mask_upscore8(mask_fuse_pool3)
mask_upscore8 = h # 1/1
# mask_score
h = mask_upscore8[:, :, 31:31 + rgb.shape[2], 31:31 + rgb.shape[3]]
mask_score = h # 1/1
if return_pool5:
return mask_score, rgb_pool5
else:
return mask_score
def __call__(self, x):
h1 = F.relu(self.l1(x))
h2 = F.relu(self.l2(h1))
return self.l3(h2)
def __call__(self, x):
# Using max_pooling -> ave_pooling
# 1 Layer
h = F.relu(self.conv1_1(x))
h1 = F.relu(self.conv1_2(h))
# 2 Layer
# h = F.max_pooling_2d(h1, ksize=2)
h = F.average_pooling_2d(h1, ksize=2)
h = F.relu(self.conv2_1(h))
h2 = F.relu(self.conv2_2(h))
# 3 Layer
# h = F.max_pooling_2d(h2, ksize=2)
h = F.average_pooling_2d(h2, ksize=2)
h = F.relu(self.conv3_1(h))
h = F.relu(self.conv3_2(h))
h3 = F.relu(self.conv3_3(h))
# 4 Layer
# h = F.max_pooling_2d(h3, ksize=2)
h = F.average_pooling_2d(h3, ksize=2)
h = F.relu(self.conv4_1(h))
h = F.relu(self.conv4_2(h))
h4 = F.relu(self.conv4_3(h))
# 5 Layer
# h = F.max_pooling_2d(h4, ksize=2)
h = F.average_pooling_2d(h4, ksize=2)
h = F.relu(self.conv5_1(h))
h = F.relu(self.conv5_2(h))
h5 = F.relu(self.conv5_3(h))
return h5
def __call__(self, x):
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
h = F.relu(self.conv3(h))
return h
def __call__(self, x):
# データを受け取った際のforward計算を書く
h1 = F.relu(self.l1(x))
h2 = F.relu(self.l2(h1))
return self.l3(h2)
def __call__(self, x, t=None):
self.x = x
self.t = t
# conv1
h = F.relu(self.conv1_1(x))
conv1_1 = h
h = F.relu(self.conv1_2(conv1_1))
conv1_2 = h
h = F.max_pooling_2d(conv1_2, 2, stride=2, pad=0)
pool1 = h # 1/2
# conv2
h = F.relu(self.conv2_1(pool1))
conv2_1 = h
h = F.relu(self.conv2_2(conv2_1))
conv2_2 = h
h = F.max_pooling_2d(conv2_2, 2, stride=2, pad=0)
pool2 = h # 1/4
# conv3
h = F.relu(self.conv3_1(pool2))
conv3_1 = h
h = F.relu(self.conv3_2(conv3_1))
conv3_2 = h
h = F.relu(self.conv3_3(conv3_2))
conv3_3 = h
h = F.max_pooling_2d(conv3_3, 2, stride=2, pad=0)
pool3 = h # 1/8
# conv4
h = F.relu(self.conv4_1(pool3))
h = F.relu(self.conv4_2(h))
h = F.relu(self.conv4_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
pool4 = h # 1/16
# conv5
h = F.relu(self.conv5_1(pool4))
h = F.relu(self.conv5_2(h))
h = F.relu(self.conv5_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
pool5 = h # 1/32
# fc6
h = F.relu(self.fc6(pool5))
h = F.dropout(h, ratio=.5)
fc6 = h # 1/32
# fc7
h = F.relu(self.fc7(fc6))
h = F.dropout(h, ratio=.5)
fc7 = h # 1/32
# score_fr
h = self.score_fr(fc7)
score_fr = h # 1/32
# score_pool3
h = self.score_pool3(pool3)
score_pool3 = h # 1/8
# score_pool4
h = self.score_pool4(pool4)
score_pool4 = h # 1/16
# upscore2
h = self.upscore2(score_fr)
upscore2 = h # 1/16
# score_pool4c
h = score_pool4[:, :, 5:5 + upscore2.data.shape[2],
5:5 + upscore2.data.shape[3]]
score_pool4c = h # 1/16
# fuse_pool4
h = upscore2 + score_pool4c
fuse_pool4 = h # 1/16
# upscore_pool4
h = self.upscore_pool4(fuse_pool4)
upscore_pool4 = h # 1/8
# score_pool4c
h = score_pool3[:, :, 9:9 + upscore_pool4.data.shape[2],
9:9 + upscore_pool4.data.shape[3]]
score_pool3c = h # 1/8
# fuse_pool3
h = upscore_pool4 + score_pool3c
fuse_pool3 = h # 1/8
# upscore8
h = self.upscore8(fuse_pool3)
upscore8 = h # 1/1
# score
h = upscore8[:, :, 31:31 + x.data.shape[2], 31:31 + x.data.shape[3]]
self.score = h # 1/1
if t is None:
assert not chainer.config.train
return
# testing with t or training
self.loss = F.softmax_cross_entropy(self.score, t, normalize=False)
chainer.report({'loss': self.loss})
return self.loss
def __call__(self, x, t=None):
# conv1
h = F.relu(self.conv1_1(x))
conv1_1 = h
h = F.relu(self.conv1_2(conv1_1))
conv1_2 = h
h = F.max_pooling_2d(conv1_2, 2, stride=2, pad=0)
pool1 = h # 1/2
# conv2
h = F.relu(self.conv2_1(pool1))
conv2_1 = h
h = F.relu(self.conv2_2(conv2_1))
conv2_2 = h
h = F.max_pooling_2d(conv2_2, 2, stride=2, pad=0)
pool2 = h # 1/4
# conv3
h = F.relu(self.conv3_1(pool2))
conv3_1 = h
h = F.relu(self.conv3_2(conv3_1))
conv3_2 = h
h = F.relu(self.conv3_3(conv3_2))
conv3_3 = h
h = F.max_pooling_2d(conv3_3, 2, stride=2, pad=0)
pool3 = h # 1/8
# conv4
h = F.relu(self.conv4_1(pool3))
h = F.relu(self.conv4_2(h))
h = F.relu(self.conv4_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
pool4 = h # 1/16
# conv5
h = F.relu(self.conv5_1(pool4))
h = F.relu(self.conv5_2(h))
h = F.relu(self.conv5_3(h))
h = F.max_pooling_2d(h, 2, stride=2, pad=0)
pool5 = h # 1/32
# fc6
h = F.relu(self.fc6(pool5))
h = F.dropout(h, ratio=.5)
fc6 = h # 1/32
# fc7
h = F.relu(self.fc7(fc6))
h = F.dropout(h, ratio=.5)
fc7 = h # 1/32
# score_fr
h = self.score_fr(fc7)
score_fr = h # 1/32
# upscore2
h = self.upscore2(score_fr)
upscore2 = h # 1/16
# score_pool4
h = self.score_pool4(pool4)
score_pool4 = h # 1/16
# score_pool4c
h = score_pool4[:, :, 5:5 + upscore2.shape[2], 5:5 + upscore2.shape[3]]
score_pool4c = h # 1/16
# fuse_pool4
h = upscore2 + score_pool4c
fuse_pool4 = h # 1/16
# upscore16
h = self.upscore16(fuse_pool4)
upscore16 = h # 1/1
# score
h = upscore16[:, :, 27:27 + x.shape[2], 27:27 + x.shape[3]]
score = h # 1/1
self.score = score
if t is None:
assert not chainer.configuration.config.train
return
loss = F.softmax_cross_entropy(self.score,
t,
normalize=False,
class_weight=self.class_weight)
if np.isnan(float(loss.data)):
raise ValueError('Loss value is nan.')
chainer.report({'loss': loss}, self)
return loss
def forward(self, x):
"""Compute Q-values of actions for given observations."""
h = F.relu(self.l0(x))
h = F.relu(self.l1(h))
return self.l2(h)
def __call__(self, x):
h = self.base(x, layers=["pool5"])["pool5"]
h = F.dropout(F.relu(self.fc_1(h)))
h = self.fc_2(h)
return h
def __call__(self, *x_list):
#print (x_list.data.shape)
#print (x_list.data[0])
#print (x_list.data[1])
##１フレームごとfc7までを検出しLSTMに対してループ
self.l8.reset_state()
y_list = []
for i in range(5):
x = x_list[i]
#x.volatile = 'on'
#print (xd.shape)
#xd = xd.asarray(xd).astype("f")
#x = np.asarray(xd, dtype=np.float32)
#print (x.data.shape)
#print (x.shape)
#x = np.asarray(xd).astype(np.float32)
#print (i)
h = F.dropout(F.relu(self.conv1_1(x)))
h = F.dropout(F.relu(self.conv1_2(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(F.relu(self.conv2_1(h)))
h = F.dropout(F.relu(self.conv2_2(h)))
h = F.max_pooling_2d(h, 2, stride=1)
#h = F.dropout(F.relu(self.conv3_1_b(h)))
#h = F.relu(self.conv3_2_b(h))
#h = F.relu(self.conv3_3_b(h))
#h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(F.relu(self.conv4_1_b(h)))
#h = F.relu(self.conv4_2(h))
#h = F.relu(self.conv4_3(h))
#h = F.max_pooling_2d(h, 2, stride=2)
#h = F.relu(self.conv5_1(h))
#h = F.relu(self.conv5_2(h))
#h = F.relu(self.conv5_3(h))
#h = F.max_pooling_2d(h, 2, stride=2)
#h = F.dropout(F.relu(self.fc6(h)), train=self.train, ratio=0.5)
#h = F.dropout(F.relu(self.fc7(h)), train=self.train, ratio=0.5)
#h.volatile = 'off'
#h = F.relu(self.fc6_b(h))
#h = F.relu(self.fc7(h))
l = self.l8(h)
#print h.data
#h = F.sigmoid(self.fc9(h))
#h = (h*4)+1
#print (h.data)
h = self.fc9(l)
if self.train:
#self.loss = F.softmax_cross_entropy(h, t)
#chainer.report({'loss': loss, 'accuracy': F.accuracy(h, t)}, self)
#print ("return loss")
return h
#return self.loss
else:
self.pred = F.softmax(l)
return self.pred
def __call__(self, x, train):
h = F.relu(self.bn1(self.conv1(x), test=not train))
h = F.relu(self.bn2(self.conv2(h), test=not train))
h = self.bn3(self.conv3(h), test=not train)
return F.relu(h + x)
def __call__(self, x):
h = F.relu(self.mid(x))
y = self.out(h)
return y
def __call__(self, x):
h1 = F.max_pooling_2d(
F.relu(self.conv1(F.reshape(x, (-1, 1, 28, 28)))), 2)
h2 = F.max_pooling_2d(F.relu(self.conv2(h1)), 2)
h3 = F.relu(self.l1(h2))
return self.l2(h3)
def __call__(self, x):
for l in self.layers[:-1]:
x = F.relu(l(x))
return self.layers[-1](x)
def fwd( self, x ): h1 = F.dropout( F.relu( self.l1(x) ) ) h2 = F.dropout( F.relu( self.l2(h1) ) ) h3 = self.l3(h2) return h3
def __call__(self, x):
sentences = x
h = F.relu(F.dropout(self.l_input(sentences), 0.5))
h = F.relu(F.dropout(self.l_merge(h), 0.5))
h = F.relu(F.dropout(self.l_merge2(h), 0.5))
return h
def __call__(self, x):
h = F.relu(self.fc1(x))
h = F.relu(self.fc2(h))
y = self.fc3(h)
return y
def forward(x_data, t_data, train=True):
# Neural net architecture
x = chainer.Variable(x_data)
t = chainer.Variable(t_data)
y = F.relu(model.l1(x))
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def forward(self, x):
y1 = F.relu(x)
return y1
def __call__(self, img, sparse_inputs):
# Downsampling layers
h = F.relu(self.conv1(np.concatenate((img, sparse_inputs), axis=1)))
# Dense1
x = np.resize(sparse_inputs, (1, 2, 240, 320))
h = F.relu(self.d1_conv1(h))
h = self.dense1(h, x)
h = F.relu(self.d1_conv2(h))
# Skip Dense 2
_, _, H, W = x.shape
x120 = np.resize(x, (1, 2, H // 2, W // 2))
skip_h2 = F.relu(self.d2c_conv1(h))
skip_h2 = self.dense2c(skip_h2, x120)
skip_h2 = F.relu(self.d2c_conv2(skip_h2))
h = F.relu(self.d2_conv1(h))
h = self.dense2(h, x120)
h = F.relu(self.d2_conv2(h))
# Skip Dense 3
_, _, H, W = x120.shape
x60 = np.resize(x120, (1, 2, H // 2, W // 2))
skip_h3 = F.relu(self.d3c_conv1(h))
skip_h3 = self.dense3c(skip_h3, x60)
skip_h3 = F.relu(self.d3c_conv2(skip_h3))
h = F.relu(self.d3_conv1(h))
h = self.dense3(h, x60)
h = F.relu(self.d3_conv2(h))
# Skip Dense 4
_, _, H, W = x60.shape
x30 = np.resize(x60, (1, 2, H // 2, W // 2))
skip_h4 = F.relu(self.d4c_conv1(h))
skip_h4 = self.dense4c(skip_h4, x30)
skip_h4 = F.relu(self.d4c_conv2(skip_h4))
h = F.relu(self.d4_conv1(h))
h = self.dense4(h, x30)
h = F.relu(self.d4_conv2(h))
# Dense 5
_, _, H, W = x30.shape
x15 = np.resize(x30, (1, 2, H // 2, W // 2))
b, c, H, W = x15.shape
x16 = np.empty((b, c, H + 1, W), dtype=np.float32)
x16[:, :, :-1, ...] = x15
x16[:, :, -1, ...] = x15[:, :, -1]
h = F.relu(self.d5_conv1(h))
h = self.dense5(h, x16)
h = F.relu(self.d5_conv2(h))
# UpDense 1
print(h.shape)
h = F.relu(self.up1_conv1(h))
print(h.shape)
h = self.up_dense1(h, x16)
h = F.relu(self.up1_conv2(h))
h = F.concat((h, skip_h4), axis=1)
# UpDense 2
print(h.shape)
h = F.relu(self.up2_conv1(h))
print(h.shape)
# h = self.up_dense2(h, new_x)
# h = F.relu(self.up2_conv2(h))
# print(h.shape)
# # UpDense 3
# h = F.relu(self.up3_conv1(h))
# h = self.up_dense3(h, new_x)
# h = F.relu(self.up3_conv2(h))
# # UpDense 4
# h = F.relu(self.up4_conv1(h))
# h = self.up_dense4(h, new_x)
# h = F.relu(self.up4_conv2(h))
return h
def forward(self, x, train, action, init_l=None):
if init_l is None:
# Location Net @t-1
m = F.tanh(self.fc_hl(self.h))
if train:
eps = (self.xp.random.normal(0, 1, size=m.data.shape)).astype(
np.float32)
l = m.data + np.sqrt(self.var) * eps
# do not backward reinforce loss via l
# log(location policy)
ln_pi = -0.5 * F.sum((l - m) * (l - m), axis=1) / self.var
with chainer.using_config('enable_backprop', train):
l = chainer.Variable(l)
else:
l = m
ln_pi = None
else:
l = init_l
ln_pi = None
# Retina Encoding
x.volatile = 'on' # do not backward
if self.xp == np:
loc = l.data
else:
loc = self.xp.asnumpy(l.data)
rho = crop(x, center=loc, size=self.g_size)
# multi-scale glimpse
for k in range(1, self.n_scales):
s = np.power(2, k)
patch = crop(x, center=loc, size=self.g_size * s)
patch = F.average_pooling_2d(patch, ksize=s)
rho = F.concat((rho, patch), axis=1)
if train: rho.volatile = 'off' # backward up to link emb_x
hg = F.relu(self.emb_x(rho))
# Location Encoding
hl = F.relu(self.emb_l(l))
# Glimpse Net
g = F.relu(self.fc_lg(hl) + self.fc_xg(hg))
# Core Net
if self.use_lstm:
self.h = self.core_lstm(g)
else:
self.h = F.relu(self.core_hh(self.h) + self.core_gh(g))
# Action Net
if action:
y = self.fc_ha(self.h)
else:
y = None
# Baseline
if train and action:
b = F.sigmoid(self.fc_hb(self.h))
b = F.reshape(b, (-1, ))
else:
b = None
return l, ln_pi, y, b
def __call__(self, x, t):
h1 = self.l1(x)
y = self.l3(F.relu(self.l2(F.relu(self.l1(x)))))
# self.loss = self.listwise_cost(y_data, t_data)
self.loss = self.jsd(t, y)
return self.loss
def __call__(self, x):
h = F.max_pooling_2d(F.relu(self.conv1(x)), 2, 2)
h = F.max_pooling_2d(F.relu(self.conv2(h)), 2, 2)
h = F.dropout(F.relu(self.conv3(h)), ratio=self.dropout)
h = F.dropout(F.relu(self.fc4(h)), ratio=self.dropout)
return h
def __call__(self, input_blob, test_mode=False):
# explicit and very flexible DAG!
#################################
data = input_blob[0]
labels = input_blob[1]
if (len(input_blob) >= 3):
weights_classes = input_blob[2]
else:
weights_classes = chainer.Variable(
cuda.cupy.ones((self.classes, 1), dtype='float32'))
# ---- CONTRACTION BLOCKS ---- #
# B1
#ipdb.set_trace()
cblob = self.conv1_1(data)
cblob = F.relu(cblob)
cblob = self.conv1_2(cblob)
cblob = F.relu(cblob)
cblob = F.max_pooling_2d(cblob, (2, 2), stride=(2, 2), pad=(0, 0))
# B2
cblob = self.conv2_1(cblob)
cblob = F.relu(cblob)
cblob = self.conv2_2(cblob)
cblob = F.relu(cblob)
cblob = F.max_pooling_2d(cblob, (2, 2), stride=(2, 2), pad=(0, 0))
# B3
cblob = self.conv3_1(cblob)
cblob = F.relu(cblob)
cblob = self.conv3_2(cblob)
cblob = F.relu(cblob)
cblob = self.conv3_3(cblob)
cblob = F.relu(cblob)
cblob_mp3 = F.max_pooling_2d(cblob, (2, 2), stride=(2, 2), pad=(0, 0))
# B4
cblob = self.conv4_1(cblob_mp3)
cblob = F.relu(cblob)
cblob = self.conv4_2(cblob)
cblob = F.relu(cblob)
cblob = self.conv4_3(cblob)
cblob = F.relu(cblob)
cblob_mp4 = F.max_pooling_2d(cblob, (2, 2), stride=(2, 2), pad=(0, 0))
# B5
cblob = self.conv5_1(cblob_mp4)
cblob = F.relu(cblob)
cblob = self.conv5_2(cblob)
cblob = F.relu(cblob)
cblob = self.conv5_3(cblob)
cblob = F.relu(cblob)
cblob = F.max_pooling_2d(cblob, (2, 2), stride=(2, 2), pad=(0, 0))
# FCs
#cblob = self.conv_aux(cblob)
#cblob = self.conv_auxb(cblob)
cblob = self.fc_6(cblob)
cblob = F.relu(cblob)
cblob = F.dropout(cblob, ratio=0.5, train=not test_mode)
cblob = self.fc_7(cblob)
cblob = F.relu(cblob)
cblob = F.dropout(cblob, ratio=0.5, train=not test_mode)
cblob = self.conv_aux2(cblob)
# ---- EXPANSION BLOCKS ---- #
cblob = self.score2(cblob)
cblob_aux1 = self.score_pool4(cblob_mp4)
cblob = F.sum_binary(cblob, cblob_aux1)
scoreup = self.score4(cblob)
cblob_aux2 = self.score_pool3(cblob_mp3)
cblob = F.sum_binary(scoreup, cblob_aux2)
cblob = self.upsample(cblob)
# ---- SOFTMAX CLASSIFIER ---- #
self.blob_class = self.classi(cblob)
self.probs = F.softmax(self.blob_class)
# ---- WEIGHTED CROSS-ENTROPY LOSS ---- #
#ipdb.set_trace()
self.output_point = self.probs
if (test_mode != 2):
self.loss = F.weighted_cross_entropy(self.probs,
labels,
weights_classes,
normalize=True)
return self.loss
else:
return 0
def predict(self, x):
h1 = F.relu(self.l1(x))
h2 = F.relu(self.l2(h1))
h = F.relu(self.l3(h2))
return h.data
def __call__(self, x, train=False):
return F.relu(self.bn(self.conv(x), test=not train))
def __call__(self, x, y):
if y is not None:
data_ab_ss = self.data_ab_ss(y)
gt_ab_313 = self.nn_enc_layer.forward(data_ab_ss.data)
gt_ab_313_va = chainer.Variable(gt_ab_313)
# non_gray_mask = self.non_gray_mask_layer.forward(data_ab_ss)
# prior_boost = self.prior_boost_layer.forward(gt_ab_313)
# prior_boost_nongray = prior_boost * non_gray_mask
h = F.relu(self.conv1_1(x))
h = self.conv1_2norm(F.relu(self.conv1_2(h)))
h = F.relu(self.conv2_1(h))
h = self.conv2_2norm(F.relu(self.conv2_2(h)))
h = F.relu(self.conv3_1(h))
h = F.relu(self.conv3_2(h))
h = F.relu(self.conv3_3(h))
h = self.conv3_3norm(h)
h = F.relu(self.conv4_1(h))
h = F.relu(self.conv4_2(h))
h = F.relu(self.conv4_3(h))
h = self.conv4_3norm(h)
h = F.relu(self.conv5_1(h))
h = F.relu(self.conv5_2(h))
h = F.relu(self.conv5_3(h))
h = self.conv5_3norm(h)
h = F.relu(self.conv6_1(h))
h = F.relu(self.conv6_2(h))
h = F.relu(self.conv6_3(h))
h = self.conv6_3norm(h)
h = F.relu(self.conv7_1(h))
h = F.relu(self.conv7_2(h))
h = F.relu(self.conv7_3(h))
h = self.conv7_3norm(h)
h = F.relu(self.conv8_1(h))
h = F.relu(self.conv8_2(h))
h = F.relu(self.conv8_3(h))
h = F.relu(self.conv313(h))
if y is not None:
print(h.shape, gt_ab_313_va.shape)
loss = F.softmax_cross_entropy(h, gt_ab_313_va)
chainer.report({"main/loss": loss})
else:
return gt_ab_313_va
