def exp_proto(mode, batchsize, T, exp_T, question_vocab_size, exp_vocab_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
n.exp_att_feature, n.exp, n.exp_out, n.exp_cont_1, n.exp_cont_2 = \
L.Python(module='exp_data_provider_layer', layer='ExpDataProviderLayer', param_str=mode_str, ntop=5)
n.exp_embed_ba = L.Embed(n.exp, input_dim=exp_vocab_size, num_output=300, \
weight_filler=dict(type='uniform', min=-0.08, max=0.08))
n.exp_embed = L.TanH(n.exp_embed_ba)
# LSTM1 for Explanation
n.exp_lstm1 = L.LSTM(\
n.exp_embed, n.exp_cont_1,\
recurrent_param=dict(\
num_output=2048,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
n.exp_lstm1_dropped = L.Dropout(n.exp_lstm1,
dropout_param={'dropout_ratio': 0.3})
# Merge with LSTM1 for explanation
n.exp_att_resh = L.Reshape(
n.exp_att_feature, reshape_param=dict(shape=dict(dim=[1, -1, 2048])))
n.exp_att_tiled = L.Tile(n.exp_att_resh, axis=0, tiles=exp_T)
n.exp_eltwise_all = L.Eltwise(n.exp_lstm1_dropped,
n.exp_att_tiled,
eltwise_param={'operation': P.Eltwise.PROD})
n.exp_eltwise_all_sqrt = L.SignedSqrt(n.exp_eltwise_all)
n.exp_eltwise_all_l2 = L.L2Normalize(n.exp_eltwise_all_sqrt)
n.exp_eltwise_all_drop = L.Dropout(n.exp_eltwise_all_l2,
dropout_param={'dropout_ratio': 0.3})
# LSTM2 for Explanation
n.exp_lstm2 = L.LSTM(\
n.exp_eltwise_all_drop, n.exp_cont_2,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
n.exp_lstm2_dropped = L.Dropout(n.exp_lstm2,
dropout_param={'dropout_ratio': 0.3})
n.exp_prediction = L.InnerProduct(n.exp_lstm2_dropped,
num_output=exp_vocab_size,
weight_filler=dict(type='xavier'),
axis=2)
n.silence_exp_prediction = L.Silence(n.exp_prediction, ntop=0)
return n.to_proto()
def bilinear_distance(self, vec1, vec2):
reshape_vec1 = L.Reshape(vec1,
shape=dict(dim=[self.batch_size, -1, 1, 1]))
reshape_vec2 = L.Reshape(vec2,
shape=dict(dim=[self.batch_size, -1, 1, 1]))
bilinear = L.CompactBilinear(reshape_vec1, reshape_vec2)
signed = L.SignedSqrt(bilinear)
l2_normalize = L.L2Normalize(signed)
score = L.InnerProduct(
l2_normalize,
num_output=1,
weight_filler=self.uniform_weight_filler(-0.08, .08),
param=self.learning_params([[1, 1], [2, 0]],
['bilinear_dist', 'bilinear_dist_b']))
return score
def qlstm(mode, batchsize, T, question_vocab_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
n.data, n.cont, n.img_feature, n.label, n.glove = L.Python(\
module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=5 )
n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08))
n.embed = L.TanH(n.embed_ba)
concat_word_embed = [n.embed, n.glove]
n.concat_embed = L.Concat(*concat_word_embed,
concat_param={'axis': 2}) # T x N x 600
# LSTM1
n.lstm1 = L.LSTM(\
n.concat_embed, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
tops1 = L.Slice(n.lstm1, ntop=T, slice_param={'axis': 0})
for i in xrange(T - 1):
n.__setattr__('slice_first' + str(i), tops1[int(i)])
n.__setattr__('silence_data_first' + str(i),
L.Silence(tops1[int(i)], ntop=0))
n.lstm1_out = tops1[T - 1]
n.lstm1_reshaped = L.Reshape(n.lstm1_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm1_reshaped_droped = L.Dropout(n.lstm1_reshaped,
dropout_param={'dropout_ratio': 0.3})
n.lstm1_droped = L.Dropout(n.lstm1, dropout_param={'dropout_ratio': 0.3})
# LSTM2
n.lstm2 = L.LSTM(\
n.lstm1_droped, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
tops2 = L.Slice(n.lstm2, ntop=T, slice_param={'axis': 0})
for i in xrange(T - 1):
n.__setattr__('slice_second' + str(i), tops2[int(i)])
n.__setattr__('silence_data_second' + str(i),
L.Silence(tops2[int(i)], ntop=0))
n.lstm2_out = tops2[T - 1]
n.lstm2_reshaped = L.Reshape(n.lstm2_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm2_reshaped_droped = L.Dropout(n.lstm2_reshaped,
dropout_param={'dropout_ratio': 0.3})
concat_botom = [n.lstm1_reshaped_droped, n.lstm2_reshaped_droped]
n.lstm_12 = L.Concat(*concat_botom)
n.q_emb_tanh_droped_resh = L.Reshape(
n.lstm_12, reshape_param=dict(shape=dict(dim=[-1, 2048, 1, 1])))
n.q_emb_tanh_droped_resh_tiled_1 = L.Tile(n.q_emb_tanh_droped_resh,
axis=2,
tiles=14)
n.q_emb_tanh_droped_resh_tiled = L.Tile(n.q_emb_tanh_droped_resh_tiled_1,
axis=3,
tiles=14)
n.i_emb_tanh_droped_resh = L.Reshape(
n.img_feature, reshape_param=dict(shape=dict(dim=[-1, 2048, 14, 14])))
n.blcf = L.CompactBilinear(n.q_emb_tanh_droped_resh_tiled,
n.i_emb_tanh_droped_resh,
compact_bilinear_param=dict(num_output=16000,
sum_pool=False))
n.blcf_sign_sqrt = L.SignedSqrt(n.blcf)
n.blcf_sign_sqrt_l2 = L.L2Normalize(n.blcf_sign_sqrt)
n.blcf_droped = L.Dropout(n.blcf_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
# multi-channel attention
n.att_conv1 = L.Convolution(n.blcf_droped,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu,
kernel_size=1,
stride=1,
num_output=2,
pad=0,
weight_filler=dict(type='xavier'))
n.att_reshaped = L.Reshape(
n.att_conv2, reshape_param=dict(shape=dict(dim=[-1, 2, 14 * 14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att = L.Reshape(n.att_softmax,
reshape_param=dict(shape=dict(dim=[-1, 2, 14, 14])))
att_maps = L.Slice(n.att, ntop=2, slice_param={'axis': 1})
n.att_map0 = att_maps[0]
n.att_map1 = att_maps[1]
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.att_feature0 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map0,
dummy)
n.att_feature1 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map1,
dummy)
n.att_feature0_resh = L.Reshape(
n.att_feature0, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature1_resh = L.Reshape(
n.att_feature1, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature = L.Concat(n.att_feature0_resh, n.att_feature1_resh)
# merge attention and lstm with compact bilinear pooling
n.att_feature_resh = L.Reshape(
n.att_feature, reshape_param=dict(shape=dict(dim=[-1, 4096, 1, 1])))
n.lstm_12_resh = L.Reshape(
n.lstm_12, reshape_param=dict(shape=dict(dim=[-1, 2048, 1, 1])))
n.bc_att_lstm = L.CompactBilinear(n.att_feature_resh,
n.lstm_12_resh,
compact_bilinear_param=dict(
num_output=16000, sum_pool=False))
n.bc_sign_sqrt = L.SignedSqrt(n.bc_att_lstm)
n.bc_sign_sqrt_l2 = L.L2Normalize(n.bc_sign_sqrt)
n.bc_dropped = L.Dropout(n.bc_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
n.bc_dropped_resh = L.Reshape(
n.bc_dropped, reshape_param=dict(shape=dict(dim=[-1, 16000])))
n.prediction = L.InnerProduct(n.bc_dropped_resh,
num_output=3000,
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
return n.to_proto()
def mfb_coatt(mode, batchsize, T, question_vocab_size, folder):
n = caffe.NetSpec()
mode_str = json.dumps({'mode':mode, 'batchsize':batchsize,'folder':folder})
if mode == 'val':
n.data, n.cont, n.img_feature, n.label, n.glove = L.Python( \
module='vqa_data_layer_hdf5', layer='VQADataProviderLayer', \
param_str=mode_str, ntop=5 )
else:
n.data, n.cont, n.img_feature, n.label, n.glove = L.Python(\
module='vqa_data_layer_kld_hdf5', layer='VQADataProviderLayer', \
param_str=mode_str, ntop=5 )
n.embed = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
weight_filler=dict(type='xavier'))
n.embed_tanh = L.TanH(n.embed)
concat_word_embed = [n.embed_tanh, n.glove]
n.concat_embed = L.Concat(*concat_word_embed, concat_param={'axis': 2}) # T x N x 600
# LSTM
n.lstm1 = L.LSTM(\
n.concat_embed, n.cont,\
recurrent_param=dict(\
num_output=config.LSTM_UNIT_NUM,\
weight_filler=dict(type='xavier')))
n.lstm1_droped = L.Dropout(n.lstm1,dropout_param={'dropout_ratio':config.LSTM_DROPOUT_RATIO})
n.lstm1_resh = L.Permute(n.lstm1_droped, permute_param=dict(order=[1,2,0]))
n.lstm1_resh2 = L.Reshape(n.lstm1_resh, \
reshape_param=dict(shape=dict(dim=[0,0,0,1])))
'''
Question Attention
'''
n.qatt_conv1 = L.Convolution(n.lstm1_resh2, kernel_size=1, stride=1, num_output=512, pad=0,
weight_filler=dict(type='xavier'))
n.qatt_relu = L.ReLU(n.qatt_conv1)
n.qatt_conv2 = L.Convolution(n.qatt_relu, kernel_size=1, stride=1, num_output=config.NUM_QUESTION_GLIMPSE, pad=0,
weight_filler=dict(type='xavier'))
n.qatt_reshape = L.Reshape(n.qatt_conv2, reshape_param=dict(shape=dict(dim=[-1,config.NUM_QUESTION_GLIMPSE,config.MAX_WORDS_IN_QUESTION,1]))) # N*NUM_QUESTION_GLIMPSE*15
n.qatt_softmax = L.Softmax(n.qatt_reshape, axis=2)
qatt_maps = L.Slice(n.qatt_softmax,ntop=config.NUM_QUESTION_GLIMPSE,slice_param={'axis':1})
dummy_lstm = L.DummyData(shape=dict(dim=[batchsize, 1]), data_filler=dict(type='constant', value=1), ntop=1)
qatt_feature_list = []
for i in xrange(config.NUM_QUESTION_GLIMPSE):
if config.NUM_QUESTION_GLIMPSE == 1:
n.__setattr__('qatt_feat%d'%i, L.SoftAttention(n.lstm1_resh2, qatt_maps, dummy_lstm))
else:
n.__setattr__('qatt_feat%d'%i, L.SoftAttention(n.lstm1_resh2, qatt_maps[i], dummy_lstm))
qatt_feature_list.append(n.__getattr__('qatt_feat%d'%i))
n.qatt_feat_concat = L.Concat(*qatt_feature_list)
'''
Image Attention with MFB
'''
n.q_feat_resh = L.Reshape(n.qatt_feat_concat,reshape_param=dict(shape=dict(dim=[0,-1,1,1])))
n.i_feat_resh = L.Reshape(n.img_feature,reshape_param=dict(shape=dict(dim=[0,-1,config.IMG_FEAT_WIDTH,config.IMG_FEAT_WIDTH])))
n.iatt_q_proj = L.InnerProduct(n.q_feat_resh, num_output = config.JOINT_EMB_SIZE,
weight_filler=dict(type='xavier'))
n.iatt_q_resh = L.Reshape(n.iatt_q_proj, reshape_param=dict(shape=dict(dim=[-1,config.JOINT_EMB_SIZE,1,1])))
n.iatt_q_tile1 = L.Tile(n.iatt_q_resh, axis=2, tiles=config.IMG_FEAT_WIDTH)
n.iatt_q_tile2 = L.Tile(n.iatt_q_tile1, axis=3, tiles=config.IMG_FEAT_WIDTH)
n.iatt_i_conv = L.Convolution(n.i_feat_resh, kernel_size=1, stride=1, num_output=config.JOINT_EMB_SIZE, pad=0,
weight_filler=dict(type='xavier'))
n.iatt_i_resh1 = L.Reshape(n.iatt_i_conv, reshape_param=dict(shape=dict(dim=[-1,config.JOINT_EMB_SIZE,
config.IMG_FEAT_WIDTH,config.IMG_FEAT_WIDTH])))
n.iatt_iq_eltwise = L.Eltwise(n.iatt_q_tile2, n.iatt_i_resh1, eltwise_param=dict(operation=0))
n.iatt_iq_droped = L.Dropout(n.iatt_iq_eltwise, dropout_param={'dropout_ratio':config.MFB_DROPOUT_RATIO})
n.iatt_iq_resh2 = L.Reshape(n.iatt_iq_droped, reshape_param=dict(shape=dict(dim=[-1,config.JOINT_EMB_SIZE,config.IMG_FEAT_SIZE,1])))
n.iatt_iq_permute1 = L.Permute(n.iatt_iq_resh2, permute_param=dict(order=[0,2,1,3]))
n.iatt_iq_resh2 = L.Reshape(n.iatt_iq_permute1, reshape_param=dict(shape=dict(dim=[-1,config.IMG_FEAT_SIZE,
config.MFB_OUT_DIM,config.MFB_FACTOR_NUM])))
n.iatt_iq_sumpool = L.Pooling(n.iatt_iq_resh2, pool=P.Pooling.SUM, \
pooling_param=dict(kernel_w=config.MFB_FACTOR_NUM, kernel_h=1))
n.iatt_iq_permute2 = L.Permute(n.iatt_iq_sumpool, permute_param=dict(order=[0,2,1,3]))
n.iatt_iq_sqrt = L.SignedSqrt(n.iatt_iq_permute2)
n.iatt_iq_l2 = L.L2Normalize(n.iatt_iq_sqrt)
## 2 conv layers 1000 -> 512 -> 2
n.iatt_conv1 = L.Convolution(n.iatt_iq_l2, kernel_size=1, stride=1, num_output=512, pad=0,
weight_filler=dict(type='xavier'))
n.iatt_relu = L.ReLU(n.iatt_conv1)
n.iatt_conv2 = L.Convolution(n.iatt_relu, kernel_size=1, stride=1, num_output=config.NUM_IMG_GLIMPSE, pad=0,
weight_filler=dict(type='xavier'))
n.iatt_resh = L.Reshape(n.iatt_conv2, reshape_param=dict(shape=dict(dim=[-1,config.NUM_IMG_GLIMPSE,config.IMG_FEAT_SIZE])))
n.iatt_softmax = L.Softmax(n.iatt_resh, axis=2)
n.iatt_softmax_resh = L.Reshape(n.iatt_softmax,reshape_param=dict(shape=dict(dim=[-1,config.NUM_IMG_GLIMPSE,config.IMG_FEAT_WIDTH,config.IMG_FEAT_WIDTH])))
iatt_maps = L.Slice(n.iatt_softmax_resh, ntop=config.NUM_IMG_GLIMPSE,slice_param={'axis':1})
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]), data_filler=dict(type='constant', value=1), ntop=1)
iatt_feature_list = []
for i in xrange(config.NUM_IMG_GLIMPSE):
if config.NUM_IMG_GLIMPSE == 1:
n.__setattr__('iatt_feat%d'%i, L.SoftAttention(n.i_feat_resh, iatt_maps, dummy))
else:
n.__setattr__('iatt_feat%d'%i, L.SoftAttention(n.i_feat_resh, iatt_maps[i], dummy))
n.__setattr__('iatt_feat%d_resh'%i, L.Reshape(n.__getattr__('iatt_feat%d'%i), \
reshape_param=dict(shape=dict(dim=[0,-1]))))
iatt_feature_list.append(n.__getattr__('iatt_feat%d_resh'%i))
n.iatt_feat_concat = L.Concat(*iatt_feature_list)
n.iatt_feat_concat_resh = L.Reshape(n.iatt_feat_concat, reshape_param=dict(shape=dict(dim=[0,-1,1,1])))
'''
Fine-grained Image-Question MFB fusion
'''
n.mfb_q_proj = L.InnerProduct(n.q_feat_resh, num_output=config.JOINT_EMB_SIZE,
weight_filler=dict(type='xavier'))
n.mfb_i_proj = L.InnerProduct(n.iatt_feat_concat_resh, num_output=config.JOINT_EMB_SIZE,
weight_filler=dict(type='xavier'))
n.mfb_iq_eltwise = L.Eltwise(n.mfb_q_proj, n.mfb_i_proj, eltwise_param=dict(operation=0))
n.mfb_iq_drop = L.Dropout(n.mfb_iq_eltwise, dropout_param={'dropout_ratio':config.MFB_DROPOUT_RATIO})
n.mfb_iq_resh = L.Reshape(n.mfb_iq_drop, reshape_param=dict(shape=dict(dim=[-1,1,config.MFB_OUT_DIM,config.MFB_FACTOR_NUM])))
n.mfb_iq_sumpool = L.Pooling(n.mfb_iq_resh, pool=P.Pooling.SUM, \
pooling_param=dict(kernel_w=config.MFB_FACTOR_NUM, kernel_h=1))
n.mfb_out = L.Reshape(n.mfb_iq_sumpool,\
reshape_param=dict(shape=dict(dim=[-1,config.MFB_OUT_DIM])))
n.mfb_sign_sqrt = L.SignedSqrt(n.mfb_out)
n.mfb_l2 = L.L2Normalize(n.mfb_sign_sqrt)
n.prediction = L.InnerProduct(n.mfb_l2, num_output=config.NUM_OUTPUT_UNITS,
weight_filler=dict(type='xavier'))
if mode == 'val':
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
else:
n.loss = L.SoftmaxKLDLoss(n.prediction, n.label)
return n.to_proto()
def qlstm(mode, batchsize, T, T_c, question_c_vocab_size, question_vocab_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
n.data, n.cont, n.data1, n.cont1, n.img_feature, n.label = L.Python(\
module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=6)#5 )
# char embedding
n.embed_c = L.Embed(n.data1, input_dim=question_c_vocab_size, num_output=15, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08))
n.embed_c_scale = L.Scale(n.embed_c,
n.cont1,
scale_param=dict(dict(axis=0)))
n.embed_c_scale_resh = L.Reshape(
n.embed_c_scale,
reshape_param=dict(shape=dict(dim=[batchsize, 1, T_c *
T, -1]))) # N x 1 x T_c x d_c
tops = L.Slice(n.embed_c_scale_resh, ntop=T, slice_param={'axis': 2})
for i in xrange(T):
n.__setattr__('slice_' + str(i + 1), tops[int(i)])
# char conv
n.c_feature_1 = L.Convolution(
n.slice_1,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_2 = L.Convolution(
n.slice_2,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_3 = L.Convolution(
n.slice_3,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_4 = L.Convolution(
n.slice_4,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_5 = L.Convolution(
n.slice_5,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_6 = L.Convolution(
n.slice_6,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_7 = L.Convolution(
n.slice_7,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_8 = L.Convolution(
n.slice_8,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_9 = L.Convolution(
n.slice_9,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_10 = L.Convolution(
n.slice_10,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_11 = L.Convolution(
n.slice_11,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_12 = L.Convolution(
n.slice_12,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_13 = L.Convolution(
n.slice_13,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_14 = L.Convolution(
n.slice_14,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_15 = L.Convolution(
n.slice_15,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_16 = L.Convolution(
n.slice_16,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_17 = L.Convolution(
n.slice_17,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_18 = L.Convolution(
n.slice_18,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_19 = L.Convolution(
n.slice_19,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_20 = L.Convolution(
n.slice_20,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_21 = L.Convolution(
n.slice_21,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_feature_22 = L.Convolution(
n.slice_22,
convolution_param={
'kernel_h': 3,
'kernel_w': 15,
'stride': 1,
'num_output': 150,
'pad_h': 1,
'pad_w': 0,
'weight_filler': dict(type='xavier')
},
param=[dict(name="conv_c_w"),
dict(name="conv_c_b")])
n.c_vec_1 = L.Pooling(n.c_feature_1,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_2 = L.Pooling(n.c_feature_2,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_3 = L.Pooling(n.c_feature_3,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_4 = L.Pooling(n.c_feature_4,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_5 = L.Pooling(n.c_feature_5,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_6 = L.Pooling(n.c_feature_6,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_7 = L.Pooling(n.c_feature_7,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_8 = L.Pooling(n.c_feature_8,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_9 = L.Pooling(n.c_feature_9,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_10 = L.Pooling(n.c_feature_10,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_11 = L.Pooling(n.c_feature_11,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_12 = L.Pooling(n.c_feature_12,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_13 = L.Pooling(n.c_feature_13,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_14 = L.Pooling(n.c_feature_14,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_15 = L.Pooling(n.c_feature_15,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_16 = L.Pooling(n.c_feature_16,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_17 = L.Pooling(n.c_feature_17,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_18 = L.Pooling(n.c_feature_18,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_19 = L.Pooling(n.c_feature_19,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_20 = L.Pooling(n.c_feature_20,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_21 = L.Pooling(n.c_feature_21,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_vec_22 = L.Pooling(n.c_feature_22,
kernel_h=T_c,
kernel_w=1,
stride=T_c,
pool=P.Pooling.MAX)
n.c_embed_1 = L.Reshape(
n.c_vec_1, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_2 = L.Reshape(
n.c_vec_2, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_3 = L.Reshape(
n.c_vec_3, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_4 = L.Reshape(
n.c_vec_4, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_5 = L.Reshape(
n.c_vec_5, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_6 = L.Reshape(
n.c_vec_6, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_7 = L.Reshape(
n.c_vec_7, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_8 = L.Reshape(
n.c_vec_8, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_9 = L.Reshape(
n.c_vec_9, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_10 = L.Reshape(
n.c_vec_10, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_11 = L.Reshape(
n.c_vec_11, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_12 = L.Reshape(
n.c_vec_12, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_13 = L.Reshape(
n.c_vec_13, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_14 = L.Reshape(
n.c_vec_14, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_15 = L.Reshape(
n.c_vec_15, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_16 = L.Reshape(
n.c_vec_16, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_17 = L.Reshape(
n.c_vec_17, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_18 = L.Reshape(
n.c_vec_18, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_19 = L.Reshape(
n.c_vec_19, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_20 = L.Reshape(
n.c_vec_20, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_21 = L.Reshape(
n.c_vec_21, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
n.c_embed_22 = L.Reshape(
n.c_vec_22, reshape_param=dict(shape=dict(dim=[batchsize, 1, 150])))
concat_c_embed = [n.c_embed_1, n.c_embed_2, n.c_embed_3, n.c_embed_4, n.c_embed_5, n.c_embed_6, n.c_embed_7, n.c_embed_8, n.c_embed_9, n.c_embed_10,\
n.c_embed_11, n.c_embed_12, n.c_embed_13, n.c_embed_14, n.c_embed_15, n.c_embed_16, n.c_embed_17, n.c_embed_18, n.c_embed_19, n.c_embed_20, n.c_embed_21, n.c_embed_22]
n.concat_char_embed = L.Concat(*concat_c_embed,
concat_param={'axis': 1}) # N x T x d_c
# word embedding
n.embed_w = L.Embed(n.data, input_dim=question_vocab_size, num_output=150, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08)) # N x T x d_w
# combine word and char embedding
concat_word_embed = [n.embed_w, n.concat_char_embed]
n.concat_embed = L.Concat(*concat_word_embed,
concat_param={'axis': 2}) # N x T x (d_c+d_w)
n.embed_scale = L.Scale(n.concat_embed,
n.cont,
scale_param=dict(dict(axis=0)))
n.embed_scale_resh = L.Reshape(
n.embed_scale,
reshape_param=dict(shape=dict(
dim=[batchsize, 1, T, -1]))) # N x 1 x T x (d_c+d_w)
# n.glove_scale = L.Scale(n.glove, n.cont, scale_param=dict(dict(axis=0)))
# n.glove_scale_resh = L.Reshape(n.glove_scale,\
# reshape_param=dict(\
# shape=dict(dim=[batchsize,1,T,300])))
# concat_word_embed = [n.embed_scale_resh, n.glove_scale_resh]
# n.concat_embed = L.Concat(*concat_word_embed, concat_param={'axis': 1}) # N x 2 x T x 300
# convolution
n.word_feature_2 = L.Convolution(
n.embed_scale_resh,
kernel_h=2,
kernel_w=300,
stride=1,
num_output=512,
pad_h=1,
pad_w=0,
weight_filler=dict(type='xavier')) # N x C x ? x 1
n.word_feature_3 = L.Convolution(n.embed_scale_resh,
kernel_h=3,
kernel_w=300,
stride=1,
num_output=512,
pad_h=2,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_4 = L.Convolution(n.embed_scale_resh,
kernel_h=4,
kernel_w=300,
stride=1,
num_output=512,
pad_h=3,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_5 = L.Convolution(n.embed_scale_resh,
kernel_h=5,
kernel_w=300,
stride=1,
num_output=512,
pad_h=4,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_relu_2 = L.ReLU(n.word_feature_2)
n.word_relu_3 = L.ReLU(n.word_feature_3)
n.word_relu_4 = L.ReLU(n.word_feature_4)
n.word_relu_5 = L.ReLU(n.word_feature_5)
n.word_vec_2 = L.Pooling(n.word_relu_2,
kernel_h=T + 1,
kernel_w=1,
stride=T + 1,
pool=P.Pooling.MAX) # N x C x 1 x 1
n.word_vec_3 = L.Pooling(n.word_relu_3,
kernel_h=T + 2,
kernel_w=1,
stride=T + 2,
pool=P.Pooling.MAX)
n.word_vec_4 = L.Pooling(n.word_relu_4,
kernel_h=T + 3,
kernel_w=1,
stride=T + 3,
pool=P.Pooling.MAX)
n.word_vec_5 = L.Pooling(n.word_relu_5,
kernel_h=T + 4,
kernel_w=1,
stride=T + 4,
pool=P.Pooling.MAX)
word_vec = [n.word_vec_2, n.word_vec_3, n.word_vec_4, n.word_vec_5]
n.concat_vec = L.Concat(*word_vec, concat_param={'axis':
1}) # N x 4C x 1 x 1
n.concat_vec_dropped = L.Dropout(n.concat_vec,
dropout_param={'dropout_ratio': 0.5})
n.q_emb_tanh_droped_resh_tiled_1 = L.Tile(n.concat_vec_dropped,
axis=2,
tiles=14)
n.q_emb_tanh_droped_resh_tiled = L.Tile(n.q_emb_tanh_droped_resh_tiled_1,
axis=3,
tiles=14)
n.i_emb_tanh_droped_resh = L.Reshape(
n.img_feature, reshape_param=dict(shape=dict(dim=[-1, 2048, 14, 14])))
n.blcf = L.CompactBilinear(n.q_emb_tanh_droped_resh_tiled,
n.i_emb_tanh_droped_resh,
compact_bilinear_param=dict(num_output=16000,
sum_pool=False))
n.blcf_sign_sqrt = L.SignedSqrt(n.blcf)
n.blcf_sign_sqrt_l2 = L.L2Normalize(n.blcf_sign_sqrt)
n.blcf_droped = L.Dropout(n.blcf_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
# multi-channel attention
n.att_conv1 = L.Convolution(n.blcf_droped,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu,
kernel_size=1,
stride=1,
num_output=2,
pad=0,
weight_filler=dict(type='xavier'))
n.att_reshaped = L.Reshape(
n.att_conv2, reshape_param=dict(shape=dict(dim=[-1, 2, 14 * 14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att = L.Reshape(n.att_softmax,
reshape_param=dict(shape=dict(dim=[-1, 2, 14, 14])))
att_maps = L.Slice(n.att, ntop=2, slice_param={'axis': 1})
n.att_map0 = att_maps[0]
n.att_map1 = att_maps[1]
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.att_feature0 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map0,
dummy)
n.att_feature1 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map1,
dummy)
n.att_feature0_resh = L.Reshape(
n.att_feature0, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature1_resh = L.Reshape(
n.att_feature1, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature = L.Concat(n.att_feature0_resh, n.att_feature1_resh)
# merge attention and lstm with compact bilinear pooling
n.att_feature_resh = L.Reshape(
n.att_feature, reshape_param=dict(shape=dict(dim=[-1, 4096, 1, 1])))
#n.lstm_12_resh = L.Reshape(n.lstm_12, reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.bc_att_lstm = L.CompactBilinear(n.att_feature_resh,
n.concat_vec_dropped,
compact_bilinear_param=dict(
num_output=16000, sum_pool=False))
n.bc_sign_sqrt = L.SignedSqrt(n.bc_att_lstm)
n.bc_sign_sqrt_l2 = L.L2Normalize(n.bc_sign_sqrt)
n.bc_dropped = L.Dropout(n.bc_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
n.bc_dropped_resh = L.Reshape(
n.bc_dropped, reshape_param=dict(shape=dict(dim=[-1, 16000])))
n.prediction = L.InnerProduct(n.bc_dropped_resh,
num_output=3000,
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
return n.to_proto()
def generator_proto(mode, batchsize, T, exp_T, question_vocab_size, exp_vocab_size, use_gt=True):
n = caffe.NetSpec()
mode_str = json.dumps({'mode':mode, 'batchsize':batchsize})
n.data, n.cont, n.img_feature, n.label, n.exp, n.exp_out, n.exp_cont_1, n.exp_cont_2 = \
L.Python(module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=8)
n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08), param=fixed_weights)
n.embed = L.TanH(n.embed_ba)
# LSTM1
n.lstm1 = L.LSTM(\
n.embed, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)),
param=fixed_weights_lstm)
tops1 = L.Slice(n.lstm1, ntop=T, slice_param={'axis':0})
for i in range(T-1):
n.__setattr__('slice_first'+str(i), tops1[int(i)])
n.__setattr__('silence_data_first'+str(i), L.Silence(tops1[int(i)],ntop=0))
n.lstm1_out = tops1[T-1]
n.lstm1_reshaped = L.Reshape(n.lstm1_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm1_reshaped_droped = L.Dropout(n.lstm1_reshaped,dropout_param={'dropout_ratio':0.3})
n.lstm1_droped = L.Dropout(n.lstm1,dropout_param={'dropout_ratio':0.3})
# LSTM2
n.lstm2 = L.LSTM(\
n.lstm1_droped, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)),
param=fixed_weights_lstm)
tops2 = L.Slice(n.lstm2, ntop=T, slice_param={'axis':0})
for i in range(T-1):
n.__setattr__('slice_second'+str(i), tops2[int(i)])
n.__setattr__('silence_data_second'+str(i), L.Silence(tops2[int(i)],ntop=0))
n.lstm2_out = tops2[T-1]
n.lstm2_reshaped = L.Reshape(n.lstm2_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm2_reshaped_droped = L.Dropout(n.lstm2_reshaped,dropout_param={'dropout_ratio':0.3})
concat_botom = [n.lstm1_reshaped_droped, n.lstm2_reshaped_droped]
n.lstm_12 = L.Concat(*concat_botom)
# Tile question feature
n.q_emb_resh = L.Reshape(n.lstm_12, reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.q_emb_tiled_1 = L.Tile(n.q_emb_resh, axis=2, tiles=14)
n.q_emb_resh_tiled = L.Tile(n.q_emb_tiled_1, axis=3, tiles=14)
# Embed image feature
n.i_emb = L.Convolution(n.img_feature, kernel_size=1, stride=1,
num_output=2048, pad=0, weight_filler=dict(type='xavier'),
param=fixed_weights)
# Eltwise product and normalization
n.eltwise = L.Eltwise(n.q_emb_resh_tiled, n.i_emb, eltwise_param={'operation': P.Eltwise.PROD})
n.eltwise_sqrt = L.SignedSqrt(n.eltwise)
n.eltwise_l2 = L.L2Normalize(n.eltwise_sqrt)
n.eltwise_drop = L.Dropout(n.eltwise_l2, dropout_param={'dropout_ratio': 0.3})
# Attention for VQA
n.att_conv1 = L.Convolution(n.eltwise_drop, kernel_size=1, stride=1, num_output=512, pad=0, weight_filler=dict(type='xavier'), param=fixed_weights)
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu, kernel_size=1, stride=1, num_output=1, pad=0, weight_filler=dict(type='xavier'), param=fixed_weights)
n.att_reshaped = L.Reshape(n.att_conv2,reshape_param=dict(shape=dict(dim=[-1,1,14*14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att_map = L.Reshape(n.att_softmax,reshape_param=dict(shape=dict(dim=[-1,1,14,14])))
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]), data_filler=dict(type='constant', value=1), ntop=1)
n.att_feature = L.SoftAttention(n.img_feature, n.att_map, dummy)
n.att_feature_resh = L.Reshape(n.att_feature, reshape_param=dict(shape=dict(dim=[-1,2048])))
# eltwise product + normalization again for VQA
n.i_emb2 = L.InnerProduct(n.att_feature_resh, num_output=2048, weight_filler=dict(type='xavier'), param=fixed_weights)
n.eltwise2 = L.Eltwise(n.lstm_12, n.i_emb2, eltwise_param={'operation': P.Eltwise.PROD})
n.eltwise2_sqrt = L.SignedSqrt(n.eltwise2)
n.eltwise2_l2 = L.L2Normalize(n.eltwise2_sqrt)
n.eltwise2_drop = L.Dropout(n.eltwise2_l2, dropout_param={'dropout_ratio': 0.3})
n.prediction = L.InnerProduct(n.eltwise2_drop, num_output=3000, weight_filler=dict(type='xavier'), param=fixed_weights)
# Take GT answer or Take the logits of the VQA model and get predicted answer to embed
if use_gt:
n.exp_emb_ans = L.Embed(n.label, input_dim=3000, num_output=300,
weight_filler=dict(type='uniform', min=-0.08, max=0.08))
else:
n.vqa_ans = L.ArgMax(n.prediction, axis=1)
n.exp_emb_ans = L.Embed(n.vqa_ans, input_dim=3000, num_output=300,
weight_filler=dict(type='uniform', min=-0.08, max=0.08))
n.exp_emb_ans_tanh = L.TanH(n.exp_emb_ans)
n.exp_emb_ans2 = L.InnerProduct(n.exp_emb_ans_tanh, num_output=2048, weight_filler=dict(type='xavier'))
# Merge VQA answer and visual+textual feature
n.exp_emb_resh = L.Reshape(n.exp_emb_ans2, reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.exp_emb_tiled_1 = L.Tile(n.exp_emb_resh, axis=2, tiles=14)
n.exp_emb_tiled = L.Tile(n.exp_emb_tiled_1, axis=3, tiles=14)
#n.exp_eltwise = L.Eltwise(n.eltwise_drop, n.exp_emb_tiled, eltwise_param={'operation': P.Eltwise.PROD})
n.eltwise_emb = L.Convolution(n.eltwise, kernel_size=1, stride=1, num_output=2048, pad=0, weight_filler=dict(type='xavier'))
n.exp_eltwise = L.Eltwise(n.eltwise_emb, n.exp_emb_tiled, eltwise_param={'operation': P.Eltwise.PROD})
n.exp_eltwise_sqrt = L.SignedSqrt(n.exp_eltwise)
n.exp_eltwise_l2 = L.L2Normalize(n.exp_eltwise_sqrt)
n.exp_eltwise_drop = L.Dropout(n.exp_eltwise_l2, dropout_param={'dropout_ratio': 0.3})
# Attention for Explanation
n.exp_att_conv1 = L.Convolution(n.exp_eltwise_drop, kernel_size=1, stride=1, num_output=512, pad=0, weight_filler=dict(type='xavier'))
n.exp_att_conv1_relu = L.ReLU(n.exp_att_conv1)
n.exp_att_conv2 = L.Convolution(n.exp_att_conv1_relu, kernel_size=1, stride=1, num_output=1, pad=0, weight_filler=dict(type='xavier'))
n.exp_att_reshaped = L.Reshape(n.exp_att_conv2,reshape_param=dict(shape=dict(dim=[-1,1,14*14])))
n.exp_att_softmax = L.Softmax(n.exp_att_reshaped, axis=2)
n.exp_att_map = L.Reshape(n.exp_att_softmax,reshape_param=dict(shape=dict(dim=[-1,1,14,14])))
exp_dummy = L.DummyData(shape=dict(dim=[batchsize, 1]), data_filler=dict(type='constant', value=1), ntop=1)
n.exp_att_feature_prev = L.SoftAttention(n.img_feature, n.exp_att_map, exp_dummy)
n.exp_att_feature_resh = L.Reshape(n.exp_att_feature_prev, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.exp_att_feature_embed = L.InnerProduct(n.exp_att_feature_resh, num_output=2048, weight_filler=dict(type='xavier'))
n.exp_lstm12_embed = L.InnerProduct(n.lstm_12, num_output=2048, weight_filler=dict(type='xavier'))
n.exp_eltwise2 = L.Eltwise(n.exp_lstm12_embed, n.exp_att_feature_embed, eltwise_param={'operation': P.Eltwise.PROD})
n.exp_att_feature = L.Eltwise(n.exp_emb_ans2, n.exp_eltwise2, eltwise_param={'operation': P.Eltwise.PROD})
n.silence_exp_att = L.Silence(n.exp_att_feature, ntop=0)
return n.to_proto()
def qlstm(mode, batchsize, T, question_vocab_size):
#prototxt 없이 network 생성시 사용
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
#지정된 Python 모듈 형식
#https://stackoverflow.com/questions/41344168/what-is-a-python-layer-in-caffe
#해당 클래스를 바탕으로 Layer를 생성하며
#리턴된 변수에 값을 채워넣으면 자동으로 Run된다.
#여기서 만들어진 Class 내부에서 실질적인 databatch load가 이루어짐.
#Glove = Global vectors for word representation
#https://www.aclweb.org/anthology/D14-1162
#Pretrained 된 GloveVector를 Concat에 사용.
#img_feature는 이미 Resnet512 통과후 L2를 적용한 Preprocessing이 끝난 상태의 Feature Vector.
n.data, n.cont, n.img_feature, n.label, n.glove = L.Python(\
module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=5 )
#module = python 파일이름
#layer = layer형식이 맞춰진 python class
#param_str = json으로 Data Load시 사용된 파라미터, 내부 class에 self.param_str = modestr 로 저장된다
#ntop = 각 setup , forward backward의 top 변수의 크기
#보통 textual Embed의 뜻은 => texture -> number
#Embed 3000개의 Vector종류를
#300개로 compact하게 표현함
n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08))
#Tanh 적용
n.embed = L.TanH(n.embed_ba)
#Glove Data와 Concat
concat_word_embed = [n.embed, n.glove]
n.concat_embed = L.Concat(*concat_word_embed,
concat_param={'axis': 2}) # T x N x 600
# LSTM1
n.lstm1 = L.LSTM(\
n.concat_embed, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
tops1 = L.Slice(n.lstm1, ntop=T, slice_param={'axis': 0})
for i in xrange(T - 1):
n.__setattr__('slice_first' + str(i), tops1[int(i)])
n.__setattr__('silence_data_first' + str(i),
L.Silence(tops1[int(i)], ntop=0))
n.lstm1_out = tops1[T - 1]
n.lstm1_reshaped = L.Reshape(n.lstm1_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm1_reshaped_droped = L.Dropout(n.lstm1_reshaped,
dropout_param={'dropout_ratio': 0.3})
n.lstm1_droped = L.Dropout(n.lstm1, dropout_param={'dropout_ratio': 0.3})
# LSTM2
n.lstm2 = L.LSTM(\
n.lstm1_droped, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
tops2 = L.Slice(n.lstm2, ntop=T, slice_param={'axis': 0})
#https://www.programcreek.com/python/example/107865/caffe.NetSpec 참조.
# give top2[~] the name specified by argument `slice_second`
#변수 부여 기능
for i in xrange(T - 1):
n.__setattr__('slice_second' + str(i), tops2[int(i)])
n.__setattr__('silence_data_second' + str(i),
L.Silence(tops2[int(i)], ntop=0))
#마지막 LSTM output을 사용.
n.lstm2_out = tops2[T - 1]
n.lstm2_reshaped = L.Reshape(n.lstm2_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm2_reshaped_droped = L.Dropout(n.lstm2_reshaped,
dropout_param={'dropout_ratio': 0.3})
concat_botom = [n.lstm1_reshaped_droped, n.lstm2_reshaped_droped]
n.lstm_12 = L.Concat(*concat_botom)
#lstm1의 output => 1024 reshape뒤 dropout
#lstm2의 output => 1024 reshape뒤 dropout
#concat
n.q_emb_tanh_droped_resh = L.Reshape(
n.lstm_12, reshape_param=dict(shape=dict(dim=[-1, 2048, 1, 1])))
#L.Tile 차원을 자동으로 안맞춰주므로 차원맞춤 함수. 2048,1 (tile=14, axis=1) =>2048,14
n.q_emb_tanh_droped_resh_tiled_1 = L.Tile(n.q_emb_tanh_droped_resh,
axis=2,
tiles=14)
n.q_emb_tanh_droped_resh_tiled = L.Tile(n.q_emb_tanh_droped_resh_tiled_1,
axis=3,
tiles=14)
n.i_emb_tanh_droped_resh = L.Reshape(
n.img_feature, reshape_param=dict(shape=dict(dim=[-1, 2048, 14, 14])))
n.blcf = L.CompactBilinear(n.q_emb_tanh_droped_resh_tiled,
n.i_emb_tanh_droped_resh,
compact_bilinear_param=dict(num_output=16000,
sum_pool=False))
n.blcf_sign_sqrt = L.SignedSqrt(n.blcf)
n.blcf_sign_sqrt_l2 = L.L2Normalize(n.blcf_sign_sqrt)
#논문 그림과 달리 Dropout 추가
n.blcf_droped = L.Dropout(n.blcf_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
# multi-channel attention
n.att_conv1 = L.Convolution(n.blcf_droped,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.att_conv1_relu = L.ReLU(n.att_conv1)
#논문 그림과 달리 output dim이 2
n.att_conv2 = L.Convolution(n.att_conv1_relu,
kernel_size=1,
stride=1,
num_output=2,
pad=0,
weight_filler=dict(type='xavier'))
n.att_reshaped = L.Reshape(
n.att_conv2, reshape_param=dict(shape=dict(dim=[-1, 2, 14 * 14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
#softmax로 attentionmap 생성
#14x14 Softmax map이 2개 생성
n.att = L.Reshape(n.att_softmax,
reshape_param=dict(shape=dict(dim=[-1, 2, 14, 14])))
#두가지 att_map을 각각 Slice
att_maps = L.Slice(n.att, ntop=2, slice_param={'axis': 1})
n.att_map0 = att_maps[0]
n.att_map1 = att_maps[1]
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.att_feature0 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map0,
dummy)
n.att_feature1 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map1,
dummy)
n.att_feature0_resh = L.Reshape(
n.att_feature0, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature1_resh = L.Reshape(
n.att_feature1, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature = L.Concat(n.att_feature0_resh, n.att_feature1_resh)
#각각 ATT를 곱한값을 연산뒤 Concat한다.
# merge attention and lstm with compact bilinear pooling
n.att_feature_resh = L.Reshape(
n.att_feature, reshape_param=dict(shape=dict(dim=[-1, 4096, 1, 1])))
#그뒤 4096으로 Reshape
n.lstm_12_resh = L.Reshape(
n.lstm_12, reshape_param=dict(shape=dict(dim=[-1, 2048, 1, 1])))
#논문과 달리 가로축 세로축 inputVector크기가 다름
#논문 2048 2048
#코드 4096 2048
n.bc_att_lstm = L.CompactBilinear(n.att_feature_resh,
n.lstm_12_resh,
compact_bilinear_param=dict(
num_output=16000, sum_pool=False))
#SignedSqrt
n.bc_sign_sqrt = L.SignedSqrt(n.bc_att_lstm)
#L2_Normalize
n.bc_sign_sqrt_l2 = L.L2Normalize(n.bc_sign_sqrt)
#Dropout
n.bc_dropped = L.Dropout(n.bc_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
n.bc_dropped_resh = L.Reshape(
n.bc_dropped, reshape_param=dict(shape=dict(dim=[-1, 16000])))
#FullyConnected
n.prediction = L.InnerProduct(n.bc_dropped_resh,
num_output=3000,
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
return n.to_proto()
def mfh_baseline(mode, batchsize, T, question_vocab_size, folder):
n = caffe.NetSpec()
mode_str = json.dumps({'mode':mode, 'batchsize':batchsize,'folder':folder})
if mode == 'val':
n.data, n.cont, n.img_feature, n.label = L.Python( \
module='vqa_data_layer', layer='VQADataProviderLayer', \
param_str=mode_str, ntop=4 )
else:
n.data, n.cont, n.img_feature, n.label = L.Python(\
module='vqa_data_layer_kld', layer='VQADataProviderLayer', \
param_str=mode_str, ntop=4 )
n.embed = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
weight_filler=dict(type='xavier'))
n.embed_tanh = L.TanH(n.embed)
# LSTM
#n.lstm1 = L.LSTM(\
# n.embed_tanh, n.cont,\
# recurrent_param=dict(\
# num_output=config.LSTM_UNIT_NUM,\
# weight_filler=dict(type='xavier')))
#tops1 = L.Slice(n.lstm1, ntop=config.MAX_WORDS_IN_QUESTION, slice_param={'axis':0})
#for i in xrange(config.MAX_WORDS_IN_QUESTION-1):
# n.__setattr__('slice_first'+str(i), tops1[int(i)])
# n.__setattr__('silence_data_first'+str(i), L.Silence(tops1[int(i)],ntop=0))
#n.lstm1_out = tops1[config.MAX_WORDS_IN_QUESTION-1]
#n.lstm1_reshaped = L.Reshape(n.lstm1_out,\
# reshape_param=dict(\
# shape=dict(dim=[-1,1024])))
#n.q_feat = L.Dropout(n.lstm1_reshaped,dropout_param={'dropout_ratio':config.LSTM_DROPOUT_RATIO})
n.lstm1 = L.LSTM(\
n.embed, n.cont,\
recurrent_param=dict(\
num_output=config.LSTM_UNIT_NUM,\
weight_filler=dict(type='xavier')))
tops1 = L.Slice(n.lstm1, ntop=config.MAX_WORDS_IN_QUESTION, slice_param={'axis':0})
for i in xrange(config.MAX_WORDS_IN_QUESTION-1):
n.__setattr__('slice_first'+str(i), tops1[int(i)])
n.__setattr__('silence_data_first'+str(i), L.Silence(tops1[int(i)],ntop=0))
n.lstm1_out = tops1[config.MAX_WORDS_IN_QUESTION-1]
n.lstm1_reshaped = L.Reshape(n.lstm1_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm1_reshaped_droped = L.Dropout(n.lstm1_reshaped,dropout_param={'dropout_ratio':config.LSTM_DROPOUT_RATIO})
n.lstm1_droped = L.Dropout(n.lstm1,dropout_param={'dropout_ratio':config.LSTM_DROPOUT_RATIO})
# LSTM2
n.lstm2 = L.LSTM(\
n.lstm1_droped, n.cont,\
recurrent_param=dict(\
num_output=config.LSTM_UNIT_NUM,
weight_filler=dict(type='xavier')))
tops2 = L.Slice(n.lstm2, ntop=config.MAX_WORDS_IN_QUESTION, slice_param={'axis':0})
for i in xrange(config.MAX_WORDS_IN_QUESTION-1):
n.__setattr__('slice_second'+str(i), tops2[int(i)])
n.__setattr__('silence_data_second'+str(i), L.Silence(tops2[int(i)],ntop=0))
n.lstm2_out = tops2[config.MAX_WORDS_IN_QUESTION-1]
n.lstm2_reshaped = L.Reshape(n.lstm2_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm2_reshaped_droped = L.Dropout(n.lstm2_reshaped,dropout_param={'dropout_ratio':config.LSTM_DROPOUT_RATIO})
concat_botom = [n.lstm1_reshaped_droped, n.lstm2_reshaped_droped]
n.q_feat = L.Concat(*concat_botom)
'''
Coarse Image-Question MFH fusion
'''
n.mfb_q_o2_proj = L.InnerProduct(n.q_feat, num_output=config.JOINT_EMB_SIZE,
weight_filler=dict(type='xavier'))
n.mfb_i_o2_proj = L.InnerProduct(n.img_feature, num_output=config.JOINT_EMB_SIZE,
weight_filler=dict(type='xavier'))
n.mfb_iq_o2_eltwise = L.Eltwise(n.mfb_q_o2_proj, n.mfb_i_o2_proj, eltwise_param=dict(operation=0))
n.mfb_iq_o2_drop = L.Dropout(n.mfb_iq_o2_eltwise, dropout_param={'dropout_ratio':config.MFB_DROPOUT_RATIO})
n.mfb_iq_o2_resh = L.Reshape(n.mfb_iq_o2_drop, reshape_param=dict(shape=dict(dim=[-1,1,config.MFB_OUT_DIM,config.MFB_FACTOR_NUM])))
n.mfb_iq_o2_sumpool = L.Pooling(n.mfb_iq_o2_resh, pool=P.Pooling.SUM, \
pooling_param=dict(kernel_w=config.MFB_FACTOR_NUM, kernel_h=1))
n.mfb_o2_out = L.Reshape(n.mfb_iq_o2_sumpool,\
reshape_param=dict(shape=dict(dim=[-1,config.MFB_OUT_DIM])))
n.mfb_o2_sign_sqrt = L.SignedSqrt(n.mfb_o2_out)
n.mfb_o2_l2 = L.L2Normalize(n.mfb_o2_sign_sqrt)
n.mfb_q_o3_proj = L.InnerProduct(n.q_feat, num_output=config.JOINT_EMB_SIZE,
weight_filler=dict(type='xavier'))
n.mfb_i_o3_proj = L.InnerProduct(n.img_feature, num_output=config.JOINT_EMB_SIZE,
weight_filler=dict(type='xavier'))
n.mfb_iq_o3_eltwise = L.Eltwise(n.mfb_q_o3_proj, n.mfb_i_o3_proj,n.mfb_iq_o2_drop, eltwise_param=dict(operation=0))
n.mfb_iq_o3_drop = L.Dropout(n.mfb_iq_o3_eltwise, dropout_param={'dropout_ratio':config.MFB_DROPOUT_RATIO})
n.mfb_iq_o3_resh = L.Reshape(n.mfb_iq_o3_drop, reshape_param=dict(shape=dict(dim=[-1,1,config.MFB_OUT_DIM,config.MFB_FACTOR_NUM])))
n.mfb_iq_o3_sumpool = L.Pooling(n.mfb_iq_o3_resh, pool=P.Pooling.SUM, \
pooling_param=dict(kernel_w=config.MFB_FACTOR_NUM, kernel_h=1))
n.mfb_o3_out = L.Reshape(n.mfb_iq_o3_sumpool,\
reshape_param=dict(shape=dict(dim=[-1,config.MFB_OUT_DIM])))
n.mfb_o3_sign_sqrt = L.SignedSqrt(n.mfb_o3_out)
n.mfb_o3_l2 = L.L2Normalize(n.mfb_o3_sign_sqrt)
n.mfb_o23_l2 = L.Concat(n.mfb_o2_l2,n.mfb_o3_l2)
n.prediction = L.InnerProduct(n.mfb_o23_l2, num_output=config.NUM_OUTPUT_UNITS,
weight_filler=dict(type='xavier'))
if mode == 'val':
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
else:
n.loss = L.SoftmaxKLDLoss(n.prediction, n.label)
return n.to_proto()
def qlstm(mode, batchsize, T, question_vocab_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
# n.data, n.cont, n.img_feature, n.label, n.glove = L.Python(\
# module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=5 )
n.data, n.cont, n.img_feature, n.label = L.Python(\
module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=4 )
# word embedding
n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08))
# n.embed = L.TanH(n.embed_ba)
n.embed_scale = L.Scale(n.embed_ba, n.cont, scale_param=dict(dict(axis=0)))
n.embed_scale_resh = L.Reshape(n.embed_scale,\
reshape_param=dict(\
shape=dict(dim=[batchsize,1,T,300])))
# Convolution
n.word_feature_2 = L.Convolution(n.embed_scale_resh,
kernel_h=2,
kernel_w=300,
stride=1,
num_output=512,
pad_h=1,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_2_g = L.Convolution(n.embed_scale_resh,
kernel_h=2,
kernel_w=300,
stride=1,
num_output=512,
pad_h=1,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_3 = L.Convolution(n.embed_scale_resh,
kernel_h=3,
kernel_w=300,
stride=1,
num_output=512,
pad_h=2,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_3_g = L.Convolution(n.embed_scale_resh,
kernel_h=3,
kernel_w=300,
stride=1,
num_output=512,
pad_h=2,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_4 = L.Convolution(n.embed_scale_resh,
kernel_h=4,
kernel_w=300,
stride=1,
num_output=512,
pad_h=3,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_4_g = L.Convolution(n.embed_scale_resh,
kernel_h=4,
kernel_w=300,
stride=1,
num_output=512,
pad_h=3,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_5 = L.Convolution(n.embed_scale_resh,
kernel_h=5,
kernel_w=300,
stride=1,
num_output=512,
pad_h=4,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_feature_5_g = L.Convolution(n.embed_scale_resh,
kernel_h=5,
kernel_w=300,
stride=1,
num_output=512,
pad_h=4,
pad_w=0,
weight_filler=dict(type='xavier'))
n.word_2_acti = L.TanH(n.word_feature_2)
n.word_3_acti = L.TanH(n.word_feature_3)
n.word_4_acti = L.TanH(n.word_feature_4)
n.word_5_acti = L.TanH(n.word_feature_5)
n.word_2_gate = L.Sigmoid(n.word_feature_2_g)
n.word_3_gate = L.Sigmoid(n.word_feature_3_g)
n.word_4_gate = L.Sigmoid(n.word_feature_4_g)
n.word_5_gate = L.Sigmoid(n.word_feature_5_g)
n.word_2 = L.Eltwise(n.word_2_acti,
n.word_2_gate,
operation=P.Eltwise.PROD)
n.word_3 = L.Eltwise(n.word_3_acti,
n.word_3_gate,
operation=P.Eltwise.PROD)
n.word_4 = L.Eltwise(n.word_4_acti,
n.word_4_gate,
operation=P.Eltwise.PROD)
n.word_5 = L.Eltwise(n.word_5_acti,
n.word_5_gate,
operation=P.Eltwise.PROD)
n.word_vec_2 = L.Pooling(n.word_2,
kernel_h=T + 1,
kernel_w=1,
stride=T + 1,
pool=P.Pooling.MAX)
n.word_vec_3 = L.Pooling(n.word_3,
kernel_h=T + 2,
kernel_w=1,
stride=T + 2,
pool=P.Pooling.MAX)
n.word_vec_4 = L.Pooling(n.word_4,
kernel_h=T + 3,
kernel_w=1,
stride=T + 3,
pool=P.Pooling.MAX)
n.word_vec_5 = L.Pooling(n.word_5,
kernel_h=T + 4,
kernel_w=1,
stride=T + 4,
pool=P.Pooling.MAX)
word_vec = [n.word_vec_2, n.word_vec_3, n.word_vec_4, n.word_vec_5]
n.concat_vec = L.Concat(*word_vec, concat_param={'axis':
1}) # N x 4*d_w x 1 x 1
n.concat_vec_dropped = L.Dropout(n.concat_vec,
dropout_param={'dropout_ratio': 0.5})
n.q_emb_tanh_droped_resh_tiled_1 = L.Tile(n.concat_vec_dropped,
axis=2,
tiles=14)
n.q_emb_tanh_droped_resh_tiled = L.Tile(n.q_emb_tanh_droped_resh_tiled_1,
axis=3,
tiles=14)
n.i_emb_tanh_droped_resh = L.Reshape(
n.img_feature, reshape_param=dict(shape=dict(dim=[-1, 2048, 14, 14])))
n.blcf = L.CompactBilinear(n.q_emb_tanh_droped_resh_tiled,
n.i_emb_tanh_droped_resh,
compact_bilinear_param=dict(num_output=16000,
sum_pool=False))
n.blcf_sign_sqrt = L.SignedSqrt(n.blcf)
n.blcf_sign_sqrt_l2 = L.L2Normalize(n.blcf_sign_sqrt)
n.blcf_droped = L.Dropout(n.blcf_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
# multi-channel attention
n.att_conv1 = L.Convolution(n.blcf_droped,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu,
kernel_size=1,
stride=1,
num_output=2,
pad=0,
weight_filler=dict(type='xavier'))
n.att_reshaped = L.Reshape(
n.att_conv2, reshape_param=dict(shape=dict(dim=[-1, 2, 14 * 14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att = L.Reshape(n.att_softmax,
reshape_param=dict(shape=dict(dim=[-1, 2, 14, 14])))
att_maps = L.Slice(n.att, ntop=2, slice_param={'axis': 1})
n.att_map0 = att_maps[0]
n.att_map1 = att_maps[1]
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.att_feature0 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map0,
dummy)
n.att_feature1 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map1,
dummy)
n.att_feature0_resh = L.Reshape(
n.att_feature0, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature1_resh = L.Reshape(
n.att_feature1, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature = L.Concat(n.att_feature0_resh, n.att_feature1_resh)
# merge attention and lstm with compact bilinear pooling
n.att_feature_resh = L.Reshape(
n.att_feature, reshape_param=dict(shape=dict(dim=[-1, 4096, 1, 1])))
#n.lstm_12_resh = L.Reshape(n.lstm_12, reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.bc_att_lstm = L.CompactBilinear(n.att_feature_resh,
n.concat_vec_dropped,
compact_bilinear_param=dict(
num_output=16000, sum_pool=False))
n.bc_sign_sqrt = L.SignedSqrt(n.bc_att_lstm)
n.bc_sign_sqrt_l2 = L.L2Normalize(n.bc_sign_sqrt)
n.bc_dropped = L.Dropout(n.bc_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
n.bc_dropped_resh = L.Reshape(
n.bc_dropped, reshape_param=dict(shape=dict(dim=[-1, 16000])))
n.prediction = L.InnerProduct(n.bc_dropped_resh,
num_output=3000,
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
return n.to_proto()
def act_proto(mode, batchsize, exp_vocab_size, use_gt=True):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
n.img_feature, n.label, n.exp, n.exp_out, n.exp_cont_1, n.exp_cont_2 = \
L.Python(module='activity_data_provider_layer', layer='ActivityDataProviderLayer', param_str=mode_str, ntop=6)
# Attention
n.att_conv1 = L.Convolution(n.img_feature,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu,
kernel_size=1,
stride=1,
num_output=1,
pad=0,
weight_filler=dict(type='xavier'))
n.att_reshaped = L.Reshape(
n.att_conv2, reshape_param=dict(shape=dict(dim=[-1, 1, 14 * 14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att_map = L.Reshape(n.att_softmax,
reshape_param=dict(shape=dict(dim=[-1, 1, 14, 14])))
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.att_feature = L.SoftAttention(n.img_feature, n.att_map, dummy)
n.att_feature_resh = L.Reshape(
n.att_feature, reshape_param=dict(shape=dict(dim=[-1, 2048])))
# Prediction
n.prediction = L.InnerProduct(n.att_feature_resh,
num_output=config.NUM_OUTPUT_UNITS,
weight_filler=dict(type='xavier'),
param=fixed_weights)
# Take GT answer or Take the logits of the VQA model and get predicted answer to embed
if use_gt:
n.exp_emb_ans = L.Embed(n.label,
input_dim=config.NUM_OUTPUT_UNITS,
num_output=300,
weight_filler=dict(type='uniform',
min=-0.08,
max=0.08))
else:
n.vqa_ans = L.ArgMax(n.prediction, axis=1)
n.exp_emb_ans = L.Embed(n.vqa_ans,
input_dim=config.NUM_OUTPUT_UNITS,
num_output=300,
weight_filler=dict(type='uniform',
min=-0.08,
max=0.08))
n.exp_emb_ans_tanh = L.TanH(n.exp_emb_ans)
n.exp_emb_ans2 = L.InnerProduct(n.exp_emb_ans_tanh,
num_output=2048,
weight_filler=dict(type='xavier'))
# Merge activity answer and visual feature
n.exp_emb_resh = L.Reshape(
n.exp_emb_ans2, reshape_param=dict(shape=dict(dim=[-1, 2048, 1, 1])))
n.exp_emb_tiled_1 = L.Tile(n.exp_emb_resh, axis=2, tiles=14)
n.exp_emb_tiled = L.Tile(n.exp_emb_tiled_1, axis=3, tiles=14)
n.img_embed = L.Convolution(n.img_feature,
kernel_size=1,
stride=1,
num_output=2048,
pad=0,
weight_filler=dict(type='xavier'))
n.exp_eltwise = L.Eltwise(n.img_embed,
n.exp_emb_tiled,
eltwise_param={'operation': P.Eltwise.PROD})
n.exp_eltwise_sqrt = L.SignedSqrt(n.exp_eltwise)
n.exp_eltwise_l2 = L.L2Normalize(n.exp_eltwise_sqrt)
n.exp_eltwise_drop = L.Dropout(n.exp_eltwise_l2,
dropout_param={'dropout_ratio': 0.3})
# Attention for Explanation
n.exp_att_conv1 = L.Convolution(n.exp_eltwise_drop,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.exp_att_conv1_relu = L.ReLU(n.exp_att_conv1)
n.exp_att_conv2 = L.Convolution(n.exp_att_conv1_relu,
kernel_size=1,
stride=1,
num_output=1,
pad=0,
weight_filler=dict(type='xavier'))
n.exp_att_reshaped = L.Reshape(
n.exp_att_conv2, reshape_param=dict(shape=dict(dim=[-1, 1, 14 * 14])))
n.exp_att_softmax = L.Softmax(n.exp_att_reshaped, axis=2)
n.exp_att_map = L.Reshape(
n.exp_att_softmax, reshape_param=dict(shape=dict(dim=[-1, 1, 14, 14])))
exp_dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.exp_att_feature_prev = L.SoftAttention(n.img_feature, n.exp_att_map,
exp_dummy)
n.exp_att_feature_resh = L.Reshape(
n.exp_att_feature_prev, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.exp_att_feature_embed = L.InnerProduct(n.exp_att_feature_resh,
num_output=2048,
weight_filler=dict(type='xavier'))
n.exp_att_feature = L.Eltwise(n.exp_emb_ans2,
n.exp_att_feature_embed,
eltwise_param={'operation': P.Eltwise.PROD})
n.silence_exp_att = L.Silence(n.exp_att_feature, ntop=0)
return n.to_proto()
def pj_x(mode, batchsize, exp_T, exp_vocab_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
n.img_feature, n.label, n.exp, n.exp_out, n.exp_cont_1, n.exp_cont_2 = \
L.Python(module='activity_data_provider_layer',
layer='ActivityDataProviderLayer',
param_str=mode_str, ntop=6)
# Attention
n.att_conv1 = L.Convolution(n.img_feature,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu,
kernel_size=1,
stride=1,
num_output=1,
pad=0,
weight_filler=dict(type='xavier'))
n.att_reshaped = L.Reshape(
n.att_conv2, reshape_param=dict(shape=dict(dim=[-1, 1, 14 * 14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att_map = L.Reshape(n.att_softmax,
reshape_param=dict(shape=dict(dim=[-1, 1, 14, 14])))
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.att_feature = L.SoftAttention(n.img_feature, n.att_map, dummy)
n.att_feature_resh = L.Reshape(
n.att_feature, reshape_param=dict(shape=dict(dim=[-1, 2048])))
# Prediction
n.prediction = L.InnerProduct(n.att_feature_resh,
num_output=config.NUM_OUTPUT_UNITS,
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
n.accuracy = L.Accuracy(n.prediction, n.label)
# Embed Activity GT answer during training
n.exp_emb_ans = L.Embed(n.label, input_dim=config.NUM_OUTPUT_UNITS, num_output=300, \
weight_filler=dict(type='uniform', min=-0.08, max=0.08))
n.exp_emb_ans_tanh = L.TanH(n.exp_emb_ans)
n.exp_emb_ans2 = L.InnerProduct(n.exp_emb_ans_tanh,
num_output=2048,
weight_filler=dict(type='xavier'))
# merge activity answer and visual feature
n.exp_emb_resh = L.Reshape(
n.exp_emb_ans2, reshape_param=dict(shape=dict(dim=[-1, 2048, 1, 1])))
n.exp_emb_tiled_1 = L.Tile(n.exp_emb_resh, axis=2, tiles=14)
n.exp_emb_tiled = L.Tile(n.exp_emb_tiled_1, axis=3, tiles=14)
n.img_embed = L.Convolution(n.img_feature,
kernel_size=1,
stride=1,
num_output=2048,
pad=0,
weight_filler=dict(type='xavier'))
n.exp_eltwise = L.Eltwise(n.img_embed,
n.exp_emb_tiled,
eltwise_param={'operation': P.Eltwise.PROD})
n.exp_eltwise_sqrt = L.SignedSqrt(n.exp_eltwise)
n.exp_eltwise_l2 = L.L2Normalize(n.exp_eltwise_sqrt)
n.exp_eltwise_drop = L.Dropout(n.exp_eltwise_l2,
dropout_param={'dropout_ratio': 0.3})
# Attention for Explanation
n.exp_att_conv1 = L.Convolution(n.exp_eltwise_drop,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.exp_att_conv1_relu = L.ReLU(n.exp_att_conv1)
n.exp_att_conv2 = L.Convolution(n.exp_att_conv1_relu,
kernel_size=1,
stride=1,
num_output=1,
pad=0,
weight_filler=dict(type='xavier'))
n.exp_att_reshaped = L.Reshape(
n.exp_att_conv2, reshape_param=dict(shape=dict(dim=[-1, 1, 14 * 14])))
n.exp_att_softmax = L.Softmax(n.exp_att_reshaped, axis=2)
n.exp_att_map = L.Reshape(
n.exp_att_softmax, reshape_param=dict(shape=dict(dim=[-1, 1, 14, 14])))
exp_dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.exp_att_feature_prev = L.SoftAttention(n.img_feature, n.exp_att_map,
exp_dummy)
n.exp_att_feature_resh = L.Reshape(
n.exp_att_feature_prev, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.exp_att_feature_embed = L.InnerProduct(n.exp_att_feature_resh,
num_output=2048,
weight_filler=dict(type='xavier'))
n.exp_att_feature = L.Eltwise(n.exp_emb_ans2,
n.exp_att_feature_embed,
eltwise_param={'operation': P.Eltwise.PROD})
# Embed explanation
n.exp_embed_ba = L.Embed(n.exp, input_dim=exp_vocab_size, num_output=300, \
weight_filler=dict(type='uniform', min=-0.08, max=0.08))
n.exp_embed = L.TanH(n.exp_embed_ba)
# LSTM1 for Explanation
n.exp_lstm1 = L.LSTM(\
n.exp_embed, n.exp_cont_1,\
recurrent_param=dict(\
num_output=2048,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
n.exp_lstm1_dropped = L.Dropout(n.exp_lstm1,
dropout_param={'dropout_ratio': 0.3})
# merge with LSTM1 for explanation
n.exp_att_resh = L.Reshape(
n.exp_att_feature, reshape_param=dict(shape=dict(dim=[1, -1, 2048])))
n.exp_att_tiled = L.Tile(n.exp_att_resh, axis=0, tiles=exp_T)
n.exp_eltwise_all = L.Eltwise(n.exp_lstm1_dropped,
n.exp_att_tiled,
eltwise_param={'operation': P.Eltwise.PROD})
n.exp_eltwise_all_l2 = L.L2Normalize(n.exp_eltwise_all)
n.exp_eltwise_all_drop = L.Dropout(n.exp_eltwise_all_l2,
dropout_param={'dropout_ratio': 0.3})
# LSTM2 for Explanation
n.exp_lstm2 = L.LSTM(\
n.exp_eltwise_all_drop, n.exp_cont_2,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
n.exp_lstm2_dropped = L.Dropout(n.exp_lstm2,
dropout_param={'dropout_ratio': 0.3})
n.exp_prediction = L.InnerProduct(n.exp_lstm2_dropped,
num_output=exp_vocab_size,
weight_filler=dict(type='xavier'),
axis=2)
n.exp_loss = L.SoftmaxWithLoss(n.exp_prediction,
n.exp_out,
loss_param=dict(ignore_label=-1),
softmax_param=dict(axis=2))
n.exp_accuracy = L.Accuracy(n.exp_prediction,
n.exp_out,
axis=2,
ignore_label=-1)
return n.to_proto()
def qlstm(mode, batchsize, T, question_vocab_size, embed_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
n.data, n.cont, n.img_feature, n.label = L.Python(\
module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=4 )
# word embedding (static + dynamic)
n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=embed_size, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08))
n.embed_scale = L.Scale(n.embed_ba, n.cont,
scale_param=dict(dict(axis=0))) # N x T x d_w
n.embed_scale_resh = L.Reshape(
n.embed_scale,
reshape_param=dict(shape=dict(dim=[batchsize, T, embed_size, 1])))
# avg of word embedding
n.embed_avg = L.Convolution(n.embed_scale_resh,
convolution_param={
'kernel_size': 1,
'num_output': 1,
'bias_term': False,
'weight_filler': dict(type='constant',
value=1)
},
param=dict(lr_mult=0,
decay_mult=0)) # N x 1 x d_w x 1
n.embed_avg_resh = L.Reshape(
n.embed_avg,
reshape_param=dict(shape=dict(dim=[batchsize, embed_size, 1, 1])))
n.q_emb_tanh_droped_resh_tiled_1 = L.Tile(n.embed_avg_resh,
axis=2,
tiles=14)
n.q_emb_tanh_droped_resh_tiled = L.Tile(n.q_emb_tanh_droped_resh_tiled_1,
axis=3,
tiles=14)
n.i_emb_tanh_droped_resh = L.Reshape(
n.img_feature, reshape_param=dict(shape=dict(dim=[-1, 2048, 14, 14])))
n.blcf = L.CompactBilinear(n.q_emb_tanh_droped_resh_tiled,
n.i_emb_tanh_droped_resh,
compact_bilinear_param=dict(num_output=16000,
sum_pool=False))
n.blcf_sign_sqrt = L.SignedSqrt(n.blcf)
n.blcf_sign_sqrt_l2 = L.L2Normalize(n.blcf_sign_sqrt)
n.blcf_droped = L.Dropout(n.blcf_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
# multi-channel attention
n.att_conv1 = L.Convolution(n.blcf_droped,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu,
kernel_size=1,
stride=1,
num_output=2,
pad=0,
weight_filler=dict(type='xavier'))
n.att_reshaped = L.Reshape(
n.att_conv2, reshape_param=dict(shape=dict(dim=[-1, 2, 14 * 14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att = L.Reshape(n.att_softmax,
reshape_param=dict(shape=dict(dim=[-1, 2, 14, 14])))
att_maps = L.Slice(n.att, ntop=2, slice_param={'axis': 1})
n.att_map0 = att_maps[0]
n.att_map1 = att_maps[1]
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.att_feature0 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map0,
dummy)
n.att_feature1 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map1,
dummy)
n.att_feature0_resh = L.Reshape(
n.att_feature0, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature1_resh = L.Reshape(
n.att_feature1, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature = L.Concat(n.att_feature0_resh, n.att_feature1_resh)
# merge attention and lstm with compact bilinear pooling
n.att_feature_resh = L.Reshape(
n.att_feature, reshape_param=dict(shape=dict(dim=[-1, 4096, 1, 1])))
#n.lstm_12_resh = L.Reshape(n.lstm_12, reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.bc_att_lstm = L.CompactBilinear(n.att_feature_resh,
n.embed_avg_resh,
compact_bilinear_param=dict(
num_output=16000, sum_pool=False))
n.bc_sign_sqrt = L.SignedSqrt(n.bc_att_lstm)
n.bc_sign_sqrt_l2 = L.L2Normalize(n.bc_sign_sqrt)
n.bc_dropped = L.Dropout(n.bc_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
n.bc_dropped_resh = L.Reshape(
n.bc_dropped, reshape_param=dict(shape=dict(dim=[-1, 16000])))
n.prediction = L.InnerProduct(n.bc_dropped_resh,
num_output=3000,
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
return n.to_proto()
def qlstm(mode, batchsize, max_words_in_question, question_vocab_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
n.data, n.cont, n.img_feature, n.label, n.glove = L.Python( \
module='vqa_data_provider_layer', layer='VQADataProviderLayer', \
param_str=mode_str, ntop=5 )
n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08)) # T x N -> T x N x 300
n.embed = L.TanH(n.embed_ba)
concat_word_embed = [n.embed, n.glove]
n.concat_embed = L.Concat(*concat_word_embed,
concat_param={'axis': 2}) # T x N x 600
# LSTM1
n.lstm1 = L.LSTM(\
n.concat_embed, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
tops1 = L.Slice(n.lstm1,
ntop=max_words_in_question,
slice_param={'axis': 0})
for i in xrange(max_words_in_question - 1):
n.__setattr__('slice_first' + str(i), tops1[int(i)])
n.__setattr__('silence_data_first' + str(i),
L.Silence(tops1[int(i)], ntop=0))
n.lstm1_out = tops1[max_words_in_question - 1]
n.lstm1_reshaped = L.Reshape(n.lstm1_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm1_reshaped_drop = L.Dropout(n.lstm1_reshaped,
dropout_param={'dropout_ratio': 0.3})
n.lstm1_drop = L.Dropout(n.lstm1, dropout_param={'dropout_ratio': 0.3})
# LSTM2
n.lstm2 = L.LSTM(\
n.lstm1_drop, n.cont,\
recurrent_param=dict(\
num_output=1024,\
weight_filler=dict(type='uniform',min=-0.08,max=0.08),\
bias_filler=dict(type='constant',value=0)))
tops2 = L.Slice(n.lstm2,
ntop=max_words_in_question,
slice_param={'axis': 0})
for i in xrange(max_words_in_question - 1):
n.__setattr__('slice_second' + str(i), tops2[int(i)])
n.__setattr__('silence_data_second' + str(i),
L.Silence(tops2[int(i)], ntop=0))
n.lstm2_out = tops2[max_words_in_question - 1]
n.lstm2_reshaped = L.Reshape(n.lstm2_out,\
reshape_param=dict(\
shape=dict(dim=[-1,1024])))
n.lstm2_reshaped_drop = L.Dropout(n.lstm2_reshaped,
dropout_param={'dropout_ratio': 0.3})
concat_lstms = [n.lstm1_reshaped_drop, n.lstm2_reshaped_drop]
n.lstm_12 = L.Concat(*concat_lstms)
n.q_emb_tanh_droped_resh = L.Reshape(n.lstm_12, \
reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.i_emb_tanh_droped_resh = L.Reshape(n.img_feature, \
reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.blcf = L.CompactBilinear(n.q_emb_tanh_droped_resh, n.i_emb_tanh_droped_resh, \
compact_bilinear_param=dict(num_output=16000,sum_pool=False))
n.blcf_sign_sqrt = L.SignedSqrt(n.blcf)
n.blcf_sign_sqrt_l2 = L.L2Normalize(n.blcf_sign_sqrt)
n.blcf_droped = L.Dropout(n.blcf_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
n.blcf_droped_resh = L.Reshape(
n.blcf_droped, reshape_param=dict(shape=dict(dim=[-1, 16000])))
n.prediction = L.InnerProduct(n.blcf_droped_resh, num_output=config.NUM_OUTPUT_UNITS, \
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
return n.to_proto()
def qlstm(mode, batchsize, T, question_vocab_size):
n = caffe.NetSpec()
mode_str = json.dumps({'mode': mode, 'batchsize': batchsize})
n.data, n.cont, n.img_feature, n.label = L.Python(\
module='vqa_data_provider_layer', layer='VQADataProviderLayer', param_str=mode_str, ntop=4)#5 )
# # word embedding (static + dynamic)
# n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=300, \
# weight_filler=dict(type='uniform',min=-0.08,max=0.08))
# n.embed_scale = L.Scale(n.embed_ba, n.cont, scale_param=dict(dict(axis=0)))
# n.embed_scale_resh = L.Reshape(n.embed_scale,\
# reshape_param=dict(\
# shape=dict(dim=[batchsize,1,T,300])))
# n.glove_scale = L.Scale(n.glove, n.cont, scale_param=dict(dict(axis=0)))
# n.glove_scale_resh = L.Reshape(n.glove_scale,\
# reshape_param=dict(\
# shape=dict(dim=[batchsize,1,T,300])))
# concat_word_embed = [n.embed_scale_resh, n.glove_scale_resh]
# n.concat_embed = L.Concat(*concat_word_embed, concat_param={'axis': 1}) # N x 2 x T x 300
# char embedding
n.embed_ba = L.Embed(n.data, input_dim=question_vocab_size, num_output=50, \
weight_filler=dict(type='uniform',min=-0.08,max=0.08))
n.embed_scale = L.Scale(n.embed_ba, n.cont, scale_param=dict(dict(axis=0)))
n.embed_scale_resh = L.Reshape(n.embed_scale,\
reshape_param=dict(\
shape=dict(dim=[batchsize,1,T,50])))
# char deep convolution
n.char_conv_1 = L.Convolution(
n.embed_scale_resh,
kernel_h=5,
kernel_w=50,
stride=1,
num_output=256,
weight_filler=dict(type='gaussian',
std=0.05)) # N x 1 x 100 x 50 -> N x 256 x 96 x 1
n.char_relu_1 = L.ReLU(n.char_conv_1)
n.char_pool_1 = L.Pooling(
n.char_relu_1, kernel_h=2, kernel_w=1, stride=2,
pool=P.Pooling.MAX) # N x 256 x 96 x 1 -> N x 256 x 48 x 1
n.char_conv_2 = L.Convolution(
n.char_pool_1,
kernel_h=5,
kernel_w=1,
stride=1,
num_output=256,
weight_filler=dict(type='gaussian',
std=0.05)) # N x 256 x 48 x 1 -> N x 256 x 44 x 1
n.char_relu_2 = L.ReLU(n.char_conv_2)
n.char_pool_2 = L.Pooling(
n.char_relu_2, kernel_h=2, kernel_w=1, stride=2,
pool=P.Pooling.MAX) # N x 256 x 44 x 1 -> N x 256 x 22 x 1
n.char_conv_3 = L.Convolution(
n.char_pool_2,
kernel_h=3,
kernel_w=1,
stride=1,
num_output=256,
weight_filler=dict(type='gaussian',
std=0.05)) # N x 256 x 22 x 1 -> N x 256 x 20 x 1
n.char_relu_3 = L.ReLU(n.char_conv_3)
n.char_conv_4 = L.Convolution(
n.char_relu_3,
kernel_h=3,
kernel_w=1,
stride=1,
num_output=256,
weight_filler=dict(type='gaussian',
std=0.05)) # N x 256 x 20 x 1 -> N x 256 x 18 x 1
n.char_relu_4 = L.ReLU(n.char_conv_4)
n.char_conv_5 = L.Convolution(
n.char_relu_4,
kernel_h=3,
kernel_w=1,
stride=1,
num_output=256,
weight_filler=dict(type='gaussian',
std=0.05)) # N x 256 x 18 x 1 -> N x 256 x 16 x 1
n.char_relu_5 = L.ReLU(n.char_conv_5)
n.char_pool_3 = L.Pooling(
n.char_relu_5, kernel_h=2, kernel_w=1, stride=2,
pool=P.Pooling.MAX) # N x 256 x 16 x 1 -> N x 256 x 8 x 1
n.vec_reshape = L.Reshape(
n.char_pool_3, reshape_param=dict(shape=dict(dim=[-1, 2048, 1, 1])))
n.concat_vec_dropped = L.Dropout(n.vec_reshape,
dropout_param={'dropout_ratio': 0.5})
n.q_emb_tanh_droped_resh_tiled_1 = L.Tile(n.concat_vec_dropped,
axis=2,
tiles=14)
n.q_emb_tanh_droped_resh_tiled = L.Tile(n.q_emb_tanh_droped_resh_tiled_1,
axis=3,
tiles=14)
n.i_emb_tanh_droped_resh = L.Reshape(
n.img_feature, reshape_param=dict(shape=dict(dim=[-1, 2048, 14, 14])))
n.blcf = L.CompactBilinear(n.q_emb_tanh_droped_resh_tiled,
n.i_emb_tanh_droped_resh,
compact_bilinear_param=dict(num_output=16000,
sum_pool=False))
n.blcf_sign_sqrt = L.SignedSqrt(n.blcf)
n.blcf_sign_sqrt_l2 = L.L2Normalize(n.blcf_sign_sqrt)
n.blcf_droped = L.Dropout(n.blcf_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
# multi-channel attention
n.att_conv1 = L.Convolution(n.blcf_droped,
kernel_size=1,
stride=1,
num_output=512,
pad=0,
weight_filler=dict(type='xavier'))
n.att_conv1_relu = L.ReLU(n.att_conv1)
n.att_conv2 = L.Convolution(n.att_conv1_relu,
kernel_size=1,
stride=1,
num_output=2,
pad=0,
weight_filler=dict(type='xavier'))
n.att_reshaped = L.Reshape(
n.att_conv2, reshape_param=dict(shape=dict(dim=[-1, 2, 14 * 14])))
n.att_softmax = L.Softmax(n.att_reshaped, axis=2)
n.att = L.Reshape(n.att_softmax,
reshape_param=dict(shape=dict(dim=[-1, 2, 14, 14])))
att_maps = L.Slice(n.att, ntop=2, slice_param={'axis': 1})
n.att_map0 = att_maps[0]
n.att_map1 = att_maps[1]
dummy = L.DummyData(shape=dict(dim=[batchsize, 1]),
data_filler=dict(type='constant', value=1),
ntop=1)
n.att_feature0 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map0,
dummy)
n.att_feature1 = L.SoftAttention(n.i_emb_tanh_droped_resh, n.att_map1,
dummy)
n.att_feature0_resh = L.Reshape(
n.att_feature0, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature1_resh = L.Reshape(
n.att_feature1, reshape_param=dict(shape=dict(dim=[-1, 2048])))
n.att_feature = L.Concat(n.att_feature0_resh, n.att_feature1_resh)
# merge attention and lstm with compact bilinear pooling
n.att_feature_resh = L.Reshape(
n.att_feature, reshape_param=dict(shape=dict(dim=[-1, 4096, 1, 1])))
#n.lstm_12_resh = L.Reshape(n.lstm_12, reshape_param=dict(shape=dict(dim=[-1,2048,1,1])))
n.bc_att_lstm = L.CompactBilinear(n.att_feature_resh,
n.concat_vec_dropped,
compact_bilinear_param=dict(
num_output=16000, sum_pool=False))
n.bc_sign_sqrt = L.SignedSqrt(n.bc_att_lstm)
n.bc_sign_sqrt_l2 = L.L2Normalize(n.bc_sign_sqrt)
n.bc_dropped = L.Dropout(n.bc_sign_sqrt_l2,
dropout_param={'dropout_ratio': 0.1})
n.bc_dropped_resh = L.Reshape(
n.bc_dropped, reshape_param=dict(shape=dict(dim=[-1, 16000])))
n.prediction = L.InnerProduct(n.bc_dropped_resh,
num_output=3000,
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.prediction, n.label)
return n.to_proto()
