def image_encoder(image_feat,
hparams,
name="image_encoder",
save_weights_to=None,
make_image_summary=True):
"""A stack of self attention layers."""
x = image_feat
with tf.variable_scope(name):
for layer in range(hparams.num_encoder_layers
or hparams.num_hidden_layers):
with tf.variable_scope("layer_%d" % layer):
with tf.variable_scope("self_attention"):
y = vqa_layers.multihead_attention(
common_layers.layer_preprocess(x, hparams),
None,
None,
hparams.attention_key_channels
or hparams.image_hidden_size,
hparams.attention_value_channels
or hparams.image_hidden_size,
hparams.image_hidden_size,
hparams.num_heads,
hparams.attention_dropout,
attention_type=hparams.self_attention_type,
save_weights_to=save_weights_to,
max_relative_position=None,
make_image_summary=make_image_summary,
dropout_broadcast_dims=None,
max_length=None,
vars_3d=False,
scale_otproduct=hparams.scale_dotproduct)
utils.collect_named_outputs("norms",
"image_feat_self_attention",
tf.norm(y, axis=-1))
x = common_layers.layer_postprocess(x, y, hparams)
utils.collect_named_outputs(
"norms", "image_feat_self_attention_zero_add",
tf.norm(x, axis=-1))
with tf.variable_scope("ffn"):
y = common_layers.dense_relu_dense(
common_layers.layer_preprocess(x, hparams),
hparams.image_filter_size,
hparams.image_hidden_size,
dropout=hparams.relu_dropout,
dropout_broadcast_dims=None)
utils.collect_named_outputs("norms", "image_feat_ffn",
tf.norm(y, axis=-1))
x = common_layers.layer_postprocess(x, y, hparams)
utils.collect_named_outputs("norms",
"image_feat_ffn_zero_add",
tf.norm(x, axis=-1))
# if normalization is done in layer_preprocess, then it should also be done
# on the output, since the output can grow very large, being the sum of
# a whole stack of unnormalized layer outputs.
return common_layers.layer_preprocess(x, hparams)
def question_encoder(question,
question_self_attention_bias,
hparams,
name="question_encoder",
save_weights_to=None,
make_image_summary=True):
"""A stack of self attention layers."""
x = question
with tf.variable_scope(name):
for layer in range(hparams.num_encoder_layers
or hparams.num_hidden_layers):
with tf.variable_scope("layer_%d" % layer):
with tf.variable_scope("self_attention"):
y = vqa_layers.multihead_attention(
common_layers.layer_preprocess(x, hparams),
None,
question_self_attention_bias,
hparams.attention_key_channels or hparams.hidden_size,
hparams.attention_value_channels
or hparams.hidden_size,
hparams.hidden_size,
hparams.num_heads,
hparams.attention_dropout,
attention_type=hparams.question_self_attention_type,
block_length=hparams.block_length,
save_weights_to=save_weights_to,
make_image_summary=make_image_summary,
scale_dotproduct=hparams.scale_dotproduct,
)
utils.collect_named_outputs(
"norms", "query_self_attention_%d" % (layer),
tf.norm(y, axis=-1))
x = common_layers.layer_postprocess(x, y, hparams)
utils.collect_named_outputs(
"norms",
"query_self_attention_postprocess_%d" % (layer),
tf.norm(x, axis=-1))
with tf.variable_scope("ffn"):
y = common_layers.dense_relu_dense(
common_layers.layer_preprocess(x, hparams),
hparams.filter_size,
hparams.hidden_size,
dropout=hparams.relu_dropout,
)
utils.collect_named_outputs("norms",
"query_ffn_%d" % (layer),
tf.norm(y, axis=-1))
x = common_layers.layer_postprocess(x, y, hparams)
utils.collect_named_outputs(
"norms", "query_ffn_postprocess_%d" % (layer),
tf.norm(x, axis=-1))
# if normalization is done in layer_preprocess, then it should also be done
# on the output, since the output can grow very large, being the sum of
# a whole stack of unnormalized layer outputs.
return common_layers.layer_preprocess(x, hparams)
def image_encoder(image_feat,
hparams,
name="image_encoder",
save_weights_to=None,
make_image_summary=True):
"""A stack of self attention layers."""
x = image_feat
image_hidden_size = hparams.image_hidden_size or hparams.hidden_size
image_filter_size = hparams.image_filter_size or hparams.filter_size
with tf.variable_scope(name):
for layer in range(hparams.num_encoder_layers or hparams.num_hidden_layers):
with tf.variable_scope("layer_%d" % layer):
with tf.variable_scope("self_attention"):
y = vqa_layers.multihead_attention(
common_layers.layer_preprocess(x, hparams),
None,
None,
hparams.attention_key_channels or image_hidden_size,
hparams.attention_value_channels or image_hidden_size,
image_hidden_size,
hparams.num_heads,
hparams.attention_dropout,
attention_type=hparams.image_self_attention_type,
save_weights_to=save_weights_to,
make_image_summary=make_image_summary,
scale_dotproduct=hparams.scale_dotproduct,
)
utils.collect_named_outputs(
"norms", "image_feat_self_attention_%d"%(layer),
tf.norm(y, axis=-1))
x = common_layers.layer_postprocess(x, y, hparams)
utils.collect_named_outputs(
"norms", "image_feat_self_attention_postprocess_%d"%(layer),
tf.norm(x, axis=-1))
with tf.variable_scope("ffn"):
y = common_layers.dense_relu_dense(
common_layers.layer_preprocess(x, hparams),
image_filter_size,
image_hidden_size,
dropout=hparams.relu_dropout,
)
utils.collect_named_outputs(
"norms", "image_feat_ffn_%d"%(layer), tf.norm(y, axis=-1))
x = common_layers.layer_postprocess(x, y, hparams)
utils.collect_named_outputs(
"norms", "image_feat_ffn_postprocess_%d"%(layer),
tf.norm(x, axis=-1))
# if normalization is done in layer_preprocess, then it should also be done
# on the output, since the output can grow very large, being the sum of
# a whole stack of unnormalized layer outputs.
return common_layers.layer_preprocess(x, hparams)
