def get_nets():
import vg.defn.three_way2 as D
net_base = load_best_run('{}/s2-t.-s2i2-s2t.-t2s.-t2i.'.format(PREFIX),
cond='')
net_mt = load_best_run('{}/s2-t1-s2i2-s2t0-t2s0-t2i1'.format(PREFIX),
cond='joint')
config = dict(TextImage=dict(ImageEncoder=dict(size=1024,
size_target=4096),
lr=0.0002,
margin_size=0.2,
max_norm=2.0,
TextEncoderTop=dict(size=1024,
size_feature=1024,
depth=1,
size_attn=128)),
SpeechImage=dict(ImageEncoder=dict(size=1024,
size_target=4096),
lr=0.0002,
margin_size=0.2,
max_norm=2.0,
SpeechEncoderTop=dict(size=1024,
size_input=1024,
depth=2,
size_attn=128)),
SpeechText=dict(TextEncoderTop=dict(size_feature=1024,
size=1024,
depth=0,
size_attn=128),
SpeechEncoderTop=dict(size=1024,
size_input=1024,
depth=0,
size_attn=128),
lr=0.0002,
margin_size=0.2,
max_norm=2.0),
SpeechEncoderBottom=dict(size=1024,
depth=2,
size_vocab=13,
filter_length=6,
filter_size=64,
stride=2),
TextEncoderBottom=dict(
size_feature=net_mt.TextEncoderBottom.size_feature,
size_embed=128,
size=1024,
depth=1))
net_base = load_best_run('{}/s2-t.-s2i2-s2t.-t2s.-t2i.'.format(PREFIX),
cond='')
net_mt = load_best_run('{}/s2-t1-s2i2-s2t0-t2s0-t2i1'.format(PREFIX),
cond='joint')
net_init = D.Net(config).cuda()
return [('m6_init', net_init), ('m1', net_base), ('m6', net_mt)]
SpeechEncoderBottom=dict(size=1024, depth=2, size_vocab=13, filter_length=6, filter_size=64, stride=2),
TextEncoderBottom=dict(size_feature=data_flickr.mapper.size(),
size_embed=128,
size=1024,
depth=1)
)
def audio(sent):
return sent['audio']
net = D.Net(model_config)
net.batcher = None
net.mapper = None
scorer = vg.scorer.Scorer(prov_flickr,
dict(split='val',
tokenize=audio,
batch_size=batch_size
))
run_config = dict(epochs=epochs,
validate_period=400,
tasks=[ ('SpeechText', net.SpeechText),
('SpeechImage', net.SpeechImage),
('TextImage', net.TextImage)],
