"sentence": (True, True),
"reiss": (False, False),
"maslow": (False, False),
"plutchik": (False, False),
"plutchik16": (False, False),
"entity": (True, False)
}
splits = ["train", "dev", "test"]
for experiment in ["emotion", "motivation"]:
print("Making {} data for {} class of models".format(
experiment, sys.argv[1]))
opt = DD()
opt.data = DD()
opt.data.pruned = True
# Make a memory model (EntNet, NPN) data loader for generation
if sys.argv[1] == "memory":
# Make save name
name = "processed/{}/{}_{}_data_loader.pth".format(
experiment, "gen_memory", "-".join(splits))
print(name)
# Initialize data loader and load vocabs and raw data
data_loader = data.MemoryGenModelDataLoader()
data_loader.load_vocabs(vocab_paths, vocab_text)
data_loader.load_data(opt,
splits,
type_=experiment,
def get_parameters(opt, exp="class"):
params = DD()
params.net = DD()
params.net.enc = DD()
get_encoder_parameters(params, opt)
params.net.gendec = DD()
params.net.classdec = DD()
if exp == "class":
get_class_parameters(params, opt)
elif exp == "gen":
get_gendec_parameters(params, opt)
else:
get_class_parameters(params, opt)
params.net.enc.pt = "none"
params.train = DD()
params.train.static = DD()
params.train.dynamic = DD()
params.eval = DD()
params.data = DD()
# Which experiment to run: {class = classification; gen = generation}
params.exp = opt.experiment
# Task to run: {emotion, motivation}
params.task = opt.task
if params.exp == "class":
# % of development set stories to keep as training data
params.train.static.tr = opt.train_ratio
# granularity of labels
# motivation: {maslow, reiss}
# emotion: {plutchik, plutchik16}
params.granularity = opt.granularity
# Labeling type (default = majority)
params.data.label = opt.label
if params.exp == "gen":
# Number of positive examples per negative example
params.train.static.pnr = opt.pos_neg_ratio
# Loss to use during training
params.train.static.wcrit = "nll"
# Prune useless words in motivation sequences such as
# "to be" in "to be famous"
params.data.pruned = opt.pruned
# Max norm at which to clip gradients
params.train.static.gc = opt.grad_clip
# Random seed
params.train.static.seed = opt.random_seed
# learning rate
params.train.dynamic.lr = opt.learning_rate
# batch size
params.train.dynamic.bs = opt.batch_size
# optimizer to use {adam, rmsprop, etc.}
# Only Adam is actually implemented
params.train.dynamic.optim = opt.optimizer
# Default parameters for the CNN model from
# 2014 Yoon Kim paper
if params.net.enc.model == "cnn+stock":
params.net.enc.ks = "3,4,5"
params.net.enc.kn = 100
params.net.classdec.dpt = 0.5
params.train.dynamic.lr = 0.001
params.train.dynamic.bs = 64
params.data.shuffle = False
params.train.static.l2 = 3
params.net.enc.iSize = 128
params.net.classdec.hSize = 300
meta = DD()
meta.iterations = opt.iterations
meta.epochs = opt.epochs
meta.mark = opt.mark
return params, meta