def build_vocab_from_pcfg(pcfg, min_freq=0, top_k=np.infty)->Vocab:
vocab = Vocab()
vocab.add_token("(")
vocab.add_token(")")
for rule in pcfg.productions():
vocab.add_token(str(rule.lhs()))
for rhse in rule.rhs():
vocab.add_token(str(rhse))
vocab.finalize(min_freq=min_freq, top_k=top_k)
return vocab
def load_ds(dataset="scan/random",
validfrac=0.1,
recompute=False,
bertname="bert-base-uncased"):
tt = q.ticktock("data")
tt.tick(f"loading '{dataset}'")
if bertname.startswith("none"):
bertname = "bert" + bertname[4:]
if dataset.startswith("cfq/") or dataset.startswith("scan/mcd"):
key = f"{dataset}|bertname={bertname}"
print(f"validfrac is ineffective with dataset '{dataset}'")
else:
key = f"{dataset}|validfrac={validfrac}|bertname={bertname}"
shelfname = os.path.basename(__file__) + ".cache.shelve"
if not recompute:
tt.tick(f"loading from shelf (key '{key}')")
with shelve.open(shelfname) as shelf:
if key not in shelf:
recompute = True
tt.tock("couldn't load from shelf")
else:
shelved = shelf[key]
trainex, validex, testex, fldic = shelved["trainex"], shelved[
"validex"], shelved["testex"], shelved["fldic"]
inpdic = shelved["inpdic"] if "inpdic" in shelved else None
trainds, validds, testds = Dataset(trainex), Dataset(
validex), Dataset(testex)
tt.tock("loaded from shelf")
if recompute:
tt.tick("loading data")
splits = dataset.split("/")
dataset, splits = splits[0], splits[1:]
split = "/".join(splits)
if dataset == "scan":
ds = SCANDatasetLoader().load(split, validfrac=validfrac)
elif dataset == "cfq":
ds = CFQDatasetLoader().load(split + "/modent")
else:
raise Exception(f"Unknown dataset: '{dataset}'")
tt.tock("loaded data")
tt.tick("creating tokenizer")
tokenizer = Tokenizer(bertname=bertname)
tt.tock("created tokenizer")
print(len(ds))
tt.tick("dictionaries")
inpdic = Vocab()
inplens, outlens = [0], []
fldic = Vocab()
for x in ds:
outtoks = tokenizer.get_out_toks(x[1])
outlens.append(len(outtoks))
for tok in outtoks:
fldic.add_token(tok, seen=x[2] == "train")
inptoks = tokenizer.get_toks(x[0])
for tok in inptoks:
inpdic.add_token(tok, seen=x[2] == "train")
inpdic.finalize(min_freq=0, top_k=np.infty)
fldic.finalize(min_freq=0, top_k=np.infty)
print(
f"input avg/max length is {np.mean(inplens):.1f}/{max(inplens)}, output avg/max length is {np.mean(outlens):.1f}/{max(outlens)}"
)
print(
f"output vocabulary size: {len(fldic.D)} at output, {len(inpdic.D)} at input"
)
tt.tock()
tt.tick("tensorizing")
tokenizer.inpvocab = inpdic
tokenizer.outvocab = fldic
trainds = ds.filter(lambda x: x[-1] == "train").map(
lambda x: x[:-1]).map(
lambda x: tokenizer.tokenize(x[0], x[1])).cache(True)
validds = ds.filter(lambda x: x[-1] == "valid").map(
lambda x: x[:-1]).map(
lambda x: tokenizer.tokenize(x[0], x[1])).cache(True)
testds = ds.filter(lambda x: x[-1] == "test").map(
lambda x: x[:-1]).map(
lambda x: tokenizer.tokenize(x[0], x[1])).cache(True)
# ds = ds.map(lambda x: tokenizer.tokenize(x[0], x[1]) + (x[2],)).cache(True)
tt.tock("tensorized")
tt.tick("shelving")
with shelve.open(shelfname) as shelf:
shelved = {
"trainex": trainds.examples,
"validex": validds.examples,
"testex": testds.examples,
"fldic": fldic,
"inpdic": inpdic,
}
shelf[key] = shelved
tt.tock("shelved")
tt.tock(f"loaded '{dataset}'")
tt.msg(
f"#train={len(trainds)}, #valid={len(validds)}, #test={len(testds)}")
tt.msg("Overlap of validation with train:")
overlaps = compute_overlaps(trainds, validds)
print(json.dumps(overlaps, indent=4))
tt.msg("Overlap of test with train:")
overlaps = compute_overlaps(trainds, testds)
print(json.dumps(overlaps, indent=4))
return trainds, validds, testds, fldic, inpdic
