def benchmark_clevr_sketch(share_funcs,
batch_norm,
batch_size,
arch,
log_dir,
lr,
arch_opt,
sample,
conversions=std_conversions(),
**kwargs):
all_arch = archs.MLPNet
sample_args = {'pbatch_norm': int(batch_norm)}
functypes = genfuns.func_types(*std_types())
import pdb
pdb.set_trace()
neu_clevr = clevrarch.funcs(all_arch, arch_opt, sample, sample_args,
**functypes)
neuclevr = asl.modules.modules.ModuleDict(neu_clevr)
refclevr = ref_clevr
clevr_sketch = ClevrSketch(neuclevr, refclevr)
util.cuda(clevr_sketch)
data_itr = data_iter(batch_size, conversions)
def loss_gen():
nonlocal data_itr
try:
progs, objsets, rels, answers = next(data_itr)
except StopIteration:
data_itr = data_iter(batch_size, conversions)
progs, objsets, rels, answers = next(data_itr)
outputs = clevr_sketch(progs, objsets, rels)
anstensors = [convert(parse(ans), conversions) for ans in answers]
acc, ncorrect = accuracy(outputs, anstensors)
asl.log("accuracy", acc)
asl.log("ncorrect", ncorrect)
asl.log("outof", len(answers))
deltas = [
asl.dist(outputs[i], anstensors[i]) for i in range(len(outputs))
]
return sum(deltas) / len(deltas)
optimizer = optim.Adam(clevr_sketch.parameters(), lr)
asl.train(
loss_gen,
optimizer,
cont=asl.converged(100),
callbacks=[
asl.print_loss(10),
print_accuracy(10),
asl.nancancel,
# every_n(plot_sketch, 500),
# common.plot_empty,
# common.plot_observes,
asl.save_checkpoint(100, clevr_sketch)
],
log_dir=log_dir)
def benchmark_clevr_sketch(share_funcs,
batch_norm,
batch_size,
arch,
log_dir,
lr,
arch_opt,
sample,
conversions=std_conversions(),
**kwargs):
all_arch = archs.MLPNet
sample_args = {'pbatch_norm': int(batch_norm)}
functypes = genfuns.func_types(*std_types())
import pdb; pdb.set_trace()
neu_clevr = clevrarch.funcs(all_arch, arch_opt, sample, sample_args, **functypes)
neuclevr = asl.modules.modules.ModuleDict(neu_clevr)
refclevr = ref_clevr
clevr_sketch = ClevrSketch(neuclevr, refclevr)
util.cuda(clevr_sketch)
data_itr = data_iter(batch_size, conversions)
def loss_gen():
nonlocal data_itr
try:
progs, objsets, rels, answers = next(data_itr)
except StopIteration:
data_itr = data_iter(batch_size, conversions)
progs, objsets, rels, answers = next(data_itr)
outputs = clevr_sketch(progs, objsets, rels)
anstensors = [convert(parse(ans), conversions) for ans in answers]
acc, ncorrect = accuracy(outputs, anstensors)
asl.log("accuracy", acc)
asl.log("ncorrect", ncorrect)
asl.log("outof", len(answers))
deltas = [asl.dist(outputs[i], anstensors[i]) for i in range(len(outputs))]
return sum(deltas) / len(deltas)
optimizer = optim.Adam(clevr_sketch.parameters(), lr)
asl.train(loss_gen,
optimizer,
cont=asl.converged(100),
callbacks=[asl.print_loss(10),
print_accuracy(10),
asl.nancancel,
# every_n(plot_sketch, 500),
# common.plot_empty,
# common.plot_observes,
asl.save_checkpoint(100, clevr_sketch)],
log_dir=log_dir)
def from_clevr_object(clevr_object):
tensor = Variable(util.cuda(util.onehotmany([clevr_object.color.value,
clevr_object.size.value,
clevr_object.material.value,
clevr_object.shape.value],
TensorClevrObject.max_prop_len)))
return TensorClevrObject(tensor)
def from_clevr_object(clevr_object):
tensor = Variable(
util.cuda(
util.onehotmany([
clevr_object.color.value, clevr_object.size.value,
clevr_object.material.value, clevr_object.shape.value
], TensorClevrObject.max_prop_len)))
return TensorClevrObject(tensor)
def from_relations(relations):
nrels = 4
maxnobjs = 10
rel_ten = torch.zeros(nrels, maxnobjs, maxnobjs)
for (i, rel) in enumerate(['behind', 'front', 'left', 'right']):
for (j, obj1rels) in enumerate(relations.listform[rel]):
for obj2 in obj1rels:
rel_ten[i, j, obj2] = 1.0
return TensorRelations(Variable(util.cuda(rel_ten)))
def from_relations(relations):
nrels = 4
maxnobjs = 10
rel_ten = torch.zeros(nrels, maxnobjs, maxnobjs)
for (i, rel) in enumerate(['behind', 'front', 'left', 'right']):
for (j, obj1rels) in enumerate(relations.listform[rel]):
for obj2 in obj1rels:
rel_ten[i, j, obj2] = 1.0
return TensorRelations(Variable(util.cuda(rel_ten)))
def std_conversions():
"Options sampler"
conversions = {}
conversions[clevr.ClevrObjectSet] = clevr.TensorClevrObjectSet.from_clevr_object_set
conversions[clevr.ClevrObject] = clevr.TensorClevrObject.from_clevr_object
conversions[clevr.Relations] = clevr.TensorRelations.from_relations
conversions[int] = lambda x: clevr.IntegerOneHot1D(Variable(util.cuda(util.onehot(x, 11, 1))))
conversions[clevr.BooleanEnum] = asl.onehot1d
conversions[clevr.ColorEnum] = asl.onehot1d
conversions[clevr.ColorEnum] = asl.onehot1d
conversions[clevr.MaterialEnum] = asl.onehot1d
conversions[clevr.ShapeEnum] = asl.onehot1d
conversions[clevr.SizeEnum] = asl.onehot1d
return conversions
def from_clevr_object_set(clevr_object_set, max_n_objects=10):
obj_tensors = [
TensorClevrObject.from_clevr_object(obj).value
for obj in clevr_object_set.objects
]
ndummies = max_n_objects - len(obj_tensors)
assert ndummies >= 0
dummies = [
Variable(
util.cuda(torch.zeros(1, 4,
TensorClevrObjectSet.max_prop_len)))
for i in range(ndummies)
]
return TensorClevrObjectSet(torch.cat(obj_tensors + dummies, 0))
def std_conversions():
"Options sampler"
conversions = {}
conversions[
clevr.
ClevrObjectSet] = clevr.TensorClevrObjectSet.from_clevr_object_set
conversions[clevr.ClevrObject] = clevr.TensorClevrObject.from_clevr_object
conversions[clevr.Relations] = clevr.TensorRelations.from_relations
conversions[int] = lambda x: clevr.IntegerOneHot1D(
Variable(util.cuda(util.onehot(x, 11, 1))))
conversions[clevr.BooleanEnum] = asl.onehot1d
conversions[clevr.ColorEnum] = asl.onehot1d
conversions[clevr.ColorEnum] = asl.onehot1d
conversions[clevr.MaterialEnum] = asl.onehot1d
conversions[clevr.ShapeEnum] = asl.onehot1d
conversions[clevr.SizeEnum] = asl.onehot1d
return conversions
def from_clevr_object_set(clevr_object_set, max_n_objects=10): obj_tensors = [TensorClevrObject.from_clevr_object(obj).value for obj in clevr_object_set.objects] ndummies = max_n_objects - len(obj_tensors) assert ndummies >= 0 dummies = [Variable(util.cuda(torch.zeros(1, 4, TensorClevrObjectSet.max_prop_len))) for i in range(ndummies)] return TensorClevrObjectSet(torch.cat(obj_tensors + dummies, 0))