def test_save_sup_load_rl(self):
pass
model_to_save = MockModel(spinn.spinn_core_model.BaseModel,
default_args())
# Parse command line flags.
get_flags()
FLAGS(sys.argv)
log_temp = tempfile.NamedTemporaryFile()
ckpt_temp = tempfile.NamedTemporaryFile()
logger = afs_safe_logger.ProtoLogger(log_temp.name)
FLAGS.ckpt_path = '.'
trainer_to_save = ModelTrainer(model_to_save, logger, FLAGS)
model_to_load = MockModel(spinn.rl_spinn.BaseModel, default_args())
trainer_to_load = ModelTrainer(model_to_load, logger, FLAGS)
# Save to and load from temporary file.
trainer_to_save.save(ckpt_temp.name)
trainer_to_load.load(ckpt_temp.name, cpu=True)
compare_models(model_to_save, model_to_load)
# Cleanup temporary file.
ckpt_temp.close()
log_temp.close()
def test_init_models(self):
MockModel(spinn.spinn_core_model.BaseModel, default_args())
MockModel(spinn.rl_spinn.BaseModel, default_args())
MockModel(spinn.spinn_core_model.BaseModel,
default_args(use_sentence_pair=True))
MockModel(spinn.rl_spinn.BaseModel,
default_args(use_sentence_pair=True))
def test_soft_eval(self):
args = default_args()
args['test_temperature_multiplier'] = 0.0000001
model = MockModel(Pyramid, args)
X, transitions = get_batch()
class Projection(nn.Module):
def forward(self, x):
return x[:, :default_args()['model_dim']]
class mlp(nn.Module):
def forward(self, x):
return x
def selection_fn(selection_input):
# Compose in random order
return Variable(torch.rand(selection_input.data.size()[0], 1))
class Reduce(nn.Module):
def forward(self, lefts, rights):
return lefts + rights
model.encode = Projection()
model.composition_fn = Reduce()
model.selection_fn = selection_fn
model.mlp = mlp()
model.eval()
outputs = model(X, transitions)
assert outputs[0][0].data[0] == (3 + 1 + 2 + 1)
assert outputs[1][0].data[0] == (3 + 2 + 4 + 5)
def test_pair_rnn(self):
model = MockModel(RNNModel, default_args(use_sentence_pair=True))
X, transitions = get_batch_pair()
assert X.shape == (2, 4, 2)
assert transitions.shape == (2, 7, 2)
outputs = model(X, transitions)
assert outputs.size() == (2, 3)
def test_basic_stack(self):
model = MockModel(SentenceModel, default_args())
train = False
X, transitions = get_batch()
class Projection(nn.Module):
def forward(self, x):
return x[:, :default_args()['model_dim']]
class Reduce(nn.Module):
def forward(self, lefts, rights, tracking):
batch_size = len(lefts)
return torch.chunk(
torch.cat(lefts, 0) - torch.cat(rights, 0), batch_size, 0)
model.embed.projection = Projection()
model.spinn.reduce = Reduce()
model(X, transitions)
outputs = model.spinn_outp[0]
assert outputs[0][0].data[0] == (3 - (1 - (2 - 1)))
assert outputs[1][0].data[0] == ((3 - 2) - (4 - 5))
def test_save_load_model(self):
model_to_save = MockModel(BaseModel, default_args())
model_to_load = MockModel(BaseModel, default_args())
# Save to and load from temporary file.
temp = tempfile.NamedTemporaryFile()
torch.save(model_to_save.state_dict(), temp.name)
model_to_load.load_state_dict(torch.load(temp.name))
compare_models(model_to_save, model_to_load)
# Cleanup temporary file.
temp.close()
def test_validate_transitions_cantskip(self):
model = MockModel(BaseModel, default_args())
train = False
# To Test:
# 1. Cant SKIP
# 2. Cant SHIFT
# 3. Cant REDUCE
# 4. No change SHIFT
# 5. No change REDUCE
bufs = [
[None],
[],
[None],
[None],
[None],
[],
]
stacks = [
[None],
[None],
[None],
[],
[],
[None, None],
]
transitions = [
2, 1, 0, 0, 0, 1
]
preds = np.array([
0, 0, 1, 1, 0, 1
]).astype(np.int32)
ret, _ = model.spinn.validate(transitions, preds, stacks, bufs, zero_padded=False)
expected = np.array([
2, 1, 0, 0, 0, 1
], dtype=np.int32)
assert all(p == e for p, e in zip(ret.ravel().tolist(), expected.ravel().tolist())), \
"gold: {}\npreds: {}".format(expected, ret)
def test_save_sup_load_rl(self):
pass
model_to_save = MockModel(spinn.pyramid.Pyramid, default_args())
opt_to_save = optim.SGD(model_to_save.parameters(), lr=0.1)
trainer_to_save = ModelTrainer(model_to_save, opt_to_save)
model_to_load = MockModel(spinn.pyramid.Pyramid, default_args())
opt_to_load = optim.SGD(model_to_load.parameters(), lr=0.1)
trainer_to_load = ModelTrainer(model_to_load, opt_to_load)
# Save to and load from temporary file.
temp = tempfile.NamedTemporaryFile()
trainer_to_save.save(temp.name, 0, 0)
trainer_to_load.load(temp.name)
compare_models(model_to_save, model_to_load)
# Cleanup temporary file.
temp.close()
def test_save_sup_load_rl(self):
scalar = 11
other_scalar = 0
model_to_save = MockModel(spinn.fat_stack.BaseModel, default_args())
opt_to_save = optim.SGD(model_to_save.parameters(), lr=0.1)
trainer_to_save = ModelTrainer(model_to_save, opt_to_save)
model_to_load = MockModel(spinn.rl_spinn.BaseModel, default_args())
opt_to_load = optim.SGD(model_to_load.parameters(), lr=0.1)
trainer_to_load = ModelTrainer(model_to_load, opt_to_load)
# Save to and load from temporary file.
temp = tempfile.NamedTemporaryFile()
trainer_to_save.save(temp.name, 0, 0)
trainer_to_load.load(temp.name)
compare_models(model_to_save, model_to_load)
# Cleanup temporary file.
temp.close()
def test_single_cbow(self):
model = MockModel(SentenceModel, default_args())
X, transitions = get_batch()
outputs = model(X, transitions)
assert outputs.size() == (2, 3)
def test_init_models(self):
MockModel(spinn.pyramid.Pyramid, default_args())
MockModel(spinn.pyramid.Pyramid, default_args(use_sentence_pair=True))