[Flax] Add docstrings & model outputs (#11498)
* add attentions & hidden states * add model outputs + docs * finish docs * finish tests * finish impl * del @ * finish * finish * correct test * apply sylvains suggestions * Update src/transformers/models/bert/modeling_flax_bert.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * simplify more Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
This commit is contained in:
Patrick von Platen
committed by
GitHub
GitHub
parent
3f6add8bab
commit
f748bd4242
@@ -13,8 +13,10 @@
|
||||
# limitations under the License.
|
||||
|
||||
import copy
|
||||
import inspect
|
||||
import random
|
||||
import tempfile
|
||||
from typing import List, Tuple
|
||||
|
||||
import numpy as np
|
||||
|
||||
@@ -28,6 +30,7 @@ if is_flax_available():
|
||||
|
||||
import jax
|
||||
import jax.numpy as jnp
|
||||
import jaxlib.xla_extension as jax_xla
|
||||
from transformers.modeling_flax_pytorch_utils import (
|
||||
convert_pytorch_state_dict_to_flax,
|
||||
load_flax_weights_in_pytorch_model,
|
||||
@@ -77,6 +80,7 @@ class FlaxModelTesterMixin:
|
||||
inputs_dict = {
|
||||
k: jnp.broadcast_to(v[:, None], (v.shape[0], self.model_tester.num_choices, v.shape[-1]))
|
||||
for k, v in inputs_dict.items()
|
||||
if isinstance(v, (jax_xla.DeviceArray, np.ndarray))
|
||||
}
|
||||
|
||||
return inputs_dict
|
||||
@@ -85,6 +89,41 @@ class FlaxModelTesterMixin:
|
||||
diff = np.abs((a - b)).max()
|
||||
self.assertLessEqual(diff, tol, f"Difference between torch and flax is {diff} (>= {tol}).")
|
||||
|
||||
def test_model_outputs_equivalence(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
def set_nan_tensor_to_zero(t):
|
||||
t[t != t] = 0
|
||||
return t
|
||||
|
||||
def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
|
||||
tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
|
||||
dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs).to_tuple()
|
||||
|
||||
def recursive_check(tuple_object, dict_object):
|
||||
if isinstance(tuple_object, (List, Tuple)):
|
||||
for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
|
||||
recursive_check(tuple_iterable_value, dict_iterable_value)
|
||||
elif tuple_object is None:
|
||||
return
|
||||
else:
|
||||
self.assert_almost_equals(
|
||||
set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), 1e-5
|
||||
)
|
||||
|
||||
recursive_check(tuple_output, dict_output)
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
model = model_class(config)
|
||||
|
||||
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
|
||||
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
|
||||
check_equivalence(model, tuple_inputs, dict_inputs)
|
||||
|
||||
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
|
||||
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
|
||||
check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
|
||||
|
||||
@is_pt_flax_cross_test
|
||||
def test_equivalence_pt_to_flax(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
@@ -108,7 +147,7 @@ class FlaxModelTesterMixin:
|
||||
with torch.no_grad():
|
||||
pt_outputs = pt_model(**pt_inputs).to_tuple()
|
||||
|
||||
fx_outputs = fx_model(**prepared_inputs_dict)
|
||||
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
|
||||
self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
|
||||
for fx_output, pt_output in zip(fx_outputs, pt_outputs):
|
||||
self.assert_almost_equals(fx_output, pt_output.numpy(), 1e-3)
|
||||
@@ -117,7 +156,7 @@ class FlaxModelTesterMixin:
|
||||
pt_model.save_pretrained(tmpdirname)
|
||||
fx_model_loaded = model_class.from_pretrained(tmpdirname, from_pt=True)
|
||||
|
||||
fx_outputs_loaded = fx_model_loaded(**prepared_inputs_dict)
|
||||
fx_outputs_loaded = fx_model_loaded(**prepared_inputs_dict).to_tuple()
|
||||
self.assertEqual(
|
||||
len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch"
|
||||
)
|
||||
@@ -149,7 +188,7 @@ class FlaxModelTesterMixin:
|
||||
with torch.no_grad():
|
||||
pt_outputs = pt_model(**pt_inputs).to_tuple()
|
||||
|
||||
fx_outputs = fx_model(**prepared_inputs_dict)
|
||||
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
|
||||
self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
|
||||
for fx_output, pt_output in zip(fx_outputs, pt_outputs):
|
||||
self.assert_almost_equals(fx_output, pt_output.numpy(), 1e-3)
|
||||
@@ -171,17 +210,20 @@ class FlaxModelTesterMixin:
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
if model_class.__name__ != "FlaxBertModel":
|
||||
continue
|
||||
|
||||
with self.subTest(model_class.__name__):
|
||||
model = model_class(config)
|
||||
|
||||
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
|
||||
outputs = model(**prepared_inputs_dict)
|
||||
outputs = model(**prepared_inputs_dict).to_tuple()
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
model_loaded = model_class.from_pretrained(tmpdirname)
|
||||
|
||||
outputs_loaded = model_loaded(**prepared_inputs_dict)
|
||||
outputs_loaded = model_loaded(**prepared_inputs_dict).to_tuple()
|
||||
for output_loaded, output in zip(outputs_loaded, outputs):
|
||||
self.assert_almost_equals(output_loaded, output, 1e-3)
|
||||
|
||||
@@ -195,19 +237,47 @@ class FlaxModelTesterMixin:
|
||||
|
||||
@jax.jit
|
||||
def model_jitted(input_ids, attention_mask=None, token_type_ids=None):
|
||||
return model(input_ids, attention_mask, token_type_ids)
|
||||
return model(
|
||||
input_ids=input_ids,
|
||||
attention_mask=attention_mask,
|
||||
token_type_ids=token_type_ids,
|
||||
).to_tuple()
|
||||
|
||||
with self.subTest("JIT Enabled"):
|
||||
jitted_outputs = model_jitted(**prepared_inputs_dict)
|
||||
|
||||
with self.subTest("JIT Disabled"):
|
||||
with jax.disable_jit():
|
||||
outputs = model_jitted(**prepared_inputs_dict)
|
||||
|
||||
with self.subTest("JIT Enabled"):
|
||||
jitted_outputs = model_jitted(**prepared_inputs_dict)
|
||||
|
||||
self.assertEqual(len(outputs), len(jitted_outputs))
|
||||
for jitted_output, output in zip(jitted_outputs, outputs):
|
||||
self.assertEqual(jitted_output.shape, output.shape)
|
||||
|
||||
@jax.jit
|
||||
def model_jitted_return_dict(input_ids, attention_mask=None, token_type_ids=None):
|
||||
return model(
|
||||
input_ids=input_ids,
|
||||
attention_mask=attention_mask,
|
||||
token_type_ids=token_type_ids,
|
||||
)
|
||||
|
||||
# jitted function cannot return OrderedDict
|
||||
with self.assertRaises(TypeError):
|
||||
model_jitted_return_dict(**prepared_inputs_dict)
|
||||
|
||||
def test_forward_signature(self):
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
model = model_class(config)
|
||||
signature = inspect.signature(model.__call__)
|
||||
# signature.parameters is an OrderedDict => so arg_names order is deterministic
|
||||
arg_names = [*signature.parameters.keys()]
|
||||
|
||||
expected_arg_names = ["input_ids", "attention_mask"]
|
||||
self.assertListEqual(arg_names[:2], expected_arg_names)
|
||||
|
||||
def test_naming_convention(self):
|
||||
for model_class in self.all_model_classes:
|
||||
model_class_name = model_class.__name__
|
||||
@@ -218,3 +288,30 @@ class FlaxModelTesterMixin:
|
||||
module_cls = getattr(bert_modeling_flax_module, module_class_name)
|
||||
|
||||
self.assertIsNotNone(module_cls)
|
||||
|
||||
def test_hidden_states_output(self):
|
||||
def check_hidden_states_output(inputs_dict, config, model_class):
|
||||
model = model_class(config)
|
||||
|
||||
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
|
||||
hidden_states = outputs.hidden_states
|
||||
|
||||
self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
|
||||
seq_length = self.model_tester.seq_length
|
||||
|
||||
self.assertListEqual(
|
||||
list(hidden_states[0].shape[-2:]),
|
||||
[seq_length, self.model_tester.hidden_size],
|
||||
)
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
inputs_dict["output_hidden_states"] = True
|
||||
check_hidden_states_output(inputs_dict, config, model_class)
|
||||
|
||||
# check that output_hidden_states also work using config
|
||||
del inputs_dict["output_hidden_states"]
|
||||
config.output_hidden_states = True
|
||||
|
||||
check_hidden_states_output(inputs_dict, config, model_class)
|
||||
|
||||
Reference in New Issue
Block a user