SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Test model outputs equivalence (#6445)

Browse Source 
* Test model outputs equivalence

* Fix failing tests

* From dict to kwargs

* DistilBERT

* Addressing @sgugger and @patrickvonplaten's comments
This commit is contained in:
Lysandre Debut
2020-08-13 11:59:35 -04:00
committed by GitHub
parent 54c687e97c
commit f7cbc13db7
4 changed files with 197 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

94
tests/test_modeling_common.py
View File

@@ -18,7 +18,7 @@ import os.path
import random
import tempfile
import unittest
from typing import List
from typing import List, Tuple
from transformers import is_torch_available
from transformers.testing_utils import require_multigpu, require_torch, slow, torch_device
@@ -37,6 +37,11 @@ if is_torch_available():
BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
MODEL_FOR_QUESTION_ANSWERING_MAPPING,
MODEL_FOR_CAUSAL_LM_MAPPING,
MODEL_FOR_MASKED_LM_MAPPING,
MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
top_k_top_p_filtering,
)
@@ -63,14 +68,39 @@ class ModelTesterMixin:
test_chunking = False
is_encoder_decoder = False
def _prepare_for_class(self, inputs_dict, model_class):
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
inputs_dict = copy.deepcopy(inputs_dict)
if model_class in MODEL_FOR_MULTIPLE_CHOICE_MAPPING.values():
return {
inputs_dict = {
k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
if isinstance(v, torch.Tensor) and v.ndim > 1
else v
for k, v in inputs_dict.items()
}
if return_labels:
if model_class in MODEL_FOR_MULTIPLE_CHOICE_MAPPING.values():
inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
elif model_class in MODEL_FOR_QUESTION_ANSWERING_MAPPING.values():
inputs_dict["start_positions"] = torch.zeros(
self.model_tester.batch_size, dtype=torch.long, device=torch_device
)
inputs_dict["end_positions"] = torch.zeros(
self.model_tester.batch_size, dtype=torch.long, device=torch_device
)
elif model_class in MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING.values():
inputs_dict["labels"] = torch.zeros(
self.model_tester.batch_size, dtype=torch.long, device=torch_device
)
elif model_class in [
*MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING.values(),
*MODEL_FOR_CAUSAL_LM_MAPPING.values(),
*MODEL_FOR_MASKED_LM_MAPPING.values(),
*MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING.values(),
]:
inputs_dict["labels"] = torch.zeros(
(self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
)
return inputs_dict
def test_save_load(self):
@@ -663,6 +693,64 @@ class ModelTesterMixin:
# self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape)
# self.assertTrue(check_same_values(model.transformer.wte, model.lm_head))
def test_model_outputs_equivalence(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
with torch.no_grad():
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.assertTrue(
torch.allclose(tuple_object, dict_object, atol=1e-5),
msg=f"Tuple and dict output are not equal. Difference: {torch.max(torch.abs(tuple_object - dict_object))}",
)
recursive_check(tuple_output, dict_output)
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
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, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
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})
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_attentions": True})
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(
model, tuple_inputs, dict_inputs, {"output_hidden_states": True, "output_attentions": True}
)
def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()