[Tests] Add Common Test for Training + Fix a couple of bugs (#8415)

* add training tests

* correct longformer

* fix docs

* fix some tests

* fix some more train tests

* remove ipdb

* fix multiple edge case model training

* fix funnel and prophetnet

* clean gpt models

* undo renaming of albert
This commit is contained in:
Patrick von Platen
2020-11-09 18:24:41 +01:00
committed by GitHub
parent 52040517b8
commit 9c83b96e62
30 changed files with 445 additions and 34 deletions

View File

@@ -14,6 +14,7 @@
# limitations under the License.
import copy
import unittest
from transformers import is_torch_available
@@ -26,7 +27,14 @@ from .test_modeling_common import ModelTesterMixin, ids_tensor
if is_torch_available():
import torch
from transformers import LxmertConfig, LxmertForPreTraining, LxmertForQuestionAnswering, LxmertModel
from transformers import (
MODEL_FOR_PRETRAINING_MAPPING,
MODEL_FOR_QUESTION_ANSWERING_MAPPING,
LxmertConfig,
LxmertForPreTraining,
LxmertForQuestionAnswering,
LxmertModel,
)
from transformers.modeling_lxmert import LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST
@@ -533,6 +541,22 @@ class LxmertModelTest(ModelTesterMixin, unittest.TestCase):
test_pruning = False
test_torchscript = False
# overwrite function because qa models takes different input label shape
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
inputs_dict = copy.deepcopy(inputs_dict)
if return_labels:
if model_class in MODEL_FOR_QUESTION_ANSWERING_MAPPING.values():
inputs_dict["labels"] = torch.zeros(
self.model_tester.batch_size, dtype=torch.long, device=torch_device
)
elif model_class in MODEL_FOR_PRETRAINING_MAPPING.values():
# special case for models like BERT that use multi-loss training for PreTraining
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 setUp(self):
self.model_tester = LxmertModelTester(self)
self.config_tester = ConfigTester(self, config_class=LxmertConfig, hidden_size=37)