2019-09-05 08:27:39 +08:00
|
|
|
# coding=utf-8
|
|
|
|
# Copyright 2019 HuggingFace Inc.
|
|
|
|
#
|
|
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
|
|
# you may not use this file except in compliance with the License.
|
|
|
|
# You may obtain a copy of the License at
|
|
|
|
#
|
|
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
#
|
|
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
|
|
# See the License for the specific language governing permissions and
|
|
|
|
# limitations under the License.
|
2019-12-22 23:20:32 +08:00
|
|
|
|
2019-09-05 08:27:39 +08:00
|
|
|
|
|
|
|
import copy
|
2020-06-24 23:37:20 +08:00
|
|
|
import inspect
|
2021-02-15 20:55:10 +08:00
|
|
|
import json
|
2019-12-21 22:57:32 +08:00
|
|
|
import os
|
2019-09-05 08:27:39 +08:00
|
|
|
import random
|
2019-12-21 22:57:32 +08:00
|
|
|
import tempfile
|
2020-03-03 22:42:15 +08:00
|
|
|
import unittest
|
2020-03-03 23:15:30 +08:00
|
|
|
from importlib import import_module
|
2020-08-13 23:59:35 +08:00
|
|
|
from typing import List, Tuple
|
2019-09-05 08:27:39 +08:00
|
|
|
|
2021-11-03 06:58:42 +08:00
|
|
|
from huggingface_hub import delete_repo, login
|
2021-04-23 21:17:37 +08:00
|
|
|
from requests.exceptions import HTTPError
|
2020-10-23 21:58:19 +08:00
|
|
|
from transformers import is_tf_available
|
2021-04-09 06:41:36 +08:00
|
|
|
from transformers.models.auto import get_values
|
2021-11-18 04:24:39 +08:00
|
|
|
from transformers.testing_utils import tooslow # noqa: F401
|
2021-03-16 05:28:01 +08:00
|
|
|
from transformers.testing_utils import (
|
2021-04-23 21:17:37 +08:00
|
|
|
PASS,
|
|
|
|
USER,
|
2021-07-13 22:15:15 +08:00
|
|
|
CaptureLogger,
|
2021-03-16 05:28:01 +08:00
|
|
|
_tf_gpu_memory_limit,
|
|
|
|
is_pt_tf_cross_test,
|
2021-04-23 21:17:37 +08:00
|
|
|
is_staging_test,
|
2021-03-16 05:28:01 +08:00
|
|
|
require_tf,
|
2022-01-15 01:35:39 +08:00
|
|
|
require_tf2onnx,
|
2021-03-16 05:28:01 +08:00
|
|
|
slow,
|
2022-03-14 20:31:32 +08:00
|
|
|
torch_device,
|
2021-03-16 05:28:01 +08:00
|
|
|
)
|
2021-07-13 22:15:15 +08:00
|
|
|
from transformers.utils import logging
|
2019-12-07 02:57:38 +08:00
|
|
|
|
2019-12-21 22:57:32 +08:00
|
|
|
|
2022-03-14 20:31:32 +08:00
|
|
|
logger = logging.get_logger(__name__)
|
|
|
|
|
|
|
|
|
2019-09-09 17:04:03 +08:00
|
|
|
if is_tf_available():
|
2019-09-24 04:08:08 +08:00
|
|
|
import numpy as np
|
2020-08-24 23:03:01 +08:00
|
|
|
import tensorflow as tf
|
2019-12-21 22:46:46 +08:00
|
|
|
|
2020-06-11 22:31:26 +08:00
|
|
|
from transformers import (
|
2020-08-24 23:03:01 +08:00
|
|
|
TF_MODEL_FOR_CAUSAL_LM_MAPPING,
|
2021-11-09 20:54:37 +08:00
|
|
|
TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
|
2020-08-24 23:03:01 +08:00
|
|
|
TF_MODEL_FOR_MASKED_LM_MAPPING,
|
2020-06-11 22:31:26 +08:00
|
|
|
TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
|
2020-11-11 22:02:06 +08:00
|
|
|
TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
|
2020-11-13 03:08:26 +08:00
|
|
|
TF_MODEL_FOR_PRETRAINING_MAPPING,
|
2020-06-24 23:37:20 +08:00
|
|
|
TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING,
|
2020-08-24 23:03:01 +08:00
|
|
|
TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
|
2020-06-24 23:37:20 +08:00
|
|
|
TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
|
2022-02-09 00:27:23 +08:00
|
|
|
TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
|
2020-06-24 23:37:20 +08:00
|
|
|
TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
|
2021-04-23 21:17:37 +08:00
|
|
|
BertConfig,
|
2021-10-22 00:31:29 +08:00
|
|
|
TFAutoModel,
|
2021-07-13 22:15:15 +08:00
|
|
|
TFAutoModelForSequenceClassification,
|
2021-04-23 21:17:37 +08:00
|
|
|
TFBertModel,
|
2020-08-24 23:03:01 +08:00
|
|
|
TFSharedEmbeddings,
|
|
|
|
tf_top_k_top_p_filtering,
|
2020-06-11 22:31:26 +08:00
|
|
|
)
|
Add output in a dictionary for TF `generate` method (#12139)
* Add output args to greedy search
* Fix critical typo + make style quality
* Handle generate_beam_search
* Add dict_specific tests and fix the placement of encoder outputs
* Add specific outputs
* Update doc
* Fix typo
* Adjust handling encoder_outputs + Fix generating for T5
* Fix generate for RAG
* Fix handling ouptut_attentions when target_mapping is not None
Take care of situations when target_mapping is provided
as there are 2-tuple of attentions
Change from:
if inputs["output_attentions"]:
attentions = tuple(tf.transpose(t, perm(2, 3, 0, 1)) for t in attentions)
to:
if inputs["output_attentions"]:
if inputs["target_mapping"] is not None:
# when target_mapping is provided, there are 2-tuple of attentions
attentions = tuple(
tuple(tf.transpose(attn_stream, perm=(2, 3, 0, 1)) for attn_stream in t) for t in attentions
)
else:
attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
* Rename kwargs to model_kwargs
* make style quality
* Move imports in test_modeling_tf_common.py
Move ModelOutput-related imports in test_modeling_tf_common.py
into the `is_tf_available():` statement.
* Rewrite nested if-statements
* Fix added tests
2021-06-23 17:52:11 +08:00
|
|
|
from transformers.generation_tf_utils import (
|
|
|
|
TFBeamSampleDecoderOnlyOutput,
|
|
|
|
TFBeamSampleEncoderDecoderOutput,
|
|
|
|
TFBeamSearchDecoderOnlyOutput,
|
|
|
|
TFBeamSearchEncoderDecoderOutput,
|
|
|
|
TFGreedySearchDecoderOnlyOutput,
|
|
|
|
TFGreedySearchEncoderDecoderOutput,
|
|
|
|
TFSampleDecoderOnlyOutput,
|
|
|
|
TFSampleEncoderDecoderOutput,
|
|
|
|
)
|
2020-03-03 22:42:15 +08:00
|
|
|
|
2020-03-03 04:45:25 +08:00
|
|
|
if _tf_gpu_memory_limit is not None:
|
|
|
|
gpus = tf.config.list_physical_devices("GPU")
|
|
|
|
for gpu in gpus:
|
|
|
|
# Restrict TensorFlow to only allocate x GB of memory on the GPUs
|
|
|
|
try:
|
2021-01-05 16:58:26 +08:00
|
|
|
tf.config.set_logical_device_configuration(
|
|
|
|
gpu, [tf.config.LogicalDeviceConfiguration(memory_limit=_tf_gpu_memory_limit)]
|
2020-03-03 04:45:25 +08:00
|
|
|
)
|
2021-01-05 16:58:26 +08:00
|
|
|
logical_gpus = tf.config.list_logical_devices("GPU")
|
2020-03-03 04:45:25 +08:00
|
|
|
print("Logical GPUs", logical_gpus)
|
|
|
|
except RuntimeError as e:
|
|
|
|
# Virtual devices must be set before GPUs have been initialized
|
|
|
|
print(e)
|
2019-09-05 08:27:39 +08:00
|
|
|
|
2019-12-21 22:46:46 +08:00
|
|
|
|
2019-09-05 08:27:39 +08:00
|
|
|
def _config_zero_init(config):
|
|
|
|
configs_no_init = copy.deepcopy(config)
|
|
|
|
for key in configs_no_init.__dict__.keys():
|
2019-12-21 22:46:46 +08:00
|
|
|
if "_range" in key or "_std" in key:
|
2019-09-05 08:27:39 +08:00
|
|
|
setattr(configs_no_init, key, 0.0)
|
|
|
|
return configs_no_init
|
|
|
|
|
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
@require_tf
|
|
|
|
class TFModelTesterMixin:
|
2019-10-10 19:08:24 +08:00
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
model_tester = None
|
|
|
|
all_model_classes = ()
|
2020-03-03 22:42:15 +08:00
|
|
|
all_generative_model_classes = ()
|
2021-10-22 00:31:29 +08:00
|
|
|
test_mismatched_shapes = True
|
2019-12-22 21:57:20 +08:00
|
|
|
test_resize_embeddings = True
|
2021-01-28 23:09:13 +08:00
|
|
|
test_head_masking = True
|
2019-12-22 21:57:20 +08:00
|
|
|
is_encoder_decoder = False
|
2019-10-10 19:08:24 +08:00
|
|
|
|
2020-10-30 22:25:48 +08:00
|
|
|
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False) -> dict:
|
2020-08-13 23:59:35 +08:00
|
|
|
inputs_dict = copy.deepcopy(inputs_dict)
|
|
|
|
|
2021-04-09 06:41:36 +08:00
|
|
|
if model_class in get_values(TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
|
2020-06-24 23:37:20 +08:00
|
|
|
inputs_dict = {
|
2020-07-30 02:26:26 +08:00
|
|
|
k: tf.tile(tf.expand_dims(v, 1), (1, self.model_tester.num_choices) + (1,) * (v.ndim - 1))
|
|
|
|
if isinstance(v, tf.Tensor) and v.ndim > 0
|
2020-06-11 22:31:26 +08:00
|
|
|
else v
|
|
|
|
for k, v in inputs_dict.items()
|
|
|
|
}
|
2020-06-24 23:37:20 +08:00
|
|
|
|
|
|
|
if return_labels:
|
2021-04-09 06:41:36 +08:00
|
|
|
if model_class in get_values(TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
|
2020-08-13 23:59:35 +08:00
|
|
|
inputs_dict["labels"] = tf.ones(self.model_tester.batch_size, dtype=tf.int32)
|
2021-04-09 06:41:36 +08:00
|
|
|
elif model_class in get_values(TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING):
|
2020-08-13 23:59:35 +08:00
|
|
|
inputs_dict["start_positions"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)
|
|
|
|
inputs_dict["end_positions"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)
|
2021-11-09 20:54:37 +08:00
|
|
|
elif model_class in [
|
|
|
|
*get_values(TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING),
|
|
|
|
*get_values(TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING),
|
|
|
|
]:
|
2020-08-13 23:59:35 +08:00
|
|
|
inputs_dict["labels"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)
|
2021-04-09 06:41:36 +08:00
|
|
|
elif model_class in get_values(TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING):
|
2020-11-11 22:02:06 +08:00
|
|
|
inputs_dict["next_sentence_label"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)
|
2020-08-13 23:59:35 +08:00
|
|
|
elif model_class in [
|
2021-04-09 06:41:36 +08:00
|
|
|
*get_values(TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING),
|
|
|
|
*get_values(TF_MODEL_FOR_CAUSAL_LM_MAPPING),
|
|
|
|
*get_values(TF_MODEL_FOR_MASKED_LM_MAPPING),
|
|
|
|
*get_values(TF_MODEL_FOR_PRETRAINING_MAPPING),
|
|
|
|
*get_values(TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING),
|
2022-02-09 00:27:23 +08:00
|
|
|
*get_values(TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING),
|
2020-08-13 23:59:35 +08:00
|
|
|
]:
|
|
|
|
inputs_dict["labels"] = tf.zeros(
|
|
|
|
(self.model_tester.batch_size, self.model_tester.seq_length), dtype=tf.int32
|
|
|
|
)
|
2020-06-11 22:31:26 +08:00
|
|
|
return inputs_dict
|
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
def test_initialization(self):
|
|
|
|
pass
|
2019-10-10 19:08:24 +08:00
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
def test_save_load(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
2019-10-10 19:08:24 +08:00
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
2020-06-11 22:31:26 +08:00
|
|
|
outputs = model(self._prepare_for_class(inputs_dict, model_class))
|
2019-10-10 19:08:24 +08:00
|
|
|
|
2019-12-23 01:12:11 +08:00
|
|
|
with tempfile.TemporaryDirectory() as tmpdirname:
|
2021-01-07 18:48:49 +08:00
|
|
|
model.save_pretrained(tmpdirname, saved_model=False)
|
2019-12-22 21:57:20 +08:00
|
|
|
model = model_class.from_pretrained(tmpdirname)
|
2020-06-11 22:31:26 +08:00
|
|
|
after_outputs = model(self._prepare_for_class(inputs_dict, model_class))
|
2019-12-22 21:57:20 +08:00
|
|
|
|
2020-03-03 23:15:30 +08:00
|
|
|
self.assert_outputs_same(after_outputs, outputs)
|
2019-12-22 21:57:20 +08:00
|
|
|
|
2021-11-11 22:21:50 +08:00
|
|
|
def test_save_load_config(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
outputs = model(self._prepare_for_class(inputs_dict, model_class))
|
2021-11-18 04:24:39 +08:00
|
|
|
model_config = model.get_config()
|
|
|
|
# make sure that returned config is jsonifiable, which is required by keras
|
|
|
|
json.dumps(model_config)
|
2021-11-11 22:21:50 +08:00
|
|
|
new_model = model_class.from_config(model.get_config())
|
2021-11-18 04:24:39 +08:00
|
|
|
# make sure it also accepts a normal config
|
|
|
|
_ = model_class.from_config(model.config)
|
2021-11-11 22:21:50 +08:00
|
|
|
_ = new_model(self._prepare_for_class(inputs_dict, model_class)) # Build model
|
|
|
|
new_model.set_weights(model.get_weights())
|
|
|
|
after_outputs = new_model(self._prepare_for_class(inputs_dict, model_class))
|
|
|
|
|
|
|
|
self.assert_outputs_same(after_outputs, outputs)
|
|
|
|
|
2020-10-01 23:38:50 +08:00
|
|
|
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()]
|
|
|
|
|
|
|
|
if model.config.is_encoder_decoder:
|
|
|
|
expected_arg_names = [
|
2020-11-25 02:55:00 +08:00
|
|
|
"input_ids",
|
2020-10-01 23:38:50 +08:00
|
|
|
"attention_mask",
|
|
|
|
"decoder_input_ids",
|
|
|
|
"decoder_attention_mask",
|
|
|
|
]
|
2021-01-20 05:50:25 +08:00
|
|
|
expected_arg_names.extend(
|
2021-04-26 20:16:21 +08:00
|
|
|
["head_mask", "decoder_head_mask"] if "head_mask" and "decoder_head_mask" in arg_names else []
|
|
|
|
)
|
|
|
|
# Necessary to handle BART with newly added cross_attn_head_mask
|
|
|
|
expected_arg_names.extend(
|
|
|
|
["cross_attn_head_mask", "encoder_outputs"]
|
|
|
|
if "cross_attn_head_mask" in arg_names
|
2021-01-20 05:50:25 +08:00
|
|
|
else ["encoder_outputs"]
|
|
|
|
)
|
|
|
|
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
|
2020-10-01 23:38:50 +08:00
|
|
|
|
|
|
|
else:
|
2020-11-25 02:55:00 +08:00
|
|
|
expected_arg_names = ["input_ids"]
|
2020-10-01 23:38:50 +08:00
|
|
|
self.assertListEqual(arg_names[:1], expected_arg_names)
|
|
|
|
|
2021-02-15 20:55:10 +08:00
|
|
|
def test_onnx_compliancy(self):
|
|
|
|
if not self.test_onnx:
|
|
|
|
return
|
|
|
|
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
INTERNAL_OPS = [
|
|
|
|
"Assert",
|
|
|
|
"AssignVariableOp",
|
|
|
|
"EmptyTensorList",
|
|
|
|
"ReadVariableOp",
|
|
|
|
"ResourceGather",
|
|
|
|
"TruncatedNormal",
|
|
|
|
"VarHandleOp",
|
|
|
|
"VarIsInitializedOp",
|
|
|
|
]
|
|
|
|
onnx_ops = []
|
|
|
|
|
|
|
|
with open(os.path.join(".", "utils", "tf_ops", "onnx.json")) as f:
|
|
|
|
onnx_opsets = json.load(f)["opsets"]
|
|
|
|
|
|
|
|
for i in range(1, self.onnx_min_opset + 1):
|
|
|
|
onnx_ops.extend(onnx_opsets[str(i)])
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model_op_names = set()
|
|
|
|
|
|
|
|
with tf.Graph().as_default() as g:
|
|
|
|
model = model_class(config)
|
|
|
|
model(model.dummy_inputs)
|
|
|
|
|
|
|
|
for op in g.get_operations():
|
|
|
|
model_op_names.add(op.node_def.op)
|
|
|
|
|
|
|
|
model_op_names = sorted(model_op_names)
|
|
|
|
incompatible_ops = []
|
|
|
|
|
|
|
|
for op in model_op_names:
|
|
|
|
if op not in onnx_ops and op not in INTERNAL_OPS:
|
|
|
|
incompatible_ops.append(op)
|
|
|
|
|
|
|
|
self.assertEqual(len(incompatible_ops), 0, incompatible_ops)
|
|
|
|
|
2022-01-15 01:35:39 +08:00
|
|
|
@require_tf2onnx
|
2021-02-15 20:55:10 +08:00
|
|
|
@slow
|
|
|
|
def test_onnx_runtime_optimize(self):
|
|
|
|
if not self.test_onnx:
|
|
|
|
return
|
|
|
|
|
|
|
|
import onnxruntime
|
2022-01-15 01:35:39 +08:00
|
|
|
import tf2onnx
|
2021-02-15 20:55:10 +08:00
|
|
|
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
model(model.dummy_inputs)
|
|
|
|
|
2022-01-15 01:35:39 +08:00
|
|
|
onnx_model_proto, _ = tf2onnx.convert.from_keras(model, opset=self.onnx_min_opset)
|
2021-02-15 20:55:10 +08:00
|
|
|
|
2022-01-15 01:35:39 +08:00
|
|
|
onnxruntime.InferenceSession(onnx_model_proto.SerializeToString())
|
2021-02-15 20:55:10 +08:00
|
|
|
|
2020-03-03 23:15:30 +08:00
|
|
|
def test_keras_save_load(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
tf_main_layer_classes = set(
|
|
|
|
module_member
|
|
|
|
for model_class in self.all_model_classes
|
|
|
|
for module in (import_module(model_class.__module__),)
|
|
|
|
for module_member_name in dir(module)
|
2020-03-04 06:31:38 +08:00
|
|
|
if module_member_name.endswith("MainLayer")
|
2021-12-24 00:19:44 +08:00
|
|
|
# This condition is required, since `modeling_tf_clip.py` has 3 classes whose names end with `MainLayer`.
|
|
|
|
and module_member_name[: -len("MainLayer")] == model_class.__name__[: -len("Model")]
|
2020-03-03 23:15:30 +08:00
|
|
|
for module_member in (getattr(module, module_member_name),)
|
2020-03-04 06:57:05 +08:00
|
|
|
if isinstance(module_member, type)
|
|
|
|
and tf.keras.layers.Layer in module_member.__bases__
|
|
|
|
and getattr(module_member, "_keras_serializable", False)
|
2020-03-03 23:15:30 +08:00
|
|
|
)
|
|
|
|
for main_layer_class in tf_main_layer_classes:
|
2020-06-05 07:45:53 +08:00
|
|
|
# T5MainLayer needs an embed_tokens parameter when called without the inputs_embeds parameter
|
|
|
|
if "T5" in main_layer_class.__name__:
|
|
|
|
# Take the same values than in TFT5ModelTester for this shared layer
|
|
|
|
shared = TFSharedEmbeddings(99, 32, name="shared")
|
2020-12-15 01:47:00 +08:00
|
|
|
config.use_cache = inputs_dict.pop("use_cache", None)
|
2020-06-05 07:45:53 +08:00
|
|
|
main_layer = main_layer_class(config, embed_tokens=shared)
|
|
|
|
else:
|
|
|
|
main_layer = main_layer_class(config)
|
2020-12-15 01:47:00 +08:00
|
|
|
|
2020-03-03 23:15:30 +08:00
|
|
|
symbolic_inputs = {
|
|
|
|
name: tf.keras.Input(tensor.shape[1:], dtype=tensor.dtype) for name, tensor in inputs_dict.items()
|
|
|
|
}
|
2020-06-05 07:45:53 +08:00
|
|
|
|
2020-03-03 23:15:30 +08:00
|
|
|
model = tf.keras.Model(symbolic_inputs, outputs=main_layer(symbolic_inputs))
|
|
|
|
outputs = model(inputs_dict)
|
|
|
|
|
|
|
|
with tempfile.TemporaryDirectory() as tmpdirname:
|
|
|
|
filepath = os.path.join(tmpdirname, "keras_model.h5")
|
|
|
|
model.save(filepath)
|
2020-06-05 07:45:53 +08:00
|
|
|
if "T5" in main_layer_class.__name__:
|
|
|
|
model = tf.keras.models.load_model(
|
|
|
|
filepath,
|
|
|
|
custom_objects={
|
|
|
|
main_layer_class.__name__: main_layer_class,
|
|
|
|
"TFSharedEmbeddings": TFSharedEmbeddings,
|
|
|
|
},
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
model = tf.keras.models.load_model(
|
|
|
|
filepath, custom_objects={main_layer_class.__name__: main_layer_class}
|
|
|
|
)
|
2020-03-03 23:15:30 +08:00
|
|
|
assert isinstance(model, tf.keras.Model)
|
|
|
|
after_outputs = model(inputs_dict)
|
|
|
|
self.assert_outputs_same(after_outputs, outputs)
|
|
|
|
|
|
|
|
def assert_outputs_same(self, after_outputs, outputs):
|
|
|
|
# Make sure we don't have nans
|
2020-06-05 07:45:53 +08:00
|
|
|
if isinstance(after_outputs, tf.Tensor):
|
|
|
|
out_1 = after_outputs.numpy()
|
2020-08-05 23:34:39 +08:00
|
|
|
elif isinstance(after_outputs, dict):
|
2020-11-17 00:43:00 +08:00
|
|
|
out_1 = after_outputs[list(after_outputs.keys())[0]].numpy()
|
2020-06-05 07:45:53 +08:00
|
|
|
else:
|
|
|
|
out_1 = after_outputs[0].numpy()
|
2020-03-03 23:15:30 +08:00
|
|
|
out_2 = outputs[0].numpy()
|
2020-03-04 06:31:38 +08:00
|
|
|
self.assertEqual(out_1.shape, out_2.shape)
|
2020-03-03 23:15:30 +08:00
|
|
|
out_1 = out_1[~np.isnan(out_1)]
|
|
|
|
out_2 = out_2[~np.isnan(out_2)]
|
|
|
|
max_diff = np.amax(np.abs(out_1 - out_2))
|
|
|
|
self.assertLessEqual(max_diff, 1e-5)
|
2019-10-10 19:08:24 +08:00
|
|
|
|
2020-10-23 21:58:19 +08:00
|
|
|
@is_pt_tf_cross_test
|
2019-12-22 21:57:20 +08:00
|
|
|
def test_pt_tf_model_equivalence(self):
|
|
|
|
import torch
|
2020-08-24 23:03:01 +08:00
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
import transformers
|
2019-09-24 04:08:08 +08:00
|
|
|
|
2022-03-14 20:31:32 +08:00
|
|
|
def prepare_pt_inputs_from_tf_inputs(tf_inputs_dict):
|
2019-09-05 18:02:14 +08:00
|
|
|
|
2020-12-15 01:47:00 +08:00
|
|
|
pt_inputs_dict = {}
|
2022-03-14 20:31:32 +08:00
|
|
|
for name, key in tf_inputs_dict.items():
|
2020-12-15 01:47:00 +08:00
|
|
|
if type(key) == bool:
|
|
|
|
pt_inputs_dict[name] = key
|
2021-06-15 01:58:54 +08:00
|
|
|
elif name == "input_values":
|
|
|
|
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.float32)
|
2021-11-09 20:54:37 +08:00
|
|
|
elif name == "pixel_values":
|
|
|
|
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.float32)
|
2022-02-09 00:27:23 +08:00
|
|
|
elif name == "input_features":
|
|
|
|
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.float32)
|
2020-12-15 01:47:00 +08:00
|
|
|
else:
|
|
|
|
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.long)
|
|
|
|
|
2022-03-14 20:31:32 +08:00
|
|
|
return pt_inputs_dict
|
|
|
|
|
|
|
|
def check_outputs(tf_outputs, pt_outputs, model_class, names):
|
|
|
|
"""
|
|
|
|
Args:
|
|
|
|
model_class: The class of the model that is currently testing. For example, `TFBertModel`,
|
|
|
|
TFBertForMaskedLM`, `TFBertForSequenceClassification`, etc. Currently unused, but it could make
|
|
|
|
debugging easier and faster.
|
|
|
|
|
|
|
|
names: A string, or a tuple of strings. These specify what tf_outputs/pt_outputs represent in the model outputs.
|
|
|
|
Currently unused, but in the future, we could use this information to make the error message clearer
|
|
|
|
by giving the name(s) of the output tensor(s) with large difference(s) between PT and TF.
|
|
|
|
"""
|
|
|
|
|
|
|
|
# Some issue (`about past_key_values`) to solve (e.g. `TFPegasusForConditionalGeneration`) in a separate PR.
|
|
|
|
if names == "past_key_values":
|
|
|
|
return
|
|
|
|
|
|
|
|
# Allow `list` because `(TF)TransfoXLModelOutput.mems` is a list of tensors.
|
|
|
|
if type(tf_outputs) in [tuple, list]:
|
|
|
|
self.assertEqual(type(tf_outputs), type(pt_outputs))
|
|
|
|
self.assertEqual(len(tf_outputs), len(pt_outputs))
|
|
|
|
if type(names) == tuple:
|
|
|
|
for tf_output, pt_output, name in zip(tf_outputs, pt_outputs, names):
|
|
|
|
check_outputs(tf_output, pt_output, model_class, names=name)
|
|
|
|
elif type(names) == str:
|
|
|
|
for idx, (tf_output, pt_output) in enumerate(zip(tf_outputs, pt_outputs)):
|
|
|
|
check_outputs(tf_output, pt_output, model_class, names=f"{names}_{idx}")
|
|
|
|
else:
|
|
|
|
raise ValueError(f"`names` should be a `tuple` or a string. Got {type(names)} instead.")
|
|
|
|
elif isinstance(tf_outputs, tf.Tensor):
|
|
|
|
self.assertTrue(isinstance(pt_outputs, torch.Tensor))
|
|
|
|
|
|
|
|
tf_outputs = tf_outputs.numpy()
|
|
|
|
pt_outputs = pt_outputs.detach().to("cpu").numpy()
|
|
|
|
|
|
|
|
tf_nans = np.isnan(tf_outputs)
|
|
|
|
pt_nans = np.isnan(pt_outputs)
|
|
|
|
|
|
|
|
pt_outputs[tf_nans] = 0
|
|
|
|
tf_outputs[tf_nans] = 0
|
|
|
|
pt_outputs[pt_nans] = 0
|
|
|
|
tf_outputs[pt_nans] = 0
|
|
|
|
|
|
|
|
max_diff = np.amax(np.abs(tf_outputs - pt_outputs))
|
|
|
|
self.assertLessEqual(max_diff, 1e-5)
|
|
|
|
else:
|
|
|
|
raise ValueError(
|
|
|
|
f"`tf_outputs` should be a `tuple` or an instance of `tf.Tensor`. Got {type(tf_outputs)} instead."
|
|
|
|
)
|
|
|
|
|
|
|
|
def check_pt_tf_models(tf_model, pt_model):
|
|
|
|
|
|
|
|
# send pytorch model to the correct device
|
|
|
|
pt_model.to(torch_device)
|
|
|
|
|
|
|
|
# Check predictions on first output (logits/hidden-states) are close enough given low-level computational differences
|
|
|
|
pt_model.eval()
|
|
|
|
|
|
|
|
pt_inputs_dict = prepare_pt_inputs_from_tf_inputs(tf_inputs_dict)
|
|
|
|
pt_inputs_dict_maybe_with_labels = prepare_pt_inputs_from_tf_inputs(tf_inputs_dict_maybe_with_labels)
|
|
|
|
|
|
|
|
# send pytorch inputs to the correct device
|
|
|
|
pt_inputs_dict = {
|
|
|
|
k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs_dict.items()
|
|
|
|
}
|
|
|
|
pt_inputs_dict_maybe_with_labels = {
|
|
|
|
k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v
|
|
|
|
for k, v in pt_inputs_dict_maybe_with_labels.items()
|
|
|
|
}
|
|
|
|
|
|
|
|
# Original test: check without `labels`
|
2019-12-22 21:57:20 +08:00
|
|
|
with torch.no_grad():
|
2022-03-14 20:31:32 +08:00
|
|
|
pt_outputs = pt_model(**pt_inputs_dict)
|
|
|
|
tf_outputs = tf_model(tf_inputs_dict)
|
|
|
|
|
|
|
|
tf_keys = tuple([k for k, v in tf_outputs.items() if v is not None])
|
|
|
|
pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])
|
|
|
|
|
|
|
|
self.assertEqual(tf_keys, pt_keys)
|
|
|
|
check_outputs(tf_outputs.to_tuple(), pt_outputs.to_tuple(), model_class, names=tf_keys)
|
|
|
|
|
|
|
|
# check the case where `labels` is passed
|
|
|
|
has_labels = any(
|
|
|
|
x in tf_inputs_dict_maybe_with_labels for x in ["labels", "next_sentence_label", "start_positions"]
|
|
|
|
)
|
|
|
|
if has_labels:
|
|
|
|
|
|
|
|
with torch.no_grad():
|
|
|
|
pt_outputs = pt_model(**pt_inputs_dict_maybe_with_labels)
|
|
|
|
tf_outputs = tf_model(tf_inputs_dict_maybe_with_labels)
|
|
|
|
|
|
|
|
# Some models' output class don't have `loss` attribute despite `labels` is used.
|
|
|
|
# TODO: identify which models
|
|
|
|
tf_loss = getattr(tf_outputs, "loss", None)
|
|
|
|
pt_loss = getattr(pt_outputs, "loss", None)
|
|
|
|
|
|
|
|
# Some PT models return loss while the corresponding TF models don't (i.e. `None` for `loss`).
|
|
|
|
# - TFFlaubertWithLMHeadModel
|
|
|
|
# - TFFunnelForPreTraining
|
|
|
|
# - TFElectraForPreTraining
|
|
|
|
# - TFXLMWithLMHeadModel
|
|
|
|
# TODO: Fix PT/TF diff -> remove this condition to fail the test if a diff occurs
|
|
|
|
if not ((tf_loss is None and pt_loss is None) or (tf_loss is not None and pt_loss is not None)):
|
|
|
|
if model_class.__name__ not in [
|
|
|
|
"TFFlaubertWithLMHeadModel",
|
|
|
|
"TFFunnelForPreTraining",
|
|
|
|
"TFElectraForPreTraining",
|
|
|
|
"TFXLMWithLMHeadModel",
|
2022-03-17 21:03:07 +08:00
|
|
|
"TFTransfoXLLMHeadModel",
|
2022-03-14 20:31:32 +08:00
|
|
|
]:
|
|
|
|
self.assertEqual(tf_loss is None, pt_loss is None)
|
|
|
|
|
|
|
|
tf_keys = tuple([k for k, v in tf_outputs.items() if v is not None])
|
|
|
|
pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])
|
|
|
|
|
|
|
|
# TODO: remove these 2 conditions once the above TODOs (above loss) are implemented
|
|
|
|
# (Also, `TFTransfoXLLMHeadModel` has no `loss` while `TransfoXLLMHeadModel` return `losses`)
|
|
|
|
if tf_keys != pt_keys:
|
|
|
|
if model_class.__name__ not in [
|
|
|
|
"TFFlaubertWithLMHeadModel",
|
|
|
|
"TFFunnelForPreTraining",
|
|
|
|
"TFElectraForPreTraining",
|
|
|
|
"TFXLMWithLMHeadModel",
|
2022-03-17 21:03:07 +08:00
|
|
|
"TFTransfoXLLMHeadModel",
|
|
|
|
]:
|
2022-03-14 20:31:32 +08:00
|
|
|
self.assertEqual(tf_keys, pt_keys)
|
|
|
|
|
|
|
|
# Since we deliberately make some tests pass above (regarding the `loss`), let's still try to test
|
|
|
|
# some remaining attributes in the outputs.
|
|
|
|
# TODO: remove this block of `index` computing once the above TODOs (above loss) are implemented
|
|
|
|
# compute the 1st `index` where `tf_keys` and `pt_keys` is different
|
|
|
|
index = 0
|
|
|
|
for _ in range(min(len(tf_keys), len(pt_keys))):
|
|
|
|
if tf_keys[index] == pt_keys[index]:
|
|
|
|
index += 1
|
|
|
|
else:
|
|
|
|
break
|
|
|
|
if tf_keys[:index] != pt_keys[:index]:
|
|
|
|
self.assertEqual(tf_keys, pt_keys)
|
|
|
|
|
|
|
|
# Some models require extra condition to return loss. For example, `(TF)BertForPreTraining` requires
|
|
|
|
# both`labels` and `next_sentence_label`.
|
|
|
|
if tf_loss is not None and pt_loss is not None:
|
|
|
|
|
|
|
|
# check anything else than `loss`
|
|
|
|
keys = tuple([k for k in tf_keys])
|
|
|
|
check_outputs(tf_outputs[1:index], pt_outputs[1:index], model_class, names=keys[1:index])
|
|
|
|
|
|
|
|
# check `loss`
|
|
|
|
|
|
|
|
# tf models returned loss is usually a tensor rather than a scalar.
|
|
|
|
# (see `hf_compute_loss`: it uses `tf.keras.losses.Reduction.NONE`)
|
|
|
|
# Change it here to a scalar to match PyTorch models' loss
|
|
|
|
tf_loss = tf.math.reduce_mean(tf_loss).numpy()
|
|
|
|
pt_loss = pt_loss.detach().to("cpu").numpy()
|
|
|
|
|
|
|
|
tf_nans = np.isnan(tf_loss)
|
|
|
|
pt_nans = np.isnan(pt_loss)
|
|
|
|
# the 2 losses need to be both nan or both not nan
|
|
|
|
self.assertEqual(tf_nans, pt_nans)
|
|
|
|
|
|
|
|
if not tf_nans:
|
|
|
|
max_diff = np.amax(np.abs(tf_loss - pt_loss))
|
|
|
|
self.assertLessEqual(max_diff, 1e-5)
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
2021-06-15 01:58:54 +08:00
|
|
|
|
2022-03-14 20:31:32 +08:00
|
|
|
# Output all for aggressive testing
|
|
|
|
config.output_hidden_states = True
|
|
|
|
# Pure convolutional models have no attention
|
|
|
|
# TODO: use a better and general criteria
|
|
|
|
if "TFConvNext" not in model_class.__name__:
|
|
|
|
config.output_attentions = True
|
|
|
|
|
|
|
|
for k in ["attention_mask", "encoder_attention_mask", "decoder_attention_mask"]:
|
|
|
|
if k in inputs_dict:
|
|
|
|
attention_mask = inputs_dict[k]
|
|
|
|
# make sure no all 0s attention masks - to avoid failure at this moment.
|
|
|
|
# TODO: remove this line once the TODO below is implemented.
|
|
|
|
attention_mask = tf.ones_like(attention_mask, dtype=tf.int32)
|
|
|
|
# Here we make the first sequence with all 0s as attention mask.
|
|
|
|
# Currently, this will fail for `TFWav2Vec2Model`. This is caused by the different large negative
|
|
|
|
# values, like `1e-4`, `1e-9`, `1e-30` and `-inf` for attention mask across models/frameworks.
|
|
|
|
# TODO: enable this block once the large negative values thing is cleaned up.
|
|
|
|
# (see https://github.com/huggingface/transformers/issues/14859)
|
|
|
|
# attention_mask = tf.concat(
|
|
|
|
# [
|
|
|
|
# tf.zeros_like(attention_mask[:1], dtype=tf.int32),
|
|
|
|
# tf.cast(attention_mask[1:], dtype=tf.int32)
|
|
|
|
# ],
|
|
|
|
# axis=0
|
|
|
|
# )
|
|
|
|
inputs_dict[k] = attention_mask
|
|
|
|
|
|
|
|
pt_model_class_name = model_class.__name__[2:] # Skip the "TF" at the beginning
|
|
|
|
pt_model_class = getattr(transformers, pt_model_class_name)
|
|
|
|
|
|
|
|
config.output_hidden_states = True
|
|
|
|
|
|
|
|
tf_model = model_class(config)
|
|
|
|
pt_model = pt_model_class(config)
|
2020-01-28 01:09:58 +08:00
|
|
|
|
2022-03-14 20:31:32 +08:00
|
|
|
tf_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
|
|
|
|
tf_inputs_dict_maybe_with_labels = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
2020-01-28 10:57:23 +08:00
|
|
|
|
2022-03-14 20:31:32 +08:00
|
|
|
# Check we can load pt model in tf and vice-versa with model => model functions
|
|
|
|
tf_model = transformers.load_pytorch_model_in_tf2_model(tf_model, pt_model, tf_inputs=tf_inputs_dict)
|
|
|
|
pt_model = transformers.load_tf2_model_in_pytorch_model(pt_model, tf_model)
|
2020-01-28 01:09:58 +08:00
|
|
|
|
2022-03-14 20:31:32 +08:00
|
|
|
check_pt_tf_models(tf_model, pt_model)
|
2019-12-22 21:57:20 +08:00
|
|
|
|
|
|
|
# Check we can load pt model in tf and vice-versa with checkpoint => model functions
|
2019-12-23 01:12:11 +08:00
|
|
|
with tempfile.TemporaryDirectory() as tmpdirname:
|
2019-12-22 21:57:20 +08:00
|
|
|
pt_checkpoint_path = os.path.join(tmpdirname, "pt_model.bin")
|
|
|
|
torch.save(pt_model.state_dict(), pt_checkpoint_path)
|
|
|
|
tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(tf_model, pt_checkpoint_path)
|
|
|
|
|
|
|
|
tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
|
|
|
|
tf_model.save_weights(tf_checkpoint_path)
|
|
|
|
pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(pt_model, tf_checkpoint_path)
|
|
|
|
|
2022-03-14 20:31:32 +08:00
|
|
|
check_pt_tf_models(tf_model, pt_model)
|
2019-12-22 21:57:20 +08:00
|
|
|
|
|
|
|
def test_compile_tf_model(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
2020-11-14 06:07:17 +08:00
|
|
|
max_input = getattr(self.model_tester, "max_position_embeddings", 512)
|
2019-12-22 21:57:20 +08:00
|
|
|
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
|
|
|
|
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
|
|
|
|
metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
2022-02-09 00:27:23 +08:00
|
|
|
if model_class.__name__ in ["TFSpeech2TextModel", "TFSpeech2TextForConditionalGeneration"]:
|
|
|
|
inputs = {
|
|
|
|
"decoder_input_ids": tf.keras.Input(
|
|
|
|
batch_shape=(2, max_input),
|
|
|
|
name="decoder_input_ids",
|
|
|
|
dtype="int32",
|
|
|
|
),
|
|
|
|
"input_features": tf.keras.Input(
|
|
|
|
batch_shape=(
|
|
|
|
2,
|
|
|
|
max_input,
|
|
|
|
self.model_tester.input_feat_per_channel * self.model_tester.input_channels,
|
|
|
|
),
|
|
|
|
name="input_features",
|
|
|
|
dtype="float32",
|
|
|
|
),
|
|
|
|
}
|
|
|
|
elif self.is_encoder_decoder:
|
2021-11-09 20:54:37 +08:00
|
|
|
inputs = {
|
2020-06-11 22:31:26 +08:00
|
|
|
"decoder_input_ids": tf.keras.Input(
|
2020-11-14 06:07:17 +08:00
|
|
|
batch_shape=(2, max_input),
|
|
|
|
name="decoder_input_ids",
|
|
|
|
dtype="int32",
|
2020-06-11 22:31:26 +08:00
|
|
|
),
|
2020-11-14 06:07:17 +08:00
|
|
|
"input_ids": tf.keras.Input(batch_shape=(2, max_input), name="input_ids", dtype="int32"),
|
2020-06-11 22:31:26 +08:00
|
|
|
}
|
2022-02-26 01:19:16 +08:00
|
|
|
# `pixel_values` implies that the input is an image
|
|
|
|
elif model_class.main_input_name == "pixel_values":
|
2021-11-09 20:54:37 +08:00
|
|
|
inputs = tf.keras.Input(
|
|
|
|
batch_shape=(
|
|
|
|
3,
|
|
|
|
self.model_tester.num_channels,
|
|
|
|
self.model_tester.image_size,
|
|
|
|
self.model_tester.image_size,
|
|
|
|
),
|
|
|
|
name="pixel_values",
|
|
|
|
dtype="float32",
|
|
|
|
)
|
2021-12-24 00:19:44 +08:00
|
|
|
elif model_class.__name__ in ["TFCLIPModel"]:
|
|
|
|
inputs = {
|
|
|
|
"input_ids": tf.keras.Input(batch_shape=(3, max_input), name="input_ids", dtype="int32"),
|
|
|
|
"pixel_values": tf.keras.Input(
|
|
|
|
batch_shape=(
|
|
|
|
3,
|
|
|
|
self.model_tester.vision_model_tester.num_channels,
|
|
|
|
self.model_tester.vision_model_tester.image_size,
|
|
|
|
self.model_tester.vision_model_tester.image_size,
|
|
|
|
),
|
|
|
|
name="pixel_values",
|
|
|
|
dtype="float32",
|
|
|
|
),
|
|
|
|
}
|
2021-04-09 06:41:36 +08:00
|
|
|
elif model_class in get_values(TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
|
2021-11-09 20:54:37 +08:00
|
|
|
inputs = tf.keras.Input(batch_shape=(4, 2, max_input), name="input_ids", dtype="int32")
|
2020-06-11 22:31:26 +08:00
|
|
|
else:
|
2021-11-09 20:54:37 +08:00
|
|
|
inputs = tf.keras.Input(batch_shape=(2, max_input), name="input_ids", dtype="int32")
|
2020-06-11 22:31:26 +08:00
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
# Prepare our model
|
|
|
|
model = model_class(config)
|
2020-10-21 19:10:16 +08:00
|
|
|
model(self._prepare_for_class(inputs_dict, model_class)) # Model must be called before saving.
|
2019-12-22 21:57:20 +08:00
|
|
|
# Let's load it from the disk to be sure we can use pretrained weights
|
2019-12-23 01:12:11 +08:00
|
|
|
with tempfile.TemporaryDirectory() as tmpdirname:
|
2021-01-07 18:48:49 +08:00
|
|
|
model.save_pretrained(tmpdirname, saved_model=False)
|
2019-12-22 21:57:20 +08:00
|
|
|
model = model_class.from_pretrained(tmpdirname)
|
|
|
|
|
2021-11-09 20:54:37 +08:00
|
|
|
outputs_dict = model(inputs)
|
2019-12-22 21:57:20 +08:00
|
|
|
hidden_states = outputs_dict[0]
|
|
|
|
|
2020-07-10 23:36:53 +08:00
|
|
|
# Add a dense layer on top to test integration with other keras modules
|
2019-12-22 21:57:20 +08:00
|
|
|
outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
|
|
|
|
|
|
|
|
# Compile extended model
|
2021-11-09 20:54:37 +08:00
|
|
|
extended_model = tf.keras.Model(inputs=[inputs], outputs=[outputs])
|
2019-12-22 21:57:20 +08:00
|
|
|
extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
|
|
|
|
|
|
|
|
def test_keyword_and_dict_args(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
2020-10-21 19:10:16 +08:00
|
|
|
inputs = self._prepare_for_class(inputs_dict, model_class)
|
|
|
|
|
|
|
|
outputs_dict = model(inputs)
|
2019-12-22 21:57:20 +08:00
|
|
|
|
2020-06-11 22:31:26 +08:00
|
|
|
inputs_keywords = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
|
2022-02-09 00:27:23 +08:00
|
|
|
outputs_keywords = model(**inputs_keywords)
|
2019-12-22 21:57:20 +08:00
|
|
|
output_dict = outputs_dict[0].numpy()
|
|
|
|
output_keywords = outputs_keywords[0].numpy()
|
|
|
|
|
|
|
|
self.assertLess(np.sum(np.abs(output_dict - output_keywords)), 1e-6)
|
|
|
|
|
|
|
|
def test_attention_outputs(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
2020-11-06 04:10:43 +08:00
|
|
|
config.return_dict = True
|
2020-10-21 19:10:16 +08:00
|
|
|
decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", self.model_tester.seq_length)
|
|
|
|
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", self.model_tester.seq_length)
|
|
|
|
decoder_key_length = getattr(self.model_tester, "key_length", decoder_seq_length)
|
|
|
|
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
|
2019-12-22 21:57:20 +08:00
|
|
|
|
2020-11-14 06:07:17 +08:00
|
|
|
def check_decoder_attentions_output(outputs):
|
|
|
|
out_len = len(outputs)
|
2021-04-26 20:16:21 +08:00
|
|
|
self.assertEqual(min(out_len % 2, out_len % 5), 0) # differentiation due to newly added cross_attentions
|
2020-11-14 06:07:17 +08:00
|
|
|
decoder_attentions = outputs.decoder_attentions
|
|
|
|
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
|
|
|
|
self.assertListEqual(
|
|
|
|
list(decoder_attentions[0].shape[-3:]),
|
|
|
|
[self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
|
|
|
|
)
|
|
|
|
|
|
|
|
def check_encoder_attentions_output(outputs):
|
2020-11-06 04:10:43 +08:00
|
|
|
attentions = [
|
|
|
|
t.numpy() for t in (outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions)
|
|
|
|
]
|
2019-12-22 21:57:20 +08:00
|
|
|
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
|
|
|
|
self.assertListEqual(
|
|
|
|
list(attentions[0].shape[-3:]),
|
|
|
|
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
|
2019-12-21 22:46:46 +08:00
|
|
|
)
|
2020-11-14 06:07:17 +08:00
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
inputs_dict["output_attentions"] = True
|
|
|
|
inputs_dict["use_cache"] = False
|
|
|
|
config.output_hidden_states = False
|
|
|
|
model = model_class(config)
|
|
|
|
outputs = model(self._prepare_for_class(inputs_dict, model_class))
|
2019-12-22 21:57:20 +08:00
|
|
|
out_len = len(outputs)
|
2020-11-14 06:07:17 +08:00
|
|
|
self.assertEqual(config.output_hidden_states, False)
|
|
|
|
check_encoder_attentions_output(outputs)
|
2019-12-10 05:13:57 +08:00
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
if self.is_encoder_decoder:
|
2020-11-14 06:07:17 +08:00
|
|
|
model = model_class(config)
|
|
|
|
outputs = model(self._prepare_for_class(inputs_dict, model_class))
|
|
|
|
self.assertEqual(config.output_hidden_states, False)
|
|
|
|
check_decoder_attentions_output(outputs)
|
2019-09-05 08:27:39 +08:00
|
|
|
|
2020-06-10 05:39:06 +08:00
|
|
|
# Check that output attentions can also be changed via the config
|
|
|
|
del inputs_dict["output_attentions"]
|
2019-12-22 21:57:20 +08:00
|
|
|
config.output_attentions = True
|
2020-06-10 05:39:06 +08:00
|
|
|
model = model_class(config)
|
2020-06-11 22:31:26 +08:00
|
|
|
outputs = model(self._prepare_for_class(inputs_dict, model_class))
|
2020-11-14 06:07:17 +08:00
|
|
|
self.assertEqual(config.output_hidden_states, False)
|
|
|
|
check_encoder_attentions_output(outputs)
|
2020-06-10 05:39:06 +08:00
|
|
|
|
|
|
|
# Check attention is always last and order is fine
|
|
|
|
inputs_dict["output_attentions"] = True
|
2019-12-22 21:57:20 +08:00
|
|
|
config.output_hidden_states = True
|
|
|
|
model = model_class(config)
|
2020-06-11 22:31:26 +08:00
|
|
|
outputs = model(self._prepare_for_class(inputs_dict, model_class))
|
2020-11-14 06:07:17 +08:00
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
|
|
|
|
self.assertEqual(model.config.output_hidden_states, True)
|
2020-11-14 06:07:17 +08:00
|
|
|
check_encoder_attentions_output(outputs)
|
2019-09-25 02:59:10 +08:00
|
|
|
|
2021-01-26 16:50:00 +08:00
|
|
|
def test_headmasking(self):
|
|
|
|
if not self.test_head_masking:
|
|
|
|
return
|
|
|
|
|
|
|
|
random.Random().seed(42)
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
random.Random().seed()
|
|
|
|
|
|
|
|
inputs_dict["output_attentions"] = True
|
|
|
|
config.output_hidden_states = True
|
|
|
|
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config=configs_no_init)
|
|
|
|
|
|
|
|
# Prepare head_mask
|
|
|
|
def prepare_layer_head_mask(i, attention_heads, num_hidden_layers):
|
|
|
|
if i == 0:
|
|
|
|
return tf.concat(
|
|
|
|
(tf.zeros(1, dtype=tf.float32), tf.ones(attention_heads - 1, dtype=tf.float32)), 0
|
|
|
|
)
|
|
|
|
elif i == num_hidden_layers - 1:
|
|
|
|
return tf.concat(
|
|
|
|
(tf.zeros(attention_heads - 1, dtype=tf.float32), tf.ones(1, dtype=tf.float32)), 0
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
return tf.ones(attention_heads, dtype=tf.float32)
|
|
|
|
|
|
|
|
head_mask = tf.stack(
|
|
|
|
[
|
|
|
|
prepare_layer_head_mask(i, config.num_attention_heads, config.num_hidden_layers)
|
|
|
|
for i in range(config.num_hidden_layers)
|
|
|
|
],
|
|
|
|
0,
|
|
|
|
)
|
|
|
|
|
|
|
|
inputs = self._prepare_for_class(inputs_dict, model_class).copy()
|
|
|
|
inputs["head_mask"] = head_mask
|
|
|
|
if model.config.is_encoder_decoder:
|
|
|
|
signature = inspect.signature(model.call)
|
|
|
|
arg_names = [*signature.parameters.keys()]
|
|
|
|
if "decoder_head_mask" in arg_names: # necessary diferentiation because of T5 model
|
|
|
|
inputs["decoder_head_mask"] = head_mask
|
2021-04-26 20:16:21 +08:00
|
|
|
if "cross_attn_head_mask" in arg_names:
|
|
|
|
inputs["cross_attn_head_mask"] = head_mask
|
2021-01-26 16:50:00 +08:00
|
|
|
|
|
|
|
outputs = model(**inputs, return_dict=True)
|
|
|
|
|
|
|
|
def check_attentions_validity(attentions):
|
|
|
|
# Remove Nan
|
|
|
|
for t in attentions:
|
|
|
|
self.assertLess(
|
|
|
|
(tf.math.reduce_sum(tf.cast(tf.math.is_nan(t), tf.float32))).numpy(), (tf.size(t) / 4).numpy()
|
|
|
|
) # Check we don't have more than 25% nans (arbitrary)
|
|
|
|
|
|
|
|
attentions = [
|
|
|
|
tf.where(tf.math.is_nan(t), 0.0, t) for t in attentions
|
|
|
|
] # remove them (the test is less complete)
|
|
|
|
|
|
|
|
self.assertAlmostEqual(tf.math.reduce_sum(attentions[0][..., 0, :, :]).numpy(), 0.0)
|
|
|
|
self.assertNotEqual(tf.math.reduce_sum(attentions[0][..., -1, :, :]).numpy(), 0.0)
|
|
|
|
if len(attentions) > 2: # encoder-decodere models have only 2 layers in each modules
|
|
|
|
self.assertNotEqual(tf.math.reduce_sum(attentions[1][..., 0, :, :]).numpy(), 0.0)
|
|
|
|
self.assertAlmostEqual(tf.math.reduce_sum(attentions[-1][..., -2, :, :]).numpy(), 0.0)
|
|
|
|
self.assertNotEqual(tf.math.reduce_sum(attentions[-1][..., -1, :, :]).numpy(), 0.0)
|
|
|
|
|
|
|
|
if model.config.is_encoder_decoder:
|
|
|
|
check_attentions_validity(outputs.encoder_attentions)
|
|
|
|
check_attentions_validity(outputs.decoder_attentions)
|
2021-04-26 20:16:21 +08:00
|
|
|
if "cross_attn_head_mask" in arg_names:
|
|
|
|
check_attentions_validity(outputs.cross_attentions)
|
2021-01-26 16:50:00 +08:00
|
|
|
else:
|
|
|
|
check_attentions_validity(outputs.attentions)
|
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
def test_hidden_states_output(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
2020-06-22 22:10:45 +08:00
|
|
|
def check_hidden_states_output(config, inputs_dict, model_class):
|
2019-12-22 21:57:20 +08:00
|
|
|
model = model_class(config)
|
2020-06-11 22:31:26 +08:00
|
|
|
outputs = model(self._prepare_for_class(inputs_dict, model_class))
|
2020-09-10 22:41:56 +08:00
|
|
|
expected_num_layers = getattr(
|
|
|
|
self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
|
|
|
|
)
|
2020-11-14 06:07:17 +08:00
|
|
|
|
2020-12-15 01:47:00 +08:00
|
|
|
if model.config.is_encoder_decoder:
|
|
|
|
encoder_hidden_states = outputs.encoder_hidden_states
|
|
|
|
decoder_hidden_states = outputs.decoder_hidden_states
|
|
|
|
|
|
|
|
self.assertEqual(config.output_attentions, False)
|
|
|
|
self.assertEqual(len(encoder_hidden_states), expected_num_layers)
|
|
|
|
self.assertListEqual(
|
|
|
|
list(encoder_hidden_states[0].shape[-2:]),
|
|
|
|
[self.model_tester.seq_length, self.model_tester.hidden_size],
|
|
|
|
)
|
|
|
|
self.assertEqual(len(decoder_hidden_states), expected_num_layers)
|
|
|
|
self.assertListEqual(
|
|
|
|
list(decoder_hidden_states[0].shape[-2:]),
|
|
|
|
[self.model_tester.seq_length, self.model_tester.hidden_size],
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
hidden_states = outputs.hidden_states
|
|
|
|
self.assertEqual(config.output_attentions, False)
|
|
|
|
self.assertEqual(len(hidden_states), expected_num_layers)
|
|
|
|
self.assertListEqual(
|
|
|
|
list(hidden_states[0].shape[-2:]),
|
|
|
|
[self.model_tester.seq_length, self.model_tester.hidden_size],
|
|
|
|
)
|
2019-09-25 02:59:10 +08:00
|
|
|
|
2020-06-22 22:10:45 +08:00
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
inputs_dict["output_hidden_states"] = True
|
|
|
|
check_hidden_states_output(config, inputs_dict, model_class)
|
|
|
|
|
|
|
|
del inputs_dict["output_hidden_states"]
|
|
|
|
config.output_hidden_states = True
|
|
|
|
check_hidden_states_output(config, inputs_dict, model_class)
|
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
def test_model_common_attributes(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
2022-02-09 00:27:23 +08:00
|
|
|
text_in_text_out_models = (
|
2021-04-09 06:41:36 +08:00
|
|
|
get_values(TF_MODEL_FOR_CAUSAL_LM_MAPPING)
|
|
|
|
+ get_values(TF_MODEL_FOR_MASKED_LM_MAPPING)
|
|
|
|
+ get_values(TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING)
|
2020-12-14 12:05:24 +08:00
|
|
|
)
|
2022-02-09 00:27:23 +08:00
|
|
|
speech_in_text_out_models = get_values(TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING)
|
2019-12-22 21:57:20 +08:00
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
2021-01-11 19:27:28 +08:00
|
|
|
assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
|
2022-02-09 00:27:23 +08:00
|
|
|
if model_class in text_in_text_out_models:
|
2021-01-11 19:27:28 +08:00
|
|
|
x = model.get_output_embeddings()
|
2020-12-14 12:05:24 +08:00
|
|
|
assert isinstance(x, tf.keras.layers.Layer)
|
2021-01-11 19:27:28 +08:00
|
|
|
name = model.get_bias()
|
|
|
|
assert isinstance(name, dict)
|
|
|
|
for k, v in name.items():
|
|
|
|
assert isinstance(v, tf.Variable)
|
2022-02-09 00:27:23 +08:00
|
|
|
elif model_class in speech_in_text_out_models:
|
|
|
|
x = model.get_output_embeddings()
|
|
|
|
assert isinstance(x, tf.keras.layers.Layer)
|
|
|
|
name = model.get_bias()
|
|
|
|
assert name is None
|
2020-12-14 12:05:24 +08:00
|
|
|
else:
|
2021-01-11 19:27:28 +08:00
|
|
|
x = model.get_output_embeddings()
|
2020-12-14 12:05:24 +08:00
|
|
|
assert x is None
|
2021-01-11 19:27:28 +08:00
|
|
|
name = model.get_bias()
|
|
|
|
assert name is None
|
2019-12-22 21:57:20 +08:00
|
|
|
|
|
|
|
def test_determinism(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
2020-03-03 22:42:15 +08:00
|
|
|
first, second = (
|
2020-06-11 22:31:26 +08:00
|
|
|
model(self._prepare_for_class(inputs_dict, model_class), training=False)[0],
|
|
|
|
model(self._prepare_for_class(inputs_dict, model_class), training=False)[0],
|
2020-03-03 22:42:15 +08:00
|
|
|
)
|
2019-12-22 21:57:20 +08:00
|
|
|
out_1 = first.numpy()
|
|
|
|
out_2 = second.numpy()
|
|
|
|
out_1 = out_1[~np.isnan(out_1)]
|
|
|
|
out_2 = out_2[~np.isnan(out_2)]
|
|
|
|
max_diff = np.amax(np.abs(out_1 - out_2))
|
|
|
|
self.assertLessEqual(max_diff, 1e-5)
|
|
|
|
|
2020-08-13 23:59:35 +08:00
|
|
|
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={}):
|
|
|
|
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(
|
|
|
|
all(tf.equal(tuple_object, dict_object)),
|
|
|
|
msg=f"Tuple and dict output are not equal. Difference: {tf.math.reduce_max(tf.abs(tuple_object - dict_object))}",
|
|
|
|
)
|
|
|
|
|
|
|
|
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, 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}
|
|
|
|
)
|
|
|
|
|
2019-12-22 21:57:20 +08:00
|
|
|
def test_inputs_embeds(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
|
2021-02-08 19:36:30 +08:00
|
|
|
inputs = copy.deepcopy(inputs_dict)
|
|
|
|
|
2020-06-11 22:31:26 +08:00
|
|
|
if not self.is_encoder_decoder:
|
|
|
|
input_ids = inputs["input_ids"]
|
|
|
|
del inputs["input_ids"]
|
|
|
|
else:
|
2020-07-08 00:15:53 +08:00
|
|
|
encoder_input_ids = inputs["input_ids"]
|
2020-06-11 22:31:26 +08:00
|
|
|
decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
|
2020-07-08 00:15:53 +08:00
|
|
|
del inputs["input_ids"]
|
2020-06-11 22:31:26 +08:00
|
|
|
inputs.pop("decoder_input_ids", None)
|
|
|
|
|
2019-12-10 22:11:07 +08:00
|
|
|
if not self.is_encoder_decoder:
|
2021-01-20 19:08:12 +08:00
|
|
|
inputs["inputs_embeds"] = model.get_input_embeddings()(input_ids)
|
2019-12-10 22:11:07 +08:00
|
|
|
else:
|
2021-01-20 19:08:12 +08:00
|
|
|
inputs["inputs_embeds"] = model.get_input_embeddings()(encoder_input_ids)
|
|
|
|
inputs["decoder_inputs_embeds"] = model.get_input_embeddings()(decoder_input_ids)
|
2019-12-22 21:57:20 +08:00
|
|
|
|
2021-02-08 19:36:30 +08:00
|
|
|
inputs = self._prepare_for_class(inputs, model_class)
|
|
|
|
|
2020-06-11 22:31:26 +08:00
|
|
|
model(inputs)
|
2019-11-12 11:19:14 +08:00
|
|
|
|
2021-01-08 21:23:29 +08:00
|
|
|
def test_numpy_arrays_inputs(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
def prepare_numpy_arrays(inputs_dict):
|
|
|
|
inputs_np_dict = {}
|
|
|
|
for k, v in inputs_dict.items():
|
|
|
|
if tf.is_tensor(v):
|
|
|
|
inputs_np_dict[k] = v.numpy()
|
|
|
|
else:
|
|
|
|
inputs_np_dict[k] = np.array(k)
|
|
|
|
|
|
|
|
return inputs_np_dict
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
|
|
|
|
inputs = self._prepare_for_class(inputs_dict, model_class)
|
|
|
|
inputs_np = prepare_numpy_arrays(inputs)
|
|
|
|
|
2022-01-14 23:19:04 +08:00
|
|
|
output_for_dict_input = model(inputs_np)
|
|
|
|
output_for_kw_input = model(**inputs_np)
|
|
|
|
self.assert_outputs_same(output_for_dict_input, output_for_kw_input)
|
2021-01-08 21:23:29 +08:00
|
|
|
|
2020-06-19 06:41:26 +08:00
|
|
|
def test_resize_token_embeddings(self):
|
|
|
|
if not self.test_resize_embeddings:
|
|
|
|
return
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
2021-01-11 19:27:28 +08:00
|
|
|
|
|
|
|
def _get_word_embedding_weight(model, embedding_layer):
|
2021-01-20 19:08:12 +08:00
|
|
|
embeds = getattr(embedding_layer, "weight", None)
|
|
|
|
if embeds is not None:
|
|
|
|
return embeds
|
|
|
|
|
|
|
|
embeds = getattr(embedding_layer, "decoder", None)
|
|
|
|
if embeds is not None:
|
|
|
|
return embeds
|
|
|
|
|
|
|
|
model(model.dummy_inputs)
|
|
|
|
|
|
|
|
embeds = getattr(embedding_layer, "weight", None)
|
|
|
|
if embeds is not None:
|
|
|
|
return embeds
|
|
|
|
|
|
|
|
embeds = getattr(embedding_layer, "decoder", None)
|
|
|
|
if embeds is not None:
|
|
|
|
return embeds
|
|
|
|
|
|
|
|
return None
|
2021-01-11 19:27:28 +08:00
|
|
|
|
2020-06-19 06:41:26 +08:00
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
for size in [config.vocab_size - 10, config.vocab_size + 10, None]:
|
|
|
|
# build the embeddings
|
|
|
|
model = model_class(config=config)
|
2021-01-11 19:27:28 +08:00
|
|
|
old_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
|
|
|
|
old_bias = model.get_bias()
|
|
|
|
old_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
|
2020-06-19 06:41:26 +08:00
|
|
|
# reshape the embeddings
|
2021-01-11 19:27:28 +08:00
|
|
|
model.resize_token_embeddings(size)
|
|
|
|
new_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
|
|
|
|
new_bias = model.get_bias()
|
|
|
|
new_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
|
|
|
|
|
|
|
|
# check that the resized embeddings size matches the desired size.
|
2020-06-19 06:41:26 +08:00
|
|
|
assert_size = size if size is not None else config.vocab_size
|
2021-01-11 19:27:28 +08:00
|
|
|
self.assertEqual(new_input_embeddings.shape[0], assert_size)
|
|
|
|
|
2020-06-19 06:41:26 +08:00
|
|
|
# check that weights remain the same after resizing
|
|
|
|
models_equal = True
|
2021-01-11 19:27:28 +08:00
|
|
|
for p1, p2 in zip(old_input_embeddings.value(), new_input_embeddings.value()):
|
|
|
|
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
2020-06-19 06:41:26 +08:00
|
|
|
models_equal = False
|
|
|
|
self.assertTrue(models_equal)
|
|
|
|
|
2021-01-11 19:27:28 +08:00
|
|
|
if old_bias is not None and new_bias is not None:
|
|
|
|
for old_weight, new_weight in zip(old_bias.values(), new_bias.values()):
|
|
|
|
self.assertEqual(new_weight.shape[0], assert_size)
|
|
|
|
|
|
|
|
models_equal = True
|
|
|
|
for p1, p2 in zip(old_weight.value(), new_weight.value()):
|
|
|
|
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
|
|
|
models_equal = False
|
|
|
|
self.assertTrue(models_equal)
|
|
|
|
|
|
|
|
if old_output_embeddings is not None and new_output_embeddings is not None:
|
|
|
|
self.assertEqual(new_output_embeddings.shape[0], assert_size)
|
|
|
|
self.assertEqual(new_output_embeddings.shape[1], old_output_embeddings.shape[1])
|
|
|
|
|
|
|
|
models_equal = True
|
|
|
|
for p1, p2 in zip(old_output_embeddings.value(), new_output_embeddings.value()):
|
|
|
|
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
|
|
|
models_equal = False
|
|
|
|
self.assertTrue(models_equal)
|
|
|
|
|
2020-04-06 16:37:05 +08:00
|
|
|
def test_lm_head_model_random_no_beam_search_generate(self):
|
2020-03-03 22:42:15 +08:00
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
2021-06-15 01:58:54 +08:00
|
|
|
input_ids = inputs_dict.get("input_ids", None)
|
2020-03-20 06:18:23 +08:00
|
|
|
|
2020-04-06 16:37:05 +08:00
|
|
|
# iterate over all generative models
|
2020-03-03 22:42:15 +08:00
|
|
|
for model_class in self.all_generative_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
|
|
|
|
if config.bos_token_id is None:
|
2022-02-09 00:27:23 +08:00
|
|
|
# if bos token id is not defined model needs input_ids
|
2022-03-03 00:13:54 +08:00
|
|
|
with self.assertRaises(ValueError):
|
2020-03-17 22:52:37 +08:00
|
|
|
model.generate(do_sample=True, max_length=5)
|
2020-04-06 16:37:05 +08:00
|
|
|
# num_return_sequences = 1
|
2020-04-01 00:42:31 +08:00
|
|
|
self._check_generated_ids(model.generate(input_ids, do_sample=True))
|
2022-02-09 00:27:23 +08:00
|
|
|
elif model_class.__name__ not in ["TFSpeech2TextForConditionalGeneration"]:
|
|
|
|
# Models with non-text inputs won't work here; num_return_sequences = 1
|
2020-04-01 00:42:31 +08:00
|
|
|
self._check_generated_ids(model.generate(do_sample=True, max_length=5))
|
2020-03-04 07:32:07 +08:00
|
|
|
|
2022-02-16 00:54:43 +08:00
|
|
|
with self.assertRaises(ValueError):
|
2020-04-06 16:37:05 +08:00
|
|
|
# generating multiple sequences when no beam search generation
|
2020-03-04 07:32:07 +08:00
|
|
|
# is not allowed as it would always generate the same sequences
|
|
|
|
model.generate(input_ids, do_sample=False, num_return_sequences=2)
|
|
|
|
|
2020-04-06 16:37:05 +08:00
|
|
|
# num_return_sequences > 1, sample
|
|
|
|
self._check_generated_ids(model.generate(input_ids, do_sample=True, num_return_sequences=2))
|
2020-04-01 00:42:31 +08:00
|
|
|
|
|
|
|
# check bad words tokens language generation
|
2020-04-06 16:37:05 +08:00
|
|
|
# create list of 1-seq bad token and list of 2-seq of bad tokens
|
|
|
|
bad_words_ids = [self._generate_random_bad_tokens(1, model), self._generate_random_bad_tokens(2, model)]
|
2020-04-01 00:42:31 +08:00
|
|
|
output_tokens = model.generate(
|
2020-04-06 16:37:05 +08:00
|
|
|
input_ids, do_sample=True, bad_words_ids=bad_words_ids, num_return_sequences=2
|
2020-03-04 07:32:07 +08:00
|
|
|
)
|
2020-04-06 16:37:05 +08:00
|
|
|
# only count generated tokens
|
2020-04-01 00:42:31 +08:00
|
|
|
generated_ids = output_tokens[:, input_ids.shape[-1] :]
|
|
|
|
self.assertFalse(self._check_match_tokens(generated_ids.numpy().tolist(), bad_words_ids))
|
2020-03-03 22:42:15 +08:00
|
|
|
|
Add output in a dictionary for TF `generate` method (#12139)
* Add output args to greedy search
* Fix critical typo + make style quality
* Handle generate_beam_search
* Add dict_specific tests and fix the placement of encoder outputs
* Add specific outputs
* Update doc
* Fix typo
* Adjust handling encoder_outputs + Fix generating for T5
* Fix generate for RAG
* Fix handling ouptut_attentions when target_mapping is not None
Take care of situations when target_mapping is provided
as there are 2-tuple of attentions
Change from:
if inputs["output_attentions"]:
attentions = tuple(tf.transpose(t, perm(2, 3, 0, 1)) for t in attentions)
to:
if inputs["output_attentions"]:
if inputs["target_mapping"] is not None:
# when target_mapping is provided, there are 2-tuple of attentions
attentions = tuple(
tuple(tf.transpose(attn_stream, perm=(2, 3, 0, 1)) for attn_stream in t) for t in attentions
)
else:
attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
* Rename kwargs to model_kwargs
* make style quality
* Move imports in test_modeling_tf_common.py
Move ModelOutput-related imports in test_modeling_tf_common.py
into the `is_tf_available():` statement.
* Rewrite nested if-statements
* Fix added tests
2021-06-23 17:52:11 +08:00
|
|
|
def test_lm_head_model_no_beam_search_generate_dict_outputs(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
input_ids = inputs_dict.get("input_ids", None)
|
2022-02-09 00:27:23 +08:00
|
|
|
if input_ids is None:
|
|
|
|
input_ids = inputs_dict.get("input_features", None)
|
Add output in a dictionary for TF `generate` method (#12139)
* Add output args to greedy search
* Fix critical typo + make style quality
* Handle generate_beam_search
* Add dict_specific tests and fix the placement of encoder outputs
* Add specific outputs
* Update doc
* Fix typo
* Adjust handling encoder_outputs + Fix generating for T5
* Fix generate for RAG
* Fix handling ouptut_attentions when target_mapping is not None
Take care of situations when target_mapping is provided
as there are 2-tuple of attentions
Change from:
if inputs["output_attentions"]:
attentions = tuple(tf.transpose(t, perm(2, 3, 0, 1)) for t in attentions)
to:
if inputs["output_attentions"]:
if inputs["target_mapping"] is not None:
# when target_mapping is provided, there are 2-tuple of attentions
attentions = tuple(
tuple(tf.transpose(attn_stream, perm=(2, 3, 0, 1)) for attn_stream in t) for t in attentions
)
else:
attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
* Rename kwargs to model_kwargs
* make style quality
* Move imports in test_modeling_tf_common.py
Move ModelOutput-related imports in test_modeling_tf_common.py
into the `is_tf_available():` statement.
* Rewrite nested if-statements
* Fix added tests
2021-06-23 17:52:11 +08:00
|
|
|
|
|
|
|
# iterate over all generative models
|
|
|
|
for model_class in self.all_generative_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
output_greedy = model.generate(
|
|
|
|
input_ids,
|
|
|
|
do_sample=False,
|
|
|
|
output_scores=True,
|
|
|
|
output_hidden_states=True,
|
|
|
|
output_attentions=True,
|
|
|
|
return_dict_in_generate=True,
|
|
|
|
)
|
|
|
|
output_sample = model.generate(
|
|
|
|
input_ids,
|
|
|
|
do_sample=True,
|
|
|
|
output_scores=True,
|
|
|
|
output_hidden_states=True,
|
|
|
|
output_attentions=True,
|
|
|
|
return_dict_in_generate=True,
|
|
|
|
)
|
|
|
|
|
|
|
|
if model.config.is_encoder_decoder:
|
|
|
|
self.assertIsInstance(output_greedy, TFGreedySearchEncoderDecoderOutput)
|
|
|
|
self.assertIsInstance(output_sample, TFSampleEncoderDecoderOutput)
|
|
|
|
else:
|
|
|
|
self.assertIsInstance(output_greedy, TFGreedySearchDecoderOnlyOutput)
|
|
|
|
self.assertIsInstance(output_sample, TFSampleDecoderOnlyOutput)
|
|
|
|
|
2020-04-06 16:37:05 +08:00
|
|
|
def test_lm_head_model_random_beam_search_generate(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
2021-06-15 01:58:54 +08:00
|
|
|
input_ids = inputs_dict.get("input_ids", None)
|
2020-04-06 16:37:05 +08:00
|
|
|
|
|
|
|
for model_class in self.all_generative_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
|
|
|
|
if config.bos_token_id is None:
|
2022-02-09 00:27:23 +08:00
|
|
|
# if bos token id is not defined model needs input_ids, num_return_sequences = 1
|
2020-04-06 16:37:05 +08:00
|
|
|
self._check_generated_ids(model.generate(input_ids, do_sample=True, num_beams=2))
|
|
|
|
else:
|
|
|
|
# num_return_sequences = 1
|
|
|
|
self._check_generated_ids(model.generate(do_sample=True, max_length=5, num_beams=2))
|
|
|
|
|
|
|
|
with self.assertRaises(AssertionError):
|
|
|
|
# generating more sequences than having beams leads is not possible
|
|
|
|
model.generate(input_ids, do_sample=False, num_return_sequences=3, num_beams=2)
|
|
|
|
|
|
|
|
# num_return_sequences > 1, sample
|
2020-08-26 23:20:22 +08:00
|
|
|
self._check_generated_ids(
|
|
|
|
model.generate(
|
|
|
|
input_ids,
|
|
|
|
do_sample=True,
|
|
|
|
num_beams=2,
|
|
|
|
num_return_sequences=2,
|
|
|
|
)
|
|
|
|
)
|
2020-04-06 16:37:05 +08:00
|
|
|
# num_return_sequences > 1, greedy
|
|
|
|
self._check_generated_ids(model.generate(input_ids, do_sample=False, num_beams=2, num_return_sequences=2))
|
|
|
|
|
|
|
|
# check bad words tokens language generation
|
|
|
|
# create list of 1-seq bad token and list of 2-seq of bad tokens
|
|
|
|
bad_words_ids = [self._generate_random_bad_tokens(1, model), self._generate_random_bad_tokens(2, model)]
|
2020-04-01 00:42:31 +08:00
|
|
|
output_tokens = model.generate(
|
2020-04-06 16:37:05 +08:00
|
|
|
input_ids, do_sample=False, bad_words_ids=bad_words_ids, num_beams=2, num_return_sequences=2
|
2020-04-01 00:42:31 +08:00
|
|
|
)
|
2020-04-06 16:37:05 +08:00
|
|
|
# only count generated tokens
|
2020-04-01 00:42:31 +08:00
|
|
|
generated_ids = output_tokens[:, input_ids.shape[-1] :]
|
|
|
|
self.assertFalse(self._check_match_tokens(generated_ids.numpy().tolist(), bad_words_ids))
|
|
|
|
|
Add output in a dictionary for TF `generate` method (#12139)
* Add output args to greedy search
* Fix critical typo + make style quality
* Handle generate_beam_search
* Add dict_specific tests and fix the placement of encoder outputs
* Add specific outputs
* Update doc
* Fix typo
* Adjust handling encoder_outputs + Fix generating for T5
* Fix generate for RAG
* Fix handling ouptut_attentions when target_mapping is not None
Take care of situations when target_mapping is provided
as there are 2-tuple of attentions
Change from:
if inputs["output_attentions"]:
attentions = tuple(tf.transpose(t, perm(2, 3, 0, 1)) for t in attentions)
to:
if inputs["output_attentions"]:
if inputs["target_mapping"] is not None:
# when target_mapping is provided, there are 2-tuple of attentions
attentions = tuple(
tuple(tf.transpose(attn_stream, perm=(2, 3, 0, 1)) for attn_stream in t) for t in attentions
)
else:
attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
* Rename kwargs to model_kwargs
* make style quality
* Move imports in test_modeling_tf_common.py
Move ModelOutput-related imports in test_modeling_tf_common.py
into the `is_tf_available():` statement.
* Rewrite nested if-statements
* Fix added tests
2021-06-23 17:52:11 +08:00
|
|
|
def test_lm_head_model_beam_search_generate_dict_outputs(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
input_ids = inputs_dict.get("input_ids", None)
|
2022-02-09 00:27:23 +08:00
|
|
|
if input_ids is None:
|
|
|
|
input_ids = inputs_dict.get("input_features", None)
|
Add output in a dictionary for TF `generate` method (#12139)
* Add output args to greedy search
* Fix critical typo + make style quality
* Handle generate_beam_search
* Add dict_specific tests and fix the placement of encoder outputs
* Add specific outputs
* Update doc
* Fix typo
* Adjust handling encoder_outputs + Fix generating for T5
* Fix generate for RAG
* Fix handling ouptut_attentions when target_mapping is not None
Take care of situations when target_mapping is provided
as there are 2-tuple of attentions
Change from:
if inputs["output_attentions"]:
attentions = tuple(tf.transpose(t, perm(2, 3, 0, 1)) for t in attentions)
to:
if inputs["output_attentions"]:
if inputs["target_mapping"] is not None:
# when target_mapping is provided, there are 2-tuple of attentions
attentions = tuple(
tuple(tf.transpose(attn_stream, perm=(2, 3, 0, 1)) for attn_stream in t) for t in attentions
)
else:
attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
* Rename kwargs to model_kwargs
* make style quality
* Move imports in test_modeling_tf_common.py
Move ModelOutput-related imports in test_modeling_tf_common.py
into the `is_tf_available():` statement.
* Rewrite nested if-statements
* Fix added tests
2021-06-23 17:52:11 +08:00
|
|
|
|
|
|
|
# iterate over all generative models
|
|
|
|
for model_class in self.all_generative_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
output_beam_search = model.generate(
|
|
|
|
input_ids,
|
|
|
|
num_beams=2,
|
|
|
|
do_sample=False,
|
|
|
|
output_scores=True,
|
|
|
|
output_hidden_states=True,
|
|
|
|
output_attentions=True,
|
|
|
|
return_dict_in_generate=True,
|
|
|
|
)
|
|
|
|
output_beam_sample = model.generate(
|
|
|
|
input_ids,
|
|
|
|
num_beams=2,
|
|
|
|
do_sample=True,
|
|
|
|
output_scores=True,
|
|
|
|
output_hidden_states=True,
|
|
|
|
output_attentions=True,
|
|
|
|
return_dict_in_generate=True,
|
|
|
|
)
|
|
|
|
|
|
|
|
if model.config.is_encoder_decoder:
|
|
|
|
self.assertIsInstance(output_beam_search, TFBeamSearchEncoderDecoderOutput)
|
|
|
|
self.assertIsInstance(output_beam_sample, TFBeamSampleEncoderDecoderOutput)
|
|
|
|
else:
|
|
|
|
self.assertIsInstance(output_beam_search, TFBeamSearchDecoderOnlyOutput)
|
|
|
|
self.assertIsInstance(output_beam_sample, TFBeamSampleDecoderOnlyOutput)
|
|
|
|
|
2020-06-24 23:37:20 +08:00
|
|
|
def test_loss_computation(self):
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model = model_class(config)
|
2022-01-19 21:29:07 +08:00
|
|
|
if getattr(model, "hf_compute_loss", None):
|
2020-06-24 23:37:20 +08:00
|
|
|
# The number of elements in the loss should be the same as the number of elements in the label
|
|
|
|
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
|
2020-11-13 03:08:26 +08:00
|
|
|
added_label = prepared_for_class[
|
|
|
|
sorted(list(prepared_for_class.keys() - inputs_dict.keys()), reverse=True)[0]
|
|
|
|
]
|
2020-06-24 23:37:20 +08:00
|
|
|
loss_size = tf.size(added_label)
|
|
|
|
|
2021-04-09 06:41:36 +08:00
|
|
|
if model.__class__ in get_values(TF_MODEL_FOR_CAUSAL_LM_MAPPING):
|
2020-07-08 00:15:53 +08:00
|
|
|
# if loss is causal lm loss, labels are shift, so that one label per batch
|
|
|
|
# is cut
|
|
|
|
loss_size = loss_size - self.model_tester.batch_size
|
|
|
|
|
2020-06-24 23:37:20 +08:00
|
|
|
# Test that model correctly compute the loss with kwargs
|
|
|
|
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
|
2022-02-09 00:27:23 +08:00
|
|
|
possible_input_names = {"input_ids", "pixel_values", "input_features"}
|
|
|
|
input_name = possible_input_names.intersection(set(prepared_for_class)).pop()
|
|
|
|
model_input = prepared_for_class.pop(input_name)
|
2020-07-08 00:15:53 +08:00
|
|
|
|
2022-02-09 00:27:23 +08:00
|
|
|
loss = model(model_input, **prepared_for_class)[0]
|
2020-06-24 23:37:20 +08:00
|
|
|
self.assertEqual(loss.shape, [loss_size])
|
|
|
|
|
|
|
|
# Test that model correctly compute the loss with a dict
|
|
|
|
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
|
|
|
|
loss = model(prepared_for_class)[0]
|
|
|
|
self.assertEqual(loss.shape, [loss_size])
|
|
|
|
|
|
|
|
# Test that model correctly compute the loss with a tuple
|
|
|
|
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
|
|
|
|
|
|
|
|
# Get keys that were added with the _prepare_for_class function
|
|
|
|
label_keys = prepared_for_class.keys() - inputs_dict.keys()
|
2020-11-13 03:08:26 +08:00
|
|
|
signature = inspect.signature(model.call).parameters
|
|
|
|
signature_names = list(signature.keys())
|
2020-06-24 23:37:20 +08:00
|
|
|
|
|
|
|
# Create a dictionary holding the location of the tensors in the tuple
|
2021-11-09 20:54:37 +08:00
|
|
|
tuple_index_mapping = {0: input_name}
|
2020-06-24 23:37:20 +08:00
|
|
|
for label_key in label_keys:
|
2020-11-13 03:08:26 +08:00
|
|
|
label_key_index = signature_names.index(label_key)
|
2020-06-24 23:37:20 +08:00
|
|
|
tuple_index_mapping[label_key_index] = label_key
|
|
|
|
sorted_tuple_index_mapping = sorted(tuple_index_mapping.items())
|
2020-11-13 03:08:26 +08:00
|
|
|
# Initialize a list with their default values, update the values and convert to a tuple
|
|
|
|
list_input = []
|
|
|
|
|
|
|
|
for name in signature_names:
|
|
|
|
if name != "kwargs":
|
|
|
|
list_input.append(signature[name].default)
|
2020-06-24 23:37:20 +08:00
|
|
|
|
|
|
|
for index, value in sorted_tuple_index_mapping:
|
2020-11-13 03:08:26 +08:00
|
|
|
list_input[index] = prepared_for_class[value]
|
|
|
|
|
2020-06-24 23:37:20 +08:00
|
|
|
tuple_input = tuple(list_input)
|
|
|
|
|
|
|
|
# Send to model
|
2020-11-13 03:08:26 +08:00
|
|
|
loss = model(tuple_input[:-1])[0]
|
|
|
|
|
2020-06-24 23:37:20 +08:00
|
|
|
self.assertEqual(loss.shape, [loss_size])
|
|
|
|
|
2021-05-26 21:02:44 +08:00
|
|
|
def test_generate_with_headmasking(self):
|
|
|
|
attention_names = ["encoder_attentions", "decoder_attentions", "cross_attentions"]
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
for model_class in self.all_generative_model_classes:
|
|
|
|
model = model_class(config)
|
|
|
|
|
|
|
|
# We want to test only encoder-decoder models
|
|
|
|
if not config.is_encoder_decoder:
|
|
|
|
continue
|
|
|
|
|
|
|
|
head_masking = {
|
|
|
|
"head_mask": tf.zeros((config.encoder_layers, config.encoder_attention_heads)),
|
|
|
|
"decoder_head_mask": tf.zeros((config.decoder_layers, config.decoder_attention_heads)),
|
|
|
|
"cross_attn_head_mask": tf.zeros((config.decoder_layers, config.decoder_attention_heads)),
|
|
|
|
}
|
|
|
|
|
|
|
|
signature = inspect.signature(model.call)
|
|
|
|
if set(head_masking.keys()) < set([*signature.parameters.keys()]):
|
|
|
|
continue
|
|
|
|
|
|
|
|
for attn_name, (name, mask) in zip(attention_names, head_masking.items()):
|
|
|
|
out = model.generate(
|
|
|
|
inputs_dict["input_ids"],
|
|
|
|
num_beams=1,
|
|
|
|
max_length=inputs_dict["input_ids"] + 5,
|
|
|
|
output_attentions=True,
|
|
|
|
return_dict_in_generate=True,
|
|
|
|
**{name: mask},
|
|
|
|
)
|
|
|
|
# We check the state of decoder_attentions and cross_attentions just from the last step
|
|
|
|
attn_weights = out[attn_name] if attn_name == attention_names[0] else out[attn_name][-1]
|
|
|
|
self.assertEqual(sum([tf.reduce_sum(w).numpy() for w in attn_weights]), 0.0)
|
|
|
|
|
2021-07-13 22:15:15 +08:00
|
|
|
def test_load_with_mismatched_shapes(self):
|
2021-10-22 00:31:29 +08:00
|
|
|
if not self.test_mismatched_shapes:
|
|
|
|
return
|
2021-07-13 22:15:15 +08:00
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
if model_class not in get_values(TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING):
|
|
|
|
continue
|
|
|
|
|
|
|
|
with self.subTest(msg=f"Testing {model_class}"):
|
|
|
|
with tempfile.TemporaryDirectory() as tmp_dir:
|
|
|
|
model = model_class(config)
|
|
|
|
inputs = self._prepare_for_class(inputs_dict, model_class)
|
|
|
|
_ = model(**inputs)
|
|
|
|
model.save_pretrained(tmp_dir)
|
|
|
|
|
|
|
|
# Fails when we don't set ignore_mismatched_sizes=True
|
|
|
|
with self.assertRaises(ValueError):
|
|
|
|
new_model = TFAutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
|
2021-10-22 00:31:29 +08:00
|
|
|
with self.assertRaises(ValueError):
|
|
|
|
new_model_without_prefix = TFAutoModel.from_pretrained(tmp_dir, vocab_size=10)
|
2021-07-13 22:15:15 +08:00
|
|
|
|
|
|
|
logger = logging.get_logger("transformers.modeling_tf_utils")
|
|
|
|
with CaptureLogger(logger) as cl:
|
|
|
|
new_model = TFAutoModelForSequenceClassification.from_pretrained(
|
|
|
|
tmp_dir, num_labels=42, ignore_mismatched_sizes=True
|
|
|
|
)
|
|
|
|
self.assertIn("the shapes did not match", cl.out)
|
|
|
|
|
|
|
|
logits = new_model(**inputs).logits
|
|
|
|
self.assertEqual(logits.shape[1], 42)
|
|
|
|
|
2021-10-22 00:31:29 +08:00
|
|
|
with CaptureLogger(logger) as cl:
|
|
|
|
new_model_without_prefix = TFAutoModel.from_pretrained(
|
|
|
|
tmp_dir, vocab_size=10, ignore_mismatched_sizes=True
|
|
|
|
)
|
|
|
|
self.assertIn("the shapes did not match", cl.out)
|
|
|
|
|
|
|
|
# Although Tf models always have a prefix pointing to `MainLayer`,
|
|
|
|
# we still add this "without prefix" test to keep a consistency between tf and pt tests.
|
|
|
|
input_ids = ids_tensor((2, 8), 10)
|
|
|
|
if self.is_encoder_decoder:
|
|
|
|
new_model_without_prefix(input_ids, decoder_input_ids=input_ids)
|
|
|
|
else:
|
|
|
|
new_model_without_prefix(input_ids)
|
|
|
|
|
2021-12-21 00:19:08 +08:00
|
|
|
def test_model_main_input_name(self):
|
|
|
|
for model_class in self.all_model_classes:
|
|
|
|
model_signature = inspect.signature(getattr(model_class, "call"))
|
|
|
|
# The main input is the name of the argument after `self`
|
|
|
|
observed_main_input_name = list(model_signature.parameters.keys())[1]
|
|
|
|
self.assertEqual(model_class.main_input_name, observed_main_input_name)
|
|
|
|
|
2020-04-06 16:37:05 +08:00
|
|
|
def _generate_random_bad_tokens(self, num_bad_tokens, model):
|
|
|
|
# special tokens cannot be bad tokens
|
|
|
|
special_tokens = []
|
|
|
|
if model.config.bos_token_id is not None:
|
|
|
|
special_tokens.append(model.config.bos_token_id)
|
|
|
|
if model.config.pad_token_id is not None:
|
|
|
|
special_tokens.append(model.config.pad_token_id)
|
|
|
|
if model.config.eos_token_id is not None:
|
|
|
|
special_tokens.append(model.config.eos_token_id)
|
|
|
|
|
|
|
|
# create random bad tokens that are not special tokens
|
|
|
|
bad_tokens = []
|
|
|
|
while len(bad_tokens) < num_bad_tokens:
|
|
|
|
token = tf.squeeze(ids_tensor((1, 1), self.model_tester.vocab_size), 0).numpy()[0]
|
|
|
|
if token not in special_tokens:
|
|
|
|
bad_tokens.append(token)
|
|
|
|
return bad_tokens
|
|
|
|
|
2020-04-01 00:42:31 +08:00
|
|
|
def _check_generated_ids(self, output_ids):
|
2020-03-03 22:42:15 +08:00
|
|
|
for token_id in output_ids[0].numpy().tolist():
|
|
|
|
self.assertGreaterEqual(token_id, 0)
|
|
|
|
self.assertLess(token_id, self.model_tester.vocab_size)
|
|
|
|
|
2020-04-01 00:42:31 +08:00
|
|
|
def _check_match_tokens(self, generated_ids, bad_words_ids):
|
|
|
|
# for all bad word tokens
|
|
|
|
for bad_word_ids in bad_words_ids:
|
|
|
|
# for all slices in batch
|
|
|
|
for generated_ids_slice in generated_ids:
|
|
|
|
# for all word idx
|
|
|
|
for i in range(len(bad_word_ids), len(generated_ids_slice)):
|
|
|
|
# if tokens match
|
|
|
|
if generated_ids_slice[i - len(bad_word_ids) : i] == bad_word_ids:
|
|
|
|
return True
|
|
|
|
return False
|
|
|
|
|
2019-09-05 08:27:39 +08:00
|
|
|
|
2019-09-18 18:17:21 +08:00
|
|
|
def ids_tensor(shape, vocab_size, rng=None, name=None, dtype=None):
|
2019-09-05 08:27:39 +08:00
|
|
|
"""Creates a random int32 tensor of the shape within the vocab size."""
|
|
|
|
if rng is None:
|
|
|
|
rng = random.Random()
|
|
|
|
|
|
|
|
total_dims = 1
|
|
|
|
for dim in shape:
|
|
|
|
total_dims *= dim
|
|
|
|
|
|
|
|
values = []
|
|
|
|
for _ in range(total_dims):
|
|
|
|
values.append(rng.randint(0, vocab_size - 1))
|
|
|
|
|
2019-12-21 22:46:46 +08:00
|
|
|
output = tf.constant(values, shape=shape, dtype=dtype if dtype is not None else tf.int32)
|
2019-09-18 18:17:21 +08:00
|
|
|
|
|
|
|
return output
|
2020-03-03 22:42:15 +08:00
|
|
|
|
|
|
|
|
2021-12-24 00:19:44 +08:00
|
|
|
def random_attention_mask(shape, rng=None, name=None, dtype=None):
|
|
|
|
attn_mask = ids_tensor(shape, vocab_size=2, rng=None, name=None, dtype=dtype)
|
|
|
|
# make sure that at least one token is attended to for each batch
|
|
|
|
attn_mask = tf.concat([tf.constant(value=1, shape=(shape[0], 1), dtype=dtype), attn_mask[:, 1:]], axis=1)
|
|
|
|
return attn_mask
|
|
|
|
|
|
|
|
|
2021-10-13 06:10:34 +08:00
|
|
|
def floats_tensor(shape, scale=1.0, rng=None, name=None, dtype=None):
|
|
|
|
"""Creates a random float32 tensor"""
|
|
|
|
if rng is None:
|
|
|
|
rng = random.Random()
|
|
|
|
|
|
|
|
total_dims = 1
|
|
|
|
for dim in shape:
|
|
|
|
total_dims *= dim
|
|
|
|
|
|
|
|
values = []
|
|
|
|
for _ in range(total_dims):
|
|
|
|
values.append(rng.random() * scale)
|
|
|
|
|
|
|
|
return tf.reshape(tf.constant(values, dtype=dtype if dtype is not None else tf.float32), shape=shape)
|
|
|
|
|
|
|
|
|
2020-03-03 22:42:15 +08:00
|
|
|
@require_tf
|
|
|
|
class UtilsFunctionsTest(unittest.TestCase):
|
|
|
|
|
|
|
|
# tests whether the top_k_top_p_filtering function behaves as expected
|
|
|
|
def test_top_k_top_p_filtering(self):
|
|
|
|
logits = tf.convert_to_tensor(
|
|
|
|
[
|
|
|
|
[
|
|
|
|
8.2220991, # 3rd highest value; idx. 0
|
|
|
|
-0.5620044,
|
|
|
|
5.23229752,
|
|
|
|
4.0386393,
|
|
|
|
-6.8798378,
|
|
|
|
-0.54785802,
|
|
|
|
-3.2012153,
|
|
|
|
2.92777176,
|
|
|
|
1.88171953,
|
|
|
|
7.35341276, # 5th highest value; idx. 9
|
|
|
|
8.43207833, # 2nd highest value; idx. 10
|
|
|
|
-9.85711836,
|
|
|
|
-5.96209236,
|
|
|
|
-1.13039161,
|
|
|
|
-7.1115294,
|
|
|
|
-0.8369633,
|
|
|
|
-5.3186408,
|
|
|
|
7.06427407,
|
|
|
|
0.81369344,
|
|
|
|
-0.82023817,
|
|
|
|
-5.9179796,
|
|
|
|
0.58813443,
|
|
|
|
-6.99778438,
|
|
|
|
4.71551189,
|
|
|
|
-0.18771637,
|
|
|
|
7.44020759, # 4th highest value; idx. 25
|
|
|
|
9.38450987, # 1st highest value; idx. 26
|
|
|
|
2.12662941,
|
|
|
|
-9.32562038,
|
|
|
|
2.35652522,
|
|
|
|
], # cummulative prob of 5 highest values <= 0.6
|
|
|
|
[
|
|
|
|
0.58425518,
|
|
|
|
4.53139238,
|
|
|
|
-5.57510464,
|
|
|
|
-6.28030699,
|
|
|
|
-7.19529503,
|
|
|
|
-4.02122551,
|
|
|
|
1.39337037,
|
|
|
|
-6.06707057,
|
|
|
|
1.59480517,
|
|
|
|
-9.643119,
|
|
|
|
0.03907799,
|
|
|
|
0.67231762,
|
|
|
|
-8.88206726,
|
|
|
|
6.27115922, # 4th highest value; idx. 13
|
|
|
|
2.28520723,
|
|
|
|
4.82767506,
|
|
|
|
4.30421368,
|
|
|
|
8.8275313, # 2nd highest value; idx. 17
|
|
|
|
5.44029958, # 5th highest value; idx. 18
|
|
|
|
-4.4735794,
|
|
|
|
7.38579536, # 3rd highest value; idx. 20
|
|
|
|
-2.91051663,
|
|
|
|
2.61946077,
|
|
|
|
-2.5674762,
|
|
|
|
-9.48959302,
|
|
|
|
-4.02922645,
|
|
|
|
-1.35416918,
|
|
|
|
9.67702323, # 1st highest value; idx. 27
|
|
|
|
-5.89478553,
|
|
|
|
1.85370467,
|
|
|
|
], # cummulative prob of 5 highest values <= 0.6
|
|
|
|
],
|
|
|
|
dtype=tf.float32,
|
|
|
|
)
|
|
|
|
|
|
|
|
non_inf_expected_idx = tf.convert_to_tensor(
|
2020-08-26 23:20:22 +08:00
|
|
|
[[0, 0], [0, 9], [0, 10], [0, 25], [0, 26], [1, 13], [1, 17], [1, 18], [1, 20], [1, 27]],
|
|
|
|
dtype=tf.int32,
|
2020-03-03 22:42:15 +08:00
|
|
|
) # expected non filtered idx as noted above
|
|
|
|
|
|
|
|
non_inf_expected_output = tf.convert_to_tensor(
|
|
|
|
[8.222099, 7.3534126, 8.432078, 7.4402075, 9.38451, 6.271159, 8.827531, 5.4402995, 7.3857956, 9.677023],
|
|
|
|
dtype=tf.float32,
|
|
|
|
) # expected non filtered values as noted above
|
|
|
|
|
|
|
|
output = tf_top_k_top_p_filtering(logits, top_k=10, top_p=0.6, min_tokens_to_keep=4)
|
|
|
|
|
|
|
|
non_inf_output = output[output != -float("inf")]
|
|
|
|
non_inf_idx = tf.cast(
|
2020-08-26 23:20:22 +08:00
|
|
|
tf.where(tf.not_equal(output, tf.constant(-float("inf"), dtype=tf.float32))),
|
|
|
|
dtype=tf.int32,
|
2020-03-03 22:42:15 +08:00
|
|
|
)
|
|
|
|
|
|
|
|
tf.debugging.assert_near(non_inf_output, non_inf_expected_output, rtol=1e-12)
|
|
|
|
tf.debugging.assert_equal(non_inf_idx, non_inf_expected_idx)
|
2021-04-23 21:17:37 +08:00
|
|
|
|
|
|
|
|
|
|
|
@require_tf
|
|
|
|
@is_staging_test
|
|
|
|
class TFModelPushToHubTester(unittest.TestCase):
|
|
|
|
@classmethod
|
|
|
|
def setUpClass(cls):
|
2021-11-03 06:58:42 +08:00
|
|
|
cls._token = login(username=USER, password=PASS)
|
2021-04-23 21:17:37 +08:00
|
|
|
|
|
|
|
@classmethod
|
|
|
|
def tearDownClass(cls):
|
|
|
|
try:
|
2021-11-03 06:58:42 +08:00
|
|
|
delete_repo(token=cls._token, name="test-model-tf")
|
2021-04-23 21:17:37 +08:00
|
|
|
except HTTPError:
|
|
|
|
pass
|
|
|
|
|
|
|
|
try:
|
2021-11-03 06:58:42 +08:00
|
|
|
delete_repo(token=cls._token, name="test-model-tf-org", organization="valid_org")
|
2021-04-23 21:17:37 +08:00
|
|
|
except HTTPError:
|
|
|
|
pass
|
|
|
|
|
|
|
|
def test_push_to_hub(self):
|
|
|
|
config = BertConfig(
|
|
|
|
vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
|
|
|
|
)
|
|
|
|
model = TFBertModel(config)
|
|
|
|
# Make sure model is properly initialized
|
|
|
|
_ = model(model.dummy_inputs)
|
|
|
|
with tempfile.TemporaryDirectory() as tmp_dir:
|
2021-06-23 22:11:19 +08:00
|
|
|
model.save_pretrained(os.path.join(tmp_dir, "test-model-tf"), push_to_hub=True, use_auth_token=self._token)
|
2021-04-23 21:17:37 +08:00
|
|
|
|
2021-04-26 23:52:23 +08:00
|
|
|
new_model = TFBertModel.from_pretrained(f"{USER}/test-model-tf")
|
2021-04-23 21:17:37 +08:00
|
|
|
models_equal = True
|
|
|
|
for p1, p2 in zip(model.weights, new_model.weights):
|
|
|
|
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
|
|
|
models_equal = False
|
|
|
|
self.assertTrue(models_equal)
|
|
|
|
|
2021-12-15 22:57:52 +08:00
|
|
|
def test_push_to_hub_with_model_card(self):
|
|
|
|
config = BertConfig(
|
|
|
|
vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
|
|
|
|
)
|
|
|
|
model = TFBertModel(config)
|
|
|
|
with tempfile.TemporaryDirectory() as tmp_dir:
|
|
|
|
model.push_to_hub(os.path.join(tmp_dir, "test-model-tf"))
|
|
|
|
self.assertTrue(os.path.isfile(os.path.join(tmp_dir, "test-model-card-tf", "README.md")))
|
|
|
|
|
2021-04-23 21:17:37 +08:00
|
|
|
def test_push_to_hub_in_organization(self):
|
|
|
|
config = BertConfig(
|
|
|
|
vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
|
|
|
|
)
|
|
|
|
model = TFBertModel(config)
|
|
|
|
with tempfile.TemporaryDirectory() as tmp_dir:
|
|
|
|
model.save_pretrained(
|
2021-06-23 22:11:19 +08:00
|
|
|
os.path.join(tmp_dir, "test-model-tf-org"),
|
2021-04-23 21:17:37 +08:00
|
|
|
push_to_hub=True,
|
|
|
|
use_auth_token=self._token,
|
|
|
|
organization="valid_org",
|
|
|
|
)
|
|
|
|
|
2021-04-26 23:52:23 +08:00
|
|
|
new_model = TFBertModel.from_pretrained("valid_org/test-model-tf-org")
|
2021-04-23 21:17:37 +08:00
|
|
|
models_equal = True
|
|
|
|
for p1, p2 in zip(model.weights, new_model.weights):
|
|
|
|
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
|
|
|
models_equal = False
|
|
|
|
self.assertTrue(models_equal)
|