chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1,13 @@
|
||||
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# 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.
|
||||
@@ -0,0 +1,504 @@
|
||||
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
|
||||
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# 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.
|
||||
|
||||
import random
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import paddle
|
||||
from parameterized import parameterized_class
|
||||
|
||||
from paddlenlp.transformers import (
|
||||
CODEGEN_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
AutoTokenizer,
|
||||
CodeGenConfig,
|
||||
CodeGenForCausalLM,
|
||||
CodeGenModel,
|
||||
)
|
||||
|
||||
from ...testing_utils import slow
|
||||
from ..test_generation_utils import GenerationTesterMixin
|
||||
from ..test_modeling_common import (
|
||||
ModelTesterMixin,
|
||||
floats_tensor,
|
||||
ids_tensor,
|
||||
random_attention_mask,
|
||||
)
|
||||
|
||||
|
||||
class CodeGenModelTester:
|
||||
test_model_name_list = False
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=14,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_input_mask=True,
|
||||
use_labels=True,
|
||||
use_mc_token_ids=True,
|
||||
vocab_size=256,
|
||||
hidden_size=32,
|
||||
rotary_dim=4,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.0,
|
||||
attention_probs_dropout_prob=0.0,
|
||||
max_position_embeddings=512,
|
||||
type_vocab_size=16,
|
||||
type_sequence_label_size=2,
|
||||
initializer_range=0.02,
|
||||
num_labels=3,
|
||||
num_choices=4,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_input_mask = use_input_mask
|
||||
self.use_labels = use_labels
|
||||
self.use_mc_token_ids = use_mc_token_ids
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.rotary_dim = rotary_dim
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.type_sequence_label_size = type_sequence_label_size
|
||||
self.initializer_range = initializer_range
|
||||
self.num_labels = num_labels
|
||||
self.num_choices = num_choices
|
||||
self.scope = None
|
||||
self.bos_token_id = vocab_size - 1
|
||||
self.eos_token_id = vocab_size - 1
|
||||
self.pad_token_id = vocab_size - 1
|
||||
|
||||
paddle.seed(128)
|
||||
np.random.seed(128)
|
||||
random.seed(128)
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size, dtype="int64")
|
||||
|
||||
input_mask = None
|
||||
if self.use_input_mask:
|
||||
input_mask = random_attention_mask([self.batch_size, self.seq_length], dtype="int64")
|
||||
|
||||
mc_token_ids = None
|
||||
if self.use_mc_token_ids:
|
||||
mc_token_ids = ids_tensor([self.batch_size, self.num_choices], self.seq_length, dtype="int64")
|
||||
|
||||
sequence_labels = None
|
||||
token_labels = None
|
||||
choice_labels = None
|
||||
if self.use_labels:
|
||||
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size, dtype="int64")
|
||||
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels, dtype="int64")
|
||||
choice_labels = ids_tensor([self.batch_size], self.num_choices, dtype="int64")
|
||||
|
||||
config = self.get_config()
|
||||
|
||||
return (
|
||||
config,
|
||||
input_ids,
|
||||
input_mask,
|
||||
mc_token_ids,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
)
|
||||
|
||||
def get_config(self):
|
||||
return CodeGenConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
n_embd=self.hidden_size,
|
||||
n_layer=self.num_hidden_layers,
|
||||
n_head=self.num_attention_heads,
|
||||
activation_function=self.hidden_act,
|
||||
resid_pdrop=self.hidden_dropout_prob,
|
||||
attn_pdrop=self.attention_probs_dropout_prob,
|
||||
n_positions=self.max_position_embeddings,
|
||||
n_ctx=self.max_position_embeddings,
|
||||
initializer_range=self.initializer_range,
|
||||
bos_token_id=self.bos_token_id,
|
||||
eos_token_id=self.eos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
rotary_dim=self.rotary_dim,
|
||||
)
|
||||
|
||||
def prepare_config_and_inputs_for_decoder(self):
|
||||
(
|
||||
config,
|
||||
input_ids,
|
||||
input_mask,
|
||||
mc_token_ids,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
) = self.prepare_config_and_inputs()
|
||||
|
||||
encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
|
||||
encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2, dtype="int64")
|
||||
|
||||
return (
|
||||
config,
|
||||
input_ids,
|
||||
input_mask,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
encoder_hidden_states,
|
||||
encoder_attention_mask,
|
||||
)
|
||||
|
||||
def create_and_check_codegen_model(self, config, input_ids, input_mask, *args):
|
||||
model = CodeGenModel(config)
|
||||
model.eval()
|
||||
|
||||
result = model(input_ids, use_cache=True, return_dict=self.parent.return_dict)
|
||||
|
||||
self.parent.assertEqual(result[0].shape, [self.batch_size, self.seq_length, self.hidden_size])
|
||||
self.parent.assertEqual(len(result[1]), config["n_layer"])
|
||||
|
||||
def create_and_check_codegen_model_past(self, config, input_ids, input_mask, *args):
|
||||
model = CodeGenModel(config)
|
||||
model.eval()
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, use_cache=True, return_dict=self.parent.return_dict)
|
||||
outputs_no_past = model(input_ids, use_cache=False, return_dict=self.parent.return_dict)
|
||||
|
||||
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
|
||||
|
||||
output, past = outputs[:2]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 1), config["vocab_size"], dtype="int64")
|
||||
|
||||
# append to next input_ids
|
||||
next_input_ids = paddle.concat([input_ids, next_tokens], axis=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, return_dict=self.parent.return_dict)[0]
|
||||
output_from_past = model(next_tokens, cache=past, return_dict=self.parent.return_dict)[0]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1], dtype="int64").item()
|
||||
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
||||
|
||||
# test that outputs are equal for slice
|
||||
self.parent.assertTrue(paddle.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
||||
|
||||
def create_and_check_codegen_model_attention_mask_past(self, config, input_ids, input_mask, *args):
|
||||
model = CodeGenModel(config)
|
||||
model.eval()
|
||||
|
||||
# create attention mask
|
||||
attn_mask = paddle.ones(input_ids.shape, dtype="int64")
|
||||
half_seq_length = self.seq_length // 2
|
||||
attn_mask[:, half_seq_length:] = 0
|
||||
|
||||
# first forward pass
|
||||
output, past = model(input_ids, attention_mask=attn_mask, use_cache=True, return_dict=self.parent.return_dict)[
|
||||
:2
|
||||
]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 1), config["vocab_size"], dtype="int64")
|
||||
|
||||
# change a random masked slice from input_ids
|
||||
random_seq_idx_to_change = ids_tensor((1,), half_seq_length, dtype="int64").item() + 1
|
||||
random_other_next_tokens = ids_tensor((self.batch_size, 1), config["vocab_size"], dtype="int64").squeeze(-1)
|
||||
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
|
||||
|
||||
# append to next input_ids and attn_mask
|
||||
next_input_ids = paddle.concat([input_ids, next_tokens], axis=-1)
|
||||
attn_mask = paddle.concat(
|
||||
[attn_mask, paddle.ones((attn_mask.shape[0], 1), dtype="int64")],
|
||||
axis=1,
|
||||
)
|
||||
|
||||
# get two different outputs
|
||||
output_from_no_past = model(next_input_ids, attention_mask=attn_mask, return_dict=self.parent.return_dict)[0]
|
||||
output_from_past = model(
|
||||
next_tokens, cache=past, attention_mask=attn_mask, return_dict=self.parent.return_dict
|
||||
)[0]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1], dtype="int64").item()
|
||||
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
||||
|
||||
# test that outputs are equal for slice
|
||||
self.parent.assertTrue(paddle.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
||||
|
||||
def create_and_check_codegen_model_past_large_inputs(self, config, input_ids, input_mask, *args):
|
||||
model = CodeGenModel(config)
|
||||
model.eval()
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=input_mask, use_cache=True, return_dict=self.parent.return_dict)
|
||||
|
||||
output, past = outputs[:2]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config["vocab_size"], dtype="int64")
|
||||
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2, dtype="int64")
|
||||
|
||||
# append to next input_ids
|
||||
next_input_ids = paddle.concat([input_ids, next_tokens], axis=-1)
|
||||
next_attention_mask = paddle.concat([input_mask, next_mask], axis=-1)
|
||||
|
||||
output_from_no_past = model(
|
||||
next_input_ids, attention_mask=next_attention_mask, return_dict=self.parent.return_dict
|
||||
)[0]
|
||||
output_from_past = model(
|
||||
next_tokens, attention_mask=next_attention_mask, cache=past, return_dict=self.parent.return_dict
|
||||
)[0]
|
||||
self.parent.assertTrue(output_from_past.shape[1] == next_tokens.shape[1])
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1], dtype="int64").item()
|
||||
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
|
||||
|
||||
# test that outputs are equal for slice
|
||||
self.parent.assertTrue(paddle.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
||||
|
||||
def create_and_check_lm_head_model(self, config, input_ids, input_mask, *args):
|
||||
model = CodeGenForCausalLM(config)
|
||||
|
||||
outputs = model(
|
||||
input_ids, labels=input_ids if self.parent.use_labels else None, return_dict=self.parent.return_dict
|
||||
)
|
||||
if self.parent.use_labels:
|
||||
loss, logits = outputs[:2]
|
||||
self.parent.assertEqual(loss.shape, [1])
|
||||
else:
|
||||
logits = outputs[0]
|
||||
self.parent.assertEqual(logits.shape, [self.batch_size, self.seq_length, self.vocab_size])
|
||||
|
||||
def create_and_check_forward_and_backwards(self, config, input_ids, input_mask, *args):
|
||||
model = CodeGenForCausalLM(config)
|
||||
|
||||
loss, logits = model(input_ids, return_dict=self.parent.return_dict, labels=input_ids)[:2]
|
||||
self.parent.assertEqual(loss.shape, [1])
|
||||
self.parent.assertEqual(logits.shape, [self.batch_size, self.seq_length, self.vocab_size])
|
||||
loss.backward()
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config_and_inputs = self.prepare_config_and_inputs()
|
||||
|
||||
(
|
||||
config,
|
||||
input_ids,
|
||||
input_mask,
|
||||
mc_token_ids,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
) = config_and_inputs
|
||||
|
||||
inputs_dict = {"input_ids": input_ids}
|
||||
|
||||
return config, inputs_dict
|
||||
|
||||
|
||||
@parameterized_class(
|
||||
("return_dict",),
|
||||
[
|
||||
[False, False],
|
||||
[False, True],
|
||||
[True, False],
|
||||
[True, True],
|
||||
],
|
||||
)
|
||||
class CodeGenModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
|
||||
base_model_class = CodeGenModel
|
||||
|
||||
all_model_classes = (CodeGenModel, CodeGenForCausalLM)
|
||||
all_generative_model_classes = {CodeGenForCausalLM: (CodeGenModel, "transformer")}
|
||||
fx_compatible = False
|
||||
test_pruning = False
|
||||
test_missing_keys = False
|
||||
test_model_parallel = False
|
||||
test_head_masking = False
|
||||
use_test_model_name_list = False
|
||||
return_dict = False
|
||||
use_labels = False
|
||||
use_test_inputs_embeds = True
|
||||
|
||||
# attention mask issue
|
||||
def _get_input_ids_and_config(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
input_ids = inputs_dict[self.input_name]
|
||||
attention_mask = paddle.zeros_like(input_ids, dtype=paddle.float32)
|
||||
|
||||
max_batch_size = 2
|
||||
sequence_length = input_ids.shape[-1] // 2
|
||||
input_ids = input_ids[:max_batch_size, :sequence_length]
|
||||
attention_mask = attention_mask[:max_batch_size, :sequence_length].unsqueeze([1, 2])
|
||||
|
||||
# generate max 3 tokens
|
||||
max_length = 3
|
||||
|
||||
if config.get("eos_token_id", None) is not None and config.get("pad_token_id", None) is None:
|
||||
# hack to allow generate for models such as GPT2 as is done in `generate()`
|
||||
config["pad_token_id"] = config["eos_token_id"]
|
||||
|
||||
return config, input_ids, attention_mask, max_length
|
||||
|
||||
# special case for DoubleHeads model
|
||||
def _prepare_for_class(self, inputs_dict, model_class):
|
||||
inputs_dict = super()._prepare_for_class(inputs_dict, model_class)
|
||||
return inputs_dict
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = CodeGenModelTester(self)
|
||||
|
||||
def test_codegen_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_codegen_model(*config_and_inputs)
|
||||
|
||||
def test_codegen_model_past(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_codegen_model_past(*config_and_inputs)
|
||||
|
||||
def test_codegen_model_att_mask_past(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_codegen_model_attention_mask_past(*config_and_inputs)
|
||||
|
||||
def test_codegen_model_past_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_codegen_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_codegen_lm_head_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_lm_head_model(*config_and_inputs)
|
||||
|
||||
@slow
|
||||
def test_batch_generation(self):
|
||||
tokenizer = AutoTokenizer.from_pretrained("Salesforce/codegen-350M-mono")
|
||||
model = CodeGenForCausalLM.from_pretrained("Salesforce/codegen-350M-mono")
|
||||
model.eval()
|
||||
|
||||
tokenizer.padding_side = "left"
|
||||
|
||||
# Define PAD Token = EOS Token = 50256
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
model.transformer.config["pad_token_id"] = model.transformer.config["eos_token_id"]
|
||||
|
||||
# use different length sentences to test batching
|
||||
sentences = ["def hellow_world():", "def greet(name):"]
|
||||
|
||||
inputs = tokenizer(sentences, return_tensors="pd", padding=True, return_attention_mask=True)
|
||||
input_ids = inputs["input_ids"]
|
||||
|
||||
outputs, _ = model.generate(
|
||||
input_ids=input_ids,
|
||||
attention_mask=inputs["attention_mask"],
|
||||
)
|
||||
|
||||
inputs_non_padded = tokenizer(sentences[0], return_tensors="pd")["input_ids"]
|
||||
output_non_padded, _ = model.generate(input_ids=inputs_non_padded)
|
||||
|
||||
inputs_padded = tokenizer(sentences[1], return_tensors="pd")["input_ids"]
|
||||
output_padded, _ = model.generate(input_ids=inputs_padded)
|
||||
|
||||
# batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
||||
|
||||
non_padded_sentence = tokenizer.decode(output_non_padded[0], skip_special_tokens=True)
|
||||
padded_sentence = tokenizer.decode(output_padded[0], skip_special_tokens=True)
|
||||
|
||||
expected_output_sentence = [
|
||||
'\n print("Hello World")\n\nhellow_world()\n\n#',
|
||||
'\n print(f"Hello {name}")\n\ngreet("Rolf")\n',
|
||||
]
|
||||
# self.assertEqual(str(expected_output_sentence), str(batch_out_sentence))
|
||||
self.assertEqual(str(expected_output_sentence), str([non_padded_sentence, padded_sentence]))
|
||||
|
||||
@slow
|
||||
def test_model_from_pretrained(self):
|
||||
for model_name in CODEGEN_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
|
||||
model = CodeGenModel.from_pretrained(model_name)
|
||||
self.assertIsNotNone(model)
|
||||
|
||||
@unittest.skip("Not implemented")
|
||||
def test_model_name_list(self):
|
||||
pass
|
||||
|
||||
@slow
|
||||
def test_auto_tokenizer(self):
|
||||
for model_name in CODEGEN_PRETRAINED_MODEL_ARCHIVE_LIST:
|
||||
AutoTokenizer.from_pretrained(model_name) # assign a tokenizer but never use
|
||||
|
||||
|
||||
class CodeGenModelLanguageGenerationTest(unittest.TestCase):
|
||||
@slow
|
||||
def test_lm_generate_codegen(self):
|
||||
tokenizer = AutoTokenizer.from_pretrained("Salesforce/codegen-350M-mono")
|
||||
model = CodeGenForCausalLM.from_pretrained("Salesforce/codegen-350M-mono")
|
||||
model.eval()
|
||||
|
||||
inputs = tokenizer(
|
||||
"def hello_world():", return_tensors="pd", return_attention_mask=True, return_token_type_ids=False
|
||||
)
|
||||
expected_output = '\n print("Hello World")\n\nhello_world()\n\n#'
|
||||
|
||||
output_ids, _ = model.generate(**inputs, decode_strategy="sampling", top_k=1)
|
||||
output_str = tokenizer.batch_decode(output_ids)[0]
|
||||
|
||||
self.assertEqual(output_str, expected_output)
|
||||
|
||||
@slow
|
||||
def test_codegen_sample(self):
|
||||
tokenizer = AutoTokenizer.from_pretrained("Salesforce/codegen-350M-mono")
|
||||
model = CodeGenForCausalLM.from_pretrained("Salesforce/codegen-350M-mono")
|
||||
model.eval()
|
||||
|
||||
tokenized = tokenizer(
|
||||
"def hello_world():", return_tensors="pd", return_token_type_ids=True, return_attention_mask=True
|
||||
)
|
||||
input_ids = tokenized["input_ids"]
|
||||
output_ids, _ = model.generate(input_ids, decode_strategy="sampling", top_k=1)
|
||||
output_str = tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
||||
|
||||
token_type_ids = tokenized.token_type_ids
|
||||
output_seq, _ = model.generate(
|
||||
input_ids=input_ids, decode_strategy="sampling", top_k=1, num_return_sequences=5
|
||||
)
|
||||
output_seq_tt, _ = model.generate(
|
||||
input_ids=input_ids,
|
||||
token_type_ids=token_type_ids,
|
||||
decode_strategy="sampling",
|
||||
top_k=1,
|
||||
num_return_sequences=5,
|
||||
)
|
||||
output_seq_strs = tokenizer.batch_decode(output_seq, skip_special_tokens=True)
|
||||
output_seq_tt_strs = tokenizer.batch_decode(output_seq_tt, skip_special_tokens=True)
|
||||
|
||||
EXPECTED_OUTPUT_STR = '\n print("Hello World")\n\nhello_world()\n\n#'
|
||||
|
||||
self.assertEqual(output_str, EXPECTED_OUTPUT_STR)
|
||||
self.assertTrue(
|
||||
all([output_seq_strs[idx] != output_seq_tt_strs[idx] for idx in range(len(output_seq_tt_strs))])
|
||||
) # token_type_ids should change output
|
||||
@@ -0,0 +1,213 @@
|
||||
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
|
||||
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# 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.
|
||||
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import unittest
|
||||
|
||||
from paddlenlp.transformers import CodeGenTokenizer
|
||||
from paddlenlp.transformers.codegen.tokenizer import VOCAB_FILES_NAMES
|
||||
|
||||
from ...testing_utils import slow
|
||||
from ..test_tokenizer_common import TokenizerTesterMixin
|
||||
|
||||
|
||||
class CodeGenTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
||||
|
||||
tokenizer_class = CodeGenTokenizer
|
||||
from_pretrained_kwargs = {"add_prefix_space": True}
|
||||
test_seq2seq = False
|
||||
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
|
||||
# Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt
|
||||
vocab = [
|
||||
"l",
|
||||
"o",
|
||||
"w",
|
||||
"e",
|
||||
"r",
|
||||
"s",
|
||||
"t",
|
||||
"i",
|
||||
"d",
|
||||
"n",
|
||||
"\u0120",
|
||||
"\u0120l",
|
||||
"\u0120n",
|
||||
"\u0120lo",
|
||||
"\u0120low",
|
||||
"er",
|
||||
"\u0120lowest",
|
||||
"\u0120newer",
|
||||
"\u0120wider",
|
||||
"<unk>",
|
||||
"<|endoftext|>",
|
||||
]
|
||||
vocab_tokens = dict(zip(vocab, range(len(vocab))))
|
||||
merges = ["#version: 0.2", "\u0120 l", "\u0120l o", "\u0120lo w", "e r", ""]
|
||||
self.special_tokens_map = {"unk_token": "<unk>"}
|
||||
|
||||
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
|
||||
self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
|
||||
with open(self.vocab_file, "w", encoding="utf-8") as fp:
|
||||
fp.write(json.dumps(vocab_tokens) + "\n")
|
||||
with open(self.merges_file, "w", encoding="utf-8") as fp:
|
||||
fp.write("\n".join(merges))
|
||||
|
||||
def get_tokenizer(self, **kwargs):
|
||||
kwargs.update(self.special_tokens_map)
|
||||
return CodeGenTokenizer.from_pretrained(self.tmpdirname, **kwargs)
|
||||
|
||||
def get_input_output_texts(self, tokenizer):
|
||||
input_text = "lower newer"
|
||||
output_text = "lower newer"
|
||||
return input_text, output_text
|
||||
|
||||
def test_full_tokenizer(self):
|
||||
tokenizer = CodeGenTokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
|
||||
text = "lower newer"
|
||||
bpe_tokens = ["\u0120low", "er", "\u0120", "n", "e", "w", "er"]
|
||||
tokens = tokenizer.tokenize(text, add_prefix_space=True)
|
||||
self.assertListEqual(tokens, bpe_tokens)
|
||||
|
||||
input_tokens = tokens + [tokenizer.unk_token]
|
||||
input_bpe_tokens = [14, 15, 10, 9, 3, 2, 15, 19]
|
||||
self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
|
||||
|
||||
def test_pretokenized_inputs(self, *args, **kwargs):
|
||||
# It's very difficult to mix/test pretokenization with byte-level
|
||||
# And get both CodeGen and Roberta to work at the same time (mostly an issue of adding a space before the string)
|
||||
pass
|
||||
|
||||
def test_padding_if_pad_token_set_slow(self):
|
||||
tokenizer = CodeGenTokenizer.from_pretrained(self.tmpdirname, pad_token="<pad>")
|
||||
|
||||
# Simple input
|
||||
s = "This is a simple input"
|
||||
s2 = ["This is a simple input looooooooong", "This is a simple input"]
|
||||
p = ("This is a simple input", "This is a pair")
|
||||
p2 = [
|
||||
("This is a simple input loooooong", "This is a simple input"),
|
||||
("This is a simple pair loooooong", "This is a simple pair"),
|
||||
]
|
||||
|
||||
pad_token_id = tokenizer.pad_token_id
|
||||
|
||||
out_s = tokenizer(s, padding="max_length", max_length=30, return_tensors="np", return_attention_mask=True)
|
||||
out_s2 = tokenizer(s2, padding=True, truncate=True, return_tensors="np", return_attention_mask=True)
|
||||
out_p = tokenizer(*p, padding="max_length", max_length=60, return_tensors="np", return_attention_mask=True)
|
||||
out_p2 = tokenizer(p2, padding=True, truncate=True, return_tensors="np", return_attention_mask=True)
|
||||
|
||||
# s
|
||||
# test single string max_length padding
|
||||
self.assertEqual(out_s["input_ids"].shape[-1], 30)
|
||||
self.assertTrue(pad_token_id in out_s["input_ids"])
|
||||
self.assertTrue(0 in out_s["attention_mask"])
|
||||
|
||||
# s2
|
||||
# test automatic padding
|
||||
self.assertEqual(out_s2["input_ids"].shape[-1], 33)
|
||||
# long slice doesn't have padding
|
||||
self.assertFalse(pad_token_id in out_s2["input_ids"][0])
|
||||
self.assertFalse(0 in out_s2["attention_mask"][0])
|
||||
# short slice does have padding
|
||||
self.assertTrue(pad_token_id in out_s2["input_ids"][1])
|
||||
self.assertTrue(0 in out_s2["attention_mask"][1])
|
||||
|
||||
# p
|
||||
# test single pair max_length padding
|
||||
self.assertEqual(out_p["input_ids"].shape[-1], 60)
|
||||
self.assertTrue(pad_token_id in out_p["input_ids"])
|
||||
self.assertTrue(0 in out_p["attention_mask"])
|
||||
|
||||
# p2
|
||||
# test automatic padding pair
|
||||
self.assertEqual(out_p2["input_ids"].shape[-1], 52)
|
||||
# long slice pair doesn't have padding
|
||||
self.assertFalse(pad_token_id in out_p2["input_ids"][0])
|
||||
self.assertFalse(0 in out_p2["attention_mask"][0])
|
||||
# short slice pair does have padding
|
||||
self.assertTrue(pad_token_id in out_p2["input_ids"][1])
|
||||
self.assertTrue(0 in out_p2["attention_mask"][1])
|
||||
|
||||
def test_add_bos_token_slow(self):
|
||||
bos_token = "$$$"
|
||||
tokenizer = CodeGenTokenizer.from_pretrained(self.tmpdirname, bos_token=bos_token, add_bos_token=True)
|
||||
|
||||
s = "This is a simple input"
|
||||
s2 = ["This is a simple input 1", "This is a simple input 2"]
|
||||
|
||||
bos_token_id = tokenizer.bos_token_id
|
||||
|
||||
out_s = tokenizer(s)
|
||||
out_s2 = tokenizer(s2)
|
||||
|
||||
self.assertEqual(out_s.input_ids[0], bos_token_id)
|
||||
self.assertTrue(all(o[0] == bos_token_id for o in out_s2.input_ids))
|
||||
|
||||
decode_s = tokenizer.decode(out_s.input_ids)
|
||||
decode_s2 = tokenizer.batch_decode(out_s2.input_ids)
|
||||
|
||||
self.assertEqual(decode_s.split()[0], bos_token)
|
||||
self.assertTrue(all(d.split()[0] == bos_token for d in decode_s2))
|
||||
|
||||
@slow
|
||||
def test_truncation(self):
|
||||
tokenizer = CodeGenTokenizer.from_pretrained("Salesforce/codegen-350M-mono")
|
||||
|
||||
text = "\nif len_a > len_b:\n result = a\nelse:\n result = b\n\n\n\n#"
|
||||
expected_trucated_text = "\nif len_a > len_b: result = a\nelse: result = b"
|
||||
|
||||
input_ids = tokenizer.encode(text)["input_ids"]
|
||||
truncation_pattern = ["^#", re.escape("<|endoftext|>"), "^'''", '^"""', "\n\n\n"]
|
||||
decoded_text = tokenizer.decode(input_ids, truncate_before_pattern=truncation_pattern)
|
||||
self.assertEqual(decoded_text, expected_trucated_text)
|
||||
|
||||
# tokenizer has no padding token
|
||||
def test_padding_different_model_input_name(self):
|
||||
pass
|
||||
|
||||
def test_pretrained_model_lists(self):
|
||||
# We should have at least one default checkpoint for each tokenizer
|
||||
# We should specify the max input length as well (used in some part to list the pretrained checkpoints)
|
||||
self.assertGreaterEqual(len(self.tokenizer_class.pretrained_resource_files_map), 1)
|
||||
self.assertEqual(
|
||||
len(list(self.tokenizer_class.pretrained_resource_files_map.values())[0]),
|
||||
len(self.tokenizer_class.max_model_input_sizes),
|
||||
)
|
||||
|
||||
weights_list = list(self.tokenizer_class.max_model_input_sizes.keys())
|
||||
weights_lists_2 = []
|
||||
for file_id, map_list in self.tokenizer_class.pretrained_resource_files_map.items():
|
||||
weights_lists_2.append(list(map_list.keys()))
|
||||
|
||||
for weights_list_2 in weights_lists_2:
|
||||
self.assertListEqual(weights_list, weights_list_2)
|
||||
|
||||
def test_consecutive_unk_string(self):
|
||||
tokenizers = self.get_tokenizers(fast=True, do_lower_case=True)
|
||||
for tokenizer in tokenizers:
|
||||
tokens = [tokenizer.unk_token for _ in range(2)]
|
||||
string = tokenizer.convert_tokens_to_string(tokens)
|
||||
encoding = tokenizer(
|
||||
text=string,
|
||||
runcation=True,
|
||||
return_offsets_mapping=True,
|
||||
)
|
||||
self.assertEqual(len(encoding["input_ids"]), 2)
|
||||
self.assertEqual(len(encoding["offset_mapping"]), 2)
|
||||
Reference in New Issue
Block a user