Spaces:
Running
Running
TedYeh
commited on
Commit
·
da060de
1
Parent(s):
2a548f2
add t5 package
Browse files- app.py +8 -6
- t5/__init__.py +9 -0
- t5/__pycache__/__init__.cpython-38.pyc +0 -0
- t5/__pycache__/copyt5_model.cpython-38.pyc +0 -0
- t5/__pycache__/copyt5_utils.cpython-38.pyc +0 -0
- t5/__pycache__/t5_model.cpython-38.pyc +0 -0
- t5/__pycache__/t5_utils.cpython-38.pyc +0 -0
- t5/config/__init__.py +5 -0
- t5/config/__pycache__/__init__.cpython-38.pyc +0 -0
- t5/config/__pycache__/model_args.cpython-38.pyc +0 -0
- t5/config/global_args.py +62 -0
- t5/config/model_args.py +464 -0
- t5/t5_model.py +1256 -0
- t5/t5_utils.py +214 -0
app.py
CHANGED
|
@@ -1,15 +1,17 @@
|
|
| 1 |
import gradio as gr
|
|
|
|
| 2 |
from transformers import AutoTokenizer, T5ForConditionalGeneration
|
| 3 |
-
tokenizer = AutoTokenizer.from_pretrained("CodeTed/CGEDit")
|
| 4 |
-
model = T5ForConditionalGeneration.from_pretrained("CodeTed/CGEDit")
|
|
|
|
| 5 |
|
| 6 |
def cged_correction(sentence, function):
|
| 7 |
prompt = {"錯別字校正":"糾正句子中的錯字:", "文法校正":"糾正句子中的錯誤:",
|
| 8 |
"文本重構":"在不改動文意的情況下改寫句子:", "文本簡化":"在不改動文意的情況下改寫句子:", "整體校正":"修改句子的錯誤或使其更通順:"}
|
| 9 |
-
input_ids = tokenizer(prompt[function] + sentence, return_tensors="pt").input_ids
|
| 10 |
-
outputs = model.
|
| 11 |
-
edited_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 12 |
-
return
|
| 13 |
|
| 14 |
with gr.Blocks() as demo:
|
| 15 |
gr.Markdown(
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
+
from t5.t5_model import T5Model
|
| 3 |
from transformers import AutoTokenizer, T5ForConditionalGeneration
|
| 4 |
+
#tokenizer = AutoTokenizer.from_pretrained("CodeTed/CGEDit")
|
| 5 |
+
#model = T5ForConditionalGeneration.from_pretrained("CodeTed/CGEDit")
|
| 6 |
+
model = T5Model('t5', "CodeTed/CGEDit", args={"eval_batch_size": 1}, cuda_device=-1, evaluate=True)
|
| 7 |
|
| 8 |
def cged_correction(sentence, function):
|
| 9 |
prompt = {"錯別字校正":"糾正句子中的錯字:", "文法校正":"糾正句子中的錯誤:",
|
| 10 |
"文本重構":"在不改動文意的情況下改寫句子:", "文本簡化":"在不改動文意的情況下改寫句子:", "整體校正":"修改句子的錯誤或使其更通順:"}
|
| 11 |
+
#input_ids = tokenizer(prompt[function] + sentence, return_tensors="pt").input_ids
|
| 12 |
+
outputs = model.predict([prompt[function] + sentence + "_輸出句:"])
|
| 13 |
+
#edited_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 14 |
+
return outputs[0]
|
| 15 |
|
| 16 |
with gr.Blocks() as demo:
|
| 17 |
gr.Markdown(
|
t5/__init__.py
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""
|
| 3 |
+
@author:XuMing([email protected])
|
| 4 |
+
@description:
|
| 5 |
+
"""
|
| 6 |
+
from textgen.config.model_args import T5Args, CopyT5Args
|
| 7 |
+
from textgen.t5.t5_model import T5Model
|
| 8 |
+
from textgen.t5.copyt5_model import CopyT5Model
|
| 9 |
+
from textgen.t5.copyt5_utils import ZHTokenizer
|
t5/__pycache__/__init__.cpython-38.pyc
ADDED
|
Binary file (448 Bytes). View file
|
|
|
t5/__pycache__/copyt5_model.cpython-38.pyc
ADDED
|
Binary file (28.5 kB). View file
|
|
|
t5/__pycache__/copyt5_utils.cpython-38.pyc
ADDED
|
Binary file (6.18 kB). View file
|
|
|
t5/__pycache__/t5_model.cpython-38.pyc
ADDED
|
Binary file (28.4 kB). View file
|
|
|
t5/__pycache__/t5_utils.cpython-38.pyc
ADDED
|
Binary file (5.93 kB). View file
|
|
|
t5/config/__init__.py
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""
|
| 3 |
+
@author:XuMing([email protected])
|
| 4 |
+
@description: refer https://github.com/ThilinaRajapakse/simpletransformers
|
| 5 |
+
"""
|
t5/config/__pycache__/__init__.cpython-38.pyc
ADDED
|
Binary file (265 Bytes). View file
|
|
|
t5/config/__pycache__/model_args.cpython-38.pyc
ADDED
|
Binary file (15.7 kB). View file
|
|
|
t5/config/global_args.py
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""
|
| 3 |
+
@author:XuMing([email protected])
|
| 4 |
+
@description: refer https://github.com/ThilinaRajapakse/simpletransformers
|
| 5 |
+
"""
|
| 6 |
+
import sys
|
| 7 |
+
from multiprocessing import cpu_count
|
| 8 |
+
|
| 9 |
+
global_args = {
|
| 10 |
+
"adam_epsilon": 1e-8,
|
| 11 |
+
"best_model_dir": "outputs/best_model",
|
| 12 |
+
"cache_dir": "cache_dir/",
|
| 13 |
+
"config": {},
|
| 14 |
+
"do_lower_case": False,
|
| 15 |
+
"early_stopping_consider_epochs": False,
|
| 16 |
+
"early_stopping_delta": 0,
|
| 17 |
+
"early_stopping_metric": "eval_loss",
|
| 18 |
+
"early_stopping_metric_minimize": True,
|
| 19 |
+
"early_stopping_patience": 3,
|
| 20 |
+
"encoding": None,
|
| 21 |
+
"eval_batch_size": 8,
|
| 22 |
+
"evaluate_during_training": False,
|
| 23 |
+
"evaluate_during_training_silent": True,
|
| 24 |
+
"evaluate_during_training_steps": 2000,
|
| 25 |
+
"evaluate_during_training_verbose": False,
|
| 26 |
+
"fp16": True,
|
| 27 |
+
"gradient_accumulation_steps": 1,
|
| 28 |
+
"learning_rate": 4e-5,
|
| 29 |
+
"local_rank": -1,
|
| 30 |
+
"logging_steps": 50,
|
| 31 |
+
"manual_seed": None,
|
| 32 |
+
"max_grad_norm": 1.0,
|
| 33 |
+
"max_seq_length": 128,
|
| 34 |
+
"multiprocessing_chunksize": 500,
|
| 35 |
+
"n_gpu": 1,
|
| 36 |
+
"no_cache": False,
|
| 37 |
+
"no_save": False,
|
| 38 |
+
"num_train_epochs": 1,
|
| 39 |
+
"output_dir": "outputs/",
|
| 40 |
+
"overwrite_output_dir": False,
|
| 41 |
+
"process_count": cpu_count() - 2 if cpu_count() > 2 else 1,
|
| 42 |
+
"reprocess_input_data": True,
|
| 43 |
+
"save_best_model": True,
|
| 44 |
+
"save_eval_checkpoints": True,
|
| 45 |
+
"save_model_every_epoch": True,
|
| 46 |
+
"save_steps": 2000,
|
| 47 |
+
"save_optimizer_and_scheduler": True,
|
| 48 |
+
"silent": False,
|
| 49 |
+
"tensorboard_dir": None,
|
| 50 |
+
"train_batch_size": 8,
|
| 51 |
+
"use_cached_eval_features": False,
|
| 52 |
+
"use_early_stopping": False,
|
| 53 |
+
"use_multiprocessing": False,
|
| 54 |
+
"wandb_kwargs": {},
|
| 55 |
+
"wandb_project": None,
|
| 56 |
+
"warmup_ratio": 0.06,
|
| 57 |
+
"warmup_steps": 0,
|
| 58 |
+
"weight_decay": 0,
|
| 59 |
+
}
|
| 60 |
+
|
| 61 |
+
if sys.platform == "win32":
|
| 62 |
+
global_args["process_count"] = min(global_args["process_count"], 61)
|
t5/config/model_args.py
ADDED
|
@@ -0,0 +1,464 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""
|
| 3 |
+
@author:XuMing([email protected])
|
| 4 |
+
@description: refer https://github.com/ThilinaRajapakse/simpletransformers
|
| 5 |
+
"""
|
| 6 |
+
import json
|
| 7 |
+
import os
|
| 8 |
+
import sys
|
| 9 |
+
from dataclasses import asdict, dataclass, field
|
| 10 |
+
from multiprocessing import cpu_count
|
| 11 |
+
from typing import Optional
|
| 12 |
+
|
| 13 |
+
from loguru import logger
|
| 14 |
+
from torch.utils.data import Dataset
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
def get_default_process_count():
|
| 18 |
+
process_count = cpu_count() - 2 if cpu_count() > 2 else 1
|
| 19 |
+
if sys.platform == "win32":
|
| 20 |
+
process_count = min(process_count, 61)
|
| 21 |
+
|
| 22 |
+
return process_count
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
def get_special_tokens():
|
| 26 |
+
return ["<s>", "<pad>", "</s>", "<unk>", "<mask>"]
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
@dataclass
|
| 30 |
+
class ModelArgs:
|
| 31 |
+
adafactor_beta1: float = None
|
| 32 |
+
adafactor_clip_threshold: float = 1.0
|
| 33 |
+
adafactor_decay_rate: float = -0.8
|
| 34 |
+
adafactor_eps: tuple = field(default_factory=lambda: (1e-30, 1e-3))
|
| 35 |
+
adafactor_relative_step: bool = True
|
| 36 |
+
adafactor_scale_parameter: bool = True
|
| 37 |
+
adafactor_warmup_init: bool = True
|
| 38 |
+
adam_epsilon: float = 1e-8
|
| 39 |
+
best_model_dir: str = "outputs/best_model"
|
| 40 |
+
cache_dir: str = "cache_dir/"
|
| 41 |
+
config: dict = field(default_factory=dict)
|
| 42 |
+
cosine_schedule_num_cycles: float = 0.5
|
| 43 |
+
custom_layer_parameters: list = field(default_factory=list)
|
| 44 |
+
custom_parameter_groups: list = field(default_factory=list)
|
| 45 |
+
dataloader_num_workers: int = 0
|
| 46 |
+
do_lower_case: bool = False
|
| 47 |
+
dynamic_quantize: bool = False
|
| 48 |
+
early_stopping_consider_epochs: bool = False
|
| 49 |
+
early_stopping_delta: float = 0
|
| 50 |
+
early_stopping_metric: str = "eval_loss"
|
| 51 |
+
early_stopping_metric_minimize: bool = True
|
| 52 |
+
early_stopping_patience: int = 3
|
| 53 |
+
encoding: str = "utf-8"
|
| 54 |
+
eval_batch_size: int = 8
|
| 55 |
+
evaluate_during_training: bool = False
|
| 56 |
+
evaluate_during_training_silent: bool = True
|
| 57 |
+
evaluate_during_training_steps: int = 6000
|
| 58 |
+
evaluate_during_training_verbose: bool = False
|
| 59 |
+
evaluate_each_epoch: bool = True
|
| 60 |
+
fp16: bool = False
|
| 61 |
+
gradient_accumulation_steps: int = 1
|
| 62 |
+
learning_rate: float = 2e-5
|
| 63 |
+
local_rank: int = -1
|
| 64 |
+
logging_steps: int = 50
|
| 65 |
+
manual_seed: int = None
|
| 66 |
+
max_grad_norm: float = 1.0
|
| 67 |
+
max_seq_length: int = 128 # max length of input sequence
|
| 68 |
+
model_name: str = None
|
| 69 |
+
model_type: str = None
|
| 70 |
+
multiprocessing_chunksize: int = -1
|
| 71 |
+
n_gpu: int = 2
|
| 72 |
+
no_cache: bool = False
|
| 73 |
+
no_save: bool = False
|
| 74 |
+
not_saved_args: list = field(default_factory=list)
|
| 75 |
+
num_train_epochs: int = 1
|
| 76 |
+
optimizer: str = "AdamW"
|
| 77 |
+
output_dir: str = "outputs/"
|
| 78 |
+
overwrite_output_dir: bool = True
|
| 79 |
+
polynomial_decay_schedule_lr_end: float = 1e-7
|
| 80 |
+
polynomial_decay_schedule_power: float = 1.0
|
| 81 |
+
process_count: int = field(default_factory=get_default_process_count)
|
| 82 |
+
quantized_model: bool = False
|
| 83 |
+
reprocess_input_data: bool = False
|
| 84 |
+
save_best_model: bool = True
|
| 85 |
+
save_eval_checkpoints: bool = True
|
| 86 |
+
save_model_every_epoch: bool = False
|
| 87 |
+
save_optimizer_and_scheduler: bool = True
|
| 88 |
+
save_steps: int = 10000
|
| 89 |
+
scheduler: str = "linear_schedule_with_warmup"
|
| 90 |
+
silent: bool = False
|
| 91 |
+
skip_special_tokens: bool = True
|
| 92 |
+
tensorboard_dir: str = None
|
| 93 |
+
thread_count: int = None
|
| 94 |
+
tokenizer_name: str = None
|
| 95 |
+
tokenizer_type: str = None
|
| 96 |
+
train_batch_size: int = 8
|
| 97 |
+
train_custom_parameters_only: bool = False
|
| 98 |
+
use_cached_eval_features: bool = False
|
| 99 |
+
use_early_stopping: bool = False
|
| 100 |
+
use_hf_datasets: bool = False
|
| 101 |
+
use_multiprocessing: bool = True
|
| 102 |
+
use_multiprocessing_for_evaluation: bool = True
|
| 103 |
+
wandb_kwargs: dict = field(default_factory=dict)
|
| 104 |
+
wandb_project: str = None
|
| 105 |
+
warmup_ratio: float = 0.06
|
| 106 |
+
warmup_steps: int = 0
|
| 107 |
+
weight_decay: float = 0.0
|
| 108 |
+
|
| 109 |
+
def update_from_dict(self, new_values):
|
| 110 |
+
if isinstance(new_values, dict):
|
| 111 |
+
for key, value in new_values.items():
|
| 112 |
+
setattr(self, key, value)
|
| 113 |
+
else:
|
| 114 |
+
raise (TypeError(f"{new_values} is not a Python dict."))
|
| 115 |
+
|
| 116 |
+
def get_args_for_saving(self):
|
| 117 |
+
args_for_saving = {key: value for key, value in asdict(self).items() if key not in self.not_saved_args}
|
| 118 |
+
return args_for_saving
|
| 119 |
+
|
| 120 |
+
def save(self, output_dir):
|
| 121 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 122 |
+
with open(os.path.join(output_dir, "model_args.json"), "w", encoding='utf-8') as f:
|
| 123 |
+
args_dict = self.get_args_for_saving()
|
| 124 |
+
if args_dict['dataset_class'] is not None and not isinstance(args_dict["dataset_class"], str):
|
| 125 |
+
args_dict['dataset_class'] = type(args_dict['dataset_class']).__name__
|
| 126 |
+
if args_dict["tokenizer_type"] is not None and not isinstance(args_dict["tokenizer_type"], str):
|
| 127 |
+
args_dict["tokenizer_type"] = type(args_dict["tokenizer_type"]).__name__
|
| 128 |
+
json.dump(args_dict, f)
|
| 129 |
+
|
| 130 |
+
def load(self, input_dir):
|
| 131 |
+
if input_dir:
|
| 132 |
+
model_args_file = os.path.join(input_dir, "model_args.json")
|
| 133 |
+
if os.path.isfile(model_args_file):
|
| 134 |
+
with open(model_args_file, "r", encoding='utf-8') as f:
|
| 135 |
+
model_args = json.load(f)
|
| 136 |
+
if model_args["dataset_class"]:
|
| 137 |
+
logger.warning(
|
| 138 |
+
"This model was trained using a custom dataset_class."
|
| 139 |
+
"This cannot be loaded automatically and must be specified in the model args"
|
| 140 |
+
"when loading the model."
|
| 141 |
+
)
|
| 142 |
+
self.update_from_dict(model_args)
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
@dataclass
|
| 146 |
+
class T5Args(ModelArgs):
|
| 147 |
+
"""
|
| 148 |
+
Model args for a T5Model
|
| 149 |
+
"""
|
| 150 |
+
|
| 151 |
+
model_class: str = "T5Model"
|
| 152 |
+
dataset_class: Dataset = None
|
| 153 |
+
do_sample: bool = False
|
| 154 |
+
early_stopping: bool = True
|
| 155 |
+
evaluate_generated_text: bool = False
|
| 156 |
+
length_penalty: float = 2.0
|
| 157 |
+
max_length: int = 180 # max length of the sequence to be generated
|
| 158 |
+
max_steps: int = -1
|
| 159 |
+
num_beams: int = 1
|
| 160 |
+
num_return_sequences: int = 1
|
| 161 |
+
preprocess_inputs: bool = True
|
| 162 |
+
repetition_penalty: float = 1.0
|
| 163 |
+
scheduler: str = "constant_schedule_with_warmup"
|
| 164 |
+
adafactor_relative_step: bool = False
|
| 165 |
+
adafactor_scale_parameter: bool = False
|
| 166 |
+
adafactor_warmup_init: bool = False
|
| 167 |
+
learning_rate: float = 5e-4
|
| 168 |
+
optimizer: str = "AdamW"
|
| 169 |
+
special_tokens_list: list = field(default_factory=list)
|
| 170 |
+
top_k: float = None
|
| 171 |
+
top_p: float = None
|
| 172 |
+
use_multiprocessed_decoding: bool = False
|
| 173 |
+
|
| 174 |
+
|
| 175 |
+
@dataclass
|
| 176 |
+
class CopyT5Args(ModelArgs):
|
| 177 |
+
"""
|
| 178 |
+
Model args for a CopyT5Model
|
| 179 |
+
"""
|
| 180 |
+
|
| 181 |
+
model_class: str = "CopyT5Model"
|
| 182 |
+
dataset_class: Dataset = None
|
| 183 |
+
do_sample: bool = False
|
| 184 |
+
early_stopping: bool = True
|
| 185 |
+
evaluate_generated_text: bool = False
|
| 186 |
+
length_penalty: float = 2.0
|
| 187 |
+
max_length: int = 128 # max length of the sequence to be generated
|
| 188 |
+
max_steps: int = -1
|
| 189 |
+
num_beams: int = 3
|
| 190 |
+
num_return_sequences: int = 1
|
| 191 |
+
preprocess_inputs: bool = True
|
| 192 |
+
repetition_penalty: float = 1.0
|
| 193 |
+
scheduler: str = "linear_schedule_with_warmup"
|
| 194 |
+
adafactor_relative_step: bool = False
|
| 195 |
+
adafactor_scale_parameter: bool = False
|
| 196 |
+
adafactor_warmup_init: bool = False
|
| 197 |
+
learning_rate: float = 1e-3
|
| 198 |
+
optimizer: str = "AdamW"
|
| 199 |
+
special_tokens_list: list = field(default_factory=list)
|
| 200 |
+
top_k: float = None
|
| 201 |
+
top_p: float = None
|
| 202 |
+
use_multiprocessed_decoding: bool = False
|
| 203 |
+
|
| 204 |
+
|
| 205 |
+
@dataclass
|
| 206 |
+
class LanguageModelingArgs(ModelArgs):
|
| 207 |
+
"""
|
| 208 |
+
Model args for a LanguageModelingModel
|
| 209 |
+
"""
|
| 210 |
+
|
| 211 |
+
model_class: str = "LanguageModelingModel"
|
| 212 |
+
block_size: int = -1
|
| 213 |
+
config_name: str = None
|
| 214 |
+
dataset_class: Dataset = None
|
| 215 |
+
dataset_type: str = "None"
|
| 216 |
+
discriminator_config: dict = field(default_factory=dict)
|
| 217 |
+
discriminator_loss_weight: float = 50.0
|
| 218 |
+
generator_config: dict = field(default_factory=dict)
|
| 219 |
+
max_steps: int = -1
|
| 220 |
+
min_frequency: int = 2
|
| 221 |
+
mlm: bool = True
|
| 222 |
+
mlm_probability: float = 0.15
|
| 223 |
+
sliding_window: bool = False
|
| 224 |
+
special_tokens: list = field(default_factory=get_special_tokens)
|
| 225 |
+
stride: float = 0.8
|
| 226 |
+
tie_generator_and_discriminator_embeddings: bool = True
|
| 227 |
+
tokenizer_name: str = None
|
| 228 |
+
vocab_size: int = None
|
| 229 |
+
clean_text: bool = True
|
| 230 |
+
handle_chinese_chars: bool = True
|
| 231 |
+
special_tokens_list: list = field(default_factory=list)
|
| 232 |
+
strip_accents: bool = True
|
| 233 |
+
local_rank: int = -1
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
@dataclass
|
| 237 |
+
class Seq2SeqArgs(ModelArgs):
|
| 238 |
+
"""
|
| 239 |
+
Model args for a Seq2SeqModel
|
| 240 |
+
"""
|
| 241 |
+
|
| 242 |
+
model_class: str = "Seq2SeqModel"
|
| 243 |
+
base_marian_model_name: str = None
|
| 244 |
+
dataset_class: Dataset = None
|
| 245 |
+
do_sample: bool = False
|
| 246 |
+
early_stopping: bool = True
|
| 247 |
+
evaluate_generated_text: bool = False
|
| 248 |
+
faiss_d: int = 768
|
| 249 |
+
faiss_m: int = 128
|
| 250 |
+
length_penalty: float = 2.0
|
| 251 |
+
max_length: int = 128 # max length of the sequence to be generated
|
| 252 |
+
max_steps: int = -1
|
| 253 |
+
num_beams: int = 1
|
| 254 |
+
num_return_sequences: int = 1
|
| 255 |
+
rag_embed_batch_size: int = 16
|
| 256 |
+
repetition_penalty: float = 1.0
|
| 257 |
+
top_k: float = None
|
| 258 |
+
top_p: float = None
|
| 259 |
+
use_multiprocessed_decoding: bool = False
|
| 260 |
+
save_knowledge_dataset: bool = True
|
| 261 |
+
save_knowledge_dataset_with_checkpoints: bool = False
|
| 262 |
+
split_text_character: str = " "
|
| 263 |
+
split_text_n: int = 100
|
| 264 |
+
src_lang: str = "en_XX"
|
| 265 |
+
tgt_lang: str = "ro_RO"
|
| 266 |
+
|
| 267 |
+
|
| 268 |
+
@dataclass
|
| 269 |
+
class LanguageGenerationArgs(ModelArgs):
|
| 270 |
+
"""
|
| 271 |
+
Model args for a LanguageGenerationModel
|
| 272 |
+
"""
|
| 273 |
+
|
| 274 |
+
model_class: str = "LanguageGenerationModel"
|
| 275 |
+
do_sample: bool = True
|
| 276 |
+
early_stopping: bool = True
|
| 277 |
+
evaluate_generated_text: bool = False
|
| 278 |
+
length_penalty: float = 2.0
|
| 279 |
+
max_length: int = 128 # max length of the sequence to be generated
|
| 280 |
+
max_steps: int = -1
|
| 281 |
+
num_beams: int = 1
|
| 282 |
+
num_return_sequences: int = 1
|
| 283 |
+
repetition_penalty: float = 1.0
|
| 284 |
+
top_k: float = 50
|
| 285 |
+
top_p: float = 0.95
|
| 286 |
+
prompt: str = ""
|
| 287 |
+
stop_token: str = None
|
| 288 |
+
temperature: float = 1.0
|
| 289 |
+
padding_text: str = ""
|
| 290 |
+
xlm_language: str = ""
|
| 291 |
+
config_name: str = None
|
| 292 |
+
tokenizer_name: str = None
|
| 293 |
+
special_tokens_list: list = field(default_factory=list)
|
| 294 |
+
|
| 295 |
+
|
| 296 |
+
@dataclass
|
| 297 |
+
class SongNetArgs(LanguageModelingArgs):
|
| 298 |
+
"""
|
| 299 |
+
Model args for a SongNetModel
|
| 300 |
+
"""
|
| 301 |
+
|
| 302 |
+
model_class: str = "SongNetModel"
|
| 303 |
+
dataset_class: Dataset = None
|
| 304 |
+
do_sample: bool = False
|
| 305 |
+
early_stopping: bool = True
|
| 306 |
+
evaluate_generated_text: bool = False
|
| 307 |
+
length_penalty: float = 2.0
|
| 308 |
+
max_length: int = 128
|
| 309 |
+
min_length: int = 10
|
| 310 |
+
max_steps: int = -1
|
| 311 |
+
num_beams: int = 3
|
| 312 |
+
num_return_sequences: int = 1
|
| 313 |
+
repetition_penalty: float = 1.0
|
| 314 |
+
scheduler: str = None
|
| 315 |
+
adafactor_relative_step: bool = False
|
| 316 |
+
adafactor_scale_parameter: bool = False
|
| 317 |
+
adafactor_warmup_init: bool = False
|
| 318 |
+
learning_rate: float = 1e-3
|
| 319 |
+
early_stopping_metric: str = "eval_ppl"
|
| 320 |
+
special_tokens_list: list = field(default_factory=list)
|
| 321 |
+
save_eval_checkpoints: bool = False
|
| 322 |
+
skip_special_tokens: bool = False
|
| 323 |
+
k: int = 16
|
| 324 |
+
use_multiprocessed_decoding: bool = False
|
| 325 |
+
embed_dim: int = 768
|
| 326 |
+
ff_embed_dim: int = 3072
|
| 327 |
+
num_heads: int = 12
|
| 328 |
+
num_layers: int = 12
|
| 329 |
+
dropout: float = 0.2
|
| 330 |
+
warmup_ratio: float = 0.05
|
| 331 |
+
weight_decay: float = 0.0
|
| 332 |
+
smoothing_factor: float = 0.1
|
| 333 |
+
|
| 334 |
+
|
| 335 |
+
@dataclass
|
| 336 |
+
class ChatGlmArgs(ModelArgs):
|
| 337 |
+
"""
|
| 338 |
+
Model args for a ChatGLMModel
|
| 339 |
+
"""
|
| 340 |
+
|
| 341 |
+
model_class: str = "ChatGlmArgs"
|
| 342 |
+
dataset_class: Dataset = None
|
| 343 |
+
learning_rate: float = 2e-5
|
| 344 |
+
fp16: bool = True
|
| 345 |
+
bf16: bool = False
|
| 346 |
+
int8: bool = False
|
| 347 |
+
int4: bool = False
|
| 348 |
+
debug: bool = False
|
| 349 |
+
max_seq_length: int = 256 # max length of input sequence
|
| 350 |
+
max_length = 384 # max length of the sequence to be generated
|
| 351 |
+
do_sample: bool = True
|
| 352 |
+
early_stopping: bool = True
|
| 353 |
+
is_train_on_prompt: bool = False # if compute loss with prompt labels
|
| 354 |
+
evaluate_generated_text: bool = True
|
| 355 |
+
report_to = "tensorboard"
|
| 356 |
+
optimizer: str = "adamw_torch"
|
| 357 |
+
save_strategy: str = "steps"
|
| 358 |
+
evaluation_strategy: str = "no"
|
| 359 |
+
eval_steps: int = 50
|
| 360 |
+
save_steps: int = 400
|
| 361 |
+
max_eval_samples: int = 20
|
| 362 |
+
length_penalty: float = 2.0
|
| 363 |
+
num_beams: int = 4
|
| 364 |
+
num_return_sequences: int = 1
|
| 365 |
+
repetition_penalty: float = 1.0
|
| 366 |
+
temperature: float = 0.1
|
| 367 |
+
special_tokens_list: list = field(default_factory=list)
|
| 368 |
+
top_k: float = 40
|
| 369 |
+
top_p: float = 0.75
|
| 370 |
+
model_name_or_path: Optional[str] = field(default="THUDM/chatglm-6b")
|
| 371 |
+
use_peft: bool = True
|
| 372 |
+
peft_type: str = "LORA"
|
| 373 |
+
peft_bin_name: str = "adapter_model.bin"
|
| 374 |
+
lora_r: int = 8
|
| 375 |
+
lora_alpha = 32
|
| 376 |
+
lora_dropout = 0.05
|
| 377 |
+
lora_target_modules = ["all"] # ["all"] or ["query_key_value"]
|
| 378 |
+
lora_bias = "none"
|
| 379 |
+
adalora_init_r: int = 12
|
| 380 |
+
adalora_tinit: int = 200
|
| 381 |
+
adalora_tfinal: int = 1000
|
| 382 |
+
adalora_delta_t: int = 10
|
| 383 |
+
lora_beta: float = 0.85
|
| 384 |
+
num_virtual_tokens: int = 20
|
| 385 |
+
prompt_encoder_hidden_size: int = 128
|
| 386 |
+
num_train_epochs = 1
|
| 387 |
+
max_steps = -1
|
| 388 |
+
per_device_train_batch_size = 2
|
| 389 |
+
eval_batch_size: int = 4
|
| 390 |
+
gradient_accumulation_steps = 1
|
| 391 |
+
gradient_checkpointing: bool = True
|
| 392 |
+
torch_compile: bool = False
|
| 393 |
+
save_total_limit = 10
|
| 394 |
+
remove_unused_columns = False
|
| 395 |
+
logging_steps = 50
|
| 396 |
+
resume_from_checkpoint: str = None
|
| 397 |
+
qlora: bool = False
|
| 398 |
+
|
| 399 |
+
|
| 400 |
+
@dataclass
|
| 401 |
+
class GptArgs(ModelArgs):
|
| 402 |
+
"""
|
| 403 |
+
Model args for a GptModel
|
| 404 |
+
"""
|
| 405 |
+
|
| 406 |
+
model_class: str = "GptArgs"
|
| 407 |
+
dataset_class: Dataset = None
|
| 408 |
+
learning_rate: float = 2e-5
|
| 409 |
+
fp16: bool = True
|
| 410 |
+
bf16: bool = False
|
| 411 |
+
int8: bool = False
|
| 412 |
+
int4: bool = False
|
| 413 |
+
debug: bool = False
|
| 414 |
+
max_seq_length: int = 256 # max length of input sequence
|
| 415 |
+
max_length = 256 # max length of the sequence to be generated
|
| 416 |
+
do_sample: bool = True
|
| 417 |
+
early_stopping: bool = True
|
| 418 |
+
evaluate_generated_text: bool = True
|
| 419 |
+
is_train_on_prompt: bool = False # if compute loss with prompt labels
|
| 420 |
+
warmup_steps: int = 50
|
| 421 |
+
report_to = "tensorboard"
|
| 422 |
+
optimizer: str = "adamw_torch"
|
| 423 |
+
save_strategy: str = "steps"
|
| 424 |
+
eval_steps: int = 200
|
| 425 |
+
save_steps: int = 400
|
| 426 |
+
pad_to_multiple_of: int = 8
|
| 427 |
+
max_eval_samples: int = 20
|
| 428 |
+
length_penalty: float = 2.0
|
| 429 |
+
num_beams: int = 1
|
| 430 |
+
num_return_sequences: int = 1
|
| 431 |
+
repetition_penalty: float = 1.3
|
| 432 |
+
temperature: float = 0.4
|
| 433 |
+
special_tokens_list: list = field(default_factory=list)
|
| 434 |
+
top_k: float = 40
|
| 435 |
+
top_p: float = 0.9
|
| 436 |
+
model_name_or_path: Optional[str] = field(default="shibing624/chinese-alpaca-plus-7b-hf")
|
| 437 |
+
use_peft: bool = True
|
| 438 |
+
peft_type: str = "LORA"
|
| 439 |
+
peft_bin_name: str = "adapter_model.bin"
|
| 440 |
+
lora_r: int = 8
|
| 441 |
+
lora_alpha = 16
|
| 442 |
+
lora_dropout = 0.05
|
| 443 |
+
lora_target_modules = ["all"] # ["all"] or ["k_proj"]
|
| 444 |
+
lora_bias = "none"
|
| 445 |
+
adalora_init_r: int = 12
|
| 446 |
+
adalora_tinit: int = 200
|
| 447 |
+
adalora_tfinal: int = 1000
|
| 448 |
+
adalora_delta_t: int = 10
|
| 449 |
+
lora_beta: float = 0.85
|
| 450 |
+
num_virtual_tokens: int = 20
|
| 451 |
+
prompt_encoder_hidden_size: int = 128
|
| 452 |
+
num_train_epochs = 3
|
| 453 |
+
max_steps = -1
|
| 454 |
+
per_device_train_batch_size = 2
|
| 455 |
+
eval_batch_size: int = 4
|
| 456 |
+
gradient_accumulation_steps = 1
|
| 457 |
+
save_total_limit = 10
|
| 458 |
+
remove_unused_columns = False
|
| 459 |
+
logging_steps = 50
|
| 460 |
+
resume_from_checkpoint: str = None
|
| 461 |
+
gradient_checkpointing: bool = True
|
| 462 |
+
torch_compile: bool = False
|
| 463 |
+
trust_remote_code: bool = True
|
| 464 |
+
qlora: bool = False
|
t5/t5_model.py
ADDED
|
@@ -0,0 +1,1256 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""
|
| 3 |
+
@author:XuMing([email protected])
|
| 4 |
+
@description: refer https://github.com/ThilinaRajapakse/simpletransformers
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import math
|
| 8 |
+
import os
|
| 9 |
+
import random
|
| 10 |
+
import warnings
|
| 11 |
+
from dataclasses import asdict
|
| 12 |
+
from multiprocessing import Pool
|
| 13 |
+
|
| 14 |
+
import numpy as np
|
| 15 |
+
import pandas as pd
|
| 16 |
+
import torch
|
| 17 |
+
from loguru import logger
|
| 18 |
+
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
|
| 19 |
+
from torch.utils.tensorboard import SummaryWriter
|
| 20 |
+
from tqdm.auto import tqdm, trange
|
| 21 |
+
from transformers import ByT5Tokenizer
|
| 22 |
+
from transformers import MT5Config, MT5ForConditionalGeneration
|
| 23 |
+
from transformers import T5Config, T5ForConditionalGeneration, T5Tokenizer, TextStreamer
|
| 24 |
+
from transformers.optimization import AdamW, Adafactor
|
| 25 |
+
from transformers.optimization import (
|
| 26 |
+
get_constant_schedule,
|
| 27 |
+
get_constant_schedule_with_warmup,
|
| 28 |
+
get_linear_schedule_with_warmup,
|
| 29 |
+
get_cosine_schedule_with_warmup,
|
| 30 |
+
get_cosine_with_hard_restarts_schedule_with_warmup,
|
| 31 |
+
get_polynomial_decay_schedule_with_warmup,
|
| 32 |
+
)
|
| 33 |
+
|
| 34 |
+
from t5.config.model_args import T5Args
|
| 35 |
+
from t5.t5_utils import T5Dataset, load_hf_dataset
|
| 36 |
+
|
| 37 |
+
try:
|
| 38 |
+
import wandb
|
| 39 |
+
|
| 40 |
+
wandb_available = True
|
| 41 |
+
except ImportError:
|
| 42 |
+
wandb_available = False
|
| 43 |
+
|
| 44 |
+
has_cuda = torch.cuda.is_available()
|
| 45 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "FALSE"
|
| 46 |
+
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def chunks(lst, n):
|
| 50 |
+
"""Yield successive n-sized chunks from lst."""
|
| 51 |
+
for i in range(0, len(lst), n):
|
| 52 |
+
yield lst[i: i + n]
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
MODEL_CLASSES = {
|
| 56 |
+
"t5": (T5Config, T5ForConditionalGeneration),
|
| 57 |
+
"mt5": (MT5Config, MT5ForConditionalGeneration),
|
| 58 |
+
"byt5": (T5Config, T5ForConditionalGeneration),
|
| 59 |
+
}
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
class T5Model:
|
| 63 |
+
def __init__(
|
| 64 |
+
self,
|
| 65 |
+
model_type,
|
| 66 |
+
model_name,
|
| 67 |
+
args=None,
|
| 68 |
+
tokenizer=None,
|
| 69 |
+
use_cuda=has_cuda,
|
| 70 |
+
cuda_device=-1,
|
| 71 |
+
evaluate=False,
|
| 72 |
+
**kwargs,
|
| 73 |
+
):
|
| 74 |
+
|
| 75 |
+
"""
|
| 76 |
+
Initializes a T5Model model.
|
| 77 |
+
|
| 78 |
+
Args:
|
| 79 |
+
model_type: The type of model (t5, mt5, byt5)
|
| 80 |
+
model_name: The exact architecture and trained weights to use. This may be a Hugging Face Transformers compatible pre-trained model, a community model, or the path to a directory containing model files.
|
| 81 |
+
args (optional): Default args will be used if this parameter is not provided. If provided, it should be a dict containing the args that should be changed in the default args.
|
| 82 |
+
use_cuda (optional): Use GPU if available. Setting to False will force model to use CPU only.
|
| 83 |
+
cuda_device (optional): Specific GPU that should be used. Will use the first available GPU by default.
|
| 84 |
+
**kwargs (optional): For providing proxies, force_download, resume_download, cache_dir and other options specific to the 'from_pretrained' implementation where this will be supplied.
|
| 85 |
+
""" # noqa: ignore flake8"
|
| 86 |
+
|
| 87 |
+
self.args = self._load_model_args(model_name)
|
| 88 |
+
|
| 89 |
+
if isinstance(args, dict):
|
| 90 |
+
self.args.update_from_dict(args)
|
| 91 |
+
elif isinstance(args, T5Args):
|
| 92 |
+
self.args = args
|
| 93 |
+
|
| 94 |
+
self.is_sweeping = False
|
| 95 |
+
|
| 96 |
+
if self.args.manual_seed:
|
| 97 |
+
random.seed(self.args.manual_seed)
|
| 98 |
+
np.random.seed(self.args.manual_seed)
|
| 99 |
+
torch.manual_seed(self.args.manual_seed)
|
| 100 |
+
if self.args.n_gpu > 0:
|
| 101 |
+
torch.cuda.manual_seed_all(self.args.manual_seed)
|
| 102 |
+
|
| 103 |
+
if use_cuda:
|
| 104 |
+
if torch.cuda.is_available():
|
| 105 |
+
if cuda_device == -1:
|
| 106 |
+
self.device = torch.device("cuda")
|
| 107 |
+
else:
|
| 108 |
+
self.device = torch.device(f"cuda:{cuda_device}")
|
| 109 |
+
else:
|
| 110 |
+
raise ValueError(
|
| 111 |
+
"'use_cuda' set to True when cuda is unavailable."
|
| 112 |
+
"Make sure CUDA is available or set `use_cuda=False`."
|
| 113 |
+
)
|
| 114 |
+
else:
|
| 115 |
+
if torch.backends.mps.is_available():
|
| 116 |
+
self.device = torch.device("mps")
|
| 117 |
+
else:
|
| 118 |
+
self.device = "cpu"
|
| 119 |
+
logger.debug(f"Device: {self.device}")
|
| 120 |
+
|
| 121 |
+
self.results = {}
|
| 122 |
+
|
| 123 |
+
config_class, model_class = MODEL_CLASSES[model_type]
|
| 124 |
+
|
| 125 |
+
if model_name is None:
|
| 126 |
+
self.config = self.args.config
|
| 127 |
+
self.model = model_class(config=self.config)
|
| 128 |
+
else:
|
| 129 |
+
self.config = config_class.from_pretrained(model_name, **self.args.config)
|
| 130 |
+
self.model = model_class.from_pretrained(model_name, config=self.config)
|
| 131 |
+
|
| 132 |
+
if isinstance(tokenizer, T5Tokenizer):
|
| 133 |
+
self.tokenizer = tokenizer
|
| 134 |
+
self.model.resize_token_embeddings(len(self.tokenizer))
|
| 135 |
+
elif model_type == "byt5":
|
| 136 |
+
self.tokenizer = ByT5Tokenizer.from_pretrained(model_name, truncate=True)
|
| 137 |
+
else:
|
| 138 |
+
self.tokenizer = T5Tokenizer.from_pretrained(model_name, truncate=True)
|
| 139 |
+
print(len(self.tokenizer))
|
| 140 |
+
if not evaluate:
|
| 141 |
+
with open('./data/字音混淆集_s13.txt', 'r', encoding='utf-8') as confusion:
|
| 142 |
+
n = 0
|
| 143 |
+
for line in confusion.readlines()+[str(chr(c+65248)) for c in range(33, 127)]:
|
| 144 |
+
token = line.split(' ')[0]
|
| 145 |
+
n+=1
|
| 146 |
+
self.tokenizer.add_tokens([token])
|
| 147 |
+
with open('./data/字音混淆集.txt', 'r', encoding='utf-8') as confusion:
|
| 148 |
+
for line in confusion.readlines():
|
| 149 |
+
token = line.split(' ')[0]
|
| 150 |
+
n+=1
|
| 151 |
+
self.tokenizer.add_tokens([token])
|
| 152 |
+
with open('./data/wordtest4.txt', 'r', encoding='utf-8') as confusion:
|
| 153 |
+
for line in confusion.readlines():
|
| 154 |
+
token = line.split(',')[0]
|
| 155 |
+
n+=1
|
| 156 |
+
self.tokenizer.add_tokens([token])
|
| 157 |
+
|
| 158 |
+
with open('./data/vocab.txt', 'r', encoding='utf-8') as confusion:
|
| 159 |
+
for line in confusion.readlines():
|
| 160 |
+
n+=1
|
| 161 |
+
self.tokenizer.add_tokens([line.replace('\n', '')])
|
| 162 |
+
|
| 163 |
+
print(n)
|
| 164 |
+
self.streamer = TextStreamer(self.tokenizer)
|
| 165 |
+
print(len(self.tokenizer))
|
| 166 |
+
self.model.resize_token_embeddings(len(self.tokenizer))
|
| 167 |
+
|
| 168 |
+
if self.args.dynamic_quantize:
|
| 169 |
+
self.model = torch.quantization.quantize_dynamic(
|
| 170 |
+
self.model, {torch.nn.Linear}, dtype=torch.qint8
|
| 171 |
+
)
|
| 172 |
+
|
| 173 |
+
if not use_cuda:
|
| 174 |
+
self.args.fp16 = False
|
| 175 |
+
|
| 176 |
+
if self.args.special_tokens_list:
|
| 177 |
+
self.tokenizer.add_tokens(
|
| 178 |
+
self.args.special_tokens_list, special_tokens=True
|
| 179 |
+
)
|
| 180 |
+
self.model.resize_token_embeddings(len(self.tokenizer))
|
| 181 |
+
|
| 182 |
+
self.args.model_type = model_type
|
| 183 |
+
if model_name is None:
|
| 184 |
+
self.args.model_name = "T5_from_scratch"
|
| 185 |
+
else:
|
| 186 |
+
self.args.model_name = model_name
|
| 187 |
+
|
| 188 |
+
if self.args.wandb_project and not wandb_available:
|
| 189 |
+
warnings.warn(
|
| 190 |
+
"wandb_project specified but wandb is not available. Wandb disabled."
|
| 191 |
+
)
|
| 192 |
+
self.args.wandb_project = None
|
| 193 |
+
|
| 194 |
+
def train_model(
|
| 195 |
+
self,
|
| 196 |
+
train_data,
|
| 197 |
+
output_dir=None,
|
| 198 |
+
show_running_loss=True,
|
| 199 |
+
args=None,
|
| 200 |
+
eval_data=None,
|
| 201 |
+
verbose=True,
|
| 202 |
+
**kwargs,
|
| 203 |
+
):
|
| 204 |
+
"""
|
| 205 |
+
Trains the model using 'train_data'
|
| 206 |
+
|
| 207 |
+
Args:
|
| 208 |
+
train_data: Pandas DataFrame containing the 3 columns - `prefix`, `input_text`, `target_text`.
|
| 209 |
+
- `prefix`: A string indicating the task to perform. (E.g. `"question"`, `"stsb"`)
|
| 210 |
+
- `input_text`: The input text sequence. `prefix` is automatically prepended to form the full input. (<prefix>: <input_text>)
|
| 211 |
+
- `target_text`: The target sequence
|
| 212 |
+
output_dir: The directory where model files will be saved. If not given, self.args.output_dir will be used.
|
| 213 |
+
show_running_loss (optional): Set to False to prevent running loss from being printed to console. Defaults to True.
|
| 214 |
+
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
|
| 215 |
+
eval_data (optional): A DataFrame against which evaluation will be performed when evaluate_during_training is enabled. Is required if evaluate_during_training is enabled.
|
| 216 |
+
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use).
|
| 217 |
+
A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs
|
| 218 |
+
will be lists of strings. Note that this will slow down training significantly as the predicted sequences need to be generated.
|
| 219 |
+
|
| 220 |
+
Returns:
|
| 221 |
+
global_step: Number of global steps trained
|
| 222 |
+
training_details: Average training loss if evaluate_during_training is False or full training progress scores if evaluate_during_training is True
|
| 223 |
+
""" # noqa: ignore flake8"
|
| 224 |
+
|
| 225 |
+
if args:
|
| 226 |
+
self.args.update_from_dict(args)
|
| 227 |
+
if self.args.evaluate_during_training and eval_data is None:
|
| 228 |
+
raise ValueError(
|
| 229 |
+
"evaluate_during_training is enabled but eval_data is not specified."
|
| 230 |
+
" Pass eval_data to model.train_model() if using evaluate_during_training."
|
| 231 |
+
)
|
| 232 |
+
|
| 233 |
+
if not output_dir:
|
| 234 |
+
output_dir = self.args.output_dir
|
| 235 |
+
|
| 236 |
+
if (
|
| 237 |
+
os.path.exists(output_dir)
|
| 238 |
+
and os.listdir(output_dir)
|
| 239 |
+
and not self.args.overwrite_output_dir
|
| 240 |
+
):
|
| 241 |
+
raise ValueError(
|
| 242 |
+
"Output directory ({}) already exists and is not empty."
|
| 243 |
+
" Set args.overwrite_output_dir = True to overcome.".format(output_dir)
|
| 244 |
+
)
|
| 245 |
+
|
| 246 |
+
self._move_model_to_device()
|
| 247 |
+
|
| 248 |
+
train_dataset = self.load_and_cache_examples(train_data, verbose=verbose)
|
| 249 |
+
|
| 250 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 251 |
+
|
| 252 |
+
global_step, training_details = self.train(
|
| 253 |
+
train_dataset,
|
| 254 |
+
output_dir,
|
| 255 |
+
show_running_loss=show_running_loss,
|
| 256 |
+
eval_data=eval_data,
|
| 257 |
+
verbose=verbose,
|
| 258 |
+
**kwargs,
|
| 259 |
+
)
|
| 260 |
+
|
| 261 |
+
self.save_model(model=self.model)
|
| 262 |
+
|
| 263 |
+
if verbose:
|
| 264 |
+
logger.info(
|
| 265 |
+
" Training of {} model complete. Saved to {}.".format(
|
| 266 |
+
self.args.model_name, output_dir
|
| 267 |
+
)
|
| 268 |
+
)
|
| 269 |
+
|
| 270 |
+
return global_step, training_details
|
| 271 |
+
|
| 272 |
+
def train(
|
| 273 |
+
self,
|
| 274 |
+
train_dataset,
|
| 275 |
+
output_dir,
|
| 276 |
+
show_running_loss=True,
|
| 277 |
+
eval_data=None,
|
| 278 |
+
verbose=True,
|
| 279 |
+
**kwargs,
|
| 280 |
+
):
|
| 281 |
+
"""
|
| 282 |
+
Trains the model on train_dataset.
|
| 283 |
+
|
| 284 |
+
Utility function to be used by the train_model() method. Not intended to be used directly.
|
| 285 |
+
"""
|
| 286 |
+
|
| 287 |
+
model = self.model
|
| 288 |
+
args = self.args
|
| 289 |
+
device = self.device
|
| 290 |
+
|
| 291 |
+
tb_writer = SummaryWriter(log_dir=args.tensorboard_dir)
|
| 292 |
+
train_sampler = RandomSampler(train_dataset)
|
| 293 |
+
train_dataloader = DataLoader(
|
| 294 |
+
train_dataset,
|
| 295 |
+
sampler=train_sampler,
|
| 296 |
+
batch_size=args.train_batch_size,
|
| 297 |
+
num_workers=self.args.dataloader_num_workers,
|
| 298 |
+
)
|
| 299 |
+
|
| 300 |
+
if args.max_steps > 0:
|
| 301 |
+
t_total = args.max_steps
|
| 302 |
+
args.num_train_epochs = (
|
| 303 |
+
args.max_steps
|
| 304 |
+
// (len(train_dataloader) // args.gradient_accumulation_steps)
|
| 305 |
+
+ 1
|
| 306 |
+
)
|
| 307 |
+
else:
|
| 308 |
+
t_total = (
|
| 309 |
+
len(train_dataloader)
|
| 310 |
+
// args.gradient_accumulation_steps
|
| 311 |
+
* args.num_train_epochs
|
| 312 |
+
)
|
| 313 |
+
|
| 314 |
+
no_decay = ["bias", "LayerNorm.weight"]
|
| 315 |
+
|
| 316 |
+
optimizer_grouped_parameters = []
|
| 317 |
+
custom_parameter_names = set()
|
| 318 |
+
for group in self.args.custom_parameter_groups:
|
| 319 |
+
params = group.pop("params")
|
| 320 |
+
custom_parameter_names.update(params)
|
| 321 |
+
param_group = {**group}
|
| 322 |
+
param_group["params"] = [
|
| 323 |
+
p for n, p in model.named_parameters() if n in params
|
| 324 |
+
]
|
| 325 |
+
optimizer_grouped_parameters.append(param_group)
|
| 326 |
+
|
| 327 |
+
for group in self.args.custom_layer_parameters:
|
| 328 |
+
layer_number = group.pop("layer")
|
| 329 |
+
layer = f"layer.{layer_number}."
|
| 330 |
+
group_d = {**group}
|
| 331 |
+
group_nd = {**group}
|
| 332 |
+
group_nd["weight_decay"] = 0.0
|
| 333 |
+
params_d = []
|
| 334 |
+
params_nd = []
|
| 335 |
+
for n, p in model.named_parameters():
|
| 336 |
+
if n not in custom_parameter_names and layer in n:
|
| 337 |
+
if any(nd in n for nd in no_decay):
|
| 338 |
+
params_nd.append(p)
|
| 339 |
+
else:
|
| 340 |
+
params_d.append(p)
|
| 341 |
+
custom_parameter_names.add(n)
|
| 342 |
+
group_d["params"] = params_d
|
| 343 |
+
group_nd["params"] = params_nd
|
| 344 |
+
|
| 345 |
+
optimizer_grouped_parameters.append(group_d)
|
| 346 |
+
optimizer_grouped_parameters.append(group_nd)
|
| 347 |
+
|
| 348 |
+
if not self.args.train_custom_parameters_only:
|
| 349 |
+
optimizer_grouped_parameters.extend(
|
| 350 |
+
[
|
| 351 |
+
{
|
| 352 |
+
"params": [
|
| 353 |
+
p
|
| 354 |
+
for n, p in model.named_parameters()
|
| 355 |
+
if n not in custom_parameter_names
|
| 356 |
+
and not any(nd in n for nd in no_decay)
|
| 357 |
+
],
|
| 358 |
+
"weight_decay": args.weight_decay,
|
| 359 |
+
},
|
| 360 |
+
{
|
| 361 |
+
"params": [
|
| 362 |
+
p
|
| 363 |
+
for n, p in model.named_parameters()
|
| 364 |
+
if n not in custom_parameter_names
|
| 365 |
+
and any(nd in n for nd in no_decay)
|
| 366 |
+
],
|
| 367 |
+
"weight_decay": 0.0,
|
| 368 |
+
},
|
| 369 |
+
]
|
| 370 |
+
)
|
| 371 |
+
|
| 372 |
+
warmup_steps = math.ceil(t_total * args.warmup_ratio)
|
| 373 |
+
args.warmup_steps = (
|
| 374 |
+
warmup_steps if args.warmup_steps == 0 else args.warmup_steps
|
| 375 |
+
)
|
| 376 |
+
|
| 377 |
+
if args.optimizer == "AdamW":
|
| 378 |
+
optimizer = AdamW(
|
| 379 |
+
optimizer_grouped_parameters,
|
| 380 |
+
lr=args.learning_rate,
|
| 381 |
+
eps=args.adam_epsilon,
|
| 382 |
+
)
|
| 383 |
+
elif args.optimizer == "Adafactor":
|
| 384 |
+
optimizer = Adafactor(
|
| 385 |
+
optimizer_grouped_parameters,
|
| 386 |
+
lr=args.learning_rate,
|
| 387 |
+
eps=args.adafactor_eps,
|
| 388 |
+
clip_threshold=args.adafactor_clip_threshold,
|
| 389 |
+
decay_rate=args.adafactor_decay_rate,
|
| 390 |
+
beta1=args.adafactor_beta1,
|
| 391 |
+
weight_decay=args.weight_decay,
|
| 392 |
+
scale_parameter=args.adafactor_scale_parameter,
|
| 393 |
+
relative_step=args.adafactor_relative_step,
|
| 394 |
+
warmup_init=args.adafactor_warmup_init,
|
| 395 |
+
)
|
| 396 |
+
|
| 397 |
+
else:
|
| 398 |
+
raise ValueError(
|
| 399 |
+
"{} is not a valid optimizer class. Please use one of ('AdamW', 'Adafactor') instead.".format(
|
| 400 |
+
args.optimizer
|
| 401 |
+
)
|
| 402 |
+
)
|
| 403 |
+
|
| 404 |
+
if args.scheduler == "constant_schedule":
|
| 405 |
+
scheduler = get_constant_schedule(optimizer)
|
| 406 |
+
|
| 407 |
+
elif args.scheduler == "constant_schedule_with_warmup":
|
| 408 |
+
scheduler = get_constant_schedule_with_warmup(
|
| 409 |
+
optimizer, num_warmup_steps=args.warmup_steps
|
| 410 |
+
)
|
| 411 |
+
|
| 412 |
+
elif args.scheduler == "linear_schedule_with_warmup":
|
| 413 |
+
scheduler = get_linear_schedule_with_warmup(
|
| 414 |
+
optimizer,
|
| 415 |
+
num_warmup_steps=args.warmup_steps,
|
| 416 |
+
num_training_steps=t_total,
|
| 417 |
+
)
|
| 418 |
+
|
| 419 |
+
elif args.scheduler == "cosine_schedule_with_warmup":
|
| 420 |
+
scheduler = get_cosine_schedule_with_warmup(
|
| 421 |
+
optimizer,
|
| 422 |
+
num_warmup_steps=args.warmup_steps,
|
| 423 |
+
num_training_steps=t_total,
|
| 424 |
+
num_cycles=args.cosine_schedule_num_cycles,
|
| 425 |
+
)
|
| 426 |
+
|
| 427 |
+
elif args.scheduler == "cosine_with_hard_restarts_schedule_with_warmup":
|
| 428 |
+
scheduler = get_cosine_with_hard_restarts_schedule_with_warmup(
|
| 429 |
+
optimizer,
|
| 430 |
+
num_warmup_steps=args.warmup_steps,
|
| 431 |
+
num_training_steps=t_total,
|
| 432 |
+
num_cycles=args.cosine_schedule_num_cycles,
|
| 433 |
+
)
|
| 434 |
+
|
| 435 |
+
elif args.scheduler == "polynomial_decay_schedule_with_warmup":
|
| 436 |
+
scheduler = get_polynomial_decay_schedule_with_warmup(
|
| 437 |
+
optimizer,
|
| 438 |
+
num_warmup_steps=args.warmup_steps,
|
| 439 |
+
num_training_steps=t_total,
|
| 440 |
+
lr_end=args.polynomial_decay_schedule_lr_end,
|
| 441 |
+
power=args.polynomial_decay_schedule_power,
|
| 442 |
+
)
|
| 443 |
+
|
| 444 |
+
else:
|
| 445 |
+
raise ValueError("{} is not a valid scheduler.".format(args.scheduler))
|
| 446 |
+
|
| 447 |
+
if (
|
| 448 |
+
args.model_name
|
| 449 |
+
and os.path.isfile(os.path.join(args.model_name, "optimizer.pt"))
|
| 450 |
+
and os.path.isfile(os.path.join(args.model_name, "scheduler.pt"))
|
| 451 |
+
):
|
| 452 |
+
# Load in optimizer and scheduler states
|
| 453 |
+
optimizer.load_state_dict(
|
| 454 |
+
torch.load(os.path.join(args.model_name, "optimizer.pt"))
|
| 455 |
+
)
|
| 456 |
+
scheduler.load_state_dict(
|
| 457 |
+
torch.load(os.path.join(args.model_name, "scheduler.pt"))
|
| 458 |
+
)
|
| 459 |
+
|
| 460 |
+
if args.n_gpu > 1:
|
| 461 |
+
model = torch.nn.DataParallel(model)
|
| 462 |
+
|
| 463 |
+
logger.info(" Training started")
|
| 464 |
+
|
| 465 |
+
global_step = 0
|
| 466 |
+
training_progress_scores = None
|
| 467 |
+
tr_loss, logging_loss = 0.0, 0.0
|
| 468 |
+
model.zero_grad()
|
| 469 |
+
train_iterator = trange(
|
| 470 |
+
int(args.num_train_epochs), desc="Epoch", disable=args.silent, mininterval=0
|
| 471 |
+
)
|
| 472 |
+
epoch_number = 0
|
| 473 |
+
best_eval_metric = None
|
| 474 |
+
early_stopping_counter = 0
|
| 475 |
+
steps_trained_in_current_epoch = 0
|
| 476 |
+
epochs_trained = 0
|
| 477 |
+
|
| 478 |
+
if args.model_name and os.path.exists(args.model_name):
|
| 479 |
+
try:
|
| 480 |
+
# set global_step to gobal_step of last saved checkpoint from model path
|
| 481 |
+
checkpoint_suffix = args.model_name.split("/")[-1].split("-")
|
| 482 |
+
if len(checkpoint_suffix) > 2:
|
| 483 |
+
checkpoint_suffix = checkpoint_suffix[1]
|
| 484 |
+
else:
|
| 485 |
+
checkpoint_suffix = checkpoint_suffix[-1]
|
| 486 |
+
global_step = int(checkpoint_suffix)
|
| 487 |
+
epochs_trained = global_step // (
|
| 488 |
+
len(train_dataloader) // args.gradient_accumulation_steps
|
| 489 |
+
)
|
| 490 |
+
steps_trained_in_current_epoch = global_step % (
|
| 491 |
+
len(train_dataloader) // args.gradient_accumulation_steps
|
| 492 |
+
)
|
| 493 |
+
|
| 494 |
+
logger.info(
|
| 495 |
+
" Continuing training from checkpoint, will skip to saved global_step"
|
| 496 |
+
)
|
| 497 |
+
logger.info(" Continuing training from epoch %d", epochs_trained)
|
| 498 |
+
logger.info(" Continuing training from global step %d", global_step)
|
| 499 |
+
logger.info(
|
| 500 |
+
" Will skip the first %d steps in the current epoch",
|
| 501 |
+
steps_trained_in_current_epoch,
|
| 502 |
+
)
|
| 503 |
+
except ValueError:
|
| 504 |
+
logger.info(" Starting fine-tuning.")
|
| 505 |
+
|
| 506 |
+
if args.evaluate_during_training:
|
| 507 |
+
training_progress_scores = self._create_training_progress_scores(**kwargs)
|
| 508 |
+
|
| 509 |
+
if args.wandb_project:
|
| 510 |
+
wandb.init(
|
| 511 |
+
project=args.wandb_project,
|
| 512 |
+
config={**asdict(args)},
|
| 513 |
+
**args.wandb_kwargs,
|
| 514 |
+
)
|
| 515 |
+
wandb.run._label(repo="textgen")
|
| 516 |
+
wandb.watch(self.model)
|
| 517 |
+
self.wandb_run_id = wandb.run.id
|
| 518 |
+
|
| 519 |
+
if args.fp16:
|
| 520 |
+
from torch.cuda import amp
|
| 521 |
+
|
| 522 |
+
scaler = amp.GradScaler()
|
| 523 |
+
|
| 524 |
+
for current_epoch in train_iterator:
|
| 525 |
+
model.train()
|
| 526 |
+
if epochs_trained > 0:
|
| 527 |
+
epochs_trained -= 1
|
| 528 |
+
continue
|
| 529 |
+
train_iterator.set_description(
|
| 530 |
+
f"Epoch {epoch_number + 1} of {args.num_train_epochs}"
|
| 531 |
+
)
|
| 532 |
+
batch_iterator = tqdm(
|
| 533 |
+
train_dataloader,
|
| 534 |
+
desc=f"Running Epoch {epoch_number} of {args.num_train_epochs}",
|
| 535 |
+
disable=args.silent,
|
| 536 |
+
mininterval=0,
|
| 537 |
+
)
|
| 538 |
+
for step, batch in enumerate(batch_iterator):
|
| 539 |
+
if steps_trained_in_current_epoch > 0:
|
| 540 |
+
steps_trained_in_current_epoch -= 1
|
| 541 |
+
continue
|
| 542 |
+
|
| 543 |
+
inputs = self._get_inputs_dict(batch)
|
| 544 |
+
if args.fp16:
|
| 545 |
+
with amp.autocast():
|
| 546 |
+
outputs = model(**inputs)
|
| 547 |
+
# model outputs are always tuple in pytorch-transformers (see doc)
|
| 548 |
+
loss = outputs[0]
|
| 549 |
+
else:
|
| 550 |
+
outputs = model(**inputs)
|
| 551 |
+
# model outputs are always tuple in pytorch-transformers (see doc)
|
| 552 |
+
loss = outputs[0]
|
| 553 |
+
|
| 554 |
+
if args.n_gpu > 1:
|
| 555 |
+
loss = (
|
| 556 |
+
loss.mean()
|
| 557 |
+
) # mean() to average on multi-gpu parallel training
|
| 558 |
+
|
| 559 |
+
current_loss = loss.item()
|
| 560 |
+
|
| 561 |
+
if show_running_loss:
|
| 562 |
+
batch_iterator.set_description(
|
| 563 |
+
f"Epochs {epoch_number}/{args.num_train_epochs}. Running Loss: {current_loss:9.4f}"
|
| 564 |
+
)
|
| 565 |
+
|
| 566 |
+
if args.gradient_accumulation_steps > 1:
|
| 567 |
+
loss = loss / args.gradient_accumulation_steps
|
| 568 |
+
|
| 569 |
+
if args.fp16:
|
| 570 |
+
scaler.scale(loss).backward()
|
| 571 |
+
else:
|
| 572 |
+
loss.backward()
|
| 573 |
+
|
| 574 |
+
tr_loss += loss.item()
|
| 575 |
+
if (step + 1) % args.gradient_accumulation_steps == 0:
|
| 576 |
+
if args.fp16:
|
| 577 |
+
scaler.unscale_(optimizer)
|
| 578 |
+
if args.optimizer == "AdamW":
|
| 579 |
+
torch.nn.utils.clip_grad_norm_(
|
| 580 |
+
model.parameters(), args.max_grad_norm
|
| 581 |
+
)
|
| 582 |
+
|
| 583 |
+
if args.fp16:
|
| 584 |
+
scaler.step(optimizer)
|
| 585 |
+
scaler.update()
|
| 586 |
+
else:
|
| 587 |
+
optimizer.step()
|
| 588 |
+
scheduler.step() # Update learning rate schedule
|
| 589 |
+
model.zero_grad()
|
| 590 |
+
global_step += 1
|
| 591 |
+
|
| 592 |
+
if args.logging_steps > 0 and global_step % args.logging_steps == 0:
|
| 593 |
+
# Log metrics
|
| 594 |
+
tb_writer.add_scalar(
|
| 595 |
+
"lr", scheduler.get_last_lr()[0], global_step
|
| 596 |
+
)
|
| 597 |
+
tb_writer.add_scalar(
|
| 598 |
+
"loss",
|
| 599 |
+
(tr_loss - logging_loss) / args.logging_steps,
|
| 600 |
+
global_step,
|
| 601 |
+
)
|
| 602 |
+
logging_loss = tr_loss
|
| 603 |
+
if args.wandb_project or self.is_sweeping:
|
| 604 |
+
wandb.log(
|
| 605 |
+
{
|
| 606 |
+
"Training loss": current_loss,
|
| 607 |
+
"lr": scheduler.get_last_lr()[0],
|
| 608 |
+
"global_step": global_step,
|
| 609 |
+
}
|
| 610 |
+
)
|
| 611 |
+
|
| 612 |
+
if args.save_steps > 0 and global_step % args.save_steps == 0:
|
| 613 |
+
# Save model checkpoint
|
| 614 |
+
output_dir_current = os.path.join(
|
| 615 |
+
output_dir, "checkpoint-{}".format(global_step)
|
| 616 |
+
)
|
| 617 |
+
|
| 618 |
+
self.save_model(
|
| 619 |
+
output_dir_current, optimizer, scheduler, model=model
|
| 620 |
+
)
|
| 621 |
+
|
| 622 |
+
if args.evaluate_during_training and (
|
| 623 |
+
args.evaluate_during_training_steps > 0
|
| 624 |
+
and global_step % args.evaluate_during_training_steps == 0
|
| 625 |
+
):
|
| 626 |
+
# Only evaluate when single GPU otherwise metrics may not average well
|
| 627 |
+
results = self.eval_model(
|
| 628 |
+
eval_data,
|
| 629 |
+
verbose=verbose and args.evaluate_during_training_verbose,
|
| 630 |
+
silent=args.evaluate_during_training_silent,
|
| 631 |
+
**kwargs,
|
| 632 |
+
)
|
| 633 |
+
for key, value in results.items():
|
| 634 |
+
try:
|
| 635 |
+
tb_writer.add_scalar(
|
| 636 |
+
"eval_{}".format(key), value, global_step
|
| 637 |
+
)
|
| 638 |
+
except (NotImplementedError, AssertionError):
|
| 639 |
+
pass
|
| 640 |
+
|
| 641 |
+
output_dir_current = os.path.join(
|
| 642 |
+
output_dir, "checkpoint-{}".format(global_step)
|
| 643 |
+
)
|
| 644 |
+
|
| 645 |
+
if args.save_eval_checkpoints:
|
| 646 |
+
self.save_model(
|
| 647 |
+
output_dir_current,
|
| 648 |
+
optimizer,
|
| 649 |
+
scheduler,
|
| 650 |
+
model=model,
|
| 651 |
+
results=results,
|
| 652 |
+
)
|
| 653 |
+
|
| 654 |
+
training_progress_scores["global_step"].append(global_step)
|
| 655 |
+
training_progress_scores["train_loss"].append(current_loss)
|
| 656 |
+
for key in results:
|
| 657 |
+
training_progress_scores[key].append(results[key])
|
| 658 |
+
report = pd.DataFrame(training_progress_scores)
|
| 659 |
+
report.to_csv(
|
| 660 |
+
os.path.join(
|
| 661 |
+
args.output_dir, "training_progress_scores.csv"
|
| 662 |
+
),
|
| 663 |
+
index=False,
|
| 664 |
+
)
|
| 665 |
+
|
| 666 |
+
if args.wandb_project or self.is_sweeping:
|
| 667 |
+
wandb.log(self._get_last_metrics(training_progress_scores))
|
| 668 |
+
|
| 669 |
+
if not best_eval_metric:
|
| 670 |
+
best_eval_metric = results[args.early_stopping_metric]
|
| 671 |
+
self.save_model(
|
| 672 |
+
args.best_model_dir,
|
| 673 |
+
optimizer,
|
| 674 |
+
scheduler,
|
| 675 |
+
model=model,
|
| 676 |
+
results=results,
|
| 677 |
+
)
|
| 678 |
+
if best_eval_metric and args.early_stopping_metric_minimize:
|
| 679 |
+
if (
|
| 680 |
+
results[args.early_stopping_metric] - best_eval_metric
|
| 681 |
+
< args.early_stopping_delta
|
| 682 |
+
):
|
| 683 |
+
best_eval_metric = results[args.early_stopping_metric]
|
| 684 |
+
self.save_model(
|
| 685 |
+
args.best_model_dir,
|
| 686 |
+
optimizer,
|
| 687 |
+
scheduler,
|
| 688 |
+
model=model,
|
| 689 |
+
results=results,
|
| 690 |
+
)
|
| 691 |
+
early_stopping_counter = 0
|
| 692 |
+
else:
|
| 693 |
+
if args.use_early_stopping:
|
| 694 |
+
if (
|
| 695 |
+
early_stopping_counter
|
| 696 |
+
< args.early_stopping_patience
|
| 697 |
+
):
|
| 698 |
+
early_stopping_counter += 1
|
| 699 |
+
if verbose:
|
| 700 |
+
logger.info(
|
| 701 |
+
f" No improvement in {args.early_stopping_metric}"
|
| 702 |
+
)
|
| 703 |
+
logger.info(
|
| 704 |
+
f" Current step: {early_stopping_counter}"
|
| 705 |
+
)
|
| 706 |
+
logger.info(
|
| 707 |
+
f" Early stopping patience: {args.early_stopping_patience}"
|
| 708 |
+
)
|
| 709 |
+
else:
|
| 710 |
+
if verbose:
|
| 711 |
+
logger.info(
|
| 712 |
+
f" Patience of {args.early_stopping_patience} steps reached"
|
| 713 |
+
)
|
| 714 |
+
logger.info(" Training terminated.")
|
| 715 |
+
train_iterator.close()
|
| 716 |
+
return (
|
| 717 |
+
global_step,
|
| 718 |
+
tr_loss / global_step
|
| 719 |
+
if not self.args.evaluate_during_training
|
| 720 |
+
else training_progress_scores,
|
| 721 |
+
)
|
| 722 |
+
else:
|
| 723 |
+
if (
|
| 724 |
+
results[args.early_stopping_metric] - best_eval_metric
|
| 725 |
+
> args.early_stopping_delta
|
| 726 |
+
):
|
| 727 |
+
best_eval_metric = results[args.early_stopping_metric]
|
| 728 |
+
self.save_model(
|
| 729 |
+
args.best_model_dir,
|
| 730 |
+
optimizer,
|
| 731 |
+
scheduler,
|
| 732 |
+
model=model,
|
| 733 |
+
results=results,
|
| 734 |
+
)
|
| 735 |
+
early_stopping_counter = 0
|
| 736 |
+
else:
|
| 737 |
+
if args.use_early_stopping:
|
| 738 |
+
if (
|
| 739 |
+
early_stopping_counter
|
| 740 |
+
< args.early_stopping_patience
|
| 741 |
+
):
|
| 742 |
+
early_stopping_counter += 1
|
| 743 |
+
if verbose:
|
| 744 |
+
logger.info(
|
| 745 |
+
f" No improvement in {args.early_stopping_metric}"
|
| 746 |
+
)
|
| 747 |
+
logger.info(
|
| 748 |
+
f" Current step: {early_stopping_counter}"
|
| 749 |
+
)
|
| 750 |
+
logger.info(
|
| 751 |
+
f" Early stopping patience: {args.early_stopping_patience}"
|
| 752 |
+
)
|
| 753 |
+
else:
|
| 754 |
+
if verbose:
|
| 755 |
+
logger.info(
|
| 756 |
+
f" Patience of {args.early_stopping_patience} steps reached"
|
| 757 |
+
)
|
| 758 |
+
logger.info(" Training terminated.")
|
| 759 |
+
train_iterator.close()
|
| 760 |
+
return (
|
| 761 |
+
global_step,
|
| 762 |
+
tr_loss / global_step
|
| 763 |
+
if not self.args.evaluate_during_training
|
| 764 |
+
else training_progress_scores,
|
| 765 |
+
)
|
| 766 |
+
model.train()
|
| 767 |
+
|
| 768 |
+
epoch_number += 1
|
| 769 |
+
output_dir_current = os.path.join(
|
| 770 |
+
output_dir, "checkpoint-{}-epoch-{}".format(global_step, epoch_number)
|
| 771 |
+
)
|
| 772 |
+
|
| 773 |
+
if args.save_model_every_epoch:
|
| 774 |
+
self.save_model(output_dir_current, optimizer, scheduler, model=model)
|
| 775 |
+
|
| 776 |
+
if args.evaluate_during_training and args.evaluate_each_epoch:
|
| 777 |
+
results = self.eval_model(
|
| 778 |
+
eval_data,
|
| 779 |
+
verbose=verbose and args.evaluate_during_training_verbose,
|
| 780 |
+
silent=args.evaluate_during_training_silent,
|
| 781 |
+
**kwargs,
|
| 782 |
+
)
|
| 783 |
+
|
| 784 |
+
if args.save_eval_checkpoints:
|
| 785 |
+
self.save_model(
|
| 786 |
+
output_dir_current, optimizer, scheduler, results=results
|
| 787 |
+
)
|
| 788 |
+
|
| 789 |
+
training_progress_scores["global_step"].append(global_step)
|
| 790 |
+
training_progress_scores["train_loss"].append(current_loss)
|
| 791 |
+
for key in results:
|
| 792 |
+
training_progress_scores[key].append(results[key])
|
| 793 |
+
report = pd.DataFrame(training_progress_scores)
|
| 794 |
+
report.to_csv(
|
| 795 |
+
os.path.join(args.output_dir, "training_progress_scores.csv"),
|
| 796 |
+
index=False,
|
| 797 |
+
)
|
| 798 |
+
|
| 799 |
+
if args.wandb_project or self.is_sweeping:
|
| 800 |
+
wandb.log(self._get_last_metrics(training_progress_scores))
|
| 801 |
+
|
| 802 |
+
if not best_eval_metric:
|
| 803 |
+
best_eval_metric = results[args.early_stopping_metric]
|
| 804 |
+
self.save_model(
|
| 805 |
+
args.best_model_dir,
|
| 806 |
+
optimizer,
|
| 807 |
+
scheduler,
|
| 808 |
+
model=model,
|
| 809 |
+
results=results,
|
| 810 |
+
)
|
| 811 |
+
if best_eval_metric and args.early_stopping_metric_minimize:
|
| 812 |
+
if (
|
| 813 |
+
results[args.early_stopping_metric] - best_eval_metric
|
| 814 |
+
< args.early_stopping_delta
|
| 815 |
+
):
|
| 816 |
+
best_eval_metric = results[args.early_stopping_metric]
|
| 817 |
+
self.save_model(
|
| 818 |
+
args.best_model_dir,
|
| 819 |
+
optimizer,
|
| 820 |
+
scheduler,
|
| 821 |
+
model=model,
|
| 822 |
+
results=results,
|
| 823 |
+
)
|
| 824 |
+
early_stopping_counter = 0
|
| 825 |
+
else:
|
| 826 |
+
if (
|
| 827 |
+
args.use_early_stopping
|
| 828 |
+
and args.early_stopping_consider_epochs
|
| 829 |
+
):
|
| 830 |
+
if early_stopping_counter < args.early_stopping_patience:
|
| 831 |
+
early_stopping_counter += 1
|
| 832 |
+
if verbose:
|
| 833 |
+
logger.info(
|
| 834 |
+
f" No improvement in {args.early_stopping_metric}"
|
| 835 |
+
)
|
| 836 |
+
logger.info(
|
| 837 |
+
f" Current step: {early_stopping_counter}"
|
| 838 |
+
)
|
| 839 |
+
logger.info(
|
| 840 |
+
f" Early stopping patience: {args.early_stopping_patience}"
|
| 841 |
+
)
|
| 842 |
+
else:
|
| 843 |
+
if verbose:
|
| 844 |
+
logger.info(
|
| 845 |
+
f" Patience of {args.early_stopping_patience} steps reached"
|
| 846 |
+
)
|
| 847 |
+
logger.info(" Training terminated.")
|
| 848 |
+
train_iterator.close()
|
| 849 |
+
return (
|
| 850 |
+
global_step,
|
| 851 |
+
tr_loss / global_step
|
| 852 |
+
if not self.args.evaluate_during_training
|
| 853 |
+
else training_progress_scores,
|
| 854 |
+
)
|
| 855 |
+
else:
|
| 856 |
+
if (
|
| 857 |
+
results[args.early_stopping_metric] - best_eval_metric
|
| 858 |
+
> args.early_stopping_delta
|
| 859 |
+
):
|
| 860 |
+
best_eval_metric = results[args.early_stopping_metric]
|
| 861 |
+
self.save_model(
|
| 862 |
+
args.best_model_dir,
|
| 863 |
+
optimizer,
|
| 864 |
+
scheduler,
|
| 865 |
+
model=model,
|
| 866 |
+
results=results,
|
| 867 |
+
)
|
| 868 |
+
early_stopping_counter = 0
|
| 869 |
+
else:
|
| 870 |
+
if (
|
| 871 |
+
args.use_early_stopping
|
| 872 |
+
and args.early_stopping_consider_epochs
|
| 873 |
+
):
|
| 874 |
+
if early_stopping_counter < args.early_stopping_patience:
|
| 875 |
+
early_stopping_counter += 1
|
| 876 |
+
if verbose:
|
| 877 |
+
logger.info(
|
| 878 |
+
f" No improvement in {args.early_stopping_metric}"
|
| 879 |
+
)
|
| 880 |
+
logger.info(
|
| 881 |
+
f" Current step: {early_stopping_counter}"
|
| 882 |
+
)
|
| 883 |
+
logger.info(
|
| 884 |
+
f" Early stopping patience: {args.early_stopping_patience}"
|
| 885 |
+
)
|
| 886 |
+
else:
|
| 887 |
+
if verbose:
|
| 888 |
+
logger.info(
|
| 889 |
+
f" Patience of {args.early_stopping_patience} steps reached"
|
| 890 |
+
)
|
| 891 |
+
logger.info(" Training terminated.")
|
| 892 |
+
train_iterator.close()
|
| 893 |
+
return (
|
| 894 |
+
global_step,
|
| 895 |
+
tr_loss / global_step
|
| 896 |
+
if not self.args.evaluate_during_training
|
| 897 |
+
else training_progress_scores,
|
| 898 |
+
)
|
| 899 |
+
|
| 900 |
+
return (
|
| 901 |
+
global_step,
|
| 902 |
+
tr_loss / global_step
|
| 903 |
+
if not self.args.evaluate_during_training
|
| 904 |
+
else training_progress_scores,
|
| 905 |
+
)
|
| 906 |
+
|
| 907 |
+
def eval_model(
|
| 908 |
+
self, eval_data, output_dir=None, verbose=True, silent=False, **kwargs
|
| 909 |
+
):
|
| 910 |
+
"""
|
| 911 |
+
Evaluates the model on eval_data. Saves results to output_dir.
|
| 912 |
+
|
| 913 |
+
Args:
|
| 914 |
+
eval_data: Pandas DataFrame containing the 3 columns - `prefix`, `input_text`, `target_text`.
|
| 915 |
+
- `prefix`: A string indicating the task to perform. (E.g. `"question"`, `"stsb"`)
|
| 916 |
+
- `input_text`: The input text sequence. `prefix` is automatically prepended to form the full input. (<prefix>: <input_text>)
|
| 917 |
+
- `target_text`: The target sequence
|
| 918 |
+
output_dir: The directory where model files will be saved. If not given, self.args.output_dir will be used.
|
| 919 |
+
verbose: If verbose, results will be printed to the console on completion of evaluation.
|
| 920 |
+
silent: If silent, tqdm progress bars will be hidden.
|
| 921 |
+
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use).
|
| 922 |
+
A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs
|
| 923 |
+
will be lists of strings. Note that this will slow down evaluation significantly as the predicted sequences need to be generated.
|
| 924 |
+
Returns:
|
| 925 |
+
results: Dictionary containing evaluation results.
|
| 926 |
+
""" # noqa: ignore flake8"
|
| 927 |
+
|
| 928 |
+
if not output_dir:
|
| 929 |
+
output_dir = self.args.output_dir
|
| 930 |
+
|
| 931 |
+
self._move_model_to_device()
|
| 932 |
+
|
| 933 |
+
eval_dataset = self.load_and_cache_examples(
|
| 934 |
+
eval_data, evaluate=True, verbose=verbose, silent=silent
|
| 935 |
+
)
|
| 936 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 937 |
+
|
| 938 |
+
result = self.evaluate(
|
| 939 |
+
eval_dataset, output_dir, verbose=verbose, silent=silent, **kwargs
|
| 940 |
+
)
|
| 941 |
+
self.results.update(result)
|
| 942 |
+
|
| 943 |
+
if self.args.evaluate_generated_text:
|
| 944 |
+
if self.args.preprocess_inputs:
|
| 945 |
+
to_predict = [
|
| 946 |
+
input_text
|
| 947 |
+
for prefix, input_text in zip(
|
| 948 |
+
eval_data["prefix"], eval_data["input_text"]
|
| 949 |
+
)
|
| 950 |
+
]
|
| 951 |
+
else:
|
| 952 |
+
to_predict = [
|
| 953 |
+
prefix + input_text
|
| 954 |
+
for prefix, input_text in zip(
|
| 955 |
+
eval_data["prefix"], eval_data["input_text"]
|
| 956 |
+
)
|
| 957 |
+
]
|
| 958 |
+
preds = self.predict(to_predict[:self.args.eval_batch_size*3])
|
| 959 |
+
|
| 960 |
+
result = self.compute_metrics(
|
| 961 |
+
eval_data["target_text"].tolist()[:self.args.eval_batch_size*3], preds, **kwargs
|
| 962 |
+
)
|
| 963 |
+
self.results.update(result)
|
| 964 |
+
|
| 965 |
+
if verbose:
|
| 966 |
+
logger.info(self.results)
|
| 967 |
+
|
| 968 |
+
return self.results
|
| 969 |
+
|
| 970 |
+
def evaluate(self, eval_dataset, output_dir, verbose=True, silent=False, **kwargs):
|
| 971 |
+
"""
|
| 972 |
+
Evaluates the model on eval_dataset.
|
| 973 |
+
|
| 974 |
+
Utility function to be used by the eval_model() method. Not intended to be used directly.
|
| 975 |
+
"""
|
| 976 |
+
|
| 977 |
+
model = self.model
|
| 978 |
+
args = self.args
|
| 979 |
+
eval_output_dir = output_dir
|
| 980 |
+
device = self.device
|
| 981 |
+
|
| 982 |
+
results = {}
|
| 983 |
+
|
| 984 |
+
eval_sampler = SequentialSampler(eval_dataset)
|
| 985 |
+
eval_dataloader = DataLoader(
|
| 986 |
+
eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size
|
| 987 |
+
)
|
| 988 |
+
|
| 989 |
+
if args.n_gpu > 1:
|
| 990 |
+
model = torch.nn.DataParallel(model)
|
| 991 |
+
|
| 992 |
+
eval_loss = 0.0
|
| 993 |
+
nb_eval_steps = 0
|
| 994 |
+
model.eval()
|
| 995 |
+
|
| 996 |
+
if self.args.fp16:
|
| 997 |
+
from torch.cuda import amp
|
| 998 |
+
|
| 999 |
+
for batch in tqdm(
|
| 1000 |
+
eval_dataloader, disable=args.silent or silent, desc="Running Evaluation"
|
| 1001 |
+
):
|
| 1002 |
+
inputs = self._get_inputs_dict(batch)
|
| 1003 |
+
with torch.no_grad():
|
| 1004 |
+
if self.args.fp16:
|
| 1005 |
+
with amp.autocast():
|
| 1006 |
+
outputs = model(**inputs)
|
| 1007 |
+
loss = outputs[0]
|
| 1008 |
+
else:
|
| 1009 |
+
outputs = model(**inputs)
|
| 1010 |
+
loss = outputs[0]
|
| 1011 |
+
if self.args.n_gpu > 1:
|
| 1012 |
+
loss = loss.mean()
|
| 1013 |
+
eval_loss += loss.item()
|
| 1014 |
+
nb_eval_steps += 1
|
| 1015 |
+
|
| 1016 |
+
eval_loss = eval_loss / nb_eval_steps
|
| 1017 |
+
|
| 1018 |
+
results["eval_loss"] = eval_loss
|
| 1019 |
+
|
| 1020 |
+
output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
|
| 1021 |
+
with open(output_eval_file, "w") as writer:
|
| 1022 |
+
for key in sorted(results.keys()):
|
| 1023 |
+
writer.write("{} = {}\n".format(key, str(results[key])))
|
| 1024 |
+
|
| 1025 |
+
return results
|
| 1026 |
+
|
| 1027 |
+
def predict(self, to_predict, split_on_space=False):
|
| 1028 |
+
"""
|
| 1029 |
+
Performs predictions on a list of text.
|
| 1030 |
+
|
| 1031 |
+
Args:
|
| 1032 |
+
to_predict: A python list of text (str) to be sent to the model for prediction. Note that the prefix should be prepended to the text.
|
| 1033 |
+
split_on_space (optional): If True, input is english string, if False, input is chinese string.
|
| 1034 |
+
|
| 1035 |
+
Returns:
|
| 1036 |
+
preds: A python list of the generated sequences.
|
| 1037 |
+
""" # noqa: ignore flake8"
|
| 1038 |
+
|
| 1039 |
+
self._move_model_to_device()
|
| 1040 |
+
|
| 1041 |
+
all_outputs = []
|
| 1042 |
+
# Batching
|
| 1043 |
+
for batch in tqdm(
|
| 1044 |
+
[
|
| 1045 |
+
to_predict[i: i + self.args.eval_batch_size]
|
| 1046 |
+
for i in range(0, len(to_predict), self.args.eval_batch_size)
|
| 1047 |
+
],
|
| 1048 |
+
desc="Generating outputs",
|
| 1049 |
+
disable=self.args.silent,
|
| 1050 |
+
):
|
| 1051 |
+
input_batch = self.tokenizer.prepare_seq2seq_batch(
|
| 1052 |
+
src_texts=batch,
|
| 1053 |
+
max_length=self.args.max_seq_length,
|
| 1054 |
+
padding="max_length",
|
| 1055 |
+
return_tensors="pt",
|
| 1056 |
+
truncation=True,
|
| 1057 |
+
)
|
| 1058 |
+
input_ids = input_batch["input_ids"]
|
| 1059 |
+
attention_mask = input_batch["attention_mask"]
|
| 1060 |
+
|
| 1061 |
+
input_ids = input_ids.to(self.device)
|
| 1062 |
+
attention_mask = attention_mask.to(self.device)
|
| 1063 |
+
|
| 1064 |
+
outputs = self.model.generate(
|
| 1065 |
+
input_ids=input_ids,
|
| 1066 |
+
attention_mask=attention_mask,
|
| 1067 |
+
num_beams=self.args.num_beams,
|
| 1068 |
+
max_length=self.args.max_length,
|
| 1069 |
+
length_penalty=self.args.length_penalty,
|
| 1070 |
+
early_stopping=self.args.early_stopping,
|
| 1071 |
+
repetition_penalty=self.args.repetition_penalty,
|
| 1072 |
+
do_sample=self.args.do_sample,
|
| 1073 |
+
top_k=self.args.top_k,
|
| 1074 |
+
top_p=self.args.top_p,
|
| 1075 |
+
num_return_sequences=self.args.num_return_sequences,
|
| 1076 |
+
#streamer=self.streamer,
|
| 1077 |
+
)
|
| 1078 |
+
all_outputs.extend(outputs.cpu().numpy())
|
| 1079 |
+
|
| 1080 |
+
if self.args.use_multiprocessed_decoding:
|
| 1081 |
+
self.model.to("cpu")
|
| 1082 |
+
with Pool(self.args.process_count) as p:
|
| 1083 |
+
if self.args.multiprocessing_chunksize == -1:
|
| 1084 |
+
chunksize = max(
|
| 1085 |
+
len(all_outputs) // (self.args.process_count * 2), 500
|
| 1086 |
+
)
|
| 1087 |
+
else:
|
| 1088 |
+
chunksize = self.args.multiprocessing_chunksize
|
| 1089 |
+
outputs = list(
|
| 1090 |
+
tqdm(
|
| 1091 |
+
p.imap(self._decode, all_outputs, chunksize=chunksize),
|
| 1092 |
+
total=len(all_outputs),
|
| 1093 |
+
desc="Decoding outputs",
|
| 1094 |
+
disable=self.args.silent,
|
| 1095 |
+
)
|
| 1096 |
+
)
|
| 1097 |
+
self._move_model_to_device()
|
| 1098 |
+
else:
|
| 1099 |
+
outputs = [
|
| 1100 |
+
self.tokenizer.decode(
|
| 1101 |
+
output_id,
|
| 1102 |
+
skip_special_tokens=self.args.skip_special_tokens,
|
| 1103 |
+
clean_up_tokenization_spaces=True,
|
| 1104 |
+
)
|
| 1105 |
+
for output_id in all_outputs
|
| 1106 |
+
]
|
| 1107 |
+
if not split_on_space:
|
| 1108 |
+
outputs = [''.join(gen_text.split(' ')) for gen_text in outputs]
|
| 1109 |
+
if self.args.num_return_sequences > 1:
|
| 1110 |
+
return [
|
| 1111 |
+
outputs[i: i + self.args.num_return_sequences]
|
| 1112 |
+
for i in range(0, len(outputs), self.args.num_return_sequences)
|
| 1113 |
+
]
|
| 1114 |
+
else:
|
| 1115 |
+
return outputs
|
| 1116 |
+
|
| 1117 |
+
def _decode(self, output_id):
|
| 1118 |
+
return self.tokenizer.decode(
|
| 1119 |
+
output_id,
|
| 1120 |
+
skip_special_tokens=self.args.skip_special_tokens,
|
| 1121 |
+
clean_up_tokenization_spaces=True,
|
| 1122 |
+
)
|
| 1123 |
+
|
| 1124 |
+
def compute_metrics(self, labels, preds, **kwargs):
|
| 1125 |
+
"""
|
| 1126 |
+
Computes the evaluation metrics for the model predictions.
|
| 1127 |
+
|
| 1128 |
+
Args:
|
| 1129 |
+
labels: List of target sequences
|
| 1130 |
+
preds: List of model generated outputs
|
| 1131 |
+
**kwargs: Custom metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use).
|
| 1132 |
+
A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs
|
| 1133 |
+
will be lists of strings. Note that this will slow down evaluation significantly as the predicted sequences need to be generated.
|
| 1134 |
+
|
| 1135 |
+
Returns:
|
| 1136 |
+
result: Dictionary containing evaluation results.
|
| 1137 |
+
""" # noqa: ignore flake8"
|
| 1138 |
+
assert len(labels) == len(preds)
|
| 1139 |
+
|
| 1140 |
+
results = {}
|
| 1141 |
+
for metric, func in kwargs.items():
|
| 1142 |
+
results[metric] = func(labels, preds)
|
| 1143 |
+
|
| 1144 |
+
return results
|
| 1145 |
+
|
| 1146 |
+
def _move_model_to_device(self):
|
| 1147 |
+
self.model.to(self.device)
|
| 1148 |
+
|
| 1149 |
+
def _get_inputs_dict(self, batch):
|
| 1150 |
+
if self.args.use_hf_datasets:
|
| 1151 |
+
inputs = {**batch, "labels": batch["input_ids"]}
|
| 1152 |
+
|
| 1153 |
+
return {key: value.to(self.device) for key, value in inputs.items()}
|
| 1154 |
+
else:
|
| 1155 |
+
batch = tuple(t.to(self.device) for t in batch)
|
| 1156 |
+
|
| 1157 |
+
input_ids = batch[0]
|
| 1158 |
+
attention_mask = batch[1]
|
| 1159 |
+
labels = batch[2]
|
| 1160 |
+
labels[labels == self.tokenizer.pad_token_id] = -100
|
| 1161 |
+
|
| 1162 |
+
inputs = {
|
| 1163 |
+
"input_ids": input_ids,
|
| 1164 |
+
"attention_mask": attention_mask,
|
| 1165 |
+
"labels": labels,
|
| 1166 |
+
}
|
| 1167 |
+
|
| 1168 |
+
return inputs
|
| 1169 |
+
|
| 1170 |
+
def load_and_cache_examples(
|
| 1171 |
+
self, data, evaluate=False, no_cache=False, verbose=True, silent=False
|
| 1172 |
+
):
|
| 1173 |
+
"""
|
| 1174 |
+
Creates a T5Dataset from data.
|
| 1175 |
+
|
| 1176 |
+
Utility function for train() and eval() methods. Not intended to be used directly.
|
| 1177 |
+
"""
|
| 1178 |
+
|
| 1179 |
+
tokenizer = self.tokenizer
|
| 1180 |
+
args = self.args
|
| 1181 |
+
|
| 1182 |
+
if not no_cache:
|
| 1183 |
+
no_cache = args.no_cache
|
| 1184 |
+
|
| 1185 |
+
if not no_cache:
|
| 1186 |
+
os.makedirs(self.args.cache_dir, exist_ok=True)
|
| 1187 |
+
|
| 1188 |
+
mode = "dev" if evaluate else "train"
|
| 1189 |
+
|
| 1190 |
+
if self.args.use_hf_datasets:
|
| 1191 |
+
dataset = load_hf_dataset(data, tokenizer, self.args)
|
| 1192 |
+
return dataset
|
| 1193 |
+
elif args.dataset_class:
|
| 1194 |
+
CustomDataset = args.dataset_class
|
| 1195 |
+
return CustomDataset(tokenizer, args, data, mode)
|
| 1196 |
+
else:
|
| 1197 |
+
return T5Dataset(
|
| 1198 |
+
tokenizer,
|
| 1199 |
+
self.args,
|
| 1200 |
+
data,
|
| 1201 |
+
mode,
|
| 1202 |
+
)
|
| 1203 |
+
|
| 1204 |
+
def _create_training_progress_scores(self, **kwargs):
|
| 1205 |
+
extra_metrics = {key: [] for key in kwargs}
|
| 1206 |
+
training_progress_scores = {
|
| 1207 |
+
"global_step": [],
|
| 1208 |
+
"eval_loss": [],
|
| 1209 |
+
"train_loss": [],
|
| 1210 |
+
**extra_metrics,
|
| 1211 |
+
}
|
| 1212 |
+
|
| 1213 |
+
return training_progress_scores
|
| 1214 |
+
|
| 1215 |
+
def _get_last_metrics(self, metric_values):
|
| 1216 |
+
return {metric: values[-1] for metric, values in metric_values.items()}
|
| 1217 |
+
|
| 1218 |
+
def save_model(
|
| 1219 |
+
self, output_dir=None, optimizer=None, scheduler=None, model=None, results=None
|
| 1220 |
+
):
|
| 1221 |
+
if not output_dir:
|
| 1222 |
+
output_dir = self.args.output_dir
|
| 1223 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 1224 |
+
|
| 1225 |
+
if model and not self.args.no_save:
|
| 1226 |
+
# Take care of distributed/parallel training
|
| 1227 |
+
model_to_save = model.module if hasattr(model, "module") else model
|
| 1228 |
+
model_to_save.save_pretrained(output_dir)
|
| 1229 |
+
self.tokenizer.save_pretrained(output_dir)
|
| 1230 |
+
torch.save(self.args, os.path.join(output_dir, "training_args.bin"))
|
| 1231 |
+
if optimizer and scheduler and self.args.save_optimizer_and_scheduler:
|
| 1232 |
+
torch.save(
|
| 1233 |
+
optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt")
|
| 1234 |
+
)
|
| 1235 |
+
torch.save(
|
| 1236 |
+
scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt")
|
| 1237 |
+
)
|
| 1238 |
+
self.save_model_args(output_dir)
|
| 1239 |
+
|
| 1240 |
+
if results:
|
| 1241 |
+
output_eval_file = os.path.join(output_dir, "eval_results.txt")
|
| 1242 |
+
with open(output_eval_file, "w") as writer:
|
| 1243 |
+
for key in sorted(results.keys()):
|
| 1244 |
+
writer.write("{} = {}\n".format(key, str(results[key])))
|
| 1245 |
+
|
| 1246 |
+
def save_model_args(self, output_dir):
|
| 1247 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 1248 |
+
self.args.save(output_dir)
|
| 1249 |
+
|
| 1250 |
+
def _load_model_args(self, input_dir):
|
| 1251 |
+
args = T5Args()
|
| 1252 |
+
args.load(input_dir)
|
| 1253 |
+
return args
|
| 1254 |
+
|
| 1255 |
+
def get_named_parameters(self):
|
| 1256 |
+
return [n for n, p in self.model.named_parameters()]
|
t5/t5_utils.py
ADDED
|
@@ -0,0 +1,214 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""
|
| 3 |
+
@author:XuMing([email protected])
|
| 4 |
+
@description: adjust for chinese tokenizer
|
| 5 |
+
"""
|
| 6 |
+
import os
|
| 7 |
+
import pickle
|
| 8 |
+
from multiprocessing import Pool
|
| 9 |
+
|
| 10 |
+
from datasets import Dataset as HFDataset
|
| 11 |
+
from datasets import load_dataset
|
| 12 |
+
from torch.utils.data import Dataset
|
| 13 |
+
from tqdm.auto import tqdm
|
| 14 |
+
from rouge import Rouge
|
| 15 |
+
from loguru import logger
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
def preprocess_batch_for_hf_dataset(dataset, tokenizer, args):
|
| 19 |
+
if args.preprocess_inputs:
|
| 20 |
+
return tokenizer.prepare_seq2seq_batch(
|
| 21 |
+
src_texts=[
|
| 22 |
+
prefix + ": " + input_text
|
| 23 |
+
for prefix, input_text in zip(dataset["prefix"], dataset["input_text"])
|
| 24 |
+
],
|
| 25 |
+
tgt_texts=dataset["target_text"],
|
| 26 |
+
max_length=args.max_seq_length,
|
| 27 |
+
max_target_length=args.max_length,
|
| 28 |
+
padding="max_length",
|
| 29 |
+
return_tensors="np",
|
| 30 |
+
truncation=True,
|
| 31 |
+
)
|
| 32 |
+
else:
|
| 33 |
+
return tokenizer.prepare_seq2seq_batch(
|
| 34 |
+
src_texts=[
|
| 35 |
+
prefix + input_text
|
| 36 |
+
for prefix, input_text in zip(dataset["prefix"], dataset["input_text"])
|
| 37 |
+
],
|
| 38 |
+
tgt_texts=dataset["target_text"],
|
| 39 |
+
max_length=args.max_seq_length,
|
| 40 |
+
max_target_length=args.max_length,
|
| 41 |
+
padding="max_length",
|
| 42 |
+
return_tensors="np",
|
| 43 |
+
truncation=True,
|
| 44 |
+
)
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
def load_hf_dataset(data, tokenizer, args):
|
| 48 |
+
if isinstance(data, str):
|
| 49 |
+
dataset = load_dataset(
|
| 50 |
+
"csv",
|
| 51 |
+
data_files=data,
|
| 52 |
+
delimiter="\t",
|
| 53 |
+
download_mode="force_redownload"
|
| 54 |
+
if args.reprocess_input_data
|
| 55 |
+
else "reuse_dataset_if_exists",
|
| 56 |
+
)
|
| 57 |
+
else:
|
| 58 |
+
dataset = HFDataset.from_pandas(data)
|
| 59 |
+
|
| 60 |
+
dataset = dataset.map(
|
| 61 |
+
lambda x: preprocess_batch_for_hf_dataset(x, tokenizer=tokenizer, args=args),
|
| 62 |
+
batched=True,
|
| 63 |
+
)
|
| 64 |
+
|
| 65 |
+
dataset.set_format(type="pt", columns=["input_ids", "attention_mask"])
|
| 66 |
+
|
| 67 |
+
if isinstance(data, str):
|
| 68 |
+
# This is not necessarily a train dataset. The datasets library insists on calling it train.
|
| 69 |
+
return dataset["train"]
|
| 70 |
+
else:
|
| 71 |
+
return dataset
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
def preprocess_data(data):
|
| 75 |
+
prefix, input_text, target_text, tokenizer, args = data
|
| 76 |
+
|
| 77 |
+
# Add EOS again if truncated?
|
| 78 |
+
if args.preprocess_inputs:
|
| 79 |
+
batch = tokenizer.prepare_seq2seq_batch(
|
| 80 |
+
src_texts=[prefix + ": " + input_text],
|
| 81 |
+
tgt_texts=[target_text],
|
| 82 |
+
max_length=args.max_seq_length,
|
| 83 |
+
padding="max_length",
|
| 84 |
+
return_tensors="pt",
|
| 85 |
+
truncation=True,
|
| 86 |
+
)
|
| 87 |
+
else:
|
| 88 |
+
batch = tokenizer.prepare_seq2seq_batch(
|
| 89 |
+
src_texts=[prefix + ": " + input_text],
|
| 90 |
+
tgt_texts=[target_text],
|
| 91 |
+
max_length=args.max_seq_length,
|
| 92 |
+
padding="max_length",
|
| 93 |
+
return_tensors="pt",
|
| 94 |
+
truncation=True,
|
| 95 |
+
)
|
| 96 |
+
input_ids = batch["input_ids"][0]
|
| 97 |
+
attention_mask = batch["attention_mask"][0]
|
| 98 |
+
labels = batch["labels"][0]
|
| 99 |
+
return (input_ids, attention_mask, labels)
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
class T5Dataset(Dataset):
|
| 103 |
+
def __init__(self, tokenizer, args, data, mode):
|
| 104 |
+
cached_features_file = os.path.join(
|
| 105 |
+
args.cache_dir,
|
| 106 |
+
args.model_name.replace("/", "_")
|
| 107 |
+
+ "_cached_"
|
| 108 |
+
+ str(args.max_seq_length)
|
| 109 |
+
+ str(len(data)),
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
if os.path.exists(cached_features_file) and (
|
| 113 |
+
(not args.reprocess_input_data and not args.no_cache)
|
| 114 |
+
or (mode == "dev" and args.use_cached_eval_features and not args.no_cache)
|
| 115 |
+
):
|
| 116 |
+
logger.info(" Loading features from cached file %s" % cached_features_file)
|
| 117 |
+
with open(cached_features_file, "rb") as handle:
|
| 118 |
+
self.examples = pickle.load(handle)
|
| 119 |
+
else:
|
| 120 |
+
logger.info(" Creating features from dataset file at %s" % args.cache_dir)
|
| 121 |
+
|
| 122 |
+
data = [
|
| 123 |
+
(prefix, input_text, target_text, tokenizer, args)
|
| 124 |
+
for prefix, input_text, target_text in zip(
|
| 125 |
+
data["prefix"], data["input_text"], data["target_text"]
|
| 126 |
+
)
|
| 127 |
+
]
|
| 128 |
+
|
| 129 |
+
if (mode == "train" and args.use_multiprocessing) or (
|
| 130 |
+
mode == "dev" and args.use_multiprocessing_for_evaluation
|
| 131 |
+
):
|
| 132 |
+
if args.multiprocessing_chunksize == -1:
|
| 133 |
+
chunksize = max(len(data) // (args.process_count * 2), 500)
|
| 134 |
+
else:
|
| 135 |
+
chunksize = args.multiprocessing_chunksize
|
| 136 |
+
|
| 137 |
+
with Pool(args.process_count) as p:
|
| 138 |
+
self.examples = list(
|
| 139 |
+
tqdm(
|
| 140 |
+
p.imap(preprocess_data, data, chunksize=chunksize),
|
| 141 |
+
total=len(data),
|
| 142 |
+
disable=args.silent,
|
| 143 |
+
)
|
| 144 |
+
)
|
| 145 |
+
else:
|
| 146 |
+
self.examples = [preprocess_data(d) for d in tqdm(data, disable=args.silent)]
|
| 147 |
+
if not args.no_cache:
|
| 148 |
+
logger.info(" Saving features into cached file %s" % cached_features_file)
|
| 149 |
+
with open(cached_features_file, "wb") as handle:
|
| 150 |
+
pickle.dump(self.examples, handle, protocol=pickle.HIGHEST_PROTOCOL)
|
| 151 |
+
|
| 152 |
+
def __len__(self):
|
| 153 |
+
return len(self.examples)
|
| 154 |
+
|
| 155 |
+
def __getitem__(self, index):
|
| 156 |
+
return self.examples[index]
|
| 157 |
+
|
| 158 |
+
|
| 159 |
+
def dynamic_lcs(X, Y):
|
| 160 |
+
# find the length of the strings
|
| 161 |
+
m = len(X)
|
| 162 |
+
n = len(Y)
|
| 163 |
+
|
| 164 |
+
# declaring the array for storing the dp values
|
| 165 |
+
L = [[None] * (n + 1) for i in range(m + 1)]
|
| 166 |
+
|
| 167 |
+
"""Following steps build L[m + 1][n + 1] in bottom up fashion
|
| 168 |
+
Note: L[i][j] contains length of LCS of X[0..i-1]
|
| 169 |
+
and Y[0..j-1]"""
|
| 170 |
+
for i in range(m + 1):
|
| 171 |
+
for j in range(n + 1):
|
| 172 |
+
if i == 0 or j == 0:
|
| 173 |
+
L[i][j] = 0
|
| 174 |
+
elif X[i - 1] == Y[j - 1]:
|
| 175 |
+
L[i][j] = L[i - 1][j - 1] + 1
|
| 176 |
+
else:
|
| 177 |
+
L[i][j] = max(L[i - 1][j], L[i][j - 1])
|
| 178 |
+
|
| 179 |
+
# L[m][n] contains the length of LCS of X[0..n-1] & Y[0..m-1]
|
| 180 |
+
return L[m][n]
|
| 181 |
+
|
| 182 |
+
|
| 183 |
+
def f1_sim(text_a, text_b):
|
| 184 |
+
"""F1相似度
|
| 185 |
+
说明:算出两个文本的最长公共子序列长度,然后乘2并处以两者
|
| 186 |
+
长度之和。
|
| 187 |
+
脚本见:https://github.com/CLUEbenchmark/pCLUE/blob/main/evaluate_pclue.py
|
| 188 |
+
计算pCLUE任务总分,及子分数
|
| 189 |
+
"""
|
| 190 |
+
if not text_a and not text_b:
|
| 191 |
+
return 0.
|
| 192 |
+
lcs_len = dynamic_lcs(text_a, text_b)
|
| 193 |
+
return 2. * lcs_len / (len(text_a) + len(text_b))
|
| 194 |
+
|
| 195 |
+
|
| 196 |
+
def rouge_l_zh(target, pred):
|
| 197 |
+
"""计算Rouge-l得分,Rouge-l指标常用于评估自动文本摘要及翻译任务
|
| 198 |
+
target: 真实标签
|
| 199 |
+
pred: 预测标签"""
|
| 200 |
+
|
| 201 |
+
if not (isinstance(target, str) or isinstance(pred, str)):
|
| 202 |
+
logger.error("target或pred为非字符串!请检查!")
|
| 203 |
+
return 0
|
| 204 |
+
rouge = Rouge()
|
| 205 |
+
scores = rouge.get_scores(" ".join(list(pred)), " ".join(list(target)))
|
| 206 |
+
score = scores[0]["rouge-l"]
|
| 207 |
+
return score["f"]
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
if __name__ == '__main__':
|
| 211 |
+
a = '123444'
|
| 212 |
+
b = '23411'
|
| 213 |
+
print(f1_sim(a, b))
|
| 214 |
+
print(dynamic_lcs(a, b))
|