forked from PaddlePaddle/PaddleMIX
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun_train.py
418 lines (355 loc) · 17.7 KB
/
run_train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
os.environ["FLAGS_use_cuda_managed_memory"] = "true"
import random
from dataclasses import dataclass, field
import numpy as np
import paddle
import paddle.distributed as dist
import paddle.nn as nn
from paddle.distributed import fleet
from paddle.distributed.fleet.meta_parallel import get_rng_state_tracker
from paddlenlp.ops import transfer_param
from paddlenlp.trainer import PdArgumentParser, TrainingArguments
from paddlenlp.transformers import LlamaForCausalLM
from paddlemix import (
MiniGPT4Config,
MiniGPT4ForConditionalGeneration,
MiniGPT4Processor,
MiniGPT4QFormerModel,
MiniGPT4VisionModel,
)
from paddlemix.datasets import load_dataset
from paddlemix.trainer.minigpt4_trainer import MiniGPT4Trainer as Trainer
from paddlemix.utils import paddlemix_load
from paddlemix.utils.log import logger
from paddlemix.utils.parameters import freeze_parameters
@dataclass
class DataArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
Using `PdArgumentParser` we can turn this class
into argparse arguments to be able to specify them on
the command line.
"""
task_name: str = field(
default="cc_sbu_dataset",
metadata={"help": "The name of the task to use (via the datasets library)."},
)
text_path: str = field(
default="data/texts.txt",
metadata={"help": "The text file recording text used as prompt."},
)
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
"""
pretrained_model_name_or_path: str = field(
default=None,
metadata={"help": "The directory path to save pretrained model or model identifier"},
)
@dataclass
class PreTrainingArguments(TrainingArguments):
"""
Arguments pertaining to what training options we are going to use during pretraining.
"""
pretrained_model_path: str = field(
default=None,
metadata={"help": "The path to pre-trained model that we will use for pretraining."},
)
# batch_size: int = field(
# default=12,
# metadata={"help": "Number of samples in one batch."}
# )
weight_decay: float = field(default=0.05, metadata={"help": "Weight decay if we apply some."})
learning_rate: float = field(default=3e-5, metadata={"help": "The initial learning rate."})
num_train_epochs: float = field(default=200, metadata={"help": "Total number of training epochs to perform."})
warmup_start_lr: float = field(default=1e-6, metadata={"help": "Initial learning rate of warm up."})
eta_min: float = field(default=1e-5, metadata={"help": "The minimum value of learning rate."})
# warmup_steps: int = field(
# default=200, metadata={"help": "Number of warmup steps."}
# )
warmup: int = field(default=200, metadata={"help": "warmup ratio or steps."})
lr_scheduler_name: str = field(default="CosineDecayWithWarmup", metadata={"help": "The scheduler name to use."})
per_device_train_batch_size: int = field(
default=6, metadata={"help": "Batch size per GPU core/CPU for training. (default: 8)"}
)
per_device_eval_batch_size: int = field(
default=6, metadata={"help": " Batch size per GPU core/CPU for evaluation. (default:8)"}
)
output_dir: str = field(default="./checkpoints", metadata={"help": "The directory name for saving checkpoint"})
do_eval: bool = field(default=False, metadata={"help": "Whether to evaluation."})
do_train: bool = field(default=True, metadata={"help": "Whether to train."})
logging_steps: int = field(default=50, metadata={"help": "Logging interval"})
evaluation_strategy: str = field(default="no", metadata={"help": "Evaluation strategy (epoch/steps/no)"})
fp16_opt_level: str = field(default="O1", metadata={"help": "Mixed Precision Type"})
fp16: bool = field(default=True, metadata={"help": "Whether to use mixed Precision"})
gradient_checkpointing: bool = field(
default=False, metadata={"help": "Forward recompute for saving graphics memory"}
)
tensor_parallel_degree: int = field(default=1, metadata={"help": "Set the number of tensor model parallel"})
sharding_parallel_degree: int = field(
default=1, metadata={"help": "Set the number of sharding, enable sharding parallel"}
)
pipeline_parallel_degree: int = field(default=1, metadata={"help": "Enable pipeline parallel"})
use_amp: str = field(default=True, metadata={"help": "Whether to use amp for training."})
warmup_proportion: float = field(default=0.1, metadata={"help": "The warmup rate."})
freeze_vit: float = field(default=True, metadata={"help": "Whether to freeze vit."})
freeze_qformer: float = field(default=True, metadata={"help": "Whether to freeze Qformer."})
freeze_llama: float = field(default=True, metadata={"help": "Whether to freeze Llama."})
seed: int = field(default=42, metadata={"help": "The random seed."})
log_freq: int = field(default=1, metadata={"help": "The log frequency."})
num_workers: int = field(default=0, metadata={"help": "The random seed."})
resume_from_checkpoint: str = field(
default=None,
metadata={"help": "The path to a folder with a valid checkpoint for your model."},
)
model_path: str = field(
default=None,
metadata={"help": "The path to model if you want to load weights from the specified path"},
)
class MiniGPT4Collator:
"""
Data collator that will dynamically pad the inputs to the longest sequence in the batch.
Args:
processor (`paddlemix.processors.ProcessorMixin`):
The processor used for pre-process the data.
"""
def __init__(self, processor, mode="test"):
self.processor = processor
self.mode = mode
def __call__(self, data_list):
images = [sample["image"] for sample in data_list]
target_texts = [sample["text_input"] for sample in data_list]
# random text from text_list read by processor and combine it with default prompt
batch_data = self.processor(images=images, mode="train")
target_outputs = self.processor.process_target_texts(target_texts)
batch_data.update(target_outputs)
return batch_data
def set_seed(seed):
paddle.seed(seed)
random.seed(seed)
np.random.seed(seed)
def create_model(model_args):
config = MiniGPT4Config.from_pretrained(model_args.pretrained_model_name_or_path)
model = MiniGPT4ForConditionalGeneration(config)
return model
# TODO, better to split qformer, vit and llama for config and checkpoint
def load_pretrained_model(model, pretrained_model_path, del_keys=[]):
if pretrained_model_path is None:
return
if not os.path.exists(pretrained_model_path):
raise ValueError("Cannot find pretrained model path: {}".format(pretrained_model_path))
state_dict = paddlemix_load(pretrained_model_path, map_location="cpu")
for key in del_keys:
state_dict.pop(key)
for key in model.state_dict().keys():
if key in state_dict.keys():
if state_dict[key].shape != model.state_dict()[key].shape:
logger.warning("{}'s shape in model is not equal to the pretrained model checkpoint's".format(key))
del state_dict[key]
model.set_state_dict(state_dict)
def convert_weights_to_dtype(model, dtype: str):
# trying to convert model dtype if necessary
if dtype not in ["float16", "float32", "float64"]:
raise ValueError("Not supported dtype: {}., only [float16, float32, float64] supported.".format(dtype))
dtype_mapping = {
"float16": paddle.float16,
"float32": paddle.float32,
"float64": paddle.float64,
}
def convert_for_vit(layer):
if isinstance(layer, (nn.Linear, nn.Conv1D, nn.Conv2D)):
if layer.weight.dtype != dtype_mapping[dtype]:
layer.weight = transfer_param(layer.weight, restore_data=True, dtype=dtype)
if layer.bias is not None and layer.bias.dtype != dtype_mapping[dtype]:
layer.bias = transfer_param(layer.bias, restore_data=True, dtype=dtype)
if isinstance(model, MiniGPT4VisionModel):
model.apply(convert_for_vit)
elif isinstance(model, (MiniGPT4QFormerModel, LlamaForCausalLM)):
model.to(dtype=dtype)
else:
raise TypeError("Not support model type: {}.".format(type(model)))
def setdistenv(args):
if args.tensor_parallel_degree * args.sharding_parallel_degree * args.pipeline_parallel_degree != 1:
args.use_hybrid_parallel = True
args.dp_degree = dist.get_world_size() // (
args.tensor_parallel_degree * args.sharding_parallel_degree * args.pipeline_parallel_degree
)
strategy = fleet.DistributedStrategy()
if args.tensor_parallel_degree > 1:
strategy.tensor_parallel = True
args.data_parallel_degree = args.dp_degree
logger.info("args.dp_degree:{}".format(args.dp_degree))
logger.info("args.sharding_parallel_degree):{}".format(args.sharding_parallel_degree))
strategy.hybrid_configs = {
"dp_degree": args.dp_degree,
"mp_degree": args.tensor_parallel_degree,
"sharding_degree": args.sharding_parallel_degree,
"pp_degree": args.pipeline_parallel_degree,
}
BATCH_SIZE = 128
MICRO_BATCH_SIZE = 32
strategy.pipeline_configs = {
"accumulate_steps": BATCH_SIZE // MICRO_BATCH_SIZE,
"micro_batch_size": MICRO_BATCH_SIZE,
}
strategy.find_unused_parameters = True
# set control in tensor parallel
strategy.tensor_parallel_configs = {"tensor_init_seed": args.seed}
fleet.init(is_collective=True, strategy=strategy)
args.rank = dist.get_rank()
# obtain rank message of hybrid parallel
hcg = fleet.get_hybrid_communicate_group()
args.mp_rank = hcg.get_model_parallel_rank()
args.dp_rank = hcg.get_data_parallel_rank()
args.sharding_rank = hcg.get_sharding_parallel_rank()
args.data_world_rank = args.dp_rank * args.sharding_parallel_degree + args.sharding_rank
args.data_world_size = dist.get_world_size() // abs(args.tensor_parallel_degree * args.pipeline_parallel_degree)
# seed control in hybrid parallel
set_hybrid_parallel_seed(args.seed, args.data_world_rank, args.mp_rank)
def set_hybrid_parallel_seed(basic_seed, data_world_rank, mp_rank, pp_rank=0):
device_id = paddle.device.get_device()
assert "gpu" in device_id
random.seed(basic_seed + data_world_rank)
np.random.seed(basic_seed + data_world_rank)
paddle.seed(basic_seed + data_world_rank)
# TODO add manual_seed
# local_seed/ global_seed is used to control dropout in ModelParallel
local_seed = 1024 + basic_seed + mp_rank * 100 + data_world_rank
global_seed = 2048 + basic_seed + data_world_rank
tracker = get_rng_state_tracker()
tracker.add("global_seed", global_seed)
tracker.add("local_seed", local_seed)
def main():
# load data, model and training parameters
parser = PdArgumentParser((ModelArguments, DataArguments, PreTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
training_args.print_config(model_args, "Model")
training_args.print_config(data_args, "Data")
setdistenv(training_args)
model_args.data_world_rank = training_args.data_world_rank
model_args.data_world_size = training_args.data_world_size
model_args.mp_degree = training_args.tensor_parallel_degree
model_args.gradient_checkpointing = training_args.gradient_checkpointing
paddle.set_device(training_args.device)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, world_size: {training_args.world_size}, "
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16 or training_args.bf16}"
)
# Detecting last checkpoint
# last_checkpoint = None
# if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
# last_checkpoint = get_last_checkpoint(training_args.output_dir)
# if last_checkpoint is not None and training_args.resume_from_checkpoint is None:
# logger.info(
# f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
# "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
# )
# load and convert dataset
processor = MiniGPT4Processor.from_pretrained(model_args.pretrained_model_name_or_path)
processor.read_texts(data_args.text_path)
minigpt4_collator = MiniGPT4Collator(processor)
dataset = load_dataset("cc_sbu_dataset", SPLITS=["train"])
# batch_sampler = BatchSampler(dataset, batch_size=training_args.batch_size, shuffle=True, drop_last=True)
# train_loader = DataLoader(dataset, batch_sampler=batch_sampler, collate_fn=minigpt4_collator, num_workers=training_args.num_workers)
# load MiniGPT4 model for training
model = create_model(model_args)
# if you wanna train from scratch, you can set del_keys = ["language_projection.weight", "language_projection.bias"]
# del_keys = []
# logger.info("Try to load the specified model.")
# load_pretrained_model(model, training_args.pretrained_model_path, del_keys=del_keys)
# logger.info("Try to convert the model dtype to the specified dtype.")
# convert_weights_to_dtype(model.vision_model, dtype="float16")
# convert_weights_to_dtype(model.qformer, dtype="float32")
# convert_weights_to_dtype(model.language_model, dtype="float16")
# logger.info("Try to freeze model parameters.")
if training_args.freeze_vit:
freeze_parameters(model.vision_model, enable_eval=True)
if training_args.freeze_qformer:
freeze_parameters(model.query_tokens)
freeze_parameters(model.qformer, enable_eval=True)
if training_args.freeze_llama:
freeze_parameters(model.language_model, enable_eval=False)
logger.info("Initializing the model done!")
logger.info("training_args.use_hybrid_parallel:{}".format(training_args.use_hybrid_parallel))
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset,
data_collator=minigpt4_collator,
processor=processor,
tokenizer=processor.tokenizer,
)
# Training
checkpoint = None
# if training_args.model_path is not None:
# checkpoint = training_args.model_path
# load_model(training_args, model, ckpt_dir=model_args.model_path, load_language_model=False)
# load_model(training_args, model.language_model, ckpt_dir=LLM_LIST[model_args.text_model_name_or_path])
# if training_args.resume_from_checkpoint is not None:
# checkpoint = os.path.join(training_args.resume_from_checkpoint, "model_state.pdparams")
# load_model(training_args, model, ckpt_dir=checkpoint, load_language_model=False)
# load_model(training_args, model.language_model, ckpt_dir=LLM_LIST[model_args.text_model_name_or_path])
# if training_args.do_eval:
# eval_metrics = trainer.evaluate(eval_dataset)
# trainer.log_metrics("eval", eval_metrics)
if training_args.do_train:
trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
trainer.save_state()
# training setting
# num_training_steps = training_args.num_train_epochs * len(train_loader)
# lr_scheduler = CosineDecayWithWarmup(
# learning_rate=training_args.learning_rate,
# total_steps=num_training_steps,
# eta_min=training_args.eta_min,
# warmup=training_args.warmup_steps,
# warmup_start_lr=training_args.warmup_start_lr,
# last_step=-1,
# )
# grouped_params = get_grouped_parameters(model, training_args)
# optimizer = paddle.optimizer.AdamW(
# learning_rate=lr_scheduler,
# parameters=grouped_params,
# weight_decay=training_args.weight_decay,
# )
# if training_args.use_amp:
# scaler = paddle.amp.GradScaler(init_loss_scaling=65536.0, incr_every_n_steps=2000, decr_every_n_nan_or_inf=1)
# # start to train MiniGPT4
# for epoch in range(training_args.num_train_epochs):
# for step, batch_data in enumerate(train_loader):
# with paddle.amp.auto_cast(enable=training_args.use_amp, custom_white_list={}, level="O1"):
# outputs = model(**batch_data, return_dict=True)
# loss = outputs.loss
# if step % training_args.log_freq == 0:
# print("epoch: {}, step: {}, lr: {}, loss: {}".format(epoch, step, lr_scheduler.get_lr(), loss.item()))
# if training_args.use_amp:
# scaled = scaler.scale(loss)
# scaled.backward()
# scaler.step(optimizer)
# scaler.update()
# else:
# loss.backward()
# optimizer.step()
# lr_scheduler.step()
# optimizer.clear_grad()
# # save model
# model.save_pretrained(training_args.output_dir)
# processor.tokenizer.save_pretrained(training_args.output_dir)
if __name__ == "__main__":
main()