151 lines
4.9 KiB
Python
151 lines
4.9 KiB
Python
"""Adapter utils."""
|
|
|
|
from pathlib import Path
|
|
from typing import Any, Callable, Dict, List, Optional, Type
|
|
|
|
import torch
|
|
import transformers
|
|
from sentence_transformers.util import cos_sim
|
|
from torch import Tensor, nn
|
|
from torch.optim import Optimizer
|
|
from tqdm.autonotebook import trange
|
|
|
|
from llama_index.embeddings.adapter_utils import BaseAdapter
|
|
from llama_index.utils import print_text
|
|
|
|
|
|
class MyMultipleNegativesRankingLoss(nn.Module):
|
|
"""Multiple negatives ranking loss.
|
|
|
|
This loss is similar to the one in sentence_transformers,
|
|
but optimized for our own embeddings.
|
|
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
model: BaseAdapter,
|
|
scale: float = 20.0,
|
|
similarity_fct: Optional[Callable] = None,
|
|
):
|
|
"""Define ranking loss."""
|
|
super().__init__()
|
|
self.model = model
|
|
self.scale = scale
|
|
self.similarity_fct = cos_sim if similarity_fct is None else similarity_fct
|
|
self.cross_entropy_loss = nn.CrossEntropyLoss()
|
|
|
|
def forward(self, query_embeds: Tensor, context_embeds: Tensor) -> Tensor:
|
|
"""Forward pass."""
|
|
# transform context embeds
|
|
# context_embeds_2 = self.model.forward(context_embeds)
|
|
query_embeds_2 = self.model.forward(query_embeds)
|
|
|
|
scores = self.similarity_fct(query_embeds_2, context_embeds) * self.scale
|
|
labels = torch.tensor(
|
|
range(len(scores)), dtype=torch.long, device=scores.device
|
|
)
|
|
return self.cross_entropy_loss(scores, labels)
|
|
|
|
|
|
def train_model(
|
|
model: BaseAdapter,
|
|
data_loader: torch.utils.data.DataLoader,
|
|
device: torch.device,
|
|
epochs: int = 1,
|
|
steps_per_epoch: Optional[int] = None,
|
|
warmup_steps: int = 10000,
|
|
optimizer_class: Type[Optimizer] = torch.optim.AdamW,
|
|
optimizer_params: Dict[str, Any] = {"lr": 2e-5},
|
|
output_path: str = "model_output",
|
|
max_grad_norm: float = 1,
|
|
show_progress_bar: bool = True,
|
|
verbose: bool = False,
|
|
# callback: Callable[[float, int, int], None] = None,
|
|
# scheduler: str = "WarmupLinear",
|
|
# weight_decay: float = 0.01,
|
|
# evaluation_steps: int = 0,
|
|
# save_best_model: bool = True,
|
|
# use_amp: bool = False, # disable this option for now
|
|
checkpoint_path: Optional[str] = None,
|
|
checkpoint_save_steps: int = 500,
|
|
# checkpoint_save_total_limit: int = 0,
|
|
) -> None:
|
|
"""Train model."""
|
|
model.to(device)
|
|
# TODO: hardcode loss now, make customizable later
|
|
loss_model = MyMultipleNegativesRankingLoss(model=model)
|
|
loss_model.to(device)
|
|
|
|
# prepare optimizer/scheduler
|
|
param_optimizer = list(model.named_parameters())
|
|
optimizer_grouped_parameters: List[Dict[str, Any]] = [
|
|
{
|
|
"params": [p for n, p in param_optimizer],
|
|
},
|
|
]
|
|
optimizer = optimizer_class(optimizer_grouped_parameters, **optimizer_params)
|
|
if steps_per_epoch is None or steps_per_epoch == 0:
|
|
steps_per_epoch = len(data_loader)
|
|
num_train_steps = int(steps_per_epoch * epochs)
|
|
scheduler_obj = transformers.get_linear_schedule_with_warmup(
|
|
optimizer, num_warmup_steps=warmup_steps, num_training_steps=num_train_steps
|
|
)
|
|
|
|
if verbose:
|
|
print_text("> Prepared optimizer, scheduler, and loss model.\n", color="blue")
|
|
|
|
global_step = 0
|
|
data_iterator = iter(data_loader)
|
|
|
|
# if checkpoint_path is specified, create if doesn't exist
|
|
if checkpoint_path is not None:
|
|
Path(checkpoint_path).mkdir(parents=True, exist_ok=True)
|
|
|
|
for epoch in trange(epochs, desc="Epoch", disable=not show_progress_bar):
|
|
training_steps = 0
|
|
loss_model.zero_grad()
|
|
loss_model.train()
|
|
for _ in trange(
|
|
steps_per_epoch,
|
|
desc="Iteration",
|
|
smoothing=0.05,
|
|
disable=not show_progress_bar,
|
|
):
|
|
try:
|
|
data = next(data_iterator)
|
|
except StopIteration:
|
|
data_iterator = iter(data_loader)
|
|
data = next(data_iterator)
|
|
|
|
query, context = data
|
|
context = context.to(device)
|
|
query = query.to(device)
|
|
|
|
loss_value = loss_model(query, context)
|
|
if verbose:
|
|
print_text(
|
|
f"> [Epoch {epoch}] Current loss: {loss_value}\n", color="blue"
|
|
)
|
|
loss_value.backward()
|
|
torch.nn.utils.clip_grad_norm_(loss_model.parameters(), max_grad_norm)
|
|
optimizer.step()
|
|
|
|
optimizer.zero_grad()
|
|
|
|
scheduler_obj.step()
|
|
|
|
training_steps += 1
|
|
global_step += 1
|
|
|
|
# TODO: skip eval for now
|
|
if checkpoint_path is not None and global_step % checkpoint_save_steps == 0:
|
|
full_ck_path = Path(checkpoint_path) / f"step_{global_step}"
|
|
model.save(str(full_ck_path))
|
|
|
|
if verbose:
|
|
print_text(f"> Finished training, saving to {output_path}\n", color="blue")
|
|
|
|
# save model
|
|
model.save(output_path)
|