ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

lazy image load

Browse Source 
Former-commit-id: cdd733b575411e003bc5ffd6560dd8eff8aa09cf
This commit is contained in:
hiyouga
2024-09-04 02:27:08 +08:00
parent fed7ae5661
commit 7056087e92
19 changed files with 353 additions and 366 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
src/llamafactory/data/processors/feedback.py
View File

@@ -21,10 +21,10 @@ from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from PIL.Image import Image
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..mm_plugin import ImageInput
from ..template import Template
@@ -37,12 +37,12 @@ def _encode_feedback_example(
kl_response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
images: Sequence["ImageInput"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int], List[int], List[int], bool, Dict[str, Any]]:
) -> Tuple[List[int], List[int], List[int], List[int], bool]:
if response[0]["content"]: # desired example
kto_tag = True
messages = prompt + [response[0]]
@@ -78,15 +78,7 @@ def _encode_feedback_example(
labels = [IGNORE_INDEX] * source_len + response_ids
kl_input_ids = kl_prompt_ids + kl_response_ids
kl_labels = [IGNORE_INDEX] * kl_source_len + kl_response_ids
extra_inputs = template.mm_plugin.get_mm_inputs(
images=images,
feature_seqlens={
"token_type_ids": len(input_ids),
"kl_token_type_ids": len(kl_input_ids),
},
processor=processor,
)
return input_ids, labels, kl_input_ids, kl_labels, kto_tag, extra_inputs
return input_ids, labels, kl_input_ids, kl_labels, kto_tag
def preprocess_feedback_dataset(
@@ -97,20 +89,20 @@ def preprocess_feedback_dataset(
data_args: "DataArguments",
) -> Dict[str, List[Any]]:
# create unrelated input-output pairs for estimating the KL term by flipping the matched pairs
kl_response = examples["response"][::-1]
kl_response = examples["_response"][::-1]
model_inputs = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
for i in range(len(examples["_prompt"])):
if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) < 2:
logger.warning("Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i]))
continue
input_ids, labels, kl_input_ids, kl_labels, kto_tag, extra_inputs = _encode_feedback_example(
prompt=examples["prompt"][i],
response=examples["response"][i],
input_ids, labels, kl_input_ids, kl_labels, kto_tag = _encode_feedback_example(
prompt=examples["_prompt"][i],
response=examples["_response"][i],
kl_response=kl_response[i],
system=examples["system"][i],
tools=examples["tools"][i],
images=examples["images"][i],
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -123,8 +115,7 @@ def preprocess_feedback_dataset(
model_inputs["kl_attention_mask"].append([1] * len(kl_input_ids))
model_inputs["kl_labels"].append(kl_labels)
model_inputs["kto_tags"].append(kto_tag)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
model_inputs["images"].append(examples["_images"][i])
desirable_num = sum([1 for tag in model_inputs["kto_tags"] if tag])
undesirable_num = len(model_inputs["kto_tags"]) - desirable_num

37
src/llamafactory/data/processors/pairwise.py
View File

@@ -21,10 +21,10 @@ from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from PIL.Image import Image
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..mm_plugin import ImageInput
from ..template import Template
@@ -36,12 +36,12 @@ def _encode_pairwise_example(
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
images: Sequence["ImageInput"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int], List[int], List[int], Dict[str, Any]]:
) -> Tuple[List[int], List[int], List[int], List[int]]:
chosen_messages = template.mm_plugin.process_messages(prompt + [response[0]], images, processor)
rejected_messages = template.mm_plugin.process_messages(prompt + [response[1]], images, processor)
prompt_ids, chosen_ids = template.encode_oneturn(tokenizer, chosen_messages, system, tools)
@@ -62,15 +62,7 @@ def _encode_pairwise_example(
chosen_labels = [IGNORE_INDEX] * source_len + chosen_ids
rejected_input_ids = prompt_ids + rejected_ids
rejected_labels = [IGNORE_INDEX] * source_len + rejected_ids
extra_inputs = template.mm_plugin.get_mm_inputs(
images=images,
feature_seqlens={
"chosen_token_type_ids": len(chosen_input_ids),
"rejected_token_type_ids": len(rejected_input_ids),
},
processor=processor,
)
return chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels, extra_inputs
return chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels
def preprocess_pairwise_dataset(
@@ -82,17 +74,17 @@ def preprocess_pairwise_dataset(
) -> Dict[str, List[Any]]:
# build input pairs with format `<bos> X`, `Y1 <eos>` and `Y2 <eos>`
model_inputs = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
for i in range(len(examples["_prompt"])):
if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) < 2:
logger.warning("Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i]))
continue
chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels, extra_inputs = _encode_pairwise_example(
prompt=examples["prompt"][i],
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
images=examples["images"][i],
chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels = _encode_pairwise_example(
prompt=examples["_prompt"][i],
response=examples["_response"][i],
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -104,8 +96,7 @@ def preprocess_pairwise_dataset(
model_inputs["rejected_input_ids"].append(rejected_input_ids)
model_inputs["rejected_attention_mask"].append([1] * len(rejected_input_ids))
model_inputs["rejected_labels"].append(rejected_labels)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
model_inputs["images"].append(examples["_images"][i])
return model_inputs

2
src/llamafactory/data/processors/pretrain.py
View File

@@ -30,7 +30,7 @@ def preprocess_pretrain_dataset(
) -> Dict[str, List[Any]]:
# build grouped texts with format `X1 X2 X3 ...` if packing is enabled
eos_token = "<|end_of_text|>" if data_args.template == "llama3" else tokenizer.eos_token
text_examples = [messages[0]["content"] + eos_token for messages in examples["prompt"]]
text_examples = [messages[0]["content"] + eos_token for messages in examples["_prompt"]]
if not data_args.packing:
if data_args.template == "gemma":

51
src/llamafactory/data/processors/supervised.py
View File

@@ -21,10 +21,10 @@ from .processor_utils import greedy_knapsack, infer_seqlen
if TYPE_CHECKING:
from PIL.Image import Image
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..mm_plugin import ImageInput
from ..template import Template
@@ -36,14 +36,14 @@ def _encode_supervised_example(
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
images: Sequence["ImageInput"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
train_on_prompt: bool,
mask_history: bool,
) -> Tuple[List[int], List[int], Dict[str, Any]]:
) -> Tuple[List[int], List[int]]:
messages = template.mm_plugin.process_messages(prompt + response, images, processor)
input_ids, labels = template.mm_plugin.process_token_ids([], [], images, tokenizer, processor)
encoded_pairs = template.encode_multiturn(tokenizer, messages, system, tools)
@@ -83,10 +83,7 @@ def _encode_supervised_example(
input_ids += [tokenizer.eos_token_id]
labels += [tokenizer.eos_token_id]
extra_inputs = template.mm_plugin.get_mm_inputs(
images=images, feature_seqlens={"token_type_ids": len(input_ids)}, processor=processor
)
return input_ids, labels, extra_inputs
return input_ids, labels
def preprocess_supervised_dataset(
@@ -99,17 +96,17 @@ def preprocess_supervised_dataset(
# build inputs with format `<bos> X Y <eos>` and labels with format `<ignore> ... <ignore> Y <eos>`
# for multiturn examples, we only mask the prompt part in each prompt-response pair.
model_inputs = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
for i in range(len(examples["_prompt"])):
if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) != 1:
logger.warning("Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i]))
continue
input_ids, labels, extra_inputs = _encode_supervised_example(
prompt=examples["prompt"][i],
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
images=examples["images"][i],
input_ids, labels = _encode_supervised_example(
prompt=examples["_prompt"][i],
response=examples["_response"][i],
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -120,8 +117,7 @@ def preprocess_supervised_dataset(
model_inputs["input_ids"].append(input_ids)
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
model_inputs["images"].append(examples["_images"][i])
return model_inputs
@@ -143,17 +139,17 @@ def preprocess_packed_supervised_dataset(
batch_input_ids, batch_labels = [], []
lengths = []
length2indexes = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
for i in range(len(examples["_prompt"])):
if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) != 1:
logger.warning("Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i]))
continue
input_ids, labels, _ = _encode_supervised_example(
prompt=examples["prompt"][i],
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
images=examples["images"][i],
input_ids, labels = _encode_supervised_example(
prompt=examples["_prompt"][i],
response=examples["_response"][i],
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
template=template,
tokenizer=tokenizer,
processor=None,
@@ -199,6 +195,7 @@ def preprocess_packed_supervised_dataset(
model_inputs["input_ids"].append(packed_input_ids)
model_inputs["attention_mask"].append(packed_attention_masks)
model_inputs["labels"].append(packed_labels)
model_inputs["images"].append(examples["_images"][i])
return model_inputs

32
src/llamafactory/data/processors/unsupervised.py
View File

@@ -21,10 +21,10 @@ from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from PIL.Image import Image
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..mm_plugin import ImageInput
from ..template import Template
@@ -36,12 +36,12 @@ def _encode_unsupervised_example(
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
images: Sequence["ImageInput"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int], Dict[str, Any]]:
) -> Tuple[List[int], List[int]]:
if len(response) == 1:
messages = prompt + response
else:
@@ -56,10 +56,7 @@ def _encode_unsupervised_example(
source_len, target_len = infer_seqlen(len(input_ids), len(labels), cutoff_len)
input_ids = input_ids[:source_len]
labels = labels[:target_len]
extra_inputs = template.mm_plugin.get_mm_inputs(
images=images, feature_seqlens={"token_type_ids": len(input_ids)}, processor=processor
)
return input_ids, labels, extra_inputs
return input_ids, labels
def preprocess_unsupervised_dataset(
@@ -71,17 +68,17 @@ def preprocess_unsupervised_dataset(
) -> Dict[str, List[Any]]:
# build inputs with format `<bos> X` and labels with format `Y <eos>`
model_inputs = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
for i in range(len(examples["_prompt"])):
if len(examples["_prompt"][i]) % 2 != 1:
logger.warning("Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i]))
continue
input_ids, labels, extra_inputs = _encode_unsupervised_example(
prompt=examples["prompt"][i],
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
images=examples["images"][i],
input_ids, labels = _encode_unsupervised_example(
prompt=examples["_prompt"][i],
response=examples["_response"][i],
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -90,8 +87,7 @@ def preprocess_unsupervised_dataset(
model_inputs["input_ids"].append(input_ids)
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
model_inputs["images"].append(examples["_images"][i])
return model_inputs