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

fix mixed mm inputs and rlhf-v

Browse Source 
Former-commit-id: 7c248fac20bf85d57a91132ce7a793c7f84e9218
This commit is contained in:
hiyouga
2024-09-01 20:52:47 +08:00
parent 1d8e9c7897
commit 7e4c5d4bb3
20 changed files with 306 additions and 277 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
src/llamafactory/data/processors/feedback.py
View File

@@ -21,6 +21,7 @@ from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from PIL.Image import Image
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
@@ -36,11 +37,12 @@ def _encode_feedback_example(
kl_response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int], List[int], List[int], bool]:
) -> Tuple[List[int], List[int], List[int], List[int], bool, Dict[str, Any]]:
if response[0]["content"]: # desired example
kto_tag = True
messages = prompt + [response[0]]
@@ -53,6 +55,8 @@ def _encode_feedback_example(
else:
kl_messages = prompt + [kl_response[1]]
messages = template.mm_plugin.process_messages(messages, images, processor)
kl_messages = template.mm_plugin.process_messages(kl_messages, images, processor)
prompt_ids, response_ids = template.encode_oneturn(tokenizer, messages, system, tools)
kl_prompt_ids, kl_response_ids = template.encode_oneturn(tokenizer, kl_messages, system, tools)
@@ -60,8 +64,8 @@ def _encode_feedback_example(
response_ids += [tokenizer.eos_token_id]
kl_response_ids += [tokenizer.eos_token_id]
prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, tokenizer, processor)
kl_prompt_ids, _ = template.mm_plugin.process_token_ids(kl_prompt_ids, None, tokenizer, processor)
prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, images, tokenizer, processor)
kl_prompt_ids, _ = template.mm_plugin.process_token_ids(kl_prompt_ids, None, images, tokenizer, processor)
source_len, target_len = infer_seqlen(len(prompt_ids), len(response_ids), cutoff_len)
prompt_ids = prompt_ids[:source_len]
@@ -74,8 +78,15 @@ 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
return input_ids, labels, kl_input_ids, kl_labels, kto_tag
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
def preprocess_feedback_dataset(
@@ -93,13 +104,13 @@ def preprocess_feedback_dataset(
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
input_ids, labels, kl_input_ids, kl_labels, kto_tag = _encode_feedback_example(
prompt=prompt,
input_ids, labels, kl_input_ids, kl_labels, kto_tag, extra_inputs = _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],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -112,15 +123,8 @@ 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)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={
"token_type_ids": len(input_ids),
"kl_token_type_ids": len(kl_input_ids),
},
processor=processor,
)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
desirable_num = sum([1 for tag in model_inputs["kto_tags"] if tag])
undesirable_num = len(model_inputs["kto_tags"]) - desirable_num

38
src/llamafactory/data/processors/pairwise.py
View File

@@ -21,6 +21,7 @@ from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from PIL.Image import Image
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
@@ -35,13 +36,14 @@ def _encode_pairwise_example(
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int], List[int], List[int]]:
chosen_messages = prompt + [response[0]]
rejected_messages = prompt + [response[1]]
) -> Tuple[List[int], List[int], List[int], List[int], Dict[str, Any]]:
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)
_, rejected_ids = template.encode_oneturn(tokenizer, rejected_messages, system, tools)
@@ -49,7 +51,7 @@ def _encode_pairwise_example(
chosen_ids += [tokenizer.eos_token_id]
rejected_ids += [tokenizer.eos_token_id]
prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, tokenizer, processor)
prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, images, tokenizer, processor)
# consider the response is more important
source_len, target_len = infer_seqlen(len(prompt_ids), max(len(chosen_ids), len(rejected_ids)), cutoff_len)
prompt_ids = prompt_ids[:source_len]
@@ -60,8 +62,15 @@ 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
return chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels
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
def preprocess_pairwise_dataset(
@@ -78,12 +87,12 @@ def preprocess_pairwise_dataset(
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels = _encode_pairwise_example(
prompt=prompt,
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],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -95,15 +104,8 @@ 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)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={
"chosen_token_type_ids": len(chosen_input_ids),
"rejected_token_type_ids": len(rejected_input_ids),
},
processor=processor,
)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
return model_inputs

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

@@ -21,6 +21,7 @@ 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
@@ -35,19 +36,18 @@ def _encode_supervised_example(
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
train_on_prompt: bool,
mask_history: bool,
) -> Tuple[List[int], List[int]]:
messages = prompt + response
input_ids, labels = [], []
input_ids, labels = template.mm_plugin.process_token_ids(input_ids, labels, tokenizer, processor)
) -> Tuple[List[int], List[int], Dict[str, Any]]:
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)
total_length = 1 if template.efficient_eos else 0
total_length = len(input_ids) + (1 if template.efficient_eos else 0)
if mask_history:
encoded_pairs = encoded_pairs[::-1] # high priority for last turns
@@ -83,7 +83,10 @@ def _encode_supervised_example(
input_ids += [tokenizer.eos_token_id]
labels += [tokenizer.eos_token_id]
return input_ids, labels
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
def preprocess_supervised_dataset(
@@ -101,12 +104,12 @@ def preprocess_supervised_dataset(
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
input_ids, labels = _encode_supervised_example(
prompt=prompt,
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],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -117,12 +120,8 @@ 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)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={"token_type_ids": len(input_ids)},
processor=processor,
)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
return model_inputs
@@ -131,10 +130,15 @@ def preprocess_packed_supervised_dataset(
examples: Dict[str, List[Any]],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
data_args: "DataArguments",
) -> Dict[str, List[Any]]:
# TODO: use `position_ids` to achieve packing
# build inputs with format `<bos> X1 Y1 <eos> <bos> X2 Y2 <eos>`
# and labels with format `<ignore> ... <ignore> Y1 <eos> <ignore> ... <ignore> Y2 <eos>`
if processor is not None:
raise NotImplementedError("`packing` have not been implemented for multimodal datasets.")
valid_num = 0
batch_input_ids, batch_labels = [], []
lengths = []
@@ -149,6 +153,7 @@ def preprocess_packed_supervised_dataset(
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
images=examples["images"][i],
template=template,
tokenizer=tokenizer,
processor=None,

26
src/llamafactory/data/processors/unsupervised.py
View File

@@ -21,6 +21,7 @@ from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from PIL.Image import Image
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
@@ -35,25 +36,30 @@ def _encode_unsupervised_example(
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int]]:
) -> Tuple[List[int], List[int], Dict[str, Any]]:
if len(response) == 1:
messages = prompt + response
else:
messages = prompt + [{"role": Role.ASSISTANT.value, "content": ""}]
messages = template.mm_plugin.process_messages(messages, images, processor)
input_ids, labels = template.encode_oneturn(tokenizer, messages, system, tools)
if template.efficient_eos:
labels += [tokenizer.eos_token_id]
input_ids, _ = template.mm_plugin.process_token_ids(input_ids, None, tokenizer, processor)
input_ids, _ = template.mm_plugin.process_token_ids(input_ids, None, images, tokenizer, processor)
source_len, target_len = infer_seqlen(len(input_ids), len(labels), cutoff_len)
input_ids = input_ids[:source_len]
labels = labels[:target_len]
return input_ids, labels
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
def preprocess_unsupervised_dataset(
@@ -70,12 +76,12 @@ def preprocess_unsupervised_dataset(
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
input_ids, labels = _encode_unsupervised_example(
prompt=prompt,
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],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -84,12 +90,8 @@ 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)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={"token_type_ids": len(input_ids)},
processor=processor,
)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
return model_inputs