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

video datasets

Browse Source 
Former-commit-id: 33f28ce82d9e44d2615909250dc56d6a4a03cd99
This commit is contained in:
hiyouga
2024-09-05 02:04:17 +08:00
parent 2c1eef34cb
commit 1874d579c5
25 changed files with 359 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
src/llamafactory/data/processors/feedback.py
View File

@@ -24,7 +24,7 @@ if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..mm_plugin import ImageInput
from ..mm_plugin import ImageInput, VideoInput
from ..template import Template
@@ -38,6 +38,7 @@ def _encode_feedback_example(
system: Optional[str],
tools: Optional[str],
images: Sequence["ImageInput"],
videos: Sequence["VideoInput"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
@@ -55,8 +56,8 @@ 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)
messages = template.mm_plugin.process_messages(messages, images, videos, processor)
kl_messages = template.mm_plugin.process_messages(kl_messages, images, videos, 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)
@@ -64,8 +65,8 @@ 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, images, tokenizer, processor)
kl_prompt_ids, _ = template.mm_plugin.process_token_ids(kl_prompt_ids, None, images, tokenizer, processor)
prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, images, videos, tokenizer, processor)
kl_prompt_ids, _ = template.mm_plugin.process_token_ids(kl_prompt_ids, None, images, videos, tokenizer, processor)
source_len, target_len = infer_seqlen(len(prompt_ids), len(response_ids), cutoff_len)
prompt_ids = prompt_ids[:source_len]
@@ -103,6 +104,7 @@ def preprocess_feedback_dataset(
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
videos=examples["_videos"][i] or [],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -116,6 +118,7 @@ def preprocess_feedback_dataset(
model_inputs["kl_labels"].append(kl_labels)
model_inputs["kto_tags"].append(kto_tag)
model_inputs["images"].append(examples["_images"][i])
model_inputs["videos"].append(examples["_videos"][i])
desirable_num = sum([1 for tag in model_inputs["kto_tags"] if tag])
undesirable_num = len(model_inputs["kto_tags"]) - desirable_num

11
src/llamafactory/data/processors/pairwise.py
View File

@@ -24,7 +24,7 @@ if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..mm_plugin import ImageInput
from ..mm_plugin import ImageInput, VideoInput
from ..template import Template
@@ -37,13 +37,14 @@ def _encode_pairwise_example(
system: Optional[str],
tools: Optional[str],
images: Sequence["ImageInput"],
videos: Sequence["VideoInput"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> 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)
chosen_messages = template.mm_plugin.process_messages(prompt + [response[0]], images, videos, processor)
rejected_messages = template.mm_plugin.process_messages(prompt + [response[1]], images, videos, processor)
prompt_ids, chosen_ids = template.encode_oneturn(tokenizer, chosen_messages, system, tools)
_, rejected_ids = template.encode_oneturn(tokenizer, rejected_messages, system, tools)
@@ -51,7 +52,7 @@ 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, images, tokenizer, processor)
prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, images, videos, 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]
@@ -85,6 +86,7 @@ def preprocess_pairwise_dataset(
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
videos=examples["_videos"][i] or [],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -97,6 +99,7 @@ def preprocess_pairwise_dataset(
model_inputs["rejected_attention_mask"].append([1] * len(rejected_input_ids))
model_inputs["rejected_labels"].append(rejected_labels)
model_inputs["images"].append(examples["_images"][i])
model_inputs["videos"].append(examples["_videos"][i])
return model_inputs

25
src/llamafactory/data/processors/supervised.py
View File

@@ -24,7 +24,7 @@ if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..mm_plugin import ImageInput
from ..mm_plugin import ImageInput, VideoInput
from ..template import Template
@@ -37,6 +37,7 @@ def _encode_supervised_example(
system: Optional[str],
tools: Optional[str],
images: Sequence["ImageInput"],
videos: Sequence["VideoInput"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
@@ -44,8 +45,8 @@ def _encode_supervised_example(
train_on_prompt: bool,
mask_history: bool,
) -> 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)
messages = template.mm_plugin.process_messages(prompt + response, images, videos, processor)
input_ids, labels = template.mm_plugin.process_token_ids([], [], images, videos, tokenizer, processor)
encoded_pairs = template.encode_multiturn(tokenizer, messages, system, tools)
total_length = len(input_ids) + (1 if template.efficient_eos else 0)
if mask_history:
@@ -107,6 +108,7 @@ def preprocess_supervised_dataset(
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
videos=examples["_videos"][i] or [],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -118,6 +120,7 @@ def preprocess_supervised_dataset(
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
model_inputs["images"].append(examples["_images"][i])
model_inputs["videos"].append(examples["_videos"][i])
return model_inputs
@@ -132,11 +135,8 @@ def preprocess_packed_supervised_dataset(
# 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 = [], []
batch_input_ids, batch_labels, batch_images, batch_videos = [], [], [], []
lengths = []
length2indexes = defaultdict(list)
for i in range(len(examples["_prompt"])):
@@ -150,9 +150,10 @@ def preprocess_packed_supervised_dataset(
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
videos=examples["_videos"][i] or [],
template=template,
tokenizer=tokenizer,
processor=None,
processor=processor,
cutoff_len=data_args.cutoff_len - 1, # reserved for the padding token
train_on_prompt=data_args.train_on_prompt,
mask_history=data_args.mask_history,
@@ -165,16 +166,21 @@ def preprocess_packed_supervised_dataset(
length2indexes[length].append(valid_num)
batch_input_ids.append(input_ids)
batch_labels.append(labels)
batch_images.append(examples["_images"][i] or [])
batch_videos.append(examples["_videos"][i] or [])
valid_num += 1
model_inputs = defaultdict(list)
knapsacks = greedy_knapsack(lengths, data_args.cutoff_len - 1) # reserved for the padding token
for knapsack in knapsacks:
packed_input_ids, packed_attention_masks, packed_labels = [], [], []
packed_images, packed_videos = [], []
for i, length in enumerate(knapsack):
index = length2indexes[length].pop()
packed_input_ids += batch_input_ids[index]
packed_labels += batch_labels[index]
packed_images += batch_images[index]
packed_videos += batch_videos[index]
if data_args.neat_packing:
packed_attention_masks += [i + 1] * len(batch_input_ids[index]) # start from 1
else:
@@ -195,7 +201,8 @@ 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])
model_inputs["images"].append(packed_images or None)
model_inputs["videos"].append(packed_videos or None)
return model_inputs

9
src/llamafactory/data/processors/unsupervised.py
View File

@@ -24,7 +24,7 @@ if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..mm_plugin import ImageInput
from ..mm_plugin import ImageInput, VideoInput
from ..template import Template
@@ -37,6 +37,7 @@ def _encode_unsupervised_example(
system: Optional[str],
tools: Optional[str],
images: Sequence["ImageInput"],
videos: Sequence["VideoInput"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
@@ -47,12 +48,12 @@ def _encode_unsupervised_example(
else:
messages = prompt + [{"role": Role.ASSISTANT.value, "content": ""}]
messages = template.mm_plugin.process_messages(messages, images, processor)
messages = template.mm_plugin.process_messages(messages, images, videos, 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, images, tokenizer, processor)
input_ids, _ = template.mm_plugin.process_token_ids(input_ids, None, images, videos, 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]
@@ -79,6 +80,7 @@ def preprocess_unsupervised_dataset(
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
videos=examples["_videos"][i] or [],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -88,6 +90,7 @@ def preprocess_unsupervised_dataset(
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
model_inputs["images"].append(examples["_images"][i])
model_inputs["videos"].append(examples["_videos"][i])
return model_inputs