Spaces:
Runtime error
Runtime error
Upload 4 files
Browse files
README.md
CHANGED
@@ -1,13 +0,0 @@
|
|
1 |
-
---
|
2 |
-
title: Bulk Embeddings
|
3 |
-
emoji: 🐠
|
4 |
-
colorFrom: purple
|
5 |
-
colorTo: indigo
|
6 |
-
sdk: gradio
|
7 |
-
sdk_version: 3.36.1
|
8 |
-
app_file: app.py
|
9 |
-
pinned: false
|
10 |
-
license: apache-2.0
|
11 |
-
---
|
12 |
-
|
13 |
-
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app.py
ADDED
@@ -0,0 +1,197 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
|
3 |
+
from utils import load_hf_dataset, get_model_and_tokenizer, batch_embed
|
4 |
+
|
5 |
+
|
6 |
+
# TODO: add instructor models
|
7 |
+
# "hkunlp/instructor-xl",
|
8 |
+
# "hkunlp/instructor-large",
|
9 |
+
# "hkunlp/instructor-base",
|
10 |
+
|
11 |
+
# model ids and hidden sizes
|
12 |
+
models_and_hidden_sizes = [
|
13 |
+
("intfloat/e5-small-v2", 384),
|
14 |
+
("intfloat/e5-base-v2", 768),
|
15 |
+
("intfloat/e5-large-v2", 1024),
|
16 |
+
("intfloat/multilingual-e5-small", 384),
|
17 |
+
("intfloat/multilingual-e5-base", 768),
|
18 |
+
("intfloat/multilingual-e5-large", 1024),
|
19 |
+
("sentence-transformers/all-MiniLM-L6-v2", 384),
|
20 |
+
("sentence-transformers/all-MiniLM-L12-v2", 384),
|
21 |
+
("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", 384),
|
22 |
+
]
|
23 |
+
|
24 |
+
model_options = [
|
25 |
+
f"{model_name} (hidden_size = {hidden_size})"
|
26 |
+
for model_name, hidden_size in models_and_hidden_sizes
|
27 |
+
]
|
28 |
+
|
29 |
+
|
30 |
+
opt2desc = {
|
31 |
+
"O2": "Most precise, slowest (O2: basic and extended general optimizations, transformers-specific fusions)",
|
32 |
+
"O3": "Less precise, faster (O3: O2 + gelu approx)",
|
33 |
+
"O4": "Least precise, fastest (O4: O3 + fp16/bf16)",
|
34 |
+
}
|
35 |
+
|
36 |
+
desc2opt = {v: k for k, v in opt2desc.items()}
|
37 |
+
|
38 |
+
|
39 |
+
optimization_options = list(opt2desc.values())
|
40 |
+
|
41 |
+
|
42 |
+
def run(
|
43 |
+
ds_name,
|
44 |
+
ds_config,
|
45 |
+
column_name,
|
46 |
+
ds_split,
|
47 |
+
model_choice,
|
48 |
+
opt_desc,
|
49 |
+
new_dataset_id,
|
50 |
+
num2skip,
|
51 |
+
num2embed,
|
52 |
+
progress=gr.Progress(),
|
53 |
+
):
|
54 |
+
if progress is not None:
|
55 |
+
progress(0.5, "Loading dataset...")
|
56 |
+
ds = load_hf_dataset(ds_name, ds_config, ds_split)
|
57 |
+
|
58 |
+
opt_level = desc2opt[opt_desc]
|
59 |
+
|
60 |
+
model_name = model_choice.split()[0]
|
61 |
+
|
62 |
+
if progress is not None:
|
63 |
+
progress(0.2, "Downloading model and tokenizer...")
|
64 |
+
model, tokenizer = get_model_and_tokenizer(model_name, opt_level, progress)
|
65 |
+
|
66 |
+
doc_count, seconds_taken = batch_embed(
|
67 |
+
ds,
|
68 |
+
model,
|
69 |
+
tokenizer,
|
70 |
+
model_name=model_name,
|
71 |
+
column_name=column_name,
|
72 |
+
new_dataset_id=new_dataset_id,
|
73 |
+
opt_level=opt_level,
|
74 |
+
num2skip=num2skip,
|
75 |
+
num2embed=num2embed,
|
76 |
+
progress=progress,
|
77 |
+
)
|
78 |
+
|
79 |
+
return f"Embedded {doc_count} docs in {seconds_taken/60:.2f} minutes ({doc_count/seconds_taken:.1f} docs/sec)"
|
80 |
+
|
81 |
+
|
82 |
+
with gr.Blocks(title="Bulk embeddings") as demo:
|
83 |
+
gr.Markdown(
|
84 |
+
"""
|
85 |
+
This Space allows you to embed a large dataset easily. For instance, this can easily create vectors for Wikipedia \
|
86 |
+
articles -- taking about __ hours and costing approximately $__.
|
87 |
+
|
88 |
+
|
89 |
+
This utilizes state-of-the-art open-source embedding models, \
|
90 |
+
and optimizes them for inference using Hugging Face [optimum](https://github.com/huggingface/optimum). There are various \
|
91 |
+
levels of optimizations that can be applied - the quality of the embeddings will degrade as the optimizations increase.
|
92 |
+
|
93 |
+
Currently available options: O2/O3/O4 on T4/A10 GPUs using onnx runtime.
|
94 |
+
|
95 |
+
Future options:
|
96 |
+
- OpenVino for CPU inference
|
97 |
+
- TensorRT for GPU inference
|
98 |
+
- Quantized models
|
99 |
+
- Instructor models
|
100 |
+
- Text splitting options
|
101 |
+
- More control about which rows to embed (skip some, stop early)
|
102 |
+
- Dynamic padding
|
103 |
+
|
104 |
+
## Steps
|
105 |
+
|
106 |
+
1. Upload the dataset to the Hugging Face Hub.
|
107 |
+
2. Enter dataset details into the form below.
|
108 |
+
3. Choose a model. These are taken from the top of the [MTEB leaderboard](https://huggingface.co/spaces/mteb/leaderboard).
|
109 |
+
4. Enter optimization level. See [here](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/optimization#optimization-configuration) for details.
|
110 |
+
5. Choose a name for the new dataset.
|
111 |
+
6. Hit run!
|
112 |
+
|
113 |
+
|
114 |
+
### Note:
|
115 |
+
|
116 |
+
If you have short documents, O3 will be faster than O4. If you have long documents, O4 will be faster than O3. \
|
117 |
+
O4 requires the tokenized documents to be padded to max length.
|
118 |
+
|
119 |
+
"""
|
120 |
+
)
|
121 |
+
|
122 |
+
with gr.Row():
|
123 |
+
ds_name = gr.Textbox(
|
124 |
+
lines=1,
|
125 |
+
label="Dataset to load from Hugging Face Hub",
|
126 |
+
value="nbroad/basic_text_dataset",
|
127 |
+
)
|
128 |
+
ds_config = gr.Textbox(
|
129 |
+
lines=1, label="Dataset config (leave blank to use default)", value=""
|
130 |
+
)
|
131 |
+
|
132 |
+
column_name = gr.Textbox(lines=1, label="Enter column to embed", value="text")
|
133 |
+
ds_split = gr.Dropdown(
|
134 |
+
choices=["train", "validation", "test"],
|
135 |
+
label="Dataset split",
|
136 |
+
value="train",
|
137 |
+
)
|
138 |
+
# TODO: idx column
|
139 |
+
# TODO: text splitting options
|
140 |
+
|
141 |
+
with gr.Row():
|
142 |
+
model_choice = gr.Dropdown(
|
143 |
+
choices=model_options, label="Embedding model", value=model_options[0]
|
144 |
+
)
|
145 |
+
opt_desc = gr.Dropdown(
|
146 |
+
choices=optimization_options,
|
147 |
+
label="Optimization level",
|
148 |
+
value=optimization_options[0],
|
149 |
+
)
|
150 |
+
|
151 |
+
with gr.Row():
|
152 |
+
new_dataset_id = gr.Textbox(
|
153 |
+
lines=1,
|
154 |
+
label="New dataset name, including username",
|
155 |
+
value="nbroad/test-embeds",
|
156 |
+
)
|
157 |
+
|
158 |
+
num2skip = gr.Slider(
|
159 |
+
value=0,
|
160 |
+
minimum=0,
|
161 |
+
maximum=10_000_000,
|
162 |
+
step=1,
|
163 |
+
label="Number of rows to skip",
|
164 |
+
)
|
165 |
+
|
166 |
+
num2embed = gr.Slider(
|
167 |
+
value=-1,
|
168 |
+
minimum=-1,
|
169 |
+
maximum=10_000_000,
|
170 |
+
step=1,
|
171 |
+
label="Number of rows to embed (-1 = all)",
|
172 |
+
)
|
173 |
+
|
174 |
+
with gr.Row():
|
175 |
+
btn = gr.Button(value="Embed texts!")
|
176 |
+
|
177 |
+
last = gr.Textbox(value="")
|
178 |
+
|
179 |
+
btn.click(
|
180 |
+
fn=run,
|
181 |
+
inputs=[
|
182 |
+
ds_name,
|
183 |
+
ds_config,
|
184 |
+
column_name,
|
185 |
+
ds_split,
|
186 |
+
model_choice,
|
187 |
+
opt_desc,
|
188 |
+
new_dataset_id,
|
189 |
+
num2skip,
|
190 |
+
num2embed,
|
191 |
+
],
|
192 |
+
outputs=last,
|
193 |
+
)
|
194 |
+
|
195 |
+
|
196 |
+
if __name__ == "__main__":
|
197 |
+
demo.queue(concurrency_count=20).launch(show_error=True, debug=True, share=True)
|
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
datasets==2.13.1
|
2 |
+
tokenizers>=0.11.1,!=0.11.3,<0.14
|
3 |
+
optimum[onnxruntime-gpu]==1.8.8
|
4 |
+
transformers==4.30.1
|
5 |
+
accelerate==0.20.3
|
6 |
+
gradio==3.35.2
|
7 |
+
--extra-index-url https://download.pytorch.org/whl/cu118
|
8 |
+
torch==2.0.1
|
utils.py
ADDED
@@ -0,0 +1,452 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import time
|
3 |
+
import shutil
|
4 |
+
from pathlib import Path
|
5 |
+
from typing import Union, Dict, List
|
6 |
+
|
7 |
+
import torch
|
8 |
+
import datasets
|
9 |
+
from datasets import load_dataset, Dataset
|
10 |
+
from transformers import AutoTokenizer, PreTrainedTokenizer
|
11 |
+
from huggingface_hub import Repository, create_repo, HfApi
|
12 |
+
from optimum.onnxruntime import (
|
13 |
+
AutoOptimizationConfig,
|
14 |
+
ORTModelForFeatureExtraction,
|
15 |
+
ORTOptimizer,
|
16 |
+
)
|
17 |
+
|
18 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
19 |
+
|
20 |
+
|
21 |
+
opt_configs = {
|
22 |
+
"O2": AutoOptimizationConfig.O2(),
|
23 |
+
"O3": AutoOptimizationConfig.O3(),
|
24 |
+
"O4": AutoOptimizationConfig.O4(),
|
25 |
+
}
|
26 |
+
|
27 |
+
|
28 |
+
def get_batch_size(device_name: str, model_name: str, opt_level: str):
|
29 |
+
"""
|
30 |
+
TODO: run actual tests
|
31 |
+
|
32 |
+
T4 has 16GB
|
33 |
+
A10 has 24GB
|
34 |
+
|
35 |
+
Args:
|
36 |
+
device_name (`str`):
|
37 |
+
The name of the GPU device in use.
|
38 |
+
model_name (`str`):
|
39 |
+
The name of the model in use.
|
40 |
+
opt_level (`str`):
|
41 |
+
The optimization level in use.
|
42 |
+
|
43 |
+
Returns:
|
44 |
+
`int`:
|
45 |
+
The batch size to use.
|
46 |
+
"""
|
47 |
+
|
48 |
+
if "small" in model_name:
|
49 |
+
bs = 192
|
50 |
+
elif "base" in model_name:
|
51 |
+
bs = 128
|
52 |
+
elif "large" in model_name:
|
53 |
+
bs = 64
|
54 |
+
else:
|
55 |
+
bs = 32
|
56 |
+
|
57 |
+
if "A10" in device_name:
|
58 |
+
bs *= 2
|
59 |
+
|
60 |
+
if opt_level == "O4":
|
61 |
+
bs *= 2
|
62 |
+
|
63 |
+
return bs
|
64 |
+
|
65 |
+
|
66 |
+
def mean_pooling(last_hidden_state: torch.Tensor, attention_mask: torch.Tensor):
|
67 |
+
"""
|
68 |
+
Mean pool the token embeddings.
|
69 |
+
|
70 |
+
Args:
|
71 |
+
last_hidden_state (`tuple`):
|
72 |
+
The output of the model.
|
73 |
+
attention_mask (`torch.Tensor`):
|
74 |
+
The attention mask.
|
75 |
+
|
76 |
+
Returns:
|
77 |
+
`torch.Tensor`:
|
78 |
+
The mean pooled embeddings.
|
79 |
+
"""
|
80 |
+
input_mask_expanded = (
|
81 |
+
attention_mask.unsqueeze(-1).expand(last_hidden_state.size()).float()
|
82 |
+
)
|
83 |
+
return torch.sum(last_hidden_state * input_mask_expanded, 1) / torch.clamp(
|
84 |
+
input_mask_expanded.sum(1), min=1e-9
|
85 |
+
)
|
86 |
+
|
87 |
+
|
88 |
+
def load_hf_dataset(ds_name: str, ds_config: str = None, ds_split: str = "train"):
|
89 |
+
"""
|
90 |
+
Load a dataset from the HuggingFace Hub. Will be streaming so
|
91 |
+
as to not load the whole dataset to local storage.
|
92 |
+
|
93 |
+
Args:
|
94 |
+
ds_name (`str`):
|
95 |
+
The name of the dataset to load.
|
96 |
+
ds_config (`str`, *optional*, Defaults to `None`):
|
97 |
+
The configuration of the dataset to load.
|
98 |
+
ds_split (`str`, *optional*, Defaults to `"train"`):
|
99 |
+
The split of the dataset to load.
|
100 |
+
|
101 |
+
Returns:
|
102 |
+
ds (`datasets.IterableDataset`):
|
103 |
+
The loaded dataset.
|
104 |
+
"""
|
105 |
+
|
106 |
+
if ds_config == "":
|
107 |
+
ds_config = None
|
108 |
+
|
109 |
+
ds = load_dataset(ds_name, ds_config, split=ds_split, streaming=True)
|
110 |
+
|
111 |
+
return ds
|
112 |
+
|
113 |
+
|
114 |
+
def get_model_and_tokenizer(model_name: str, optimization_level: str, progress):
|
115 |
+
"""
|
116 |
+
Load the model and tokenizer from the HuggingFace Hub.
|
117 |
+
|
118 |
+
If the model is not already optimized, optimize it and save it to the local directory.
|
119 |
+
|
120 |
+
Args:
|
121 |
+
model_name (`str`):
|
122 |
+
The name of the model to load.
|
123 |
+
optimization_level (`str`):
|
124 |
+
The optimization level to use. Should be one of `"O2"`, `"O3"`, or `"O4"`.
|
125 |
+
|
126 |
+
Returns:
|
127 |
+
model (`ORTModelForFeatureExtraction`):
|
128 |
+
The optimized model.
|
129 |
+
tokenizer (`PreTrainedTokenizer`):
|
130 |
+
The tokenizer.
|
131 |
+
"""
|
132 |
+
optimized_model_name = f"model_optimized_{optimization_level}.onnx"
|
133 |
+
|
134 |
+
model_dir = Path(model_name.replace("/", "_"))
|
135 |
+
if not (model_dir / optimized_model_name).exists():
|
136 |
+
if progress is not None:
|
137 |
+
progress(0.2, "Downloading tokenizer...")
|
138 |
+
|
139 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
140 |
+
tokenizer.save_pretrained(model_dir)
|
141 |
+
|
142 |
+
if progress is not None:
|
143 |
+
progress(0.4, "Downloading model...")
|
144 |
+
|
145 |
+
model = ORTModelForFeatureExtraction.from_pretrained(model_name, export=True)
|
146 |
+
model.save_pretrained(model_dir)
|
147 |
+
|
148 |
+
optimizer = ORTOptimizer.from_pretrained(model)
|
149 |
+
optimization_config = opt_configs[optimization_level]
|
150 |
+
|
151 |
+
if progress is not None:
|
152 |
+
progress(0.6, "Optimizing model...")
|
153 |
+
|
154 |
+
optimizer.optimize(save_dir=model_dir, optimization_config=optimization_config)
|
155 |
+
Path(model_dir / "model_optimized.onnx").rename(
|
156 |
+
model_dir / optimized_model_name
|
157 |
+
)
|
158 |
+
|
159 |
+
else:
|
160 |
+
tokenizer = AutoTokenizer.from_pretrained(model_dir)
|
161 |
+
|
162 |
+
if progress is not None:
|
163 |
+
progress(0.8, "Loading optimized model and tokenizer...")
|
164 |
+
|
165 |
+
return (
|
166 |
+
ORTModelForFeatureExtraction.from_pretrained(
|
167 |
+
model_dir,
|
168 |
+
file_name=optimized_model_name,
|
169 |
+
provider="CUDAExecutionProvider",
|
170 |
+
),
|
171 |
+
tokenizer,
|
172 |
+
)
|
173 |
+
|
174 |
+
|
175 |
+
def tokenize(
|
176 |
+
examples: Dict[str, List[str]],
|
177 |
+
tokenizer: PreTrainedTokenizer,
|
178 |
+
column_name: str = "text",
|
179 |
+
padding: Union[bool, str] = True,
|
180 |
+
max_length: int = 512,
|
181 |
+
):
|
182 |
+
"""
|
183 |
+
Tokenize the examples using the tokenizer.
|
184 |
+
|
185 |
+
Args:
|
186 |
+
examples (`Dict[str, List[str]]`):
|
187 |
+
examples to tokenize
|
188 |
+
tokenizer (`PreTrainedTokenizer`):
|
189 |
+
tokenizer to use
|
190 |
+
column_name (`str`, *optional*, defaults to `text`):
|
191 |
+
column name to use for tokenization. Defaults to `text`
|
192 |
+
padding (`bool`, *optional*, defaults to `True`):
|
193 |
+
whether to pad the examples. Defaults to `True`
|
194 |
+
Use `"max_length"` if using `O4` optimization level
|
195 |
+
If `True`, the batch will be padded to the longest in the batch.
|
196 |
+
max_length (`int`, *optional*, Defaults to `512`):
|
197 |
+
max length to use for the model. Defaults to `512`.
|
198 |
+
Any sequences longer will be truncated.
|
199 |
+
If padding is `"max_length"`, the padding will be added until the sequence
|
200 |
+
is of length `max_length`.
|
201 |
+
|
202 |
+
Returns:
|
203 |
+
`Dict[str, List[List[int]]]`:
|
204 |
+
tokenized examples
|
205 |
+
"""
|
206 |
+
# TODO: add lengths, sort by length, use dynamic padding
|
207 |
+
# TODO: option for controlling length for models that can go shorter/longer than 512
|
208 |
+
return tokenizer(
|
209 |
+
examples[column_name], truncation=True, padding=padding, max_length=max_length
|
210 |
+
)
|
211 |
+
|
212 |
+
|
213 |
+
@torch.inference_mode()
|
214 |
+
def batch_embed(
|
215 |
+
ds: datasets.IterableDataset,
|
216 |
+
model: ORTModelForFeatureExtraction,
|
217 |
+
tokenizer: PreTrainedTokenizer,
|
218 |
+
model_name: str,
|
219 |
+
column_name: str,
|
220 |
+
new_dataset_id: str,
|
221 |
+
opt_level: str,
|
222 |
+
upload_batch_size: int = 10_000,
|
223 |
+
map_batch_size: int = 2000,
|
224 |
+
num2skip: int = 0,
|
225 |
+
num2embed: int = -1,
|
226 |
+
progress=None,
|
227 |
+
):
|
228 |
+
"""
|
229 |
+
Run the model on the dataset and upload the embeddings to the hub.
|
230 |
+
|
231 |
+
Args:
|
232 |
+
ds (`datasets.Dataset`):
|
233 |
+
dataset to embed. From `load_hf_dataset`
|
234 |
+
model (`ORTModelForFeatureExtraction`):
|
235 |
+
model to use for embedding. From `get_model_and_tokenizer`
|
236 |
+
tokenizer (`AutoTokenizer`):
|
237 |
+
tokenizer to use for embedding. From `get_model_and_tokenizer`
|
238 |
+
model_name (`str`):
|
239 |
+
name of the model to use. Used to determine batch size.
|
240 |
+
column_name (`str`):
|
241 |
+
column name to use for embedding. Default option in gradio app is `text`
|
242 |
+
new_dataset_id (`str`):
|
243 |
+
id of the new dataset to create. Should include username or organization.
|
244 |
+
e.g. nbroad/new-embeddings
|
245 |
+
opt_level (`str`):
|
246 |
+
optimization level to use. Should be one of `O2`, `O3`, `O4`
|
247 |
+
See here for more details on optimization levels:
|
248 |
+
https://huggingface.co/docs/optimum/onnxruntime/usage_guides/optimization#optimization-configuration
|
249 |
+
upload_batch_size (`int`, *optional*, defaults to `10_000`):
|
250 |
+
number of embeddings to upload at once. Defaults to 10,000.
|
251 |
+
map_batch_size (`int`, *optional*, defaults to `2000`):
|
252 |
+
number of examples to tokenize at once. Defaults to 2000.
|
253 |
+
num2skip (`int`, *optional*, defaults to `0`):
|
254 |
+
number of examples to skip. Defaults to 0.
|
255 |
+
num2embed (`int`, *optional*, defaults to `-1`):
|
256 |
+
number of examples to embed. Defaults to -1, which means all examples.
|
257 |
+
|
258 |
+
Returns:
|
259 |
+
current_count (`int`):
|
260 |
+
number of examples embedded so far
|
261 |
+
time_taken (`float`):
|
262 |
+
time taken to embed the examples in seconds
|
263 |
+
|
264 |
+
"""
|
265 |
+
|
266 |
+
api = HfApi(
|
267 |
+
token=os.environ["HF_TOKEN"],
|
268 |
+
)
|
269 |
+
|
270 |
+
username = api.whoami()["name"]
|
271 |
+
|
272 |
+
if "/" in new_dataset_id:
|
273 |
+
new_dataset_id = username + "/" + new_dataset_id
|
274 |
+
|
275 |
+
repo = init_git_repo(new_dataset_id)
|
276 |
+
|
277 |
+
iterator = iter(
|
278 |
+
ds.map(
|
279 |
+
tokenize,
|
280 |
+
batched=True,
|
281 |
+
batch_size=map_batch_size,
|
282 |
+
fn_kwargs={
|
283 |
+
"tokenizer": tokenizer,
|
284 |
+
"column_name": column_name,
|
285 |
+
"padding": "max_length" if opt_level == "O4" else True,
|
286 |
+
},
|
287 |
+
remove_columns=ds.column_names,
|
288 |
+
)
|
289 |
+
)
|
290 |
+
|
291 |
+
embeds = []
|
292 |
+
texts = []
|
293 |
+
|
294 |
+
# last_count keeps track of how many had been embedded since last push
|
295 |
+
last_count = 0
|
296 |
+
# current count keeps track of how many have been embedded in total
|
297 |
+
current_count = 0
|
298 |
+
|
299 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
300 |
+
|
301 |
+
inference_bs = get_batch_size(torch.cuda.get_device_name(0), model_name, opt_level)
|
302 |
+
|
303 |
+
loop = True
|
304 |
+
|
305 |
+
# skip through some examples
|
306 |
+
if num2skip > 0:
|
307 |
+
[next(iterator) for _ in range(num2skip)]
|
308 |
+
|
309 |
+
start_time = time.time()
|
310 |
+
while loop:
|
311 |
+
batch = [next(iterator, None) for _ in range(inference_bs)]
|
312 |
+
|
313 |
+
# batch will have None values when iterator runs out
|
314 |
+
if batch[-1] is None:
|
315 |
+
batch = [x for x in batch if x is not None]
|
316 |
+
loop = False
|
317 |
+
if len(batch) == 0:
|
318 |
+
break
|
319 |
+
|
320 |
+
ids = torch.tensor([b["input_ids"] for b in batch], device=device)
|
321 |
+
mask = torch.tensor([b["attention_mask"] for b in batch], device=device)
|
322 |
+
t_ids = torch.zeros_like(ids)
|
323 |
+
|
324 |
+
outputs = model(input_ids=ids, attention_mask=mask, token_type_ids=t_ids)
|
325 |
+
|
326 |
+
embeds.extend(mean_pooling(outputs[0], mask).cpu().tolist())
|
327 |
+
texts.extend([b[column_name] for b in batch])
|
328 |
+
|
329 |
+
current_count += len(batch)
|
330 |
+
|
331 |
+
# Check if we have embedded enough examples
|
332 |
+
if current_count >= num2embed:
|
333 |
+
diff = current_count - num2embed
|
334 |
+
embeds = embeds[:-diff]
|
335 |
+
texts = texts[:-diff]
|
336 |
+
current_count = num2embed
|
337 |
+
break
|
338 |
+
|
339 |
+
# Periodically upload to the hub
|
340 |
+
if len(embeds) > upload_batch_size:
|
341 |
+
push_to_repo(repo, last_count, current_count, embeds, texts)
|
342 |
+
embeds = []
|
343 |
+
last_count = current_count
|
344 |
+
|
345 |
+
# Provide updates
|
346 |
+
if progress is not None:
|
347 |
+
progress(
|
348 |
+
(current_count, None),
|
349 |
+
"Embedding docs...",
|
350 |
+
total=None,
|
351 |
+
unit="Docs Embedded",
|
352 |
+
)
|
353 |
+
|
354 |
+
time_taken = time.time() - start_time
|
355 |
+
|
356 |
+
# If there are any remaining embeddings, upload them
|
357 |
+
if len(embeds) > 0:
|
358 |
+
push_to_repo(repo, last_count, current_count, embeds, texts)
|
359 |
+
|
360 |
+
return current_count - num2skip, time_taken
|
361 |
+
|
362 |
+
|
363 |
+
def init_git_repo(repo_id: str):
|
364 |
+
"""
|
365 |
+
Initialize a git repo for the new dataset.
|
366 |
+
|
367 |
+
***Removes existing local folder if exists***
|
368 |
+
|
369 |
+
Args:
|
370 |
+
repo_id (`str`):
|
371 |
+
id of the new dataset to create. Should include username or organization.
|
372 |
+
e.g. nbroad/new-embeddings
|
373 |
+
"""
|
374 |
+
local_dir = repo_id.replace("/", "_")
|
375 |
+
|
376 |
+
create_repo(
|
377 |
+
repo_id,
|
378 |
+
repo_type="dataset",
|
379 |
+
token=os.environ["HF_TOKEN"],
|
380 |
+
private=True,
|
381 |
+
exist_ok=True,
|
382 |
+
)
|
383 |
+
try:
|
384 |
+
repo = Repository(
|
385 |
+
local_dir=local_dir,
|
386 |
+
clone_from=repo_id,
|
387 |
+
repo_type="dataset",
|
388 |
+
token=os.environ["HF_TOKEN"],
|
389 |
+
skip_lfs_files=True,
|
390 |
+
)
|
391 |
+
except EnvironmentError:
|
392 |
+
shutil.rmtree(local_dir)
|
393 |
+
repo = Repository(
|
394 |
+
local_dir=local_dir,
|
395 |
+
clone_from=repo_id,
|
396 |
+
repo_type="dataset",
|
397 |
+
token=os.environ["HF_TOKEN"],
|
398 |
+
skip_lfs_files=True,
|
399 |
+
)
|
400 |
+
|
401 |
+
if repo is not None:
|
402 |
+
repo.git_pull()
|
403 |
+
|
404 |
+
return repo
|
405 |
+
|
406 |
+
|
407 |
+
def push_to_repo(
|
408 |
+
repo: str,
|
409 |
+
last_count: int,
|
410 |
+
current_count: int,
|
411 |
+
embeds: List[List[float]],
|
412 |
+
texts: List[str],
|
413 |
+
):
|
414 |
+
"""
|
415 |
+
Push embeddings to the repo.
|
416 |
+
|
417 |
+
Args:
|
418 |
+
repo (`huggingface_hub.Repository`):
|
419 |
+
repo to push to
|
420 |
+
last_count (`int`):
|
421 |
+
last count of embeddings.
|
422 |
+
This is the number of embeddings that have already been pushed.
|
423 |
+
current_count (`int`):
|
424 |
+
current count of embeddings.
|
425 |
+
This is the number of embeddings that have been pushed after this batch.
|
426 |
+
embeds (`List[List[float]]`):
|
427 |
+
list of embeddings to push to the repo
|
428 |
+
texts (`List[str]`):
|
429 |
+
list of texts to push to the repo
|
430 |
+
"""
|
431 |
+
|
432 |
+
# TODO: write dataset loading script as well
|
433 |
+
|
434 |
+
temp_ds = Dataset.from_dict(
|
435 |
+
{
|
436 |
+
"embedding": embeds,
|
437 |
+
"text": texts,
|
438 |
+
}
|
439 |
+
)
|
440 |
+
|
441 |
+
data_dir = Path(repo.local_dir) / "data"
|
442 |
+
data_dir.mkdir(exist_ok=True, parents=True)
|
443 |
+
|
444 |
+
temp_ds.to_parquet(
|
445 |
+
str(data_dir / f"embeddings_{last_count}_{current_count}.parquet")
|
446 |
+
)
|
447 |
+
|
448 |
+
repo.push_to_hub(
|
449 |
+
commit_message=f"Embedded examples {last_count} thru {current_count}",
|
450 |
+
blocking=False,
|
451 |
+
auto_lfs_prune=True,
|
452 |
+
)
|