Spaces:
Running
Running
caching
Browse files- .gitignore +166 -0
- app.py +66 -41
- lm_steer/__pycache__/__init__.cpython-310.pyc +0 -0
- lm_steer/__pycache__/utils.cpython-310.pyc +0 -0
- lm_steer/models/__pycache__/get_model.cpython-310.pyc +0 -0
- lm_steer/models/__pycache__/model_base.cpython-310.pyc +0 -0
- lm_steer/models/__pycache__/model_gpt_neo.cpython-310.pyc +0 -0
- lm_steer/models/__pycache__/model_gpt_neox.cpython-310.pyc +0 -0
- lm_steer/models/__pycache__/model_utils.cpython-310.pyc +0 -0
- lm_steer/models/__pycache__/steers.cpython-310.pyc +0 -0
- lm_steer/models/model_base.py +70 -5
- lm_steer/models/model_gpt_j.py +102 -198
- lm_steer/models/model_gpt_neo.py +4 -28
- lm_steer/models/model_gpt_neox.py +3 -52
.gitignore
ADDED
@@ -0,0 +1,166 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Byte-compiled / optimized / DLL files
|
2 |
+
__pycache__/
|
3 |
+
*.py[cod]
|
4 |
+
*$py.class
|
5 |
+
|
6 |
+
# C extensions
|
7 |
+
*.so
|
8 |
+
|
9 |
+
# Distribution / packaging
|
10 |
+
.Python
|
11 |
+
build/
|
12 |
+
develop-eggs/
|
13 |
+
dist/
|
14 |
+
downloads/
|
15 |
+
eggs/
|
16 |
+
.eggs/
|
17 |
+
lib/
|
18 |
+
lib64/
|
19 |
+
parts/
|
20 |
+
sdist/
|
21 |
+
var/
|
22 |
+
wheels/
|
23 |
+
share/python-wheels/
|
24 |
+
*.egg-info/
|
25 |
+
.installed.cfg
|
26 |
+
*.egg
|
27 |
+
MANIFEST
|
28 |
+
|
29 |
+
# PyInstaller
|
30 |
+
# Usually these files are written by a python script from a template
|
31 |
+
# before PyInstaller builds the exe, so as to inject date/other infos into it.
|
32 |
+
*.manifest
|
33 |
+
*.spec
|
34 |
+
|
35 |
+
# Installer logs
|
36 |
+
pip-log.txt
|
37 |
+
pip-delete-this-directory.txt
|
38 |
+
|
39 |
+
# Unit test / coverage reports
|
40 |
+
htmlcov/
|
41 |
+
.tox/
|
42 |
+
.nox/
|
43 |
+
.coverage
|
44 |
+
.coverage.*
|
45 |
+
.cache
|
46 |
+
nosetests.xml
|
47 |
+
coverage.xml
|
48 |
+
*.cover
|
49 |
+
*.py,cover
|
50 |
+
.hypothesis/
|
51 |
+
.pytest_cache/
|
52 |
+
cover/
|
53 |
+
|
54 |
+
# Translations
|
55 |
+
*.mo
|
56 |
+
*.pot
|
57 |
+
|
58 |
+
# Django stuff:
|
59 |
+
*.log
|
60 |
+
local_settings.py
|
61 |
+
db.sqlite3
|
62 |
+
db.sqlite3-journal
|
63 |
+
|
64 |
+
# Flask stuff:
|
65 |
+
instance/
|
66 |
+
.webassets-cache
|
67 |
+
|
68 |
+
# Scrapy stuff:
|
69 |
+
.scrapy
|
70 |
+
|
71 |
+
# Sphinx documentation
|
72 |
+
docs/_build/
|
73 |
+
|
74 |
+
# PyBuilder
|
75 |
+
.pybuilder/
|
76 |
+
target/
|
77 |
+
|
78 |
+
# Jupyter Notebook
|
79 |
+
.ipynb_checkpoints
|
80 |
+
|
81 |
+
# IPython
|
82 |
+
profile_default/
|
83 |
+
ipython_config.py
|
84 |
+
|
85 |
+
# pyenv
|
86 |
+
# For a library or package, you might want to ignore these files since the code is
|
87 |
+
# intended to run in multiple environments; otherwise, check them in:
|
88 |
+
# .python-version
|
89 |
+
|
90 |
+
# pipenv
|
91 |
+
# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
|
92 |
+
# However, in case of collaboration, if having platform-specific dependencies or dependencies
|
93 |
+
# having no cross-platform support, pipenv may install dependencies that don't work, or not
|
94 |
+
# install all needed dependencies.
|
95 |
+
#Pipfile.lock
|
96 |
+
|
97 |
+
# poetry
|
98 |
+
# Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
|
99 |
+
# This is especially recommended for binary packages to ensure reproducibility, and is more
|
100 |
+
# commonly ignored for libraries.
|
101 |
+
# https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
|
102 |
+
#poetry.lock
|
103 |
+
|
104 |
+
# pdm
|
105 |
+
# Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
|
106 |
+
#pdm.lock
|
107 |
+
# pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
|
108 |
+
# in version control.
|
109 |
+
# https://pdm.fming.dev/#use-with-ide
|
110 |
+
.pdm.toml
|
111 |
+
|
112 |
+
# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
|
113 |
+
__pypackages__/
|
114 |
+
|
115 |
+
# Celery stuff
|
116 |
+
celerybeat-schedule
|
117 |
+
celerybeat.pid
|
118 |
+
|
119 |
+
# SageMath parsed files
|
120 |
+
*.sage.py
|
121 |
+
|
122 |
+
# Environments
|
123 |
+
.env
|
124 |
+
.venv
|
125 |
+
env/
|
126 |
+
venv/
|
127 |
+
ENV/
|
128 |
+
env.bak/
|
129 |
+
venv.bak/
|
130 |
+
|
131 |
+
# Spyder project settings
|
132 |
+
.spyderproject
|
133 |
+
.spyproject
|
134 |
+
|
135 |
+
# Rope project settings
|
136 |
+
.ropeproject
|
137 |
+
|
138 |
+
# mkdocs documentation
|
139 |
+
/site
|
140 |
+
|
141 |
+
# mypy
|
142 |
+
.mypy_cache/
|
143 |
+
.dmypy.json
|
144 |
+
dmypy.json
|
145 |
+
|
146 |
+
# Pyre type checker
|
147 |
+
.pyre/
|
148 |
+
|
149 |
+
# pytype static type analyzer
|
150 |
+
.pytype/
|
151 |
+
|
152 |
+
# Cython debug symbols
|
153 |
+
cython_debug/
|
154 |
+
|
155 |
+
# PyCharm
|
156 |
+
# JetBrains specific template is maintained in a separate JetBrains.gitignore that can
|
157 |
+
# be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
|
158 |
+
# and can be added to the global gitignore or merged into this file. For a more nuclear
|
159 |
+
# option (not recommended) you can uncomment the following to ignore the entire idea folder.
|
160 |
+
#.idea/
|
161 |
+
|
162 |
+
# repo-specific
|
163 |
+
**/.DS_Store
|
164 |
+
_logs
|
165 |
+
_logs/
|
166 |
+
checkpoints/
|
app.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
# https://huggingface.co/spaces/Glaciohound/LM-Steer
|
2 |
|
3 |
import torch
|
|
|
4 |
import streamlit as st
|
5 |
import random
|
6 |
import numpy as np
|
@@ -23,39 +24,53 @@ def st_get_model(model_name, low_resource_mode):
|
|
23 |
return model, tokenizer
|
24 |
|
25 |
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
|
|
|
|
|
|
30 |
embeddings = model.steer.lm_head.weight
|
|
|
|
|
31 |
|
32 |
data = []
|
33 |
-
|
34 |
-
|
35 |
-
|
|
|
|
|
|
|
36 |
|
37 |
def filter_words(side_tokens):
|
38 |
output = []
|
39 |
for t in side_tokens:
|
40 |
word = tokenizer.decode([t])
|
41 |
-
if
|
42 |
-
|
43 |
-
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
data.append([
|
46 |
", ".join(filter_words(side_tokens))
|
47 |
for side_tokens in [left_tokens, right_tokens]
|
48 |
])
|
49 |
-
|
50 |
data,
|
51 |
columns=["One Direction", "Another Direction"],
|
52 |
index=[f"Dim {_i}" for _i in range(10)],
|
53 |
-
)
|
54 |
|
55 |
|
56 |
def main():
|
57 |
# set up the page
|
58 |
random.seed(0)
|
|
|
59 |
title = "LM-Steer: Word Embeddings Are Steers for Language Models"
|
60 |
st.set_page_config(
|
61 |
layout="wide",
|
@@ -92,14 +107,14 @@ def main():
|
|
92 |
'''
|
93 |
Due to resource limits, we are only able to provide a few models for
|
94 |
steering. You can also refer to the Github repository:
|
95 |
-
https://github.com/Glaciohound/LM-Steer
|
96 |
Some generated texts may contain toxic or offensive content. Please be
|
97 |
cautious when using the generated texts.
|
98 |
Note that for these smaller models, the generation quality may not be as
|
99 |
good as the larger models (GPT-4, Llama, etc.).
|
100 |
'''
|
101 |
col1, col2 = st.columns(2)
|
102 |
-
|
103 |
"Select a model to steer",
|
104 |
[
|
105 |
"gpt2",
|
@@ -108,48 +123,57 @@ def main():
|
|
108 |
"EleutherAI/pythia-70m",
|
109 |
"EleutherAI/pythia-160m",
|
110 |
"EleutherAI/pythia-410m",
|
111 |
-
# "EleutherAI/pythia-1b",
|
112 |
-
# "EleutherAI/pythia-
|
|
|
|
|
113 |
# "EleutherAI/gpt-j-6B",
|
114 |
],
|
115 |
)
|
116 |
-
low_resource_mode = True if st.session_state.model_name in (
|
117 |
-
|
118 |
-
|
119 |
-
) else False
|
|
|
120 |
model, tokenizer = st_get_model(
|
121 |
-
|
|
|
|
|
122 |
num_param = model.steer.projector1.data.shape[1] ** 2 / 1024 ** 2
|
123 |
total_param = sum(p.numel() for _, p in model.named_parameters()) / \
|
124 |
1024 ** 2
|
125 |
ratio = num_param / total_param
|
126 |
-
|
127 |
-
|
128 |
|
129 |
# steering
|
130 |
-
steer_range =
|
131 |
-
steer_interval = 0.
|
132 |
st.subheader("Enter a sentence and steer the model")
|
133 |
st.session_state.prompt = st.text_input(
|
134 |
"Enter a prompt",
|
135 |
st.session_state.get("prompt", "My life")
|
136 |
)
|
137 |
-
# col1, col2, col3 = st.columns(3, gap="medium")
|
138 |
col1, col2, col3 = st.columns([2, 2, 1], gap="medium")
|
139 |
sentiment = col1.slider(
|
140 |
"Sentiment (the larger the more positive)",
|
141 |
-
-steer_range, steer_range,
|
142 |
detoxification = col2.slider(
|
143 |
"Detoxification Strength (the larger the less toxic)",
|
144 |
-steer_range, steer_range, 0.0,
|
145 |
steer_interval)
|
146 |
-
max_length = col3.number_input("Max length",
|
147 |
col1, col2, col3, _ = st.columns(4)
|
148 |
randomness = col2.checkbox("Random sampling", value=False)
|
149 |
|
150 |
if "output" not in st.session_state:
|
151 |
st.session_state.output = ""
|
152 |
if col1.button("Steer and generate!", type="primary"):
|
|
|
|
|
|
|
|
|
|
|
153 |
with st.spinner("Generating..."):
|
154 |
steer_values = [detoxification, 0, sentiment, 0]
|
155 |
st.session_state.output = model.generate(
|
@@ -159,8 +183,9 @@ def main():
|
|
159 |
min_length=0,
|
160 |
max_length=max_length,
|
161 |
do_sample=True,
|
|
|
162 |
)
|
163 |
-
analyzed_text = \
|
164 |
st.text_area("Generated text:", st.session_state.output, height=200)
|
165 |
|
166 |
# Analysing the sentence
|
@@ -174,7 +199,7 @@ def main():
|
|
174 |
text or use your own. Please note that these two dimensions can be
|
175 |
entangled, as a negative sentiment may also detoxify the text.
|
176 |
'''
|
177 |
-
if st.session_state.get("
|
178 |
st.button("Analyze the styled text", type="primary"):
|
179 |
col1, col2 = st.columns(2)
|
180 |
for name, col, dim, color in zip(
|
@@ -187,9 +212,9 @@ def main():
|
|
187 |
col.subheader(name)
|
188 |
# classification
|
189 |
col.markdown(
|
190 |
-
"#####
|
191 |
_, dist_list, _ = model.steer_analysis(
|
192 |
-
analyzed_text,
|
193 |
dim, -steer_range, steer_range,
|
194 |
bins=2*int(steer_range)+1,
|
195 |
)
|
@@ -209,10 +234,10 @@ def main():
|
|
209 |
pos_steer[dim] = 1
|
210 |
neg_steer[dim] = -1
|
211 |
_, token_evidence = model.evidence_words(
|
212 |
-
analyzed_text,
|
213 |
[pos_steer, neg_steer],
|
214 |
)
|
215 |
-
tokens = tokenizer(analyzed_text).input_ids
|
216 |
tokens = [f"{i:3d}: {tokenizer.decode([t])}"
|
217 |
for i, t in enumerate(tokens)]
|
218 |
col.markdown("##### Token's Evidence Score in the Dimension")
|
@@ -241,13 +266,13 @@ def main():
|
|
241 |
dimension, sometimes only one side of the word embeddings is most relevant
|
242 |
to the style (can be either left or right).
|
243 |
'''
|
244 |
-
dimension
|
245 |
-
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
251 |
|
252 |
|
253 |
if __name__ == "__main__":
|
|
|
1 |
# https://huggingface.co/spaces/Glaciohound/LM-Steer
|
2 |
|
3 |
import torch
|
4 |
+
import nltk
|
5 |
import streamlit as st
|
6 |
import random
|
7 |
import numpy as np
|
|
|
24 |
return model, tokenizer
|
25 |
|
26 |
|
27 |
+
@st.cache_data()
|
28 |
+
def word_embedding_space_analysis(
|
29 |
+
model_name, dim):
|
30 |
+
model = st.session_state.model
|
31 |
+
tokenizer = st.session_state.tokenizer
|
32 |
+
projector1 = model.steer.projector1.data[dim]
|
33 |
+
projector2 = model.steer.projector2.data[dim]
|
34 |
embeddings = model.steer.lm_head.weight
|
35 |
+
matrix = projector1.matmul(projector2.transpose(0, 1))
|
36 |
+
S, V, D = torch.linalg.svd(matrix)
|
37 |
|
38 |
data = []
|
39 |
+
top = 30
|
40 |
+
select_words = 20
|
41 |
+
n_dim = 10
|
42 |
+
for _i in range(n_dim):
|
43 |
+
left_tokens = embeddings.matmul(D[_i]).argsort()[-top:].flip(0)
|
44 |
+
right_tokens = embeddings.matmul(D[_i]).argsort()[:top]
|
45 |
|
46 |
def filter_words(side_tokens):
|
47 |
output = []
|
48 |
for t in side_tokens:
|
49 |
word = tokenizer.decode([t])
|
50 |
+
if (
|
51 |
+
len(word) > 2 and not word[0].isalpha() and
|
52 |
+
word[1:].isalpha() and word[1:].lower().islower()
|
53 |
+
):
|
54 |
+
word = word[1:]
|
55 |
+
if word.lower() in nltk.corpus.words.words():
|
56 |
+
output.append(word)
|
57 |
+
return output[:select_words]
|
58 |
|
59 |
data.append([
|
60 |
", ".join(filter_words(side_tokens))
|
61 |
for side_tokens in [left_tokens, right_tokens]
|
62 |
])
|
63 |
+
return pd.DataFrame(
|
64 |
data,
|
65 |
columns=["One Direction", "Another Direction"],
|
66 |
index=[f"Dim {_i}" for _i in range(10)],
|
67 |
+
)
|
68 |
|
69 |
|
70 |
def main():
|
71 |
# set up the page
|
72 |
random.seed(0)
|
73 |
+
nltk.download('words')
|
74 |
title = "LM-Steer: Word Embeddings Are Steers for Language Models"
|
75 |
st.set_page_config(
|
76 |
layout="wide",
|
|
|
107 |
'''
|
108 |
Due to resource limits, we are only able to provide a few models for
|
109 |
steering. You can also refer to the Github repository:
|
110 |
+
https://github.com/Glaciohound/LM-Steer to host larger models.
|
111 |
Some generated texts may contain toxic or offensive content. Please be
|
112 |
cautious when using the generated texts.
|
113 |
Note that for these smaller models, the generation quality may not be as
|
114 |
good as the larger models (GPT-4, Llama, etc.).
|
115 |
'''
|
116 |
col1, col2 = st.columns(2)
|
117 |
+
model_name = col1.selectbox(
|
118 |
"Select a model to steer",
|
119 |
[
|
120 |
"gpt2",
|
|
|
123 |
"EleutherAI/pythia-70m",
|
124 |
"EleutherAI/pythia-160m",
|
125 |
"EleutherAI/pythia-410m",
|
126 |
+
# "EleutherAI/pythia-1b",
|
127 |
+
# "EleutherAI/pythia-1.4b",
|
128 |
+
# "EleutherAI/pythia-2.8b",
|
129 |
+
# "EleutherAI/pythia-6.9b",
|
130 |
# "EleutherAI/gpt-j-6B",
|
131 |
],
|
132 |
)
|
133 |
+
# low_resource_mode = True if st.session_state.model_name in (
|
134 |
+
# "EleutherAI/pythia-1.4b", "EleutherAI/pythia-2.8b",
|
135 |
+
# "EleutherAI/pythia-6.9b", "EleutherAI/gpt-j-6B",
|
136 |
+
# ) else False
|
137 |
+
low_resource_mode = False
|
138 |
model, tokenizer = st_get_model(
|
139 |
+
model_name, low_resource_mode)
|
140 |
+
st.session_state.model = model
|
141 |
+
st.session_state.tokenizer = tokenizer
|
142 |
num_param = model.steer.projector1.data.shape[1] ** 2 / 1024 ** 2
|
143 |
total_param = sum(p.numel() for _, p in model.named_parameters()) / \
|
144 |
1024 ** 2
|
145 |
ratio = num_param / total_param
|
146 |
+
st.write(f"Steered {num_param:.1f}M out of {total_param:.1f}M "
|
147 |
+
"parameters, ratio: {:.2%}".format(ratio))
|
148 |
|
149 |
# steering
|
150 |
+
steer_range = 3.
|
151 |
+
steer_interval = 0.2
|
152 |
st.subheader("Enter a sentence and steer the model")
|
153 |
st.session_state.prompt = st.text_input(
|
154 |
"Enter a prompt",
|
155 |
st.session_state.get("prompt", "My life")
|
156 |
)
|
|
|
157 |
col1, col2, col3 = st.columns([2, 2, 1], gap="medium")
|
158 |
sentiment = col1.slider(
|
159 |
"Sentiment (the larger the more positive)",
|
160 |
+
-steer_range, steer_range, 0.0, steer_interval)
|
161 |
detoxification = col2.slider(
|
162 |
"Detoxification Strength (the larger the less toxic)",
|
163 |
-steer_range, steer_range, 0.0,
|
164 |
steer_interval)
|
165 |
+
max_length = col3.number_input("Max length", 20, 200, 20, 20)
|
166 |
col1, col2, col3, _ = st.columns(4)
|
167 |
randomness = col2.checkbox("Random sampling", value=False)
|
168 |
|
169 |
if "output" not in st.session_state:
|
170 |
st.session_state.output = ""
|
171 |
if col1.button("Steer and generate!", type="primary"):
|
172 |
+
if sentiment == 0 and detoxification == 0:
|
173 |
+
'''
|
174 |
+
**The steer values are both 0, which means the steered model
|
175 |
+
is the same as the original model.**
|
176 |
+
'''
|
177 |
with st.spinner("Generating..."):
|
178 |
steer_values = [detoxification, 0, sentiment, 0]
|
179 |
st.session_state.output = model.generate(
|
|
|
183 |
min_length=0,
|
184 |
max_length=max_length,
|
185 |
do_sample=True,
|
186 |
+
top_p=0.9,
|
187 |
)
|
188 |
+
st.session_state.analyzed_text = \
|
189 |
st.text_area("Generated text:", st.session_state.output, height=200)
|
190 |
|
191 |
# Analysing the sentence
|
|
|
199 |
text or use your own. Please note that these two dimensions can be
|
200 |
entangled, as a negative sentiment may also detoxify the text.
|
201 |
'''
|
202 |
+
if st.session_state.get("analyzed_text", "") != "" and \
|
203 |
st.button("Analyze the styled text", type="primary"):
|
204 |
col1, col2 = st.columns(2)
|
205 |
for name, col, dim, color in zip(
|
|
|
212 |
col.subheader(name)
|
213 |
# classification
|
214 |
col.markdown(
|
215 |
+
"##### Sentence Classification Distribution")
|
216 |
_, dist_list, _ = model.steer_analysis(
|
217 |
+
st.session_state.analyzed_text,
|
218 |
dim, -steer_range, steer_range,
|
219 |
bins=2*int(steer_range)+1,
|
220 |
)
|
|
|
234 |
pos_steer[dim] = 1
|
235 |
neg_steer[dim] = -1
|
236 |
_, token_evidence = model.evidence_words(
|
237 |
+
st.session_state.analyzed_text,
|
238 |
[pos_steer, neg_steer],
|
239 |
)
|
240 |
+
tokens = tokenizer(st.session_state.analyzed_text).input_ids
|
241 |
tokens = [f"{i:3d}: {tokenizer.decode([t])}"
|
242 |
for i, t in enumerate(tokens)]
|
243 |
col.markdown("##### Token's Evidence Score in the Dimension")
|
|
|
266 |
dimension, sometimes only one side of the word embeddings is most relevant
|
267 |
to the style (can be either left or right).
|
268 |
'''
|
269 |
+
for dimension in ["Sentiment", "Detoxification"]:
|
270 |
+
f'##### {dimension} Dimension'
|
271 |
+
dim = 2 if dimension == "Sentiment" else 0
|
272 |
+
analysis_result = word_embedding_space_analysis(
|
273 |
+
model_name, dim)
|
274 |
+
with st.expander("Show the analysis results"):
|
275 |
+
st.table(analysis_result)
|
276 |
|
277 |
|
278 |
if __name__ == "__main__":
|
lm_steer/__pycache__/__init__.cpython-310.pyc
DELETED
Binary file (166 Bytes)
|
|
lm_steer/__pycache__/utils.cpython-310.pyc
DELETED
Binary file (1.44 kB)
|
|
lm_steer/models/__pycache__/get_model.cpython-310.pyc
DELETED
Binary file (1.48 kB)
|
|
lm_steer/models/__pycache__/model_base.cpython-310.pyc
DELETED
Binary file (4.88 kB)
|
|
lm_steer/models/__pycache__/model_gpt_neo.cpython-310.pyc
DELETED
Binary file (2.6 kB)
|
|
lm_steer/models/__pycache__/model_gpt_neox.cpython-310.pyc
DELETED
Binary file (3.7 kB)
|
|
lm_steer/models/__pycache__/model_utils.cpython-310.pyc
DELETED
Binary file (2.23 kB)
|
|
lm_steer/models/__pycache__/steers.cpython-310.pyc
DELETED
Binary file (3.07 kB)
|
|
lm_steer/models/model_base.py
CHANGED
@@ -26,8 +26,8 @@ class LMSteerBase(nn.Module):
|
|
26 |
if isinstance(comparing_steer_values, list):
|
27 |
comparing_steer_values = \
|
28 |
torch.Tensor(comparing_steer_values).to(self.device)
|
29 |
-
if (comparing_steer_values[0] - comparing_steer_values[1]
|
30 |
-
|
31 |
return [(prompt, None)]
|
32 |
tokenized = self.tokenizer(
|
33 |
prompt, return_tensors="pt",
|
@@ -162,12 +162,77 @@ class LMSteerBase(nn.Module):
|
|
162 |
self.device)
|
163 |
self.steer.set_value(steer_values[None])
|
164 |
with torch.no_grad():
|
165 |
-
|
166 |
-
prompt,
|
|
|
|
|
|
|
167 |
do_sample=do_sample, temperature=temperature, top_p=top_p,
|
168 |
min_length=min_length, max_length=max_length,
|
169 |
pad_token_id=self.tokenizer.pad_token_id,
|
170 |
)
|
171 |
-
text = text[0]
|
172 |
|
173 |
return text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
if isinstance(comparing_steer_values, list):
|
27 |
comparing_steer_values = \
|
28 |
torch.Tensor(comparing_steer_values).to(self.device)
|
29 |
+
if (comparing_steer_values[0] - comparing_steer_values[1]).abs().sum()\
|
30 |
+
<= 0.2:
|
31 |
return [(prompt, None)]
|
32 |
tokenized = self.tokenizer(
|
33 |
prompt, return_tensors="pt",
|
|
|
162 |
self.device)
|
163 |
self.steer.set_value(steer_values[None])
|
164 |
with torch.no_grad():
|
165 |
+
inputs = self.tokenizer(
|
166 |
+
prompt, return_tensors="pt").to(self.device)
|
167 |
+
text = self.model.generate(
|
168 |
+
**inputs,
|
169 |
+
num_beams=num_beams, num_beam_groups=num_beam_groups,
|
170 |
do_sample=do_sample, temperature=temperature, top_p=top_p,
|
171 |
min_length=min_length, max_length=max_length,
|
172 |
pad_token_id=self.tokenizer.pad_token_id,
|
173 |
)
|
174 |
+
text = self.tokenizer.decode(text[0], skip_special_tokens=True)
|
175 |
|
176 |
return text
|
177 |
+
|
178 |
+
def generate_low_resource(
|
179 |
+
self, prompt, steer_values, min_length=20, max_length=100,
|
180 |
+
seed=None, num_beams=1, num_beam_groups=1, do_sample=True,
|
181 |
+
temperature=1, top_p=1
|
182 |
+
):
|
183 |
+
'''
|
184 |
+
prompt: a string
|
185 |
+
steer_values
|
186 |
+
min_length: minimum generation length
|
187 |
+
max_length: maximum generation length
|
188 |
+
seed: seed for generation. None if not specified.
|
189 |
+
'''
|
190 |
+
if seed is not None:
|
191 |
+
set_seed(seed)
|
192 |
+
steer_values = torch.Tensor(steer_values).to(
|
193 |
+
self.device)
|
194 |
+
fp16 = torch.float16
|
195 |
+
steer_values = steer_values.to(fp16)
|
196 |
+
self.steer.projector1.data = self.steer.projector1.to(fp16)
|
197 |
+
self.steer.projector2.data = self.steer.projector2.to(fp16)
|
198 |
+
self.steer.set_value(steer_values[None])
|
199 |
+
with torch.no_grad():
|
200 |
+
input_ids = self.tokenizer(
|
201 |
+
prompt, return_tensors="pt").input_ids.to(self.device)
|
202 |
+
gen_tokens = self.model.generate(
|
203 |
+
input_ids,
|
204 |
+
num_beams=num_beams, num_beam_groups=num_beam_groups,
|
205 |
+
do_sample=do_sample, temperature=temperature, top_p=top_p,
|
206 |
+
min_length=min_length, max_length=max_length,
|
207 |
+
pad_token_id=self.tokenizer.pad_token_id)
|
208 |
+
text = self.tokenizer.batch_decode(gen_tokens)[0]
|
209 |
+
|
210 |
+
# recovering
|
211 |
+
fp32 = torch.float32
|
212 |
+
self.steer.projector1.data = self.steer.projector1.to(fp32)
|
213 |
+
self.steer.projector2.data = self.steer.projector2.to(fp32)
|
214 |
+
return text
|
215 |
+
|
216 |
+
def state_dict(self):
|
217 |
+
return self.steer.state_dict()
|
218 |
+
|
219 |
+
def load_state_dict(self, state_dict):
|
220 |
+
self.steer.load_state_dict(state_dict)
|
221 |
+
|
222 |
+
def parameters(self):
|
223 |
+
return self.steer.parameters()
|
224 |
+
|
225 |
+
def to_device(self, device):
|
226 |
+
self.model.to(device)
|
227 |
+
self.device = device
|
228 |
+
|
229 |
+
def regularization_term(self):
|
230 |
+
return self.steer.regularization_term()
|
231 |
+
|
232 |
+
def forward(self, input_ids, attention_mask, steer_values):
|
233 |
+
self.steer.set_value(steer_values)
|
234 |
+
output = self.model(
|
235 |
+
input_ids=input_ids,
|
236 |
+
attention_mask=attention_mask,
|
237 |
+
labels=input_ids)
|
238 |
+
return output
|
lm_steer/models/model_gpt_j.py
CHANGED
@@ -1,27 +1,14 @@
|
|
1 |
import torch
|
2 |
-
import numpy as np
|
3 |
-
import torch.nn as nn
|
4 |
import torch.nn.functional as F
|
5 |
from transformers import GPTJForCausalLM, AutoTokenizer
|
6 |
|
7 |
from .model_utils import Hack_no_grad, find_max_subspans
|
8 |
from .steers import Projected_Adaptor
|
|
|
9 |
from lm_steer.utils import set_seed
|
10 |
|
11 |
|
12 |
-
|
13 |
-
'!', '"', '$', '%', '&', "'", '(', ')', '*', '+', ',', '-', '.',
|
14 |
-
# '/', '#',
|
15 |
-
':', ';', '<', '=', '>', '?', '@',
|
16 |
-
'[', '\\', ']', '^', '_', '`',
|
17 |
-
'{', '|', '}', '~',
|
18 |
-
'¨', '©', 'ª', '«', '¬', '®', '¯', '°', '±', '²', '³', '´', 'µ', '¶', '·',
|
19 |
-
'¸', '¹', 'º', '»', '¼', '½', '¾',
|
20 |
-
'\n', ' ',
|
21 |
-
]
|
22 |
-
|
23 |
-
|
24 |
-
class Switching_GPTJModel(nn.Module):
|
25 |
def __init__(self, model_name, adapted_component, adaptor_class,
|
26 |
num_steers, rank, epsilon, init_var, low_resource_mode):
|
27 |
super().__init__()
|
@@ -67,31 +54,6 @@ class Switching_GPTJModel(nn.Module):
|
|
67 |
else:
|
68 |
raise NotImplementedError()
|
69 |
|
70 |
-
def forward(self, input_ids, attention_mask, steer_values):
|
71 |
-
self.steer.set_value(steer_values)
|
72 |
-
output = self.model(
|
73 |
-
input_ids=input_ids,
|
74 |
-
attention_mask=attention_mask,
|
75 |
-
labels=input_ids)
|
76 |
-
return output
|
77 |
-
|
78 |
-
def parameters(self):
|
79 |
-
return self.steer.parameters()
|
80 |
-
|
81 |
-
def state_dict(self):
|
82 |
-
return self.steer.state_dict()
|
83 |
-
|
84 |
-
def load_state_dict(self, state_dict):
|
85 |
-
self.steer.load_state_dict(state_dict)
|
86 |
-
|
87 |
-
def to_device(self, device):
|
88 |
-
# self.generator.device = device
|
89 |
-
self.model.to(device)
|
90 |
-
self.device = device
|
91 |
-
|
92 |
-
def regularization_term(self):
|
93 |
-
return self.steer.regularization_term()
|
94 |
-
|
95 |
def generate(self, prompt, steer_values, min_length=20, max_length=100,
|
96 |
seed=None, num_beams=1, num_beam_groups=1, do_sample=True,
|
97 |
temperature=1, top_p=1):
|
@@ -102,33 +64,9 @@ class Switching_GPTJModel(nn.Module):
|
|
102 |
max_length: maximum generation length
|
103 |
seed: seed for generation. None if not specified.
|
104 |
'''
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
self.device)
|
109 |
-
if self.low_resource_mode:
|
110 |
-
fp16 = torch.float16
|
111 |
-
steer_values = steer_values.to(fp16)
|
112 |
-
self.steer.projector1.data = self.steer.projector1.to(fp16)
|
113 |
-
self.steer.projector2.data = self.steer.projector2.to(fp16)
|
114 |
-
self.steer.set_value(steer_values[None])
|
115 |
-
with torch.no_grad():
|
116 |
-
input_ids = self.tokenizer(
|
117 |
-
prompt, return_tensors="pt").input_ids.to(self.device)
|
118 |
-
gen_tokens = self.model.generate(
|
119 |
-
input_ids,
|
120 |
-
num_beams=num_beams, num_beam_groups=num_beam_groups,
|
121 |
-
do_sample=do_sample, temperature=temperature, top_p=top_p,
|
122 |
-
min_new_tokens=min_length, max_new_tokens=max_length,
|
123 |
-
pad_token_id=self.tokenizer.pad_token_id)
|
124 |
-
text = self.tokenizer.batch_decode(gen_tokens)[0]
|
125 |
-
|
126 |
-
# recovering
|
127 |
-
if self.low_resource_mode:
|
128 |
-
fp32 = torch.float32
|
129 |
-
self.steer.projector1.data = self.steer.projector1.to(fp32)
|
130 |
-
self.steer.projector2.data = self.steer.projector2.to(fp32)
|
131 |
-
return text
|
132 |
|
133 |
def generate_multiple(
|
134 |
self, prompts, steer_values, min_length=20, max_length=100,
|
@@ -167,13 +105,14 @@ class Switching_GPTJModel(nn.Module):
|
|
167 |
self.steer.projector2.data = self.steer.projector2.to(fp32)
|
168 |
return text
|
169 |
|
170 |
-
# def evidence_words(self, prompt, original_steer_values,
|
171 |
-
# max_length=10):
|
172 |
# if isinstance(original_steer_values, list):
|
173 |
# original_steer_values = torch.Tensor(original_steer_values)
|
174 |
# if original_steer_values.abs().sum() <= 0.2:
|
175 |
# return [(prompt, None)]
|
176 |
-
# tokenized = self.tokenizer(
|
|
|
177 |
# input_ids = torch.LongTensor(tokenized["input_ids"]).to(self.device)
|
178 |
# input_ids = input_ids.expand(2, -1)
|
179 |
# attention_mask = torch.LongTensor(tokenized["attention_mask"]).to(
|
@@ -201,133 +140,98 @@ class Switching_GPTJModel(nn.Module):
|
|
201 |
# )
|
202 |
# loss_token = loss_token.reshape(2, length - 1)
|
203 |
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
loss_token = loss_token.reshape(2, length - 1)
|
238 |
-
|
239 |
-
token_evidence = (- loss_token[0] + loss_token[1])
|
240 |
-
tokens = input_ids[0]
|
241 |
-
evidence_segments = find_max_subspans(
|
242 |
-
token_evidence.cpu().numpy().tolist(), max_segments, max_length)[0]
|
243 |
-
evidence_segments = [
|
244 |
-
(_seg[0]+1, _seg[1]+1) for _seg in evidence_segments]
|
245 |
-
start = 0
|
246 |
-
output = []
|
247 |
-
color = (
|
248 |
-
"gray" if original_steer_values.shape[0] > 1
|
249 |
-
else "red" if original_steer_values[0] > 0
|
250 |
-
else "blue"
|
251 |
-
)
|
252 |
-
if len(evidence_segments) > 0:
|
253 |
-
for _segment in evidence_segments:
|
254 |
-
if _segment[0] > start:
|
255 |
-
output.append((
|
256 |
-
self.tokenizer.decode(tokens[start: _segment[0]]),
|
257 |
-
None
|
258 |
-
))
|
259 |
-
output.append((
|
260 |
-
self.tokenizer.decode(tokens[_segment[0]: _segment[1]]),
|
261 |
-
color
|
262 |
-
))
|
263 |
-
start = _segment[1]
|
264 |
-
length = tokens.shape[-1]
|
265 |
-
if _segment[1] < length:
|
266 |
-
output.append((
|
267 |
-
self.tokenizer.decode(tokens[_segment[1]: length]),
|
268 |
-
None
|
269 |
-
))
|
270 |
-
else:
|
271 |
-
output = [(prompt, None)]
|
272 |
-
|
273 |
-
if self.low_resource_mode:
|
274 |
-
fp32 = torch.float32
|
275 |
-
self.steer.projector1.data = self.steer.projector1.to(fp32)
|
276 |
-
self.steer.projector2.data = self.steer.projector2.to(fp32)
|
277 |
-
return output
|
278 |
-
|
279 |
-
def steer_analysis(self, prompt, steer_dim, min_value=-3, max_value=3,
|
280 |
-
bins=7, truncation_length=1024):
|
281 |
-
tokenized = self.tokenizer(
|
282 |
-
prompt, return_tensors="pt",
|
283 |
-
max_length=truncation_length,
|
284 |
-
truncation=True)
|
285 |
-
input_ids = torch.LongTensor(tokenized["input_ids"]).to(self.device)
|
286 |
-
input_ids = input_ids.expand(bins + 1, -1)
|
287 |
-
attention_mask = torch.LongTensor(tokenized["attention_mask"]).to(
|
288 |
-
self.device)
|
289 |
-
attention_mask = attention_mask.expand(bins + 1, -1)
|
290 |
-
steer_values = torch.zeros(bins+1, self.num_steers).to(self.device)
|
291 |
-
for bin_i in range(bins):
|
292 |
-
steer_values[bin_i, steer_dim] = (
|
293 |
-
min_value + (max_value - min_value) / (bins - 1) * bin_i
|
294 |
-
)
|
295 |
-
if self.low_resource_mode:
|
296 |
-
fp16 = torch.float16
|
297 |
-
steer_values = steer_values.to(fp16)
|
298 |
-
self.steer.projector1.data = self.steer.projector1.to(fp16)
|
299 |
-
self.steer.projector2.data = self.steer.projector2.to(fp16)
|
300 |
-
self.steer.set_value(steer_values)
|
301 |
-
with torch.no_grad():
|
302 |
-
output = self.model(
|
303 |
-
input_ids=input_ids,
|
304 |
-
attention_mask=attention_mask,
|
305 |
-
labels=input_ids)
|
306 |
-
length = input_ids.shape[1]
|
307 |
-
loss_token = F.cross_entropy(
|
308 |
-
output.logits[:, :-1].reshape((bins+1)*(length-1), -1),
|
309 |
-
input_ids[:, 1:].reshape(-1),
|
310 |
-
reduction="none"
|
311 |
-
)
|
312 |
-
loss_token = loss_token.reshape(bins + 1, length - 1)
|
313 |
-
loss = loss_token.mean(-1)[:-1]
|
314 |
-
dist = ((- loss + loss.mean()) * 100).softmax(0)
|
315 |
-
dist_list = list(zip(
|
316 |
-
[
|
317 |
-
min_value + (max_value - min_value) / (bins - 1) * bin_i
|
318 |
-
for bin_i in range(bins)
|
319 |
-
],
|
320 |
-
dist.tolist(),
|
321 |
-
))
|
322 |
-
best_guess = loss.argmin(0)
|
323 |
-
best_guess_value = min_value + \
|
324 |
-
(max_value - min_value) / (bins - 1) * best_guess.item()
|
325 |
|
326 |
-
|
327 |
-
|
328 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
329 |
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
|
|
1 |
import torch
|
|
|
|
|
2 |
import torch.nn.functional as F
|
3 |
from transformers import GPTJForCausalLM, AutoTokenizer
|
4 |
|
5 |
from .model_utils import Hack_no_grad, find_max_subspans
|
6 |
from .steers import Projected_Adaptor
|
7 |
+
from .model_base import LMSteerBase
|
8 |
from lm_steer.utils import set_seed
|
9 |
|
10 |
|
11 |
+
class Switching_GPTJModel(LMSteerBase):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
def __init__(self, model_name, adapted_component, adaptor_class,
|
13 |
num_steers, rank, epsilon, init_var, low_resource_mode):
|
14 |
super().__init__()
|
|
|
54 |
else:
|
55 |
raise NotImplementedError()
|
56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
def generate(self, prompt, steer_values, min_length=20, max_length=100,
|
58 |
seed=None, num_beams=1, num_beam_groups=1, do_sample=True,
|
59 |
temperature=1, top_p=1):
|
|
|
64 |
max_length: maximum generation length
|
65 |
seed: seed for generation. None if not specified.
|
66 |
'''
|
67 |
+
return super().generate_low_resource(
|
68 |
+
prompt, steer_values, min_length, max_length, seed,
|
69 |
+
num_beams, num_beam_groups, do_sample, temperature, top_p)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
|
71 |
def generate_multiple(
|
72 |
self, prompts, steer_values, min_length=20, max_length=100,
|
|
|
105 |
self.steer.projector2.data = self.steer.projector2.to(fp32)
|
106 |
return text
|
107 |
|
108 |
+
# def evidence_words(self, prompt, original_steer_values,
|
109 |
+
# truncation_length=1024, max_segments=4, max_length=10):
|
110 |
# if isinstance(original_steer_values, list):
|
111 |
# original_steer_values = torch.Tensor(original_steer_values)
|
112 |
# if original_steer_values.abs().sum() <= 0.2:
|
113 |
# return [(prompt, None)]
|
114 |
+
# tokenized = self.tokenizer(
|
115 |
+
# prompt, return_tensors="pt", max_length=truncation_length, truncation=True)
|
116 |
# input_ids = torch.LongTensor(tokenized["input_ids"]).to(self.device)
|
117 |
# input_ids = input_ids.expand(2, -1)
|
118 |
# attention_mask = torch.LongTensor(tokenized["attention_mask"]).to(
|
|
|
140 |
# )
|
141 |
# loss_token = loss_token.reshape(2, length - 1)
|
142 |
|
143 |
+
# token_evidence = (- loss_token[0] + loss_token[1])
|
144 |
+
# tokens = input_ids[0]
|
145 |
+
# evidence_segments = find_max_subspans(
|
146 |
+
# token_evidence.cpu().numpy().tolist(), max_segments, max_length)[0]
|
147 |
+
# evidence_segments = [
|
148 |
+
# (_seg[0]+1, _seg[1]+1) for _seg in evidence_segments]
|
149 |
+
# start = 0
|
150 |
+
# output = []
|
151 |
+
# color = (
|
152 |
+
# "gray" if original_steer_values.shape[0] > 1
|
153 |
+
# else "red" if original_steer_values[0] > 0
|
154 |
+
# else "blue"
|
155 |
+
# )
|
156 |
+
# if len(evidence_segments) > 0:
|
157 |
+
# for _segment in evidence_segments:
|
158 |
+
# if _segment[0] > start:
|
159 |
+
# output.append((
|
160 |
+
# self.tokenizer.decode(tokens[start: _segment[0]]),
|
161 |
+
# None
|
162 |
+
# ))
|
163 |
+
# output.append((
|
164 |
+
# self.tokenizer.decode(tokens[_segment[0]: _segment[1]]),
|
165 |
+
# color
|
166 |
+
# ))
|
167 |
+
# start = _segment[1]
|
168 |
+
# length = tokens.shape[-1]
|
169 |
+
# if _segment[1] < length:
|
170 |
+
# output.append((
|
171 |
+
# self.tokenizer.decode(tokens[_segment[1]: length]),
|
172 |
+
# None
|
173 |
+
# ))
|
174 |
+
# else:
|
175 |
+
# output = [(prompt, None)]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
176 |
|
177 |
+
# if self.low_resource_mode:
|
178 |
+
# fp32 = torch.float32
|
179 |
+
# self.steer.projector1.data = self.steer.projector1.to(fp32)
|
180 |
+
# self.steer.projector2.data = self.steer.projector2.to(fp32)
|
181 |
+
# return output
|
182 |
+
|
183 |
+
# def steer_analysis(self, prompt, steer_dim, min_value=-3, max_value=3,
|
184 |
+
# bins=7, truncation_length=1024):
|
185 |
+
# tokenized = self.tokenizer(
|
186 |
+
# prompt, return_tensors="pt",
|
187 |
+
# max_length=truncation_length,
|
188 |
+
# truncation=True)
|
189 |
+
# input_ids = torch.LongTensor(tokenized["input_ids"]).to(self.device)
|
190 |
+
# input_ids = input_ids.expand(bins + 1, -1)
|
191 |
+
# attention_mask = torch.LongTensor(tokenized["attention_mask"]).to(
|
192 |
+
# self.device)
|
193 |
+
# attention_mask = attention_mask.expand(bins + 1, -1)
|
194 |
+
# steer_values = torch.zeros(bins+1, self.num_steers).to(self.device)
|
195 |
+
# for bin_i in range(bins):
|
196 |
+
# steer_values[bin_i, steer_dim] = (
|
197 |
+
# min_value + (max_value - min_value) / (bins - 1) * bin_i
|
198 |
+
# )
|
199 |
+
# if self.low_resource_mode:
|
200 |
+
# fp16 = torch.float16
|
201 |
+
# steer_values = steer_values.to(fp16)
|
202 |
+
# self.steer.projector1.data = self.steer.projector1.to(fp16)
|
203 |
+
# self.steer.projector2.data = self.steer.projector2.to(fp16)
|
204 |
+
# self.steer.set_value(steer_values)
|
205 |
+
# with torch.no_grad():
|
206 |
+
# output = self.model(
|
207 |
+
# input_ids=input_ids,
|
208 |
+
# attention_mask=attention_mask,
|
209 |
+
# labels=input_ids)
|
210 |
+
# length = input_ids.shape[1]
|
211 |
+
# loss_token = F.cross_entropy(
|
212 |
+
# output.logits[:, :-1].reshape((bins+1)*(length-1), -1),
|
213 |
+
# input_ids[:, 1:].reshape(-1),
|
214 |
+
# reduction="none"
|
215 |
+
# )
|
216 |
+
# loss_token = loss_token.reshape(bins + 1, length - 1)
|
217 |
+
# loss = loss_token.mean(-1)[:-1]
|
218 |
+
# dist = ((- loss + loss.mean()) * 100).softmax(0)
|
219 |
+
# dist_list = list(zip(
|
220 |
+
# [
|
221 |
+
# min_value + (max_value - min_value) / (bins - 1) * bin_i
|
222 |
+
# for bin_i in range(bins)
|
223 |
+
# ],
|
224 |
+
# dist.tolist(),
|
225 |
+
# ))
|
226 |
+
# best_guess = loss.argmin(0)
|
227 |
+
# best_guess_value = min_value + \
|
228 |
+
# (max_value - min_value) / (bins - 1) * best_guess.item()
|
229 |
+
|
230 |
+
# token_evidence = self.evidence_words(
|
231 |
+
# prompt, steer_values[best_guess],
|
232 |
+
# )
|
233 |
|
234 |
+
# if self.low_resource_mode:
|
235 |
+
# fp32 = torch.float32
|
236 |
+
# self.steer.projector1.data = self.steer.projector1.to(fp32)
|
237 |
+
# return best_guess_value, dist_list, token_evidence
|
lm_steer/models/model_gpt_neo.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
import torch
|
2 |
from transformers import pipeline
|
3 |
|
|
|
4 |
from .model_utils import Hack_no_grad
|
5 |
from .steers import Projected_Adaptor
|
6 |
from .model_base import LMSteerBase
|
@@ -12,9 +13,9 @@ class Switching_GPTNeoModel(LMSteerBase):
|
|
12 |
low_resource_mode):
|
13 |
super().__init__()
|
14 |
self.adapted_component = adapted_component
|
15 |
-
self.
|
16 |
-
self.
|
17 |
-
self.
|
18 |
self.tokenizer.pad_token = self.tokenizer.eos_token
|
19 |
self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
|
20 |
self.init_var = init_var
|
@@ -39,28 +40,3 @@ class Switching_GPTNeoModel(LMSteerBase):
|
|
39 |
self.model.transformer.set_input_embeddings(self.steer)
|
40 |
else:
|
41 |
raise NotImplementedError()
|
42 |
-
|
43 |
-
def forward(self, input_ids, attention_mask, steer_values):
|
44 |
-
self.steer.set_value(steer_values)
|
45 |
-
output = self.model(
|
46 |
-
input_ids=input_ids,
|
47 |
-
attention_mask=attention_mask,
|
48 |
-
labels=input_ids)
|
49 |
-
return output
|
50 |
-
|
51 |
-
def parameters(self):
|
52 |
-
return self.steer.parameters()
|
53 |
-
|
54 |
-
def state_dict(self):
|
55 |
-
return self.steer.state_dict()
|
56 |
-
|
57 |
-
def load_state_dict(self, state_dict):
|
58 |
-
self.steer.load_state_dict(state_dict)
|
59 |
-
|
60 |
-
def to_device(self, device):
|
61 |
-
self.generator.device = device
|
62 |
-
self.model.to(device)
|
63 |
-
self.device = device
|
64 |
-
|
65 |
-
def regularization_term(self):
|
66 |
-
return self.steer.regularization_term()
|
|
|
1 |
import torch
|
2 |
from transformers import pipeline
|
3 |
|
4 |
+
|
5 |
from .model_utils import Hack_no_grad
|
6 |
from .steers import Projected_Adaptor
|
7 |
from .model_base import LMSteerBase
|
|
|
13 |
low_resource_mode):
|
14 |
super().__init__()
|
15 |
self.adapted_component = adapted_component
|
16 |
+
self.pipeline = pipeline('text-generation', model=model_name)
|
17 |
+
self.model = self.pipeline.model
|
18 |
+
self.tokenizer = self.pipeline.tokenizer
|
19 |
self.tokenizer.pad_token = self.tokenizer.eos_token
|
20 |
self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
|
21 |
self.init_var = init_var
|
|
|
40 |
self.model.transformer.set_input_embeddings(self.steer)
|
41 |
else:
|
42 |
raise NotImplementedError()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
lm_steer/models/model_gpt_neox.py
CHANGED
@@ -4,7 +4,6 @@ from transformers import GPTNeoXForCausalLM, AutoTokenizer
|
|
4 |
from .model_utils import Hack_no_grad
|
5 |
from .steers import Projected_Adaptor
|
6 |
from .model_base import LMSteerBase
|
7 |
-
from lm_steer.utils import set_seed
|
8 |
|
9 |
|
10 |
class Switching_GPTNeoXModel(LMSteerBase):
|
@@ -42,30 +41,6 @@ class Switching_GPTNeoXModel(LMSteerBase):
|
|
42 |
else:
|
43 |
raise NotImplementedError()
|
44 |
|
45 |
-
def forward(self, input_ids, attention_mask, steer_values):
|
46 |
-
self.steer.set_value(steer_values)
|
47 |
-
output = self.model(
|
48 |
-
input_ids=input_ids,
|
49 |
-
attention_mask=attention_mask,
|
50 |
-
labels=input_ids)
|
51 |
-
return output
|
52 |
-
|
53 |
-
def parameters(self):
|
54 |
-
return self.steer.parameters()
|
55 |
-
|
56 |
-
def state_dict(self):
|
57 |
-
return self.steer.state_dict()
|
58 |
-
|
59 |
-
def load_state_dict(self, state_dict):
|
60 |
-
self.steer.load_state_dict(state_dict)
|
61 |
-
|
62 |
-
def to_device(self, device):
|
63 |
-
self.model.to(device)
|
64 |
-
self.device = device
|
65 |
-
|
66 |
-
def regularization_term(self):
|
67 |
-
return self.steer.regularization_term()
|
68 |
-
|
69 |
def generate(self, prompt, steer_values, min_length=20, max_length=100,
|
70 |
seed=None, num_beams=1, num_beam_groups=1, do_sample=True,
|
71 |
temperature=1, top_p=1):
|
@@ -76,30 +51,6 @@ class Switching_GPTNeoXModel(LMSteerBase):
|
|
76 |
max_length: maximum generation length
|
77 |
seed: seed for generation. None if not specified.
|
78 |
'''
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
self.device)
|
83 |
-
if self.low_resource_mode:
|
84 |
-
fp16 = torch.float16
|
85 |
-
steer_values = steer_values.to(fp16)
|
86 |
-
self.steer.projector1.data = self.steer.projector1.to(fp16)
|
87 |
-
self.steer.projector2.data = self.steer.projector2.to(fp16)
|
88 |
-
self.steer.set_value(steer_values[None])
|
89 |
-
with torch.no_grad():
|
90 |
-
input_ids = self.tokenizer(
|
91 |
-
prompt, return_tensors="pt").input_ids.to(self.device)
|
92 |
-
gen_tokens = self.model.generate(
|
93 |
-
input_ids,
|
94 |
-
num_beams=num_beams, num_beam_groups=num_beam_groups,
|
95 |
-
do_sample=do_sample, temperature=temperature, top_p=top_p,
|
96 |
-
min_length=min_length, max_length=max_length,
|
97 |
-
pad_token_id=self.tokenizer.pad_token_id)
|
98 |
-
text = self.tokenizer.batch_decode(gen_tokens)[0]
|
99 |
-
|
100 |
-
# recovering
|
101 |
-
if self.low_resource_mode:
|
102 |
-
fp32 = torch.float32
|
103 |
-
self.steer.projector1.data = self.steer.projector1.to(fp32)
|
104 |
-
self.steer.projector2.data = self.steer.projector2.to(fp32)
|
105 |
-
return text
|
|
|
4 |
from .model_utils import Hack_no_grad
|
5 |
from .steers import Projected_Adaptor
|
6 |
from .model_base import LMSteerBase
|
|
|
7 |
|
8 |
|
9 |
class Switching_GPTNeoXModel(LMSteerBase):
|
|
|
41 |
else:
|
42 |
raise NotImplementedError()
|
43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
def generate(self, prompt, steer_values, min_length=20, max_length=100,
|
45 |
seed=None, num_beams=1, num_beam_groups=1, do_sample=True,
|
46 |
temperature=1, top_p=1):
|
|
|
51 |
max_length: maximum generation length
|
52 |
seed: seed for generation. None if not specified.
|
53 |
'''
|
54 |
+
return super().generate_low_resource(
|
55 |
+
prompt, steer_values, min_length, max_length, seed,
|
56 |
+
num_beams, num_beam_groups, do_sample, temperature, top_p)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|