Spaces:
Runtime error
Runtime error
Commit
·
c7e16d0
1
Parent(s):
b0721f8
fix: fixing attention visualization
Browse files- explanation/attention.py +22 -10
- explanation/interpret_captum.py +1 -3
- explanation/interpret_shap.py +1 -1
- main.py +13 -7
- model/godel.py +9 -2
- model/mistral.py +1 -0
explanation/attention.py
CHANGED
|
@@ -3,18 +3,22 @@
|
|
| 3 |
|
| 4 |
# internal imports
|
| 5 |
from utils import formatting as fmt
|
|
|
|
| 6 |
from .markup import markup_text
|
| 7 |
|
|
|
|
| 8 |
# chat function that returns an answer
|
| 9 |
# and marked text based on attention
|
| 10 |
def chat_explained(model, prompt):
|
| 11 |
|
|
|
|
|
|
|
| 12 |
# get encoded input
|
| 13 |
encoder_input_ids = model.TOKENIZER(
|
| 14 |
prompt, return_tensors="pt", add_special_tokens=True
|
| 15 |
).input_ids
|
| 16 |
# generate output together with attentions of the model
|
| 17 |
-
decoder_input_ids = model.MODEL
|
| 18 |
encoder_input_ids, output_attentions=True, generation_config=model.CONFIG
|
| 19 |
)
|
| 20 |
|
|
@@ -26,16 +30,24 @@ def chat_explained(model, prompt):
|
|
| 26 |
model.TOKENIZER.convert_ids_to_tokens(decoder_input_ids[0])
|
| 27 |
)
|
| 28 |
|
| 29 |
-
#
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
-
#
|
| 38 |
-
|
|
|
|
| 39 |
|
| 40 |
# format response text for clean output
|
| 41 |
response_text = fmt.format_output_text(decoder_text)
|
|
|
|
| 3 |
|
| 4 |
# internal imports
|
| 5 |
from utils import formatting as fmt
|
| 6 |
+
from model import godel
|
| 7 |
from .markup import markup_text
|
| 8 |
|
| 9 |
+
|
| 10 |
# chat function that returns an answer
|
| 11 |
# and marked text based on attention
|
| 12 |
def chat_explained(model, prompt):
|
| 13 |
|
| 14 |
+
model.set_config({"return_dict": True})
|
| 15 |
+
|
| 16 |
# get encoded input
|
| 17 |
encoder_input_ids = model.TOKENIZER(
|
| 18 |
prompt, return_tensors="pt", add_special_tokens=True
|
| 19 |
).input_ids
|
| 20 |
# generate output together with attentions of the model
|
| 21 |
+
decoder_input_ids = model.MODEL(
|
| 22 |
encoder_input_ids, output_attentions=True, generation_config=model.CONFIG
|
| 23 |
)
|
| 24 |
|
|
|
|
| 30 |
model.TOKENIZER.convert_ids_to_tokens(decoder_input_ids[0])
|
| 31 |
)
|
| 32 |
|
| 33 |
+
# getting attention if model is godel
|
| 34 |
+
if isinstance(model, godel):
|
| 35 |
+
print("attention.py: Model detected to be GODEL")
|
| 36 |
+
|
| 37 |
+
# get attention values for the input and output vectors
|
| 38 |
+
# using already generated input and output
|
| 39 |
+
attention_output = model.MODEL(
|
| 40 |
+
input_ids=encoder_input_ids,
|
| 41 |
+
decoder_input_ids=decoder_input_ids,
|
| 42 |
+
output_attentions=True,
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
# averaging attention across layers
|
| 46 |
+
averaged_attention = fmt.avg_attention(attention_output)
|
| 47 |
|
| 48 |
+
# getting attention is model is mistral
|
| 49 |
+
else:
|
| 50 |
+
averaged_attention = fmt.avg_attention(decoder_input_ids)
|
| 51 |
|
| 52 |
# format response text for clean output
|
| 53 |
response_text = fmt.format_output_text(decoder_text)
|
explanation/interpret_captum.py
CHANGED
|
@@ -45,11 +45,9 @@ def chat_explained(model, prompt):
|
|
| 45 |
|
| 46 |
# getting response text, graphic placeholder and marked text object
|
| 47 |
response_text = fmt.format_output_text(attribution_result.output_tokens)
|
| 48 |
-
graphic =
|
| 49 |
-
"""<div style='text-align: center; font-family:arial;'><h4>
|
| 50 |
Intepretation with Captum doesn't support an interactive graphic.</h4></div>
|
| 51 |
"""
|
| 52 |
-
)
|
| 53 |
marked_text = markup_text(input_tokens, values, variant="captum")
|
| 54 |
|
| 55 |
# return response, graphic and marked_text array
|
|
|
|
| 45 |
|
| 46 |
# getting response text, graphic placeholder and marked text object
|
| 47 |
response_text = fmt.format_output_text(attribution_result.output_tokens)
|
| 48 |
+
graphic = """<div style='text-align: center; font-family:arial;'><h4>
|
|
|
|
| 49 |
Intepretation with Captum doesn't support an interactive graphic.</h4></div>
|
| 50 |
"""
|
|
|
|
| 51 |
marked_text = markup_text(input_tokens, values, variant="captum")
|
| 52 |
|
| 53 |
# return response, graphic and marked_text array
|
explanation/interpret_shap.py
CHANGED
|
@@ -32,7 +32,7 @@ def wrap_shap(model):
|
|
| 32 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 33 |
|
| 34 |
# updating the model settings
|
| 35 |
-
model.set_config()
|
| 36 |
|
| 37 |
# (re)initialize the shap models and masker
|
| 38 |
# creating a shap text_generation model
|
|
|
|
| 32 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 33 |
|
| 34 |
# updating the model settings
|
| 35 |
+
model.set_config({})
|
| 36 |
|
| 37 |
# (re)initialize the shap models and masker
|
| 38 |
# creating a shap text_generation model
|
main.py
CHANGED
|
@@ -110,11 +110,10 @@ with gr.Blocks(
|
|
| 110 |
label="System Prompt",
|
| 111 |
info="Set the models system prompt, dictating how it answers.",
|
| 112 |
# default system prompt is set to this in the backend
|
| 113 |
-
placeholder=
|
| 114 |
You are a helpful, respectful and honest assistant. Always
|
| 115 |
answer as helpfully as possible, while being safe.
|
| 116 |
-
"""
|
| 117 |
-
),
|
| 118 |
)
|
| 119 |
# column that takes up 1/4 of the row
|
| 120 |
with gr.Column(scale=1):
|
|
@@ -122,7 +121,9 @@ with gr.Blocks(
|
|
| 122 |
xai_selection = gr.Radio(
|
| 123 |
["None", "SHAP", "Attention"],
|
| 124 |
label="Interpretability Settings",
|
| 125 |
-
info=
|
|
|
|
|
|
|
| 126 |
value="None",
|
| 127 |
interactive=True,
|
| 128 |
show_label=True,
|
|
@@ -209,10 +210,15 @@ with gr.Blocks(
|
|
| 209 |
gr.Examples(
|
| 210 |
label="Example Questions",
|
| 211 |
examples=[
|
| 212 |
-
["Does money buy happiness?", "Mistral", "SHAP"],
|
| 213 |
-
["Does money buy happiness?", "Mistral", "Attention"],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 214 |
],
|
| 215 |
-
inputs=[user_prompt, model_selection, xai_selection],
|
| 216 |
)
|
| 217 |
with gr.Accordion("GODEL Model Examples", open=False):
|
| 218 |
# examples util component
|
|
|
|
| 110 |
label="System Prompt",
|
| 111 |
info="Set the models system prompt, dictating how it answers.",
|
| 112 |
# default system prompt is set to this in the backend
|
| 113 |
+
placeholder="""
|
| 114 |
You are a helpful, respectful and honest assistant. Always
|
| 115 |
answer as helpfully as possible, while being safe.
|
| 116 |
+
""",
|
|
|
|
| 117 |
)
|
| 118 |
# column that takes up 1/4 of the row
|
| 119 |
with gr.Column(scale=1):
|
|
|
|
| 121 |
xai_selection = gr.Radio(
|
| 122 |
["None", "SHAP", "Attention"],
|
| 123 |
label="Interpretability Settings",
|
| 124 |
+
info=(
|
| 125 |
+
"Select a Interpretability Approach Implementation to use."
|
| 126 |
+
),
|
| 127 |
value="None",
|
| 128 |
interactive=True,
|
| 129 |
show_label=True,
|
|
|
|
| 210 |
gr.Examples(
|
| 211 |
label="Example Questions",
|
| 212 |
examples=[
|
| 213 |
+
["Does money buy happiness?", "", "Mistral", "SHAP"],
|
| 214 |
+
["Does money buy happiness?", "", "Mistral", "Attention"],
|
| 215 |
+
],
|
| 216 |
+
inputs=[
|
| 217 |
+
user_prompt,
|
| 218 |
+
knowledge_input,
|
| 219 |
+
model_selection,
|
| 220 |
+
xai_selection,
|
| 221 |
],
|
|
|
|
| 222 |
)
|
| 223 |
with gr.Accordion("GODEL Model Examples", open=False):
|
| 224 |
# examples util component
|
model/godel.py
CHANGED
|
@@ -13,7 +13,12 @@ MODEL = AutoModelForSeq2SeqLM.from_pretrained("microsoft/GODEL-v1_1-large-seq2se
|
|
| 13 |
|
| 14 |
# model config definition
|
| 15 |
CONFIG = GenerationConfig.from_pretrained("microsoft/GODEL-v1_1-large-seq2seq")
|
| 16 |
-
base_config_dict = {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
CONFIG.update(**base_config_dict)
|
| 18 |
|
| 19 |
|
|
@@ -59,11 +64,13 @@ def format_prompt(message: str, history: list, system_prompt: str, knowledge: st
|
|
| 59 |
# CREDIT: Copied from official interference example on Huggingface
|
| 60 |
## see https://huggingface.co/microsoft/GODEL-v1_1-large-seq2seq
|
| 61 |
def respond(prompt):
|
|
|
|
|
|
|
| 62 |
# tokenizing input string
|
| 63 |
input_ids = TOKENIZER(f"{prompt}", return_tensors="pt").input_ids
|
| 64 |
|
| 65 |
# generating using config and decoding output
|
| 66 |
-
outputs = MODEL.generate(input_ids,generation_config=CONFIG)
|
| 67 |
output = TOKENIZER.decode(outputs[0], skip_special_tokens=True)
|
| 68 |
|
| 69 |
# returns the model output string
|
|
|
|
| 13 |
|
| 14 |
# model config definition
|
| 15 |
CONFIG = GenerationConfig.from_pretrained("microsoft/GODEL-v1_1-large-seq2seq")
|
| 16 |
+
base_config_dict = {
|
| 17 |
+
"max_new_tokens": 50,
|
| 18 |
+
"min_length": 8,
|
| 19 |
+
"top_p": 0.9,
|
| 20 |
+
"do_sample": True,
|
| 21 |
+
}
|
| 22 |
CONFIG.update(**base_config_dict)
|
| 23 |
|
| 24 |
|
|
|
|
| 64 |
# CREDIT: Copied from official interference example on Huggingface
|
| 65 |
## see https://huggingface.co/microsoft/GODEL-v1_1-large-seq2seq
|
| 66 |
def respond(prompt):
|
| 67 |
+
set_config({})
|
| 68 |
+
|
| 69 |
# tokenizing input string
|
| 70 |
input_ids = TOKENIZER(f"{prompt}", return_tensors="pt").input_ids
|
| 71 |
|
| 72 |
# generating using config and decoding output
|
| 73 |
+
outputs = MODEL.generate(input_ids, generation_config=CONFIG)
|
| 74 |
output = TOKENIZER.decode(outputs[0], skip_special_tokens=True)
|
| 75 |
|
| 76 |
# returns the model output string
|
model/mistral.py
CHANGED
|
@@ -110,6 +110,7 @@ def format_answer(answer: str):
|
|
| 110 |
|
| 111 |
|
| 112 |
def respond(prompt: str):
|
|
|
|
| 113 |
|
| 114 |
# tokenizing inputs and configuring model
|
| 115 |
input_ids = TOKENIZER(f"{prompt}", return_tensors="pt")["input_ids"].to(device)
|
|
|
|
| 110 |
|
| 111 |
|
| 112 |
def respond(prompt: str):
|
| 113 |
+
set_config({})
|
| 114 |
|
| 115 |
# tokenizing inputs and configuring model
|
| 116 |
input_ids = TOKENIZER(f"{prompt}", return_tensors="pt")["input_ids"].to(device)
|