Spaces:
Running
Running
Implement KL divergence score
Browse files
app.py
CHANGED
|
@@ -3,6 +3,7 @@ from pathlib import Path
|
|
| 3 |
|
| 4 |
import streamlit as st
|
| 5 |
import streamlit.components.v1 as components
|
|
|
|
| 6 |
import torch
|
| 7 |
import torch.nn.functional as F
|
| 8 |
from transformers import AutoModelForCausalLM, AutoTokenizer, BatchEncoding
|
|
@@ -41,6 +42,28 @@ def ids_to_readable_tokens(tokenizer, ids, strip_whitespace=False):
|
|
| 41 |
result.append("")
|
| 42 |
return result
|
| 43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
compact_layout = st.experimental_get_query_params().get("compact", ["false"]) == ["true"]
|
| 45 |
|
| 46 |
if not compact_layout:
|
|
@@ -53,7 +76,7 @@ if not compact_layout:
|
|
| 53 |
)
|
| 54 |
|
| 55 |
model_name = st.selectbox("Model", ["distilgpt2", "gpt2", "EleutherAI/gpt-neo-125m"])
|
| 56 |
-
metric_name = st.selectbox("Metric", ["KL divergence", "
|
| 57 |
|
| 58 |
tokenizer = st.cache_resource(AutoTokenizer.from_pretrained, show_spinner=False)(model_name, use_fast=False)
|
| 59 |
|
|
@@ -107,10 +130,6 @@ if num_user_tokens > max_tokens:
|
|
| 107 |
)
|
| 108 |
st.stop()
|
| 109 |
|
| 110 |
-
if metric_name == "KL divergence":
|
| 111 |
-
st.error("KL divergence is not supported yet. Stay tuned!", icon="😭")
|
| 112 |
-
st.stop()
|
| 113 |
-
|
| 114 |
with st.spinner("Loading model…"):
|
| 115 |
model = st.cache_resource(AutoModelForCausalLM.from_pretrained, show_spinner=False)(model_name)
|
| 116 |
|
|
@@ -124,7 +143,7 @@ def get_logprobs(_model, _inputs, cache_key):
|
|
| 124 |
|
| 125 |
@st.cache_data(show_spinner=False)
|
| 126 |
@torch.inference_mode()
|
| 127 |
-
def run_context_length_probing(_model, _tokenizer, _inputs, window_len, cache_key):
|
| 128 |
del cache_key
|
| 129 |
|
| 130 |
inputs_sliding = get_windows_batched(
|
|
@@ -157,8 +176,13 @@ def run_context_length_probing(_model, _tokenizer, _inputs, window_len, cache_ke
|
|
| 157 |
logprobs = logprobs.view(-1, logprobs.shape[-1])[:-window_len]
|
| 158 |
logprobs = logprobs.view(window_len, len(input_ids) + window_len - 2, logprobs.shape[-1])
|
| 159 |
|
| 160 |
-
|
| 161 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 162 |
scores = scores.nan_to_num()
|
| 163 |
scores /= scores.abs().max(dim=1, keepdim=True).values + 1e-6
|
| 164 |
scores = scores.to(torch.float16)
|
|
@@ -170,6 +194,7 @@ scores = run_context_length_probing(
|
|
| 170 |
_tokenizer=tokenizer,
|
| 171 |
_inputs=inputs,
|
| 172 |
window_len=window_len,
|
|
|
|
| 173 |
cache_key=(model_name, text),
|
| 174 |
)
|
| 175 |
tokens = ids_to_readable_tokens(tokenizer, input_ids)
|
|
|
|
| 3 |
|
| 4 |
import streamlit as st
|
| 5 |
import streamlit.components.v1 as components
|
| 6 |
+
import numpy as np
|
| 7 |
import torch
|
| 8 |
import torch.nn.functional as F
|
| 9 |
from transformers import AutoModelForCausalLM, AutoTokenizer, BatchEncoding
|
|
|
|
| 42 |
result.append("")
|
| 43 |
return result
|
| 44 |
|
| 45 |
+
def nll_score(logprobs, labels):
|
| 46 |
+
return -logprobs[:, torch.arange(len(labels)), labels]
|
| 47 |
+
|
| 48 |
+
def kl_div_score(logprobs):
|
| 49 |
+
log_p = logprobs[
|
| 50 |
+
torch.arange(logprobs.shape[1]).clamp(max=logprobs.shape[0] - 1),
|
| 51 |
+
torch.arange(logprobs.shape[1])
|
| 52 |
+
]
|
| 53 |
+
# Compute things in place as much as possible
|
| 54 |
+
log_p_minus_log_q = logprobs
|
| 55 |
+
del logprobs
|
| 56 |
+
log_p_minus_log_q *= -1
|
| 57 |
+
log_p_minus_log_q += log_p
|
| 58 |
+
|
| 59 |
+
# Use np.exp because torch.exp is not implemented for float16
|
| 60 |
+
p_np = log_p.numpy()
|
| 61 |
+
del log_p
|
| 62 |
+
np.exp(p_np, out=p_np)
|
| 63 |
+
result = log_p_minus_log_q
|
| 64 |
+
result *= torch.as_tensor(p_np)
|
| 65 |
+
return result.sum(dim=-1)
|
| 66 |
+
|
| 67 |
compact_layout = st.experimental_get_query_params().get("compact", ["false"]) == ["true"]
|
| 68 |
|
| 69 |
if not compact_layout:
|
|
|
|
| 76 |
)
|
| 77 |
|
| 78 |
model_name = st.selectbox("Model", ["distilgpt2", "gpt2", "EleutherAI/gpt-neo-125m"])
|
| 79 |
+
metric_name = st.selectbox("Metric", ["KL divergence", "NLL loss"], index=1)
|
| 80 |
|
| 81 |
tokenizer = st.cache_resource(AutoTokenizer.from_pretrained, show_spinner=False)(model_name, use_fast=False)
|
| 82 |
|
|
|
|
| 130 |
)
|
| 131 |
st.stop()
|
| 132 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 133 |
with st.spinner("Loading model…"):
|
| 134 |
model = st.cache_resource(AutoModelForCausalLM.from_pretrained, show_spinner=False)(model_name)
|
| 135 |
|
|
|
|
| 143 |
|
| 144 |
@st.cache_data(show_spinner=False)
|
| 145 |
@torch.inference_mode()
|
| 146 |
+
def run_context_length_probing(_model, _tokenizer, _inputs, window_len, metric, cache_key):
|
| 147 |
del cache_key
|
| 148 |
|
| 149 |
inputs_sliding = get_windows_batched(
|
|
|
|
| 176 |
logprobs = logprobs.view(-1, logprobs.shape[-1])[:-window_len]
|
| 177 |
logprobs = logprobs.view(window_len, len(input_ids) + window_len - 2, logprobs.shape[-1])
|
| 178 |
|
| 179 |
+
if metric == "NLL loss":
|
| 180 |
+
scores = nll_score(logprobs=logprobs, labels=input_ids[1:])
|
| 181 |
+
elif metric == "KL divergence":
|
| 182 |
+
scores = kl_div_score(logprobs)
|
| 183 |
+
del logprobs # possibly destroyed by the score computation to save memory
|
| 184 |
+
|
| 185 |
+
scores = (-scores).diff(dim=0).transpose(0, 1)
|
| 186 |
scores = scores.nan_to_num()
|
| 187 |
scores /= scores.abs().max(dim=1, keepdim=True).values + 1e-6
|
| 188 |
scores = scores.to(torch.float16)
|
|
|
|
| 194 |
_tokenizer=tokenizer,
|
| 195 |
_inputs=inputs,
|
| 196 |
window_len=window_len,
|
| 197 |
+
metric=metric_name,
|
| 198 |
cache_key=(model_name, text),
|
| 199 |
)
|
| 200 |
tokens = ids_to_readable_tokens(tokenizer, input_ids)
|