leandro
commited on
Commit
·
16b19cc
1
Parent(s):
ad022dd
add app and requirements
Browse files- app.py +162 -0
- requirements.txt +4 -0
app.py
ADDED
@@ -0,0 +1,162 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
3 |
+
from transformers import pipeline
|
4 |
+
import numpy as np
|
5 |
+
import pandas as pd
|
6 |
+
import matplotlib.cm as cm
|
7 |
+
import html
|
8 |
+
from torch.nn.functional import softmax
|
9 |
+
import torch
|
10 |
+
from matplotlib.colors import LinearSegmentedColormap
|
11 |
+
|
12 |
+
cdict = {'red': [[0.0, 0.8, 0.8],
|
13 |
+
[1.0, 1.0, 1.0]],
|
14 |
+
'green': [[0.0, 0.0, 0.0],
|
15 |
+
[1.0, 1.0, 1.0]],
|
16 |
+
'blue': [[0.0, 0.0, 0.0],
|
17 |
+
[1.0, 1.0, 1.0]],
|
18 |
+
'alpha':[[0.0, 1.0, 1.0],
|
19 |
+
[1.0, 0.0, 0.0]]}
|
20 |
+
|
21 |
+
cmap = LinearSegmentedColormap('codemap', segmentdata=cdict, N=256)
|
22 |
+
|
23 |
+
def value2rgba(x, cmap=cmap, alpha_mult=1.0):
|
24 |
+
c = cmap(x)
|
25 |
+
rgb = (np.array(c[:-1]) * 255).astype(int)
|
26 |
+
a = c[-1] * alpha_mult
|
27 |
+
return tuple(rgb.tolist() + [a])
|
28 |
+
|
29 |
+
def highlight_token_scores(tokens, scores, sep=' ', **kwargs):
|
30 |
+
html_code,spans = [''], []#['<span style="font-family: monospace;">'], []
|
31 |
+
for t, s in zip(tokens, scores):
|
32 |
+
t = html.escape(t)
|
33 |
+
t = t.replace("\n", " \n")
|
34 |
+
c = str(value2rgba(s, alpha_mult=0.8, **kwargs))
|
35 |
+
spans.append(f'<span title="{s:.3f}" style="background-color: rgba{c};">{t}</span>')
|
36 |
+
html_code.append(sep.join(spans))
|
37 |
+
return '<pre><code>' + ''.join(html_code)
|
38 |
+
|
39 |
+
def color_dataframe(row):
|
40 |
+
styles = []
|
41 |
+
c = str(value2rgba(row["scores"], alpha_mult=0.8))
|
42 |
+
for key in row.index:
|
43 |
+
if key in {"tokens", "scores"}:
|
44 |
+
styles.append(f"background-color: rgba{c}")
|
45 |
+
else:
|
46 |
+
styles.append(f"background-color: None")
|
47 |
+
return styles
|
48 |
+
|
49 |
+
@st.cache(allow_output_mutation=True)
|
50 |
+
def load_tokenizer(model_ckpt):
|
51 |
+
return AutoTokenizer.from_pretrained(model_ckpt)
|
52 |
+
|
53 |
+
@st.cache(allow_output_mutation=True)
|
54 |
+
def load_model(model_ckpt):
|
55 |
+
model = AutoModelForCausalLM.from_pretrained(model_ckpt)
|
56 |
+
return model
|
57 |
+
|
58 |
+
def calculate_scores(probs, token_ids):
|
59 |
+
probs = probs[:-1]
|
60 |
+
token_ids = token_ids[1:]
|
61 |
+
sorted_ids = np.argsort(probs, axis=-1)[:, ::-1]
|
62 |
+
sorted_probs = np.sort(probs, axis=-1)[:, ::-1]
|
63 |
+
selected_token_mask = sorted_ids == token_ids[:, None]
|
64 |
+
masked_probs = np.ma.array(sorted_probs, mask=~selected_token_mask)
|
65 |
+
token_probs = masked_probs.sum(axis=1).data
|
66 |
+
|
67 |
+
masked_indices = np.cumsum(selected_token_mask[:, ::-1], axis=-1)[:, ::-1].astype(bool)
|
68 |
+
masked_probs = np.ma.array(sorted_probs, mask=~masked_indices)
|
69 |
+
token_rank = masked_indices.sum(axis=-1)
|
70 |
+
cumulative_probs = masked_probs.sum(axis=1).data/token_rank
|
71 |
+
scores = token_probs/cumulative_probs
|
72 |
+
return [1.] + list(scores), sorted_ids
|
73 |
+
|
74 |
+
def calculate_loss(logits, labels):
|
75 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
76 |
+
shift_labels = labels[..., 1:].contiguous()
|
77 |
+
loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
|
78 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
79 |
+
norm_loss = 1 - (loss/torch.max(loss))
|
80 |
+
return [1.] + list(norm_loss.numpy())
|
81 |
+
|
82 |
+
default_code = """\
|
83 |
+
from torch import nn
|
84 |
+
from transformers import Model
|
85 |
+
|
86 |
+
class Transformer:
|
87 |
+
def __init__(config):
|
88 |
+
self.model = Model(config)
|
89 |
+
|
90 |
+
def forward(inputs):
|
91 |
+
return self.model(inputs)"""
|
92 |
+
|
93 |
+
solution_code = """\
|
94 |
+
from torch import nn
|
95 |
+
from transformers import Model
|
96 |
+
|
97 |
+
class Transformer(nn.Module):
|
98 |
+
def __init__(self, config):
|
99 |
+
super(Transformer, self).__init__()
|
100 |
+
self.config = config
|
101 |
+
self.model = Model(config)
|
102 |
+
|
103 |
+
def forward(self, inputs):
|
104 |
+
return self.model(inputs)
|
105 |
+
"""
|
106 |
+
|
107 |
+
st.set_page_config(page_icon=':parrot:', layout="wide")
|
108 |
+
|
109 |
+
np.random.seed(42)
|
110 |
+
model_ckpt = "lvwerra/codeparrot"
|
111 |
+
tokenizer = load_tokenizer(model_ckpt)
|
112 |
+
model = load_model(model_ckpt)
|
113 |
+
st.markdown("<h1 style='text-align: center;'>CodeParrot 🦜</h1>", unsafe_allow_html=True)
|
114 |
+
st.markdown('##')
|
115 |
+
|
116 |
+
col1, col2 = st.columns(2)
|
117 |
+
|
118 |
+
col1.subheader("Edit code")
|
119 |
+
code = col1.text_area(label="", value=default_code, height=220,).strip()
|
120 |
+
inputs = tokenizer(code, return_tensors='pt')
|
121 |
+
token_list = [tokenizer.decode(t) for t in inputs["input_ids"][0]]
|
122 |
+
|
123 |
+
with torch.no_grad():
|
124 |
+
logits = model(input_ids=inputs["input_ids"]).logits[0]
|
125 |
+
probs = softmax(logits, dim=-1)
|
126 |
+
|
127 |
+
loss = calculate_loss(logits, inputs["input_ids"][0])
|
128 |
+
norm_probs, sorted_token_ids = calculate_scores(probs.numpy(), inputs["input_ids"][0].numpy())
|
129 |
+
|
130 |
+
if len(inputs['input_ids'])>1024:
|
131 |
+
st.warning("Your input is longer than the maximum 1024 tokens and will be truncated.")
|
132 |
+
|
133 |
+
st.sidebar.title("Settings:")
|
134 |
+
if st.sidebar.radio("Highlight mode:", ["Probability heuristics", "Scaled loss per token"]) == "Probability heuristics":
|
135 |
+
scores = norm_probs
|
136 |
+
else:
|
137 |
+
scores = loss
|
138 |
+
|
139 |
+
suggestion_threshold = st.sidebar.slider("Suggestion threshold", 0.0, 1.0, 0.2)
|
140 |
+
|
141 |
+
col2.subheader("Highlighted code")
|
142 |
+
col2.markdown('##')
|
143 |
+
html_string = highlight_token_scores(token_list, scores, sep="")
|
144 |
+
col2.markdown(html_string, unsafe_allow_html=True)
|
145 |
+
col2.markdown('##')
|
146 |
+
|
147 |
+
st.subheader("Model suggestions")
|
148 |
+
top_k = {}
|
149 |
+
for i in range(5):
|
150 |
+
top_k[f"top-{i+1}"] = ["No prediction for first token"] + [repr(tokenizer.decode(idx)) for idx in sorted_token_ids[:, i]]
|
151 |
+
df = pd.DataFrame({"tokens": [repr(t) for t in token_list], "scores": scores, **top_k})
|
152 |
+
df.index.name = "position"
|
153 |
+
df_filter = df.loc[df["scores"]<=suggestion_threshold]
|
154 |
+
df_filter.reset_index(inplace=True)
|
155 |
+
df_filter = df_filter[["tokens", "scores", "position", "top-1", "top-2", "top-3", "top-4", "top-5",]]
|
156 |
+
df_filter = df_filter.style.apply(color_dataframe, axis=1)
|
157 |
+
st.dataframe(df_filter)
|
158 |
+
|
159 |
+
st.markdown('##')
|
160 |
+
|
161 |
+
st.subheader("Possible solution")
|
162 |
+
st.code(solution_code)
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
transformers==4.12.2
|
2 |
+
pandas
|
3 |
+
matplotlib
|
4 |
+
torch
|