commit-message-editing-visualization / change_visualizer.py
Petr Tsvetkov
release
9513395
import gradio as gr
import numpy as np
import pandas as pd
import plotly.graph_objects as go
from datasets import load_dataset
from evaluate.utils import parse_readme
from scipy.stats import gaussian_kde, spearmanr
import generate_annotated_diffs
from api_wrappers import hf_data_loader
from generation_steps.metrics_analysis import AGGR_METRICS, edit_distance_fn
colors = {
"Expert-labeled": "#C19C0B",
"Synthetic Backward": "#913632",
"Synthetic Forward": "#58136a",
"Full": "#000000",
}
METRICS = {
"Edit Distance": "editdist",
"Edit Similarity": "editsim",
"BLEU": "bleu",
"METEOR": "meteor",
"ROUGE-1": "rouge1",
"ROUGE-2": "rouge2",
"ROUGE-L": "rougeL",
"BERTScore": "bertscore",
"ChrF": "chrF",
}
df_related = generate_annotated_diffs.data_with_annotated_diffs()
def golden():
return df_related.loc[(df_related["G_type"] == "initial") & (df_related["E_type"] == "expert_labeled")].reset_index(
drop=True
)
def backward():
return df_related.loc[
(df_related["G_type"] == "synthetic_backward") & (df_related["E_type"] == "expert_labeled")
].reset_index(drop=True)
def forward():
return df_related.loc[
(df_related["G_type"] == "initial") & (df_related["E_type"] == "synthetic_forward")
].reset_index(drop=True)
def forward_from_backward():
return df_related.loc[
(df_related.G_type == "synthetic_backward")
& (df_related.E_type.isin(["synthetic_forward", "synthetic_forward_from_backward"]))
].reset_index(drop=True)
n_diffs_manual = len(golden())
n_diffs_synthetic_backward = len(backward())
n_diffs_synthetic_forward = len(forward())
n_diffs_synthetic_forward_backward = len(forward_from_backward())
def update_dataset_view(diff_idx, df):
diff_idx -= 1
return (
df.iloc[diff_idx]["annotated_diff"],
df.iloc[diff_idx]["commit_msg_start"] if "commit_msg_start" in df.columns else df.iloc[diff_idx]["G_text"],
df.iloc[diff_idx]["commit_msg_end"] if "commit_msg_end" in df.columns else df.iloc[diff_idx]["E_text"],
f"https://github.com/{df.iloc[diff_idx]['repo']}/commit/{df.iloc[diff_idx]['hash']}",
)
def update_dataset_view_manual(diff_idx):
return update_dataset_view(diff_idx, golden())
def update_dataset_view_synthetic_backward(diff_idx):
return update_dataset_view(diff_idx, backward())
def update_dataset_view_synthetic_forward(diff_idx):
return update_dataset_view(diff_idx, forward())
def update_dataset_view_synthetic_forward_backward(diff_idx):
return update_dataset_view(diff_idx, forward_from_backward())
def number_of_pairs_plot():
related_plot_dict = {
"Full": df_related,
"Synthetic Backward": backward(),
"Synthetic Forward": pd.concat([forward(), forward_from_backward()], axis=0, ignore_index=True),
"Expert-labeled": golden(),
}
df_unrelated = hf_data_loader.load_synthetic_as_pandas()
df_unrelated = df_unrelated.loc[~df_unrelated.is_related].copy()
unrelated_plot_dict = {
"Full": df_unrelated,
"Synthetic Backward": df_unrelated.loc[
(df_unrelated["G_type"] == "synthetic_backward")
& (~df_unrelated.E_type.isin(["synthetic_forward", "synthetic_forward_from_backward"]))
],
"Synthetic Forward": df_unrelated.loc[
((df_unrelated["G_type"] == "initial") & (df_unrelated["E_type"] == "synthetic_forward"))
| (
(df_unrelated["G_type"] == "synthetic_backward")
& (df_unrelated["E_type"].isin(["synthetic_forward", "synthetic_forward_from_backward"]))
)
],
"Expert-labeled": df_unrelated.loc[
(df_unrelated.G_type == "initial") & (df_unrelated.E_type == "expert_labeled")
],
}
traces = []
for split in related_plot_dict.keys():
related_count = len(related_plot_dict[split])
unrelated_count = len(unrelated_plot_dict[split])
traces.append(
go.Bar(
name=f"{split} - Related pairs",
x=[split],
y=[related_count],
marker=dict(
color=colors[split],
),
)
)
traces.append(
go.Bar(
name=f"{split} - Conditionally independent pairs",
x=[split],
y=[unrelated_count],
marker=dict(
color=colors[split],
pattern=dict(
shape="/", # Crosses
fillmode="overlay",
solidity=0.5,
),
),
)
)
fig = go.Figure(data=traces)
fig.update_layout(
barmode="stack",
bargap=0.2,
xaxis=dict(title="Split", showgrid=True, gridcolor="lightgrey"),
yaxis=dict(title="Number of Examples", showgrid=True, gridcolor="lightgrey"),
legend=dict(title="Pair Type", orientation="h", yanchor="bottom", y=1.02, xanchor="right", x=1),
plot_bgcolor="rgba(0,0,0,0)",
paper_bgcolor="rgba(0,0,0,0)",
width=1100,
)
return fig
def edit_distance_plot():
df_edit_distance = {
"Full": [edit_distance_fn(pred=row["G_text"], ref=row["E_text"]) for _, row in df_related.iterrows()],
"Synthetic Backward": [
edit_distance_fn(pred=row["G_text"], ref=row["E_text"]) for _, row in backward().iterrows()
],
"Synthetic Forward": [
edit_distance_fn(pred=row["G_text"], ref=row["E_text"])
for _, row in pd.concat([forward(), forward_from_backward()], axis=0, ignore_index=True).iterrows()
],
"Expert-labeled": [edit_distance_fn(pred=row["G_text"], ref=row["E_text"]) for _, row in golden().iterrows()],
}
traces = []
for key in df_edit_distance:
kde_x = np.linspace(0, 1200, 1000)
kde = gaussian_kde(df_edit_distance[key])
kde_line = go.Scatter(x=kde_x, y=kde(kde_x), mode="lines", name=key, line=dict(color=colors[key], width=5))
traces.append(kde_line)
fig = go.Figure(data=traces)
fig.update_layout(
bargap=0.1,
xaxis=dict(title=dict(text="Edit Distance"), range=[0, 1200], showgrid=True, gridcolor="lightgrey"),
yaxis=dict(
title=dict(text="Probability Density"),
range=[0, 0.004],
showgrid=True,
gridcolor="lightgrey",
tickvals=[0.0005, 0.001, 0.0015, 0.002, 0.0025, 0.003, 0.0035, 0.004],
tickformat=".4f",
),
plot_bgcolor="rgba(0,0,0,0)",
paper_bgcolor="rgba(0,0,0,0)",
width=1100,
)
return fig
def get_correlations_table(online_metric_name: str) -> pd.DataFrame:
df = load_dataset(
"JetBrains-Research/synthetic-commit-msg-edits", "all_pairs_with_metrics_other_online_metrics", split="train"
).to_pandas()
corr_df = (
df.loc[~df.is_related]
.groupby(["G_text", "G_type", "hash", "repo"] + [f"online_{online_metric_name}"])
.apply(lambda g: g.to_dict(orient="records"), include_groups=False)
.reset_index(name="unrelated_pairs")
.copy()
)
_ = corr_df.copy()
for metric in AGGR_METRICS:
if metric in ["editdist"]:
_[metric] = _.unrelated_pairs.apply(lambda pairs: min(pair[metric] for pair in pairs))
else:
_[metric] = _.unrelated_pairs.apply(lambda pairs: max(pair[metric] for pair in pairs))
results = []
for metric in AGGR_METRICS:
x = _[metric].to_numpy()
y = _[f"online_{online_metric_name}"].to_numpy()
corr, p_value = spearmanr(x, y)
results.append({"metric": metric, "corr": corr, "p_value": p_value})
__ = pd.DataFrame(results)
__["p_value"] = ["< 0.05" if p < 0.05 else p for p in __.p_value]
__["corr_abs"] = abs(__["corr"])
__["corr"] = __["corr"].round(2)
__["metric"] = __["metric"].map({v: k for k, v in METRICS.items()})
return (
__.sort_values(by=["corr_abs"], ascending=False)
.drop(columns=["corr_abs"])
.rename(columns={"metric": "Metric m", "corr": "Correlation Q(m, m*)", "p_value": "p-value"})
)
force_light_theme_js_func = """
function refresh() {
const url = new URL(window.location);
if (url.searchParams.get('__theme') !== 'light') {
url.searchParams.set('__theme', 'light');
window.location.href = url.href;
}
}
"""
if __name__ == "__main__":
with gr.Blocks(theme=gr.themes.Soft(), js=force_light_theme_js_func) as application:
gr.Markdown(parse_readme("README.md"))
def dataset_view_tab(n_items):
slider = gr.Slider(minimum=1, maximum=n_items, step=1, value=1, label=f"Sample number (total: {n_items})")
diff_view = gr.Highlightedtext(combine_adjacent=True, color_map={"+": "green", "-": "red"})
start_view = gr.Textbox(interactive=False, label="Initial message G", container=True)
end_view = gr.Textbox(interactive=False, label="Edited message E", container=True)
link_view = gr.Markdown()
view = [diff_view, start_view, end_view, link_view]
return slider, view
with gr.Tab("Examples Exploration"):
with gr.Tab("Manual"):
slider_manual, view_manual = dataset_view_tab(n_diffs_manual)
slider_manual.change(update_dataset_view_manual, inputs=slider_manual, outputs=view_manual)
with gr.Tab("Synthetic Backward"):
slider_synthetic_backward, view_synthetic_backward = dataset_view_tab(n_diffs_synthetic_backward)
slider_synthetic_backward.change(
update_dataset_view_synthetic_backward,
inputs=slider_synthetic_backward,
outputs=view_synthetic_backward,
)
with gr.Tab("Synthetic Forward (from initial)"):
slider_synthetic_forward, view_synthetic_forward = dataset_view_tab(n_diffs_synthetic_forward)
slider_synthetic_forward.change(
update_dataset_view_synthetic_forward,
inputs=slider_synthetic_forward,
outputs=view_synthetic_forward,
)
with gr.Tab("Synthetic Forward (from backward)"):
slider_synthetic_forward_backward, view_synthetic_forward_backward = dataset_view_tab(
n_diffs_synthetic_forward_backward
)
slider_synthetic_forward_backward.change(
update_dataset_view_synthetic_forward_backward,
inputs=slider_synthetic_forward_backward,
outputs=view_synthetic_forward_backward,
)
with gr.Tab("Dataset Statistics"):
gr.Markdown("## Number of examples per split")
number_of_pairs_gr_plot = gr.Plot(number_of_pairs_plot, label=None)
gr.Markdown("## Edit Distance Distribution (w/o PyCharm Logs)")
edit_distance_gr_plot = gr.Plot(edit_distance_plot(), label=None)
with gr.Tab("Experimental Results"):
gr.Markdown(
"Here, we provide the additional experimental results with different text similarity metrics used as the target online metric, "
"in addition to edit distance between generated messages G and their edited counterparts E."
)
gr.Markdown(
"Please, select one of the available metrics **m*** below to see the correlations **Q(m, m\*)** of offline text similarity metrics with **m*** as an online metric."
)
for metric in METRICS:
with gr.Tab(metric):
gr.Markdown(
f"The table below presents the correlation coefficients **Q(m, m\*)** where {metric} is used as an online metric **m***."
)
result_df = get_correlations_table(METRICS[metric])
gr.DataFrame(result_df)
application.load(update_dataset_view_manual, inputs=slider_manual, outputs=view_manual)
application.load(
update_dataset_view_synthetic_backward, inputs=slider_synthetic_backward, outputs=view_synthetic_backward
)
application.load(
update_dataset_view_synthetic_forward, inputs=slider_synthetic_forward, outputs=view_synthetic_forward
)
application.load(
update_dataset_view_synthetic_forward_backward,
inputs=slider_synthetic_forward_backward,
outputs=view_synthetic_forward_backward,
)
application.launch()