Meehai's picture
added m2f metris analysis and moved data analysis nb in a dir
6d72a48
import sys
from vre.readers import MultiTaskDataset
from vre.representations import build_representations_from_cfg, Representation
from vre.representations.cv_representations import SemanticRepresentation
from vre.logger import vre_logger as logger
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import io
import base64
import bs4
from PIL import Image
import seaborn as sns
def extract_pil_from_b64_image(base64_buf: str) -> Image:
return Image.open(io.BytesIO(base64.b64decode(base64_buf)))
def extract_b64_image_from_fig(fig: plt.Figure) -> str:
buffer = io.BytesIO()
fig.savefig(buffer, format="png", dpi=fig.dpi)
buffer.seek(0)
base64_buf = base64.b64encode(buffer.getvalue()).decode("utf-8")
return base64_buf
def extract_b64_imgsrc_from_fig(fig: plt.Figure) -> str:
base64_buf = extract_b64_image_from_fig(fig)
return f"""<img src="data:image/png;base64,{base64_buf}" alt="Sample Plot">"""
def save_html(html_imgs: list[str], description: str, out_path: str):
html = bs4.BeautifulSoup(f"""
<!DOCTYPE html>
<html>
<head>
<title>VRE Dataset Analysis</title>
</head>
<body>
<h1 id="description">Description</h1>
<h1 id="plots">Plots</h1>
</body>
</html>""", features="lxml")
html.find(id="description").insert_after(bs4.BeautifulSoup(description.replace("\n", "<br/>"), features="lxml"))
for html_img in html_imgs[::-1]:
html.find(id="plots").insert_after(bs4.BeautifulSoup(html_img, features="lxml"))
open(out_path, "w").write(str(html))
print(f"Written html at '{out_path}'")
def histogram_from_classification_task(reader: MultiTaskDataset, classif: SemanticRepresentation,
n: int | None = None, mode: str = "sequential", **figkwargs) -> plt.Figure:
fig = plt.Figure(**figkwargs)
counts = np.zeros(len(classif.classes), dtype=np.uint64)
ixs = np.arange(len(reader)) if mode == "sequential" else np.random.permutation(len(reader))
ixs = ixs[0:n] if n is not None and n < len(reader) else ixs
assert getattr(classif, "load_mode", "binary") == "binary", classif.load_mode
for i in ixs:
item = reader.get_one_item(i.item(), subset_tasks=[classif.name])
data_cnts = item[0][classif.name].unique(return_counts=True)
item_classes, item_counts = data_cnts[0].numpy().astype(int), data_cnts[1].numpy().astype(int)
counts[item_classes] = counts[item_classes] + item_counts
df = pd.DataFrame({"Labels": classif.classes, "Values": counts})
df["Values"] = df["Values"] / df["Values"].sum()
df = df.sort_values("Values", ascending=True)
df = df[df["Values"] > 0.005]
ax = fig.gca()
sns.barplot(data=df, y="Labels", x="Values", palette="viridis", legend=True, ax=ax, width=1)
# Adjust y-axis tick positions and spacing
ax.set_title(classif.name, fontsize=14, fontweight='bold')
ax.set_ylabel("Labels", fontsize=12)
fig.set_size_inches(8, 2 if len(df) <= 2 else len(df) * 0.5)
fig.gca().set_xlim(0, 1)
fig.tight_layout()
plt.close()
return fig
def gaussian_from_statistics(reader: MultiTaskDataset, regression_task: Representation) -> plt.Figure:
_, __, mean, std = [x.numpy() for x in reader.statistics[regression_task.name]]
fig, ax = plt.subplots(1, n_ch := mean.shape[0], figsize=(10, 5))
ax = [ax] if n_ch == 1 else ax
x = np.linspace(mean - 4*std, mean + 4*std, 1000)
y = (1 / (std * np.sqrt(2 * np.pi))) * np.exp(-0.5 * ((x - mean) / std) ** 2)
for i in range(n_ch):
ax[i].plot(x[:, i], y[:, i])
fig.suptitle(regression_task.name)
return fig
if __name__ == "__main__":
data_path = sys.argv[1]
cfg_path = sys.argv[2]
representations = build_representations_from_cfg(cfg_path)
print(representations)
reader = MultiTaskDataset(data_path, task_names=list(representations),
task_types=representations, normalization="min_max")
print(reader)
imgsrcs = []
for classif_task in reader.classification_tasks:
fig = histogram_from_classification_task(reader, classif_task)
imgsrcs.append(extract_b64_imgsrc_from_fig(fig))
regression_tasks = [t for t in reader.tasks if t not in reader.classification_tasks]
for regression_task in regression_tasks:
fig = gaussian_from_statistics(reader, regression_task)
imgsrcs.append(extract_b64_imgsrc_from_fig(fig))
save_html(imgsrcs, str(reader), "plot.html")