File size: 3,326 Bytes
60b21d3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 | #!/usr/bin/env python3
# SPDX-FileCopyrightText: 2025 Stanford University, ETH Zurich, and the project authors (see CONTRIBUTORS.md)
# SPDX-FileCopyrightText: 2025 This source file is part of the OpenTSLM open-source project.
#
# SPDX-License-Identifier: MIT
"""
Alternative visualization for memory usage.
Shows one subplot per base model (Llama, Gemma, etc).
Within each subplot: bars for SoftPrompt vs Flamingo across datasets.
Uses clean, colorblind-friendly colors.
"""
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
def parse_model_name(llm_id, model_type):
if llm_id.startswith("meta-llama/"):
base_name = llm_id.replace("meta-llama/", "")
elif llm_id.startswith("google/"):
base_name = llm_id.replace("google/", "")
else:
base_name = llm_id
if model_type == "OpenTSLMSP":
type_name = "SoftPrompt"
elif model_type == "OpenTSLMFlamingo":
type_name = "Flamingo"
else:
type_name = model_type
return base_name, type_name
def plot_memory_usage(csv_file="memory_use.csv"):
df = pd.read_csv(csv_file)
# Replace -1 with 50GB
df["peak_cuda_reserved_gb"] = df["peak_cuda_reserved_gb"].replace(-1, 50.0)
# Parse into base + config
df[["base_model", "config"]] = df.apply(
lambda row: pd.Series(parse_model_name(row["llm_id"], row["model"])), axis=1
)
# Order datasets
dataset_order = ["TSQA", "HAR-CoT", "SleepEDF-CoT", "ECG-QA-CoT"]
df["dataset"] = pd.Categorical(
df["dataset"], categories=dataset_order, ordered=True
)
# Order configs
config_order = ["SoftPrompt", "Flamingo"]
df["config"] = pd.Categorical(df["config"], categories=config_order, ordered=True)
# Get unique base models
base_models = df["base_model"].unique()
# Plot
n_models = len(base_models)
fig, axes = plt.subplots(1, n_models, figsize=(6 * n_models, 6), sharey=True)
if n_models == 1:
axes = [axes]
# Colorblind-friendly palette
palette = {"SoftPrompt": "#4C78A8", "Flamingo": "#F58518"}
for ax, base_model in zip(axes, base_models):
subdf = df[df["base_model"] == base_model]
sns.barplot(
data=subdf,
x="dataset",
y="peak_cuda_reserved_gb",
hue="config",
ax=ax,
palette=palette,
)
ax.set_title(base_model, fontsize=14, fontweight="bold")
ax.set_xlabel("Dataset")
ax.set_ylabel("Peak CUDA Reserved GB")
ax.tick_params(axis="x", rotation=30)
ax.grid(axis="y", alpha=0.3, linestyle="--")
ax.set_axisbelow(True)
# Add value labels
for container in ax.containers:
ax.bar_label(container, fmt="%.1f", padding=2, fontsize=8)
# Shared legend
handles, labels = axes[0].get_legend_handles_labels()
fig.legend(
handles,
labels,
loc="upper center",
ncol=2,
title="Config",
fontsize=10,
)
fig.suptitle("Peak CUDA Reserved Memory by Model, Dataset, and Config", fontsize=16)
plt.tight_layout(rect=[0, 0, 1, 0.92])
plt.savefig(
"memory_usage_facet.png", dpi=300, bbox_inches="tight", facecolor="white"
)
plt.show()
if __name__ == "__main__":
plot_memory_usage()
|