Visualization

Training automatically generates visualizations. Additional visualizations can be created separately.

---

Visualization Outputs

topic_words_bars.png Bar charts showing top-10 words for each topic with probability weights.

topic_similarity.png Heatmap showing cosine similarity between topic-word distributions.

doc_topic_umap.png UMAP projection of documents in topic space. Points are colored by dominant topic.

topic_wordclouds.png Word clouds for each topic sized by word probability.

metrics.png Bar charts comparing evaluation metrics.

pyldavis.html Interactive visualization using pyLDAvis library. Open in web browser.

---

Generating Visualizations Separately

Generate visualizations for THETA models:

cd /root/autodl-tmp/ETM

python -m visualization.run_visualization \
    --result_dir /root/autodl-tmp/result/0.6B \
    --dataset my_dataset \
    --mode zero_shot \
    --model_size 0.6B \
    --language en \
    --dpi 300

Generate visualizations for baseline models:

python -m visualization.run_visualization \
    --baseline \
    --result_dir /root/autodl-tmp/result/baseline \
    --dataset my_dataset \
    --model lda \
    --num_topics 20 \
    --language en \
    --dpi 300

Replace lda with etm, ctm, or dtm for other baseline models.

---

Customizing Visualization

Higher resolution:

python -m visualization.run_visualization \
    --result_dir /root/autodl-tmp/result/0.6B \
    --dataset my_dataset \
    --mode zero_shot \
    --model_size 0.6B \
    --language en \
    --dpi 600

Chinese language visualizations:

python -m visualization.run_visualization \
    --result_dir /root/autodl-tmp/result/0.6B \
    --dataset chinese_dataset \
    --mode zero_shot \
    --model_size 0.6B \
    --language zh \
    --dpi 300

Chinese visualizations use appropriate fonts and handle character rendering correctly.

---

Skipping Visualization During Training

Skip automatic visualization to save time:

python run_pipeline.py \
    --dataset my_dataset \
    --models theta \
    --model_size 0.6B \
    --mode zero_shot \
    --num_topics 20 \
    --epochs 100 \
    --batch_size 64 \
    --skip-viz \
    --gpu 0 \
    --language en

Visualizations can be generated later using the separate visualization command.