20newsgroup_viz.py

# coding: utf-8

import pickle as pkl
import seaborn as sns
import pandas as pd

from sklearn.manifold import TSNE
from matplotlib import pyplot as plt
from itertools import product
from tqdm import tqdm


embedding_path = 'embeddings/20newsgroup/alpha2-l2_param0.001-epochs1-pre_epochs0.pkl'
embeddings, labels = pkl.load(open(embedding_path, 'rb'))


parameter_names = ['perplexity', 'n_iter']
# parameter_values = [[10, 30, 50], [500, 750, 1000, 1250]]
parameter_values = [[10, 20], [500, 1000]]

value_combinations = list(product(*parameter_values))

parameter_dicts = [dict(list(zip(parameter_names, values))) for values in value_combinations]


nrow, ncol = len(parameter_values[0]), len(parameter_values[1])
width = 5
fig, axes = plt.subplots(nrow, ncol, figsize=(width * ncol, width * nrow))

for i, perp in tqdm(enumerate(parameter_values[0])):
    for j, n_iter in enumerate(parameter_values[1]):
        ax = axes[i, j]
            
        pos = TSNE(n_components=2, perplexity=perp, n_iter=n_iter).fit_transform(embeddings)
        ax.scatter(pos[:, 0], pos[:, 1], c=labels)



pos = TSNE(n_components=2, perplexity=20, n_iter=500).fit_transform(embeddings)



# In[43]:


# prettier plot
df = pd.DataFrame()

df['x'] = pos[:, 0]
df['y'] = pos[:, 1]
legends = ['comp.graphics', 'rec.sport.baseball', 'talk.politics.guns']
df['class'] = [legends[l] for l in labels]

sns.set_context("notebook", font_scale=1.5)
sns.set_style("ticks")


# Create scatterplot of dataframe
sns.lmplot('x',  # Horizontal axis
           'y',  # Vertical axis
           data=df,  # Data source
           fit_reg=False,  # Don't fix a regression line
           hue="class",  # Set color,
           legend=True,
           scatter_kws={"s": 25, 'alpha': 0.5})  # S marker size

sns.despine(top=True, left=True, right=True, bottom=True)
plt.xticks([])
plt.yticks([])

plt.xlabel('')
plt.ylabel('')

plt.savefig('figs/20newsgroup_viz.png')