File size: 3,006 Bytes
4f77642 | 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 | import os
from typing import List
import fire
import numpy as np
import pandas as pd
import torch
from matplotlib import pyplot as plt
from tqdm import tqdm
def rst_weights_plot(model_name, k, forward_arcs=True, backward_arcs=True, plot_width=8, plot_height=-1, xlim=-1):
def real_coeffs(theta: torch.tensor, beta: torch.tensor, relations) -> pd.Series:
""" What are the real coefficients c_ij for specific RST relations in the final model. """
weights_dict = dict(zip(relations, torch.nn.functional.relu(theta + beta) / beta))
return pd.Series(weights_dict).map(lambda tensor: tensor.item()).sort_values()
def weights_stats(weights: List[pd.Series], backward=False) -> list:
if not weights: return []
w_stats = pd.DataFrame(weights).agg([np.mean, np.std])
w_stats.columns = [column[0].upper() + column[1:] + '⁻¹' * backward for column in w_stats.columns]
if 'Root⁻¹' in w_stats.columns: del w_stats['Root⁻¹']
return [w_stats]
pos_res = []
neg_res = []
for fold in tqdm(range(k), desc=f'Processing folds'):
weights = torch.load(f'{model_name}/fold_{fold}/best.th', map_location='cpu')
relations = [rel.strip() for rel in
open(f'{model_name}/fold_{fold}/vocabulary/rst_rels_labels.txt', 'r').readlines()]
if forward_arcs:
theta, beta = weights['ff_rst_rels_arc.weight'][0], weights['ff_rst_rels_arc.bias'][0]
pos_res.append(real_coeffs(theta, beta, relations))
if backward_arcs:
theta, beta = weights['ff_rst_rels_arc_back.weight'][0], weights['ff_rst_rels_arc_back.bias'][0]
neg_res.append(real_coeffs(theta, beta, relations))
all_rels = pd.concat(weights_stats(pos_res) + weights_stats(neg_res, backward=True), axis=1)
std = all_rels.T.max()['std']
std = 0. if np.isnan(std) else std # If k==1, there will be nan
xlim = xlim if xlim != -1 else all_rels.T.max()[
'mean'] + std # right limit for X; enforce or compute from the statistics (-1)
plot_height = plot_height if plot_height != -1 else (all_rels.shape[
1] + 1) * 0.2 # plot height; enforce or compute from the number of relations (-1)
ax = all_rels.T.sort_values('mean').plot(kind='barh', y='mean', legend=False, xerr='std', logx=False,
color='#81AD75',
xlim=(0, xlim),
figsize=(plot_width, plot_height))
ax.axvline(x=1, color='r', linestyle='--', lw=2)
ax.set_xlabel('Weight')
ax.set_ylabel('RST relation')
# ax.set_title(f'Weights for {model_name}')
plt.savefig(f"{model_name}/rst_coeff_{os.path.basename(model_name)}.pdf", format="pdf", bbox_inches="tight")
return all_rels.T.sort_values('mean')
if __name__ == '__main__':
fire.Fire(rst_weights_plot)
|