Delete unused files, refactoring for CR

bam.py

@@ -0,0 +1,17 @@
+import numpy as np
+
+def debias_proc(equality_sets, vecs, vocab):
+ A = []
+ B = []
+ for eq in equality_sets:
+ if eq[0] in vocab and eq[1] in vocab:
+ A.append(vecs[vocab[eq[0]]])
+ B.append(vecs[vocab[eq[1]]])
+ A = np.array(A); B = np.array(B);
+
+ product = np.matmul(A.transpose(), B)
+ U, s, V = np.linalg.svd(product)
+ print(U.shape, V.shape)
+ proj_mat = V #np.matmul(U, V)
+ res = np.matmul(vecs, proj_mat)
+ return (res + vecs) / 2, proj_mat

data_handler.py

@@ -1,3 +1,7 @@
+'''
+Used to prepare data for input to debiasnet, e.g. preparing training constraints
+'''
+
 import codecs
 from itertools import product
 import pickle
@@ -31,7 +35,7 @@ def split_weat_in_train_dev(input_paths, output_train, output_dev):
  :param output_train:
  :param output_dev:
  :return:
- >>> split_weat_in_train_dev(["./data/weat_1_aug_postspec_2_new.txt"], "./data/weat_1_aug_postspec_2_train.txt", "./data/weat_1_aug_postspec_2_dev.txt")
+ >>> split_weat_in_train_dev(["./data/weat_1_aug_postspec_2.txt"], "./data/weat_1_aug_postspec_2_train.txt", "./data/weat_1_aug_postspec_2_dev.txt")
  """
  t1, t2, a1, a2 = fuse_stimuli(input_paths)
  weat_dict = {}
@@ -50,7 +54,7 @@ def split_weat_in_train_dev(input_paths, output_train, output_dev):
 
 
 
-def filter_vocabulary(term_list, vocab_path="/work/gglavas/data/word_embs/yacle/fasttext/200K/npformat/ft.wiki.en.300.vocab"):
+def filter_vocabulary(term_list, vocab_path="./data/word_embs/yacle/fasttext/200K/npformat/ft.wiki.en.300.vocab"):
  vocab = pickle.load(open(vocab_path, "rb"))
  new_term_list = []
  for t in term_list:
@@ -62,7 +66,7 @@ def filter_vocabulary(term_list, vocab_path="/work/gglavas/data/word_embs/yacle/
 
 
 def prepare_input_examples(input_paths, output_path, random_attributes=False,
- sampling_vocab_path="/work/gglavas/data/word_embs/yacle/fasttext/200K/npformat/ft.wiki.en.300.vocab",
+ sampling_vocab_path="./data/word_embs/yacle/fasttext/200K/npformat/ft.wiki.en.300.vocab",
  original_weat_path="", k=60, switch_targets_and_attributes=False):
  """
  :param paths:
@@ -89,7 +93,7 @@ def prepare_input_examples(input_paths, output_path, random_attributes=False,
  random_attributes=True, original_weat_path="./data/weat_8.txt", k=60, switch_targets_and_attributes=False)
  #>>> prepare_input_examples(["./data/weat_8_aug_postspec_4_new.txt"], "./data/weat_8_prepared_filtered_postspec_4_wo_original_random_switched.txt", \
  random_attributes=True, original_weat_path="./data/weat_8.txt", k=60, switch_targets_and_attributes=True)
- >>> prepare_input_examples(["./data/weat_8_aug_postspec_4_new.txt"], "./data/weat_8_prepared_filtered_postspec_4_wo_original_switched.txt", switch_targets_and_attributes=True)
+ >>> prepare_input_examples(["./data/weat_8_aug_postspec_4.txt"], "./data/weat_8_prepared_filtered_postspec_4_wo_original_switched.txt", switch_targets_and_attributes=True)
  """
  t1, t2, a1, a2 = fuse_stimuli(input_paths)
  if switch_targets_and_attributes:

model.py → debiasnet.py

@@ -1,3 +1,6 @@
+"""
+Full debiasnet model
+"""
 import tensorflow as tf
 from flip_gradient import flip_gradient
 
@@ -24,8 +27,8 @@ def le_distance(t1, t2, asym_fact):
 def hinge_loss(true_ledists, false_ledists, margin):
  return tf.reduce_sum(tf.maximum(tf.subtract(tf.constant(margin, dtype = tf.float32), tf.subtract(false_ledists, true_ledists)), 0.0))
 
-class DebbieModel(object): # setting the adversarial grad scale to -1 turns of the flipping of the gradient
- def __init__(self, embs, mlp_layers, activation = tf.nn.tanh, scope = "debbie", reg_factor = 0.1, learning_rate = 0.0001, adversarial_grad_scale=1.0, i_factor=1.0, e_factor=1.0):
+class DebiasNetModel(object): # setting the adversarial grad scale to -1 turns of the flipping of the gradient
+ def __init__(self, embs, mlp_layers, activation = tf.nn.tanh, scope = "debie", reg_factor = 0.1, learning_rate = 0.0001, adversarial_grad_scale=1.0, i_factor=1.0, e_factor=1.0):
  self.embeddings = embs
  self.scope = scope
 

model_direct.py → debiasnet_direct.py

@@ -1,3 +1,6 @@
+"""
+Debiasnet model with direct implicit debiasing training objective (not reported in the paper)
+"""
 import tensorflow as tf
 from flip_gradient import flip_gradient
 
@@ -24,8 +27,8 @@ def le_distance(t1, t2, asym_fact):
 def hinge_loss(true_ledists, false_ledists, margin):
  return tf.reduce_sum(tf.maximum(tf.subtract(tf.constant(margin, dtype = tf.float32), tf.subtract(false_ledists, true_ledists)), 0.0))
 
-class DebbieModel(object): # setting the adversarial grad scale to -1 turns of the flipping of the gradient
- def __init__(self, embs, mlp_layers, activation = tf.nn.tanh, scope = "debbie", reg_factor = 0.1, learning_rate = 0.0001, adversarial_grad_scale=1.0, i_factor=1.0, e_factor=1.0):
+class DebiasNetModel(object): # setting the adversarial grad scale to -1 turns of the flipping of the gradient
+ def __init__(self, embs, mlp_layers, activation = tf.nn.tanh, scope = "debie", reg_factor = 0.1, learning_rate = 0.0001, adversarial_grad_scale=1.0, i_factor=1.0, e_factor=1.0):
  self.embeddings = embs
  self.scope = scope