def test_different_signature_dims_flow():
first_signature = np.array([0.0, 1.0])
second_signature = np.array([5.0, 3.0, 3.0])
distance_matrix = np.array([[0.0, 0.5, 0.0], [0.5, 0.0, 0.0],
[0.5, 0.0, 0.0]])
with pytest.raises(ValueError):
emd_with_flow(first_signature, second_signature, distance_matrix)
def test_emd_with_flow_validate_square_distance_matrix():
first_signature = np.array([0.0, 1.0])
second_signature = np.array([5.0, 3.0])
distance_matrix = np.array([[0.0, 0.5, 3.0],
[0.5, 0.0]])
with pytest.raises(ValueError):
emd_with_flow(first_signature, second_signature, distance_matrix)
def test_emd_with_flow_1():
first_signature = np.array([0.0, 1.0])
second_signature = np.array([5.0, 3.0])
distance_matrix = np.array([[0.0, 0.5], [0.5, 0.0]])
emd_flow_assert(
emd_with_flow(first_signature, second_signature, distance_matrix),
(3.5, [[0.0, 0.0], [0.0, 1.0]]))
def test_case_2_flow():
first_signature = np.array([1.0, 1.0])
second_signature = np.array([1.0, 1.0])
distance_matrix = np.array([[0.0, 1.0], [1.0, 0.0]])
assert (emd_with_flow(first_signature, second_signature,
distance_matrix) == (0.0, [[1.0, 0.0], [0.0, 1.0]]))
def word_mover_score(refs,
hyps,
idf_dict_ref,
idf_dict_hyp,
stop_words=[],
n_gram=1,
remove_subwords=True,
batch_size=256):
preds = []
for batch_start in range(0, len(refs), batch_size):
batch_refs = refs[batch_start:batch_start + batch_size]
batch_hyps = hyps[batch_start:batch_start + batch_size]
ref_embedding, ref_lens, ref_masks, ref_idf, ref_tokens = get_bert_embedding(
batch_refs, model, tokenizer, idf_dict_ref)
hyp_embedding, hyp_lens, hyp_masks, hyp_idf, hyp_tokens = get_bert_embedding(
batch_hyps, model, tokenizer, idf_dict_hyp)
ref_embedding = ref_embedding[-1]
hyp_embedding = hyp_embedding[-1]
batch_size = len(ref_tokens)
for i in range(batch_size):
ref_ids = [
k for k, w in enumerate(ref_tokens[i])
if w in stop_words or '##' in w or w in set(string.punctuation)
]
hyp_ids = [
k for k, w in enumerate(hyp_tokens[i])
if w in stop_words or '##' in w or w in set(string.punctuation)
]
ref_embedding[i, ref_ids, :] = 0
hyp_embedding[i, hyp_ids, :] = 0
ref_idf[i, ref_ids] = 0
hyp_idf[i, hyp_ids] = 0
raw = torch.cat([ref_embedding, hyp_embedding], 1)
raw.div_(torch.norm(raw, dim=-1).unsqueeze(-1) + 1e-30)
distance_matrix = batched_cdist_l2(raw, raw).double().cpu().numpy()
for i in range(batch_size):
c1 = np.zeros(raw.shape[1], dtype=np.float)
c2 = np.zeros(raw.shape[1], dtype=np.float)
c1[:len(ref_idf[i])] = ref_idf[i]
c2[len(ref_idf[i]):] = hyp_idf[i]
c1 = _safe_divide(c1, np.sum(c1))
c2 = _safe_divide(c2, np.sum(c2))
dst = distance_matrix[i]
_, flow = emd_with_flow(c1, c2, dst)
flow = np.array(flow, dtype=np.float32)
score = 1 - np.sum(flow * dst)
preds.append(score)
return preds
def test_emd_with_flow_5():
first_signature = np.array([3.0, 5.0])
second_signature = np.array([6.0, 2.0])
distance_matrix = np.array([[0.0, 0.0], [0.0, 0.0]])
emd_flow_assert(
emd_with_flow(first_signature, second_signature, distance_matrix),
(0.0, [[3.0, 0.0], [3.0, 2.0]]))
def emd_rep_loss(student_reps, teacher_reps, student_layer_weight,
teacher_layer_weight, stu_layer_num, tea_layer_num,
loss_func, device):
student_layer_weight = np.concatenate(
(student_layer_weight, np.zeros(tea_layer_num)))
teacher_layer_weight = np.concatenate(
(np.zeros(stu_layer_num), teacher_layer_weight))
total_num = stu_layer_num + tea_layer_num
distance_matrix = torch.zeros([total_num, total_num]).to(device)
for i in range(stu_layer_num):
student_rep = student_reps[i]
for j in range(tea_layer_num):
teacher_rep = teacher_reps[j]
tmp_loss = loss_func(student_rep, teacher_rep)
distance_matrix[i][j + stu_layer_num] = distance_matrix[
j + stu_layer_num][i] = tmp_loss
_, trans_matrix = emd_with_flow(
student_layer_weight, teacher_layer_weight,
distance_matrix.detach().cpu().numpy().astype('float64'))
rep_loss = torch.sum(
torch.tensor(trans_matrix).to(device) * distance_matrix)
# tmp = distance_matrix.detach().cpu().numpy()
return rep_loss, trans_matrix, distance_matrix
def attn_loss(self, student_atts, teacher_atts, return_distance=False):
if len(teacher_atts[0].shape) == 4 and len(student_atts[0].shape) == 3:
teacher_atts = [torch.mean(x, dim=1) for x in teacher_atts]
elif len(student_atts[0].shape) == 4 and len(
teacher_atts[0].shape) == 4:
if student_atts[0].shape[1] != teacher_atts[0].shape[1]:
teacher_atts = [torch.mean(x, dim=1) for x in teacher_atts]
student_atts = [torch.mean(x, dim=1) for x in student_atts]
elif len(student_atts[0].shape) == 4 and len(
teacher_atts[0].shape) == 3:
student_atts = [torch.mean(x, dim=1) for x in student_atts]
_s_weight = np.concatenate(
(self.s_weight, np.zeros_like(self.t_weight)))
_t_weight = np.concatenate(
(np.zeros_like(self.s_weight), self.t_weight))
totol_num = self.s_layer_num + self.t_layer_num
distance_matrix = torch.zeros([totol_num, totol_num]).to(self.device)
for i in range(self.s_layer_num):
student_att = student_atts[i]
for j in range(self.t_layer_num):
teacher_att = teacher_atts[j]
tmp_loss = self.d_method(student_att, teacher_att)
distance_matrix[i][j + self.s_layer_num] = distance_matrix[
j + self.s_layer_num][i] = tmp_loss
_, trans_matrix = emd_with_flow(
_s_weight, _t_weight,
distance_matrix.detach().cpu().numpy().astype('float64'))
d = torch.sum(
torch.tensor(trans_matrix).to(self.device) * distance_matrix)
if return_distance:
return d, trans_matrix, distance_matrix
else:
return d
def get_opinion_distance(model, noun_freq_polar1_model, noun_freq_polar1_terms,
noun_freq_polar2_model, noun_freq_polar2_terms,
filename1, filename2):
dictionary = Dictionary(
documents=[noun_freq_polar1_terms, noun_freq_polar2_terms])
# Compute the euclidean distance between word vectors
semantic_distance_matrix = compute_semantic_distance_matrix(
model, noun_freq_polar1_terms, noun_freq_polar2_terms, dictionary,
filename1, filename2)
if semantic_distance_matrix is None:
print("Semantic distance is none")
return np.nan, np.nan
# Get normalized frequency and it's polarity
normalized_freq1, pol1 = get_norm_freq_polarity(noun_freq_polar1_model,
dictionary)
normalized_freq2, pol2 = get_norm_freq_polarity(noun_freq_polar2_model,
dictionary)
# Change output to d1_terms, d2_terms, d_matrix[len(d1_terms),len(d2_terms)]
emd_distance, matching_matrix = emd_with_flow(normalized_freq1,
normalized_freq2,
semantic_distance_matrix)
polarity_distance = compute_polarity_distance(normalized_freq1,
normalized_freq2, dictionary,
matching_matrix,
semantic_distance_matrix,
pol1, pol2)
return polarity_distance, emd_distance
def get_one_grad(self, id1, id2):
vn = len( self.all_words )
n = len(self.tr_X)
vec1 = np.zeros(vn)
for k,w in enumerate( self.tr_words[id1][0] ):
word = w#w[0]
vec1[ self.word_id_map[word] ] = self.tr_BOW_X[id1][0][k]
vec2 = np.zeros(vn)
for k,w in enumerate( self.tr_words[id2][0] ):
word = w#w[0]
vec2[ self.word_id_map[word] ] = self.tr_BOW_X[id2][0][k]
part_grad_A = np.zeros_like( self.A )
T = emd_with_flow( np.array(vec1), np.array(vec2), self.cw_dist_mat )
for id1,t1 in enumerate(T[1]):
for id2,t2 in enumerate(t1):
if t2 > 0.:
x = np.array(self.all_X[ id1 ]) - np.array(self.all_X[ id2 ])
const = t2 / (2 * np.sqrt( np.sum( self.A.dot(x)**2 ) )) if np.sum( x**2 ) != 0 else 0
for i in range(self.vector_dim):
Ax = self.A[i,:].dot(x)
for j in range(self.vector_dim):
part_grad_A[i,j] += const * 2 * Ax * x[j]
return part_grad_A
def word_mover_score(mapping, projection, bias, model, tokenizer, src, hyps, \
n_gram=2, layer=8, dropout_rate=0.3, batch_size=256, device='cuda:0'):
idf_dict_src = defaultdict(lambda: 1.)
idf_dict_hyp = defaultdict(lambda: 1.)
preds = []
for batch_start in range(0, len(src), batch_size):
batch_src = src[batch_start:batch_start + batch_size]
batch_hyps = hyps[batch_start:batch_start + batch_size]
src_embedding, src_lens, src_masks, src_idf, src_tokens = get_bert_embedding(
batch_src, model, tokenizer, idf_dict_src, device=device)
hyp_embedding, hyp_lens, hyp_masks, hyp_idf, hyp_tokens = get_bert_embedding(
batch_hyps, model, tokenizer, idf_dict_hyp, device=device)
src_embedding = src_embedding[layer]
hyp_embedding = hyp_embedding[layer]
src_embedding = cross_lingual_mapping(mapping, src_embedding,
projection, bias[0])
batch_size = src_embedding.shape[0]
for i in range(batch_size):
src_embedding_i = get_ngram_embs(src_embedding[i, :src_lens[i], :],
ngram=n_gram)
hyp_embedding_i = get_ngram_embs(hyp_embedding[i, :hyp_lens[i], :],
ngram=n_gram)
src_idf_i = [1] * (src_lens[i] - n_gram + 1)
hyp_idf_i = [1] * (hyp_lens[i] - n_gram + 1)
W = torch.cat([src_embedding_i, hyp_embedding_i], 0)
W.div_(torch.norm(W, dim=-1).unsqueeze(-1))
c1 = list(src_idf_i) + [0] * len(hyp_idf_i)
c2 = [0] * len(src_idf_i) + list(hyp_idf_i)
c1 = c1 / np.sum(c1) + 1e-9
c2 = c2 / np.sum(c2) + 1e-9
dist = torch.cdist(W, W, p=2).double().cpu().numpy()
flow = np.stack(emd_with_flow(c1, c2, dist)[1])
flow = torch.from_numpy(flow[:len(src_idf_i), len(src_idf_i):])
dist = torch.from_numpy(dist[:len(src_idf_i), len(src_idf_i):])
# remove noisy elements in a flow
flow_flatten = flow.reshape(-1)
idx = torch.nonzero(flow_flatten)
threshold = flow_flatten[idx].topk(k=max(
int(len(idx) * dropout_rate), 1),
dim=0,
largest=False)[0][-1]
flow[flow < threshold] = 0
score = (flow * dist).sum()
preds.append(1 - score)
return preds
def test_emd_with_flow_6():
first_signature = np.array([1.0, 2.0, 1.0, 2.0])
second_signature = np.array([2.0, 1.0, 2.0, 1.0])
distance_matrix = np.array([[0.0, 1.0, 1.0, 2.0], [1.0, 0.0, 2.0, 1.0],
[1.0, 2.0, 0.0, 1.0], [2.0, 1.0, 1.0, 0.0]])
emd_flow_assert(
emd_with_flow(first_signature, second_signature, distance_matrix),
(2.0, [[1.0, 0.0, 0.0, 0.0], [1.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 1.0, 1.0]]))
def test_emd_with_flow_1():
first_signature = np.array([0.0, 1.0])
second_signature = np.array([5.0, 3.0])
distance_matrix = np.array([[0.0, 0.5],
[0.5, 0.0]])
emd_flow_assert(
emd_with_flow(first_signature, second_signature, distance_matrix),
(3.5, [[0.0, 0.0],
[0.0, 1.0]])
)
def test_emd_with_flow_5():
first_signature = np.array([3.0, 5.0])
second_signature = np.array([6.0, 2.0])
distance_matrix = np.array([[0.0, 0.0],
[0.0, 0.0]])
emd_flow_assert(
emd_with_flow(first_signature, second_signature, distance_matrix),
(0.0, [[3.0, 0.0],
[3.0, 2.0]])
)
def get_wmd(s1, s2, dists, w2i, get_flow=False):
"""
Get WMD for two input sentences
"""
s1, s2 = s1.split(), s2.split()
h1, h2, words = get_wmd_histograms(s1, s2, w2i)
D = dists[np.ix_(words, words)]
if get_flow:
return pyemd.emd_with_flow(h1, h2, D)
return pyemd.emd(h1, h2, D)
def test_case_6_flow():
first_signature = np.array([1.0, 2.0, 1.0, 2.0])
second_signature = np.array([2.0, 1.0, 2.0, 1.0])
distance_matrix = np.array([[0.0, 1.0, 1.0, 2.0], [1.0, 0.0, 2.0, 1.0],
[1.0, 2.0, 0.0, 1.0], [2.0, 1.0, 1.0, 0.0]])
emd_value, flow = emd_with_flow(first_signature, second_signature,
distance_matrix)
emd_value = round(emd_value, EMD_PRECISION)
assert emd_value == 2.0
flow = np.round(flow, FLOW_PRECISION)
assert np.array_equal(flow, [[1.0, 0.0, 0.0, 0.0], [1.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 1.0, 1.0]])
def emd_rep_loss(student_reps, teacher_reps, student_layer_weight,
teacher_layer_weight, stu_layer_num, tea_layer_num,
loss_mse, args):
student_layer_weight = np.concatenate(
(student_layer_weight, np.zeros(tea_layer_num)))
teacher_layer_weight = np.concatenate(
(np.zeros(stu_layer_num), teacher_layer_weight))
total_num = stu_layer_num + tea_layer_num
distance_matrix = torch.zeros([total_num, total_num]).to(args.device)
for i in range(stu_layer_num):
student_rep = student_reps[i]
for j in range(tea_layer_num):
teacher_rep = teacher_reps[j]
if args.emd_type == "v6":
tmp_loss = 1 - torch.sum(student_rep * teacher_rep, dim=2)
tmp_loss = tmp_loss.mean()
elif args.emd_type == "v8":
# weight = torch.norm(teacher_rep, dim=2, keepdim=True)
# weight = weight / torch.sum(weight, dim=1, keepdim=True)
tmp_loss = torch.sum(v8_weight *
(student_rep - teacher_rep)**2,
dim=1).mean()
else:
tmp_loss = loss_mse(student_rep, teacher_rep)
distance_matrix[i][j + stu_layer_num] = distance_matrix[
j + stu_layer_num][i] = tmp_loss
if args.emd_type == "v10":
tmp = distance_matrix.detach().cpu().numpy()
student_layer_weight = tmp.mean(axis=1)[:stu_layer_num]
teacher_layer_weight = tmp.mean(axis=0)[stu_layer_num:]
student_layer_weight = sum(
student_layer_weight) / student_layer_weight
teacher_layer_weight = sum(
teacher_layer_weight) / teacher_layer_weight
student_layer_weight = utils.softmax(student_layer_weight / 20)
teacher_layer_weight = utils.softmax(teacher_layer_weight / 20)
student_layer_weight = np.concatenate(
(student_layer_weight, np.zeros(tea_layer_num)))
teacher_layer_weight = np.concatenate(
(np.zeros(stu_layer_num), teacher_layer_weight))
_, trans_matrix = emd_with_flow(
student_layer_weight, teacher_layer_weight,
distance_matrix.detach().cpu().numpy().astype('float64'))
rep_loss = torch.sum(
torch.tensor(trans_matrix).to(args.device) * distance_matrix)
tmp = distance_matrix.detach().cpu().numpy()
return rep_loss, trans_matrix, distance_matrix
def test_emd_with_flow_extra_mass_penalty():
first_signature = np.array([0.0, 2.0, 1.0, 2.0])
second_signature = np.array([2.0, 1.0, 2.0, 1.0])
distance_matrix = np.array([[0.0, 1.0, 1.0, 2.0], [1.0, 0.0, 2.0, 1.0],
[1.0, 2.0, 0.0, 1.0], [2.0, 1.0, 1.0, 0.0]])
emd_flow_assert(
emd_with_flow(first_signature,
second_signature,
distance_matrix,
extra_mass_penalty=2.5),
(4.5, [[0.0, 0.0, 0.0, 0.0], [1.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 1.0, 1.0]]))
def emd_attn_loss(student_attns, teacher_attns, student_layer_weight,
teacher_layer_weight, stu_layer_num, tea_layer_num,
loss_mse, args):
student_layer_weight = np.concatenate(
(student_layer_weight, np.zeros(tea_layer_num)))
teacher_layer_weight = np.concatenate(
(np.zeros(stu_layer_num), teacher_layer_weight))
total_num = stu_layer_num + tea_layer_num
distance_matrix = torch.zeros([total_num, total_num]).to(args.device)
if args.emd_type == "v3":
t_dis_matrix = torch.zeros([total_num, total_num]).to(args.device)
for i in range(stu_layer_num):
student_attn = student_attns[i]
for j in range(tea_layer_num):
teacher_attn = teacher_attns[j]
if args.emd_type == "v6":
tmp_loss = 1 - torch.sum(teacher_attn * student_attn,
dim=2)
tmp_loss = tmp_loss.mean()
elif args.emd_type == "v8":
# weight = torch.norm(teacher_attn, dim=2, keepdim=True)
# weight = weight / torch.sum(weight, dim=1, keepdim=True)
tmp_loss = torch.sum(v8_weight *
(student_attn - teacher_attn)**2,
dim=1).mean()
else:
tmp_loss = loss_mse(student_attn, teacher_attn)
if args.emd_type == "v3":
t_dis_matrix[i][j + stu_layer_num] = t_dis_matrix[
j + stu_layer_num][i] = tmp_loss
tmp_loss *= (
1 + abs(i / stu_layer_num - j / tea_layer_num) / 5)
distance_matrix[i][j + stu_layer_num] = distance_matrix[
j + stu_layer_num][i] = tmp_loss
_, trans_matrix = emd_with_flow(
student_layer_weight, teacher_layer_weight,
distance_matrix.detach().cpu().numpy().astype('float64'))
if args.emd_type == "v3":
attn_loss = torch.sum(
torch.tensor(trans_matrix).to(args.device) * t_dis_matrix)
else:
attn_loss = torch.sum(
torch.tensor(trans_matrix).to(args.device) * distance_matrix)
tmp = distance_matrix.detach().cpu().numpy()
return attn_loss, trans_matrix, distance_matrix
def test_emd_with_flow_6():
first_signature = np.array([1.0, 2.0, 1.0, 2.0])
second_signature = np.array([2.0, 1.0, 2.0, 1.0])
distance_matrix = np.array([[0.0, 1.0, 1.0, 2.0],
[1.0, 0.0, 2.0, 1.0],
[1.0, 2.0, 0.0, 1.0],
[2.0, 1.0, 1.0, 0.0]])
emd_flow_assert(
emd_with_flow(first_signature, second_signature, distance_matrix),
(2.0, [[1.0, 0.0, 0.0, 0.0],
[1.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 1.0, 1.0]])
)
def test_extra_mass_penalty_flow():
first_signature = np.array([0.0, 2.0, 1.0, 2.0])
second_signature = np.array([2.0, 1.0, 2.0, 1.0])
distance_matrix = np.array([[0.0, 1.0, 1.0, 2.0], [1.0, 0.0, 2.0, 1.0],
[1.0, 2.0, 0.0, 1.0], [2.0, 1.0, 1.0, 0.0]])
emd_value, flow = emd_with_flow(first_signature,
second_signature,
distance_matrix,
extra_mass_penalty=2.5)
emd_value = round(emd_value, EMD_PRECISION)
assert emd_value == 4.5
flow = np.round(flow, FLOW_PRECISION)
print(flow)
assert np.array_equal(flow, [[0.0, 0.0, 0.0, 0.0], [1.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 1.0, 1.0]])
def test_emd_with_flow_extra_mass_penalty():
first_signature = np.array([0.0, 2.0, 1.0, 2.0])
second_signature = np.array([2.0, 1.0, 2.0, 1.0])
distance_matrix = np.array([[0.0, 1.0, 1.0, 2.0],
[1.0, 0.0, 2.0, 1.0],
[1.0, 2.0, 0.0, 1.0],
[2.0, 1.0, 1.0, 0.0]])
emd_flow_assert(
emd_with_flow(first_signature, second_signature, distance_matrix,
extra_mass_penalty=2.5),
(4.5, [[0.0, 0.0, 0.0, 0.0],
[1.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 1.0, 1.0]])
)
def loss(self, student_reps, teacher_reps, return_distance=False):
_s_weight = np.concatenate(
(self.s_weight, np.zeros_like(self.t_weight)))
_t_weight = np.concatenate(
(np.zeros_like(self.s_weight), self.t_weight))
totol_num = self.s_layer_num + self.t_layer_num
distance_matrix = torch.zeros([totol_num, totol_num]).to(self.device)
for i in range(self.s_layer_num):
student_rep = student_reps[i]
for j in range(self.t_layer_num):
teacher_rep = teacher_reps[j + 1]
tmp_loss = self.d_method(student_rep, teacher_rep)
distance_matrix[i][j + self.s_layer_num] = distance_matrix[
j + self.s_layer_num][i] = tmp_loss
_, trans_matrix = emd_with_flow(
_s_weight, _t_weight,
distance_matrix.detach().cpu().numpy().astype('float64'))
d = torch.sum(
torch.tensor(trans_matrix).to(self.device) * distance_matrix)
if return_distance:
return d, trans_matrix, distance_matrix
else:
return d
#from sys import argv
import csv
import numpy as np
import pandas as pd
#import os
#from math import radians, cos, sin, asin, sqrt, exp
#import copy
#import random
import pyemd
from pyemd import emd_with_flow
# Load species seasonal abundance distributions (estimated from eBird data)
abundance_BR = pd.read_table('results/output/seasonalAbundance_BR.csv', sep=",")
abundance_NB = pd.read_table('results/output/seasonalAbundance_NB.csv', sep=",")
abundance_BR = abundance_BR.apply(lambda x: x/sum(x))
abundance_NB = abundance_NB.apply(lambda x: x/sum(x))
# Load matrix of pairwise distance between every hexagons on the grid
distanceMatrix = np.loadtxt('results/output/distanceMatrix.csv', delimiter=";")
# Compute optimal redistribution using the Earth Mover's Distance algorithm
for s in abundance_BR.columns:
EMD_results = emd_with_flow(np.array(abundance_BR[s]), np.array(abundance_NB[s]), distanceMatrix)
EMD_results = np.array(EMD_results[1])[(np.array(abundance_BR[s]) > 0),:][:,(np.array(abundance_NB[s]) > 0)]
np.savetxt("results/output/ORSIM_results_" + s + ".csv", EMD_results, delimiter=',') # Save simulated migratory connectivity
from pyemd import emd_with_flow import numpy as np first_histogram = np.array([0, 1.0]) second_histogram = np.array([5.0, 3.0]) distance_matrix = np.array([[0, 2.0], [2.5, 0]]) w_dis, F = emd_with_flow(first_histogram, second_histogram, distance_matrix) print(w_dis, F) sumF = np.sum(F) norm_dis = w_dis / sumF print(norm_dis)
def emd_distill_loss(self, layer_wise_hidden, teacher_layer_wise_hidden):
# 1. Compute EMD loss
# 2. Update weight
student_weight = self.student_weights.copy() # [|S|]
teacher_weight = self.teacher_weights.copy() # [|T|]
student_weight_hist = np.concatenate(
(student_weight, np.zeros(self.teacher_layers))) # [|S|+|T|]
teacher_weight_hist = np.concatenate(
(np.zeros(self.student_layers), teacher_weight)) # [|S|+|T|]
total = self.teacher_layers + self.student_layers
distance_matrix = torch.zeros([total, total], device=self.dummy.device)
# Compute distance matrix, shape=(|S|, |T|)
for i in range(self.student_layers):
student_hidden = layer_wise_hidden[i] # [B, L, C]
for j in range(self.teacher_layers):
teacher_hidden = teacher_layer_wise_hidden[j] # [B, L, C]
# KL Div
distance_matrix[i][
j + self.student_layers] = self.layer_distill_loss(
student_hidden, teacher_hidden)
distance_matrix[
j + self.student_layers][i] = self.layer_distill_loss(
teacher_hidden, student_hidden)
# MSE symmetric
# distance = self.mse(student_hidden, teacher_hidden)
# distance_matrix[i][j + self.student_layers] = distance
# distance_matrix[j + self.student_layers][i] = distance
d_np = distance_matrix.detach().cpu().numpy().astype(
"float64") # [|S|+|T|, |S|+|T|]
_, transfer_matrix = emd_with_flow(student_weight_hist,
teacher_weight_hist, d_np)
transfer_matrix = np.array(transfer_matrix,
dtype=np.float) # [|S|+|T|, |S|+|T|]
transfer_matrix_torch = torch.tensor(transfer_matrix,
device=self.dummy.device)
kd_loss = torch.sum(transfer_matrix_torch * distance_matrix)
# Update weight
def update_weight(weight, t_mat, num_layers, bias=0):
# t_mat: [|S|+|T|, |S|+|T|]
transfer_weight = np.sum(t_mat * d_np, -1) # [|S|+|T|]
for idx in range(num_layers):
weight[idx] = transfer_weight[idx + bias] / weight[idx]
weight_sum = np.sum(weight)
for idx in range(num_layers):
if weight[idx] != 0:
weight[idx] = weight_sum / weight[idx]
weight = np_softmax(weight / self.emd_temperature)
return weight
self.student_weights = update_weight(student_weight, transfer_matrix,
self.student_layers)
self.teacher_weights = update_weight(teacher_weight,
np.transpose(transfer_matrix),
self.teacher_layers,
bias=self.student_layers)
return kd_loss
def plot_example(is_flow, reference, translation, device='cuda:0'):
idf_dict_ref = defaultdict(lambda: 1.)
idf_dict_hyp = defaultdict(lambda: 1.)
ref_embedding, ref_lens, ref_masks, ref_idf, ref_tokens = get_bert_embedding(
[reference], model, tokenizer, idf_dict_ref, device=device)
hyp_embedding, hyp_lens, hyp_masks, hyp_idf, hyp_tokens = get_bert_embedding(
[translation], model, tokenizer, idf_dict_hyp, device=device)
ref_embedding = ref_embedding[-1]
hyp_embedding = hyp_embedding[-1]
raw = torch.cat([ref_embedding, hyp_embedding], 1)
raw.div_(torch.norm(raw, dim=-1).unsqueeze(-1) + 1e-30)
distance_matrix = batched_cdist_l2(raw, raw)
masks = torch.cat([ref_masks, hyp_masks], 1)
masks = torch.einsum('bi,bj->bij', (masks, masks))
distance_matrix = masks * distance_matrix
i = 0
c1 = np.zeros(raw.shape[1], dtype=np.float)
c2 = np.zeros(raw.shape[1], dtype=np.float)
c1[:len(ref_idf[i])] = ref_idf[i]
c2[len(ref_idf[i]):] = hyp_idf[i]
c1 = _safe_divide(c1, np.sum(c1))
c2 = _safe_divide(c2, np.sum(c2))
dst = distance_matrix[i].double().cpu().numpy()
if is_flow:
_, flow = emd_with_flow(c1, c2, dst)
new_flow = np.array(flow, dtype=np.float32)
res = new_flow[:len(ref_tokens[i]),
len(ref_idf[i]):(len(ref_idf[i]) + len(hyp_tokens[i]))]
else:
res = 1 - dst[:len(ref_tokens[i]),
len(ref_idf[i]):(len(ref_idf[i]) + len(hyp_tokens[i]))]
r_tokens = ref_tokens[i]
h_tokens = hyp_tokens[i]
fig, ax = plt.subplots(figsize=(len(r_tokens) * 0.8, len(h_tokens) * 0.8))
im = ax.imshow(res, cmap='Blues')
ax.set_xticks(np.arange(len(h_tokens)))
ax.set_yticks(np.arange(len(r_tokens)))
ax.set_xticklabels(h_tokens, fontsize=10)
ax.set_yticklabels(r_tokens, fontsize=10)
plt.xlabel("System Translation", fontsize=14)
plt.ylabel("Human Reference", fontsize=14)
plt.title("Flow Matrix", fontsize=14)
plt.setp(ax.get_xticklabels(),
rotation=45,
ha="right",
rotation_mode="anchor")
# for i in range(len(r_tokens)):
# for j in range(len(h_tokens)):
# text = ax.text(j, i, '{:.2f}'.format(res[i, j].item()),
# ha="center", va="center", color="k" if res[i, j].item() < 0.6 else "w")
fig.tight_layout()
plt.show()
def wmdo(wvvecs,
ref,
cand,
ref_lang='en',
cand_lang='en',
delta=0.18,
alpha=0.1):
'''
wvvecs: word vectors -- retrieved from load_wv method
ref: reference translation
cand: candidate translation
missing: missing word dictionary -- initialise as {}
dim: word vector dimension
delta: weight of fragmentation penalty
alpha: weight of missing word penalty
'''
ref_list = get_input_words(ref)
cand_list = get_input_words(cand)
ref = ' '.join(ref_list)
cand = ' '.join(cand_list)
common_vectorizer = CountVectorizer().fit(ref_list + cand_list)
ref_count_vector, cand_count_vector = common_vectorizer.transform(
[ref, cand])
ref_count_vector = ref_count_vector.toarray().ravel()
cand_count_vector = cand_count_vector.toarray().ravel()
dim = wvvecs[ref_lang].vector_size
wvoc, missing = create_vocabulary(common_vectorizer, wvvecs, dim, ref_list,
cand_list, ref_lang, cand_lang)
distance_matrix = cosine_distances(wvoc)
vocab_words = common_vectorizer.get_feature_names()
for cand_word_idx, count in enumerate(cand_count_vector):
if count > 0:
most_similar_ref_indexes = np.argsort(
distance_matrix[cand_word_idx])
for ref_word_index in most_similar_ref_indexes[1:]:
if ref_count_vector[ref_word_index] > 0:
print('{}: {}'.format(vocab_words[cand_word_idx],
vocab_words[ref_word_index]))
break
if np.sum(distance_matrix) == 0.0:
return 0., {}
#return float('inf')
ref_count_vector = ref_count_vector.astype(np.double)
cand_count_vector = cand_count_vector.astype(np.double)
ref_count_vector /= ref_count_vector.sum()
cand_count_vector /= cand_count_vector.sum()
distance_matrix = distance_matrix.astype(np.double)
(wmd, flow) = emd_with_flow(ref_count_vector, cand_count_vector,
distance_matrix)
return wmd, {}
# adding penalty
ratio = fragmentation(ref_list, cand_list, common_vectorizer, flow)
if ratio > 1:
ratio = 1
penalty = delta * ratio
# missing words penalty
missingwords = 0
for w in cand_list:
if w not in wvvecs:
missingwords += 1
missingratio = missingwords / len(cand_list)
missing_penalty = alpha * missingratio
penalty += missing_penalty
wmd += penalty
return wmd, missing
def compute_loss(self, results_S, results_T):
losses_dict = dict()
total_loss = 0
if 'logits' in results_T and 'logits' in results_S:
logits_list_T = results_T['logits'] # list of tensor
logits_list_S = results_S['logits'] # list of tensor
total_kd_loss = 0
if 'logits_mask' in results_S:
masks_list_S = results_S['logits_mask']
logits_list_S = select_logits_with_mask(
logits_list_S, masks_list_S) #(mask_sum, num_of_class)
if 'logits_mask' in results_T:
masks_list_T = results_T['logits_mask']
logits_list_T = select_logits_with_mask(
logits_list_T, masks_list_T) #(mask_sum, num_of_class)
for l_T, l_S in zip(logits_list_T, logits_list_S):
if self.d_config.temperature_scheduler is not None:
temperature = self.d_config.temperature_scheduler(
l_S, l_T, self.d_config.temperature)
else:
temperature = self.d_config.temperature
total_kd_loss += self.kd_loss(l_S, l_T, temperature)
total_loss += total_kd_loss * self.d_config.kd_loss_weight
losses_dict['unweighted_kd_loss'] = total_kd_loss
inters_T = {
feature: results_T.get(feature, [])
for feature in FEATURES
}
inters_S = {
feature: results_S.get(feature, [])
for feature in FEATURES
}
inputs_mask_T = results_T.get('inputs_mask', None)
inputs_mask_S = results_S.get('inputs_mask', None)
#hidden states and embedding
feature = self.emd_feature
emd_loss_weight = self.emd_loss_weight
loss_type = self.emd_loss_type
match_loss = MATCH_LOSS_MAP[loss_type]
feature_maps_S = inters_S[feature][1:] # list of features
feature_maps_T = inters_T[feature][1:] # list of features
embeddings_S = inters_S[feature][0]
embeddings_T = inters_T[feature][0]
assert isinstance(feature_maps_S, (tuple, list))
assert isinstance(feature_maps_T, (tuple, list))
assert isinstance(feature_maps_S[0], torch.Tensor)
assert isinstance(feature_maps_T[0], torch.Tensor)
assert len(feature_maps_S) == self.layer_num_S - 1
assert len(feature_maps_T) == self.layer_num_T - 1
if len(self.projs) > 0:
assert len(self.projs) == self.layer_num_S
embeddings_S = self.projs[0](embeddings_S)
feature_maps_S = [
proj(s) for proj, s in zip(self.projs[1:], feature_maps_S)
]
feature_num_S = len(feature_maps_S)
feature_num_T = len(feature_maps_T)
feature_num_A = feature_num_S + feature_num_T
distance_matrix = torch.zeros([feature_num_A,
feature_num_A]).to(feature_maps_S[0])
for s in range(feature_num_S):
f_S = feature_maps_S[s]
for t in range(feature_num_T):
f_T = feature_maps_T[t]
distance_matrix[s][t + feature_num_S] = distance_matrix[
t + feature_num_S][s] = match_loss(f_S,
f_T,
mask=inputs_mask_S)
feature_weight_S = np.concatenate(
[self.feature_weight_S,
np.zeros(feature_num_T)])
feature_weight_T = np.concatenate(
[np.zeros(feature_num_S), self.feature_weight_T])
_, trans_matrix = emd_with_flow(
feature_weight_S, feature_weight_T,
distance_matrix.detach().cpu().numpy().astype('float64'))
trans_matrix = torch.tensor(trans_matrix).to(distance_matrix)
emd_loss = torch.sum(trans_matrix * distance_matrix)
total_loss += emd_loss * emd_loss_weight
losses_dict[f'unweighted_{feature}_{loss_type}_emd'] = emd_loss
if (self.feature_weight_S <= 0).any() or (self.feature_weight_T <=
0).any():
import sys
logger.info(f"{self.feature_weight_S}")
logger.info(f"{self.feature_weight_T}")
if np.isnan(self.feature_weight_S).any() or np.isnan(
self.feature_weight_T).any():
import sys
logger.info(f"{self.feature_weight_S}")
logger.info(f"{self.feature_weight_T}")
sys.exit()
#feature_weight_S = np.copy(self.feature_weight_S)
#feature_weight_T = np.copy(self.feature_weight_T)
#self.feature_weight_S, self.feature_weight_T = get_new_feature_weight(
# trans_matrix, distance_matrix.detach(), feature_weight_S, feature_weight_T, self.d_config.temperature)
#embedding matching
embedding_loss = match_loss(embeddings_S,
embeddings_T,
mask=inputs_mask_S)
total_loss += embedding_loss * emd_loss_weight #sharing the same weight
losses_dict[f'unweighted_embedding_{loss_type}'] = embedding_loss
if 'losses' in results_S:
total_hl_loss = 0
for loss in results_S['losses']:
# in case of multi-GPU
total_hl_loss += loss.mean()
total_loss += total_hl_loss * self.d_config.hard_label_weight
losses_dict['unweighted_hard_label_loss'] = total_hl_loss
return total_loss, losses_dict
import numpy as np
from pyemd import emd
from pyemd import emd_with_flow
from pyemd import emd_samples
s1 = 8
s2 = 8
np.random.seed(10)
a = np.random.rand(s1)
b = np.random.rand(s2)
d = np.random.rand(s1, s2)
result1 = emd(a, b, d)
result2 = emd_with_flow(a, b, d)
result3 = emd_samples(a, b)
print (result1)
print ("\n", result2)
print (result3)
def test_emd_with_flow_validate_square_distance_matrix():
first_signature = np.array([0.0, 1.0])
second_signature = np.array([5.0, 3.0])
distance_matrix = np.array([[0.0, 0.5, 3.0], [0.5, 0.0]])
with pytest.raises(ValueError):
emd_with_flow(first_signature, second_signature, distance_matrix)
os.chdir(
'/Volumes/Marius_SSD/American-Flyway/Connectivity_NAbirds/Redistribution-model'
)
# Load species seasonal abundance distributions (estimated from eBird data)
abundance_BR = np.loadtxt('Data/STEMs/seasonalAbundance_wlswar_BR.csv',
delimiter=';')
abundance_NB = np.loadtxt('Data/STEMs/seasonalAbundance_wlswar_NB.csv',
delimiter=';')
abundance_BR = abundance_BR / sum(abundance_BR)
abundance_NB = abundance_NB / sum(abundance_NB)
# Load matrix of pairwise distance between every hexagons on the grid
distanceMatrix = np.loadtxt('ideal-optimal-redistribution/distanceMatrix.csv',
delimiter=';')
# Compute optimal redistribution using the Earth Mover's Distance algorithm
EMD_results = emd_with_flow(abundance_BR, abundance_NB, distanceMatrix)
flow = 0
for i in range(0, len(distanceMatrix)):
flow = flow + sum(EMD_results[1][i])
EMD_results2 = EMD_results[0] / flow
print EMD_results2
# Save simulated migratory connectivity
np.savetxt("ORSIM-outputs/ORSIMresults_wlswar.csv",
EMD_results[1],
delimiter=',')
def wmdsimilarity(doc1, doc2, lang1, lang2, vecs, with_flow=False):
tok1 = list(processing.tokenize(lang1, doc1, include_stopwords=True))
tok2 = list(processing.tokenize(lang2, doc2, include_stopwords=True))
print(tok1, tok2)
dictionary = Dictionary(documents=[tok1, tok2])
vocab_len = len(dictionary)
if vocab_len == 1:
# Both documents are composed by a single unique token
return 0.0
# Sets for faster look-up.
docset1 = set(tok1)
docset2 = set(tok2)
print(dictionary, docset1, docset2)
# Compute distance matrix.
distance_matrix = np.zeros((vocab_len, vocab_len), dtype=np.double)
for i, t1 in dictionary.items():
for j, t2 in dictionary.items():
if t1 not in docset1 or t2 not in docset2:
continue
# Compute Euclidean distance between word vectors.
distance_matrix[i, j] = np.sqrt(
np.sum((vecs[lang1][t1] - vecs[lang2][t2])**2))
if np.sum(distance_matrix) == 0.0:
# `emd` gets stuck if the distance matrix contains only zeros.
print('The distance matrix is all zeros. Aborting (returning inf).')
return float('inf')
def nbow(document):
d = np.zeros(vocab_len, dtype=np.double)
nbow = dictionary.doc2bow(document) # Word frequencies.
doc_len = len(document)
for idx, freq in nbow:
d[idx] = freq / float(doc_len) # Normalized word frequencies.
return d
# Compute nBOW representation of documents.
d1 = nbow(tok1)
d2 = nbow(tok2)
# Compute WMD.
if with_flow:
emd = emd_with_flow(d1, d2, distance_matrix)
return {
'tokens': list(dictionary.values()),
'pdf1': list(d1),
'pdf2': list(d2),
'wmd': emd[0],
'flow': emd[1],
'dist_matrix': distance_matrix.tolist()
}
else:
return {
'tokens': list(dictionary.values),
'pdf1': list(d1),
'pdf2': list(d2),
'wmd': emd(d1, d2, distance_matrix),
'dist_matrix': distance_matrix.tolist()
}
def wmdo(wvvecs, ref, cand, missing, dim, delta, alpha):
'''
wvvecs: word vectors -- retrieved from load_wv method
ref: reference translation
cand: candidate translation
missing: missing word dictionary -- initialise as {}
dim: word vector dimension
delta: weight of fragmentation penalty
alpha: weight of missing word penalty
'''
ref_list = [w.lower() for w in word_tokenize(ref)]
cand_list = [w.lower() for w in word_tokenize(cand)]
vc = CountVectorizer().fit(ref_list + cand_list)
v_obj, v_cap = vc.transform([ref, cand])
v_obj = v_obj.toarray().ravel()
v_cap = v_cap.toarray().ravel()
# need to deal with missing words
wvoc = []
for w in vc.get_feature_names():
if w in wvvecs:
wvoc.append(wvvecs[w])
else:
if w not in missing:
missing[w] = np.zeros(dim)
wvoc.append(missing[w])
distance_matrix = cosine_distances(wvoc)
if np.sum(distance_matrix) == 0.0:
return float('inf')
v_obj = v_obj.astype(np.double)
v_cap = v_cap.astype(np.double)
v_obj /= v_obj.sum()
v_cap /= v_cap.sum()
distance_matrix = distance_matrix.astype(np.double)
(wmd, flow) = emd_with_flow(v_obj, v_cap, distance_matrix)
# adding penalty
penalty = 0
ratio = fragmentation(ref_list, cand_list, vc, flow)
if ratio > 1:
ratio = 1
penalty = delta * ratio
# missing words penalty
missingwords = 0
for w in cand_list:
if w not in wvvecs:
missingwords += 1
missingratio = missingwords / len(cand_list)
missing = alpha * missingratio
penalty += missing
wmd += penalty
return wmd
def test_larger_signatures_1():
first_signature = np.array([0.0, 1.0, 2.0])
second_signature = np.array([5.0, 3.0])
distance_matrix = np.array([[0.0, 0.5], [0.5, 0.0]])
with pytest.raises(ValueError):
emd_with_flow(first_signature, second_signature, distance_matrix)
