def analogy(pos1, neg1, pos2, neg2, word2idx, idx2word, W, is_printing, max_distance):
V, D = W.shape
# don't actually use pos2 in calculation, just print what's expected
for w in (pos1, neg1, pos2, neg2):
if w not in word2idx:
print("Sorry, %s not in word2idx" % w)
return max_distance
p1 = W[word2idx[pos1]]
n1 = W[word2idx[neg1]]
p2 = W[word2idx[pos2]]
n2 = W[word2idx[neg2]]
vec = p1 - n1 + n2
distances = pairwise_distances(vec.reshape(1, D), W, metric='cosine').reshape(V)
idx = distances.argsort()[:10]
# pick one that's not p1, n1, or n2
best_idx = -1
keep_out = [word2idx[w] for w in (pos1, neg1, neg2)]
for i in idx:
if i not in keep_out:
best_idx = i
break
if is_printing:
print("got: %s - %s = %s - %s" % (pos1, neg1, idx2word[idx[0]], neg2))
print("closest 10:")
for i in idx:
print(idx2word[i], distances[i])
print("dist to %s:" % pos2, cos_dist(p2, vec))
print("testing: %s - %s = %s - %s: distance = %s " % (pos1, neg1, pos2, neg2, abs(1 - cos_dist(p2, vec))))
return abs(1 - cos_dist(p2, vec))
def assign_bert_mentions(entities, doc, label_prefix = 'NER', \
max_dist = 0.10, add_unlinked_entities = True):
for t in 'io':
valid_entities = [e for e in entities if e.type == t]
valid_mentions = get_doc_spans(doc, label_prefix + '_'+ t)
valid_mentions = [m for m in valid_mentions if m.text and not m.text.isspace()]
if not valid_mentions:
print('Warning! No valid mentions for {} ({})'.format(doc.id, t))
continue
try:
entity_embs = list(encode([e.text for e in valid_entities]))
mention_embs = encode([m.text for m in valid_mentions])
except ValueError:
print(doc.id)
print(t)
print(valid_mentions)
raise
for m, m_emb in zip(valid_mentions, mention_embs):
dists = [cos_dist(m_emb, e_emb) for e_emb in entity_embs]
# explicitly sort on first element - builtin comparator breaks when dists are tied!
sorted_dists = sorted(zip(dists, valid_entities), key = itemgetter(0))
# require a minimum similarity between mention and entity
if sorted_dists[0][0] <= max_dist:
sorted_dists[0][1].mentions.append(m)
else:
if add_unlinked_entities:
# ooohhh sheeeeeit create a new entity
unlinked_e = classes.Entity(m, t)
unlinked_e.mentions.append(m)
entities.append(unlinked_e)
valid_entities.append(unlinked_e)
entity_embs.append(encode([unlinked_e.text])[0])
def analogy(pos1, neg1, pos2, neg2):
# don't actually use pos2 in calculation, just print what's expected
print("testing: %s - %s = %s - %s" % (pos1, neg1, pos2, neg2))
for w in (pos1, neg1, pos2, neg2):
if w not in word2idx:
print("Sorry, %s not in word2idx" % w)
return
p1 = W[word2idx[pos1]]
n1 = W[word2idx[neg1]]
p2 = W[word2idx[pos2]]
n2 = W[word2idx[neg2]]
vec = p1 - n1 + n2
distances = pairwise_distances(vec.reshape(1, D), W,
metric='cosine').reshape(V)
idx = distances.argsort()[:10]
# pick the best that's not p1, n1, or n2
best_idx = -1
keep_out = [word2idx[w] for w in (pos1, neg1, neg2)]
for i in idx:
if i not in keep_out:
best_idx = i
break
print("got: %s - %s = %s - %s" % (pos1, neg1, top_words[best_idx], neg2))
print("closest 10:")
for i in idx:
print(top_words[i], distances[i])
print("dist to %s:" % pos2, cos_dist(p2, vec))
def analogy(p1, n1, p2, n2, word2idx, idx2word, WordEmbeddings):
V, D = WordEmbeddings.shape
for w in (p1, n1, p2, n2):
if w not in word2idx:
print("%s not found in word2idx" % w)
return
print("testing %s - %s = %s - %s" % (p1, n1, p2, n2))
vec1 = WordEmbeddings[word2idx[p1]]
vec2 = WordEmbeddings[word2idx[n2]]
vec3 = WordEmbeddings[word2idx[p2]]
vec4 = WordEmbeddings[word2idx[n2]]
vec = vec1 - vec2 + vec4
closest_neighbours = pairwise_distances(vec.reshape(1, D),
WordEmbeddings,
metric='cosine').reshape(V)
top_ten = closest_neighbours.argsort()[:10]
best_pick = -1
keep_out = [word2idx[w] for w in (p1, n1, n2)]
for idx in top_ten:
if idx not in keep_out:
best_pick = idx
break
print("got %s - %s = %s - %s" % (p1, n1, idx2word[best_pick], n2))
print("distance to %s", p2, cos_dist(vec3, vec))
def analogy(pos1, neg1, pos2, neg2):
# don't actually use pos2 in calculation, just print what's expected
print("testing: %s - %s = %s - %s" % (pos1, neg1, pos2, neg2))
for w in (pos1, neg1, pos2, neg2):
if w not in word2idx:
print("Sorry, %s not in word2idx" % w)
return
p1 = W[word2idx[pos1]]
n1 = W[word2idx[neg1]]
p2 = W[word2idx[pos2]]
n2 = W[word2idx[neg2]]
vec = p1 - n1 + n2
distances = pairwise_distances(vec.reshape(1, D), W, metric='cosine').reshape(V)
idx = distances.argsort()[:10]
# pick the best that's not p1, n1, or n2
best_idx = -1
keep_out = [word2idx[w] for w in (pos1, neg1, neg2)]
for i in idx:
if i not in keep_out:
best_idx = i
break
print("got: %s - %s = %s - %s" % (pos1, neg1, top_words[best_idx], neg2))
print("closest 10:")
for i in idx:
print(top_words[i], distances[i])
print("dist to %s:" % pos2, cos_dist(p2, vec))
def analogy(pos1, neg1, pos2, neg2, word2idx, idx2word, W):
V, D = W.shape
print("testing: %s - %s = %s - %s" % (pos1, neg1, pos2, neg2))
for w in (pos1, neg1, pos2, neg2):
if w not in word2idx:
print("sorry, %s not in word2idx" % w)
return
p1 = W[word2idx[pos1]]
n1 = W[word2idx[neg1]]
p2 = W[word2idx[pos2]]
n2 = W[word2idx[neg2]]
vec = p1 - n1 + n2
distances = pairwise_distances(vec.reshape(1, D), W, metric='cosine').reshape(V)
idx = distances.argsort()[:10]
best_idx = -1
keep_out = [word2idx[w] for w in (pos1, neg1, neg2)]
for i in idx:
if i not in keep_out:
best_idx = i
break
print("got: %s - %s = %s - %s" % (pos1, neg1, idx2word[best_idx], neg2))
print('closest 10:')
for i in idx:
print(idx2word[i], distances[i])
print('dist to %s:' % pos2, cos_dist(p2, vec))
def cos_predict(cut_text, match_sents, match_labels, thre):
vec = bow.transform([cut_text])
vec = svd.transform(vec)
match_texts, dists = list(), list()
for sent_ind, label in zip(match_sents, match_labels):
match_texts.append(texts[sent_ind])
match_vec = sent_vec[sent_ind]
dists.append(cos_dist(vec, match_vec))
return sort(dists, match_texts, match_labels, thre, cand=5)
def weights_true_cossim(estr, X, Y, Xorig, Yorig, x_align_ref, y_align_ref,
results, **kwargs):
"""Calculates cosine-distance between estimated and true weights.
Provides outcome metrics ``x_weights_true_cossim`` and
``y_weights_true_cossim``
Parameters
----------
estr : **sklearn**-style estimator
fitted estimator
X : np.ndarray (n_samples, n_features)
dataset `X`
Y : np.ndarray (n_samples, n_features)
dataset `Y`
Xorig : ``None`` or np.ndarray (n_samples, n_orig_features)
if ``None`` set to ``X``. Allows to provide an alternative set of `X`
features for calculating loadings. I.e. an implicit assumption is that
the rows in ``X`` and ``Xorig`` correspond to the same samples
(subjects).
Yorig : ``None`` or np.ndarray (n_samples, n_orig_features)
if ``None`` set to ``Y``. Allows to provide an alternative set of `Y`
features for calculating loadings. I.e. an implicit assumption is that
the rows in ``Y`` and ``Yorig`` correspond to the same samples
(subjects).
x_align_ref : (n_features,)
the sign of `X` weights is chosen such that the cosine-distance between
fitted `X` weights and ``x_align_ref`` is positive
y_align_ref : (n_features,)
the sign of `Y` weights is chosen such that the cosine-distance between
fitted `Y` weights and ``y_align_ref`` is positive
results : xr.Dataset
containing outcome features computed so far, and is modified with
outcomes of this function
kwargs : dict
keyword arguments
"""
results['x_weights_true_cossim'] = 1 - cos_dist(
results['x_weights'].values, x_align_ref)
results['y_weights_true_cossim'] = 1 - cos_dist(
results['y_weights'].values, y_align_ref)
def get_weighted_nn(self, word):
nn_w_dists = {}
if word in self.word_vectors:
all_vectors = (self.word_vectors, )
for clue in self.cm_cluelist:
b_dist = cos_dist(self.concatenate(clue, all_vectors),
self.concatenate(word, all_vectors))
nn_w_dists[clue] = b_dist
else:
if self.configuration.verbose:
print(word, "not in word_vectors")
self.word_dists[word] = nn_w_dists
return {k: 1 - v for k, v in nn_w_dists.items()}
def get_similarities(self, data):
edge_list = []
for i, node in enumerate(data):
for j, other_node in enumerate(data):
best_node_dist = (None, 1)
qualifying_edges = []
if i != j:
dist = cos_dist(node[self.qa:], other_node[:self.qa])
if dist < self.threshold:
qualifying_edges.append((i, j, dist))
self.distances.append(dist)
return edge_list
def step(self, a):
posbefore = np.copy(self.get_body_com("torso")[:2])
self.do_simulation(a, self.frame_skip)
posafter = self.get_body_com("torso")[:2]
self.ant_pos_before = posbefore.copy()
self.ant_pos_after = posafter.copy()
# original
# forward_reward = np.sum(self.goal_direction * (posafter - posbefore))/self.dt
# ctrl_cost = .5 * np.square(a).sum()
# contact_cost = 0.5 * 1e-3 * np.sum(
# np.square(np.clip(self.sim.data.cfrc_ext, -1, 1)))
# survive_reward = 1.0
# # new try v1: cosine similarlity
# forward_reward = 1.0 - cos_dist(self.goal_direction, posafter - posbefore)
# ctrl_cost = 0.0
# contact_cost = 0.0
# survive_reward = 0.0
# new try v2: just region clipped forward cost
forward_reward = np.sum(self.goal_direction *
(posafter - posbefore)) / self.dt
ctrl_cost = 0.0
contact_cost = 0.0
survive_reward = 0.0
# clipping based on region
# for this we also have to clip the rewards from below by 0 so that the agent won't try to do weird stuff
# clip min by zero
forward_reward = max(forward_reward, 0.0)
# clip by region
if 1.0 - cos_dist(self.goal_direction, posafter) < 0.96:
forward_reward = 0.0
reward = forward_reward - ctrl_cost - contact_cost + survive_reward
state = self.state_vector()
notdone = np.isfinite(state).all() and 1.0 >= state[2] >= 0.
done = not notdone
ob = self._get_obs()
return ob, reward, done, dict(
reward_forward=forward_reward,
reward_ctrl=-ctrl_cost,
reward_contact=-contact_cost,
reward_survive=survive_reward,
goal_direction=self.goal_direction.copy(),
projected_dist=np.sum(posafter * self.goal_direction),
debug_target_dist=np.linalg.norm(posafter - self.goal_direction))
def assign_best_mention(entities, doc):
mentions = { \
'i': [m for m in get_doc_spans(doc, 'NER_i') if m.text and not m.text.isspace()],
'o': [m for m in get_doc_spans(doc, 'NER_o') if m.text and not m.text.isspace()]
}
embs = { \
'i': encode([m.text for m in mentions['i']]),
'o': encode([m.text for m in mentions['o']])
}
for e in entities:
e_emb = encode([e.text])[0]
dists = [cos_dist(e_emb, m_emb) for m_emb in embs[e.type]]
sorted_dists = sorted(zip(dists, mentions[e.type]), key = itemgetter(0))
if sorted_dists[0][0] >= 0.15:
continue
e.mentions = [sorted_dists[0][1]]
def simC(self, user_i, user_j, vector_type):
'''
:param user_i:
:param user_j:
:return: simC
'''
if vector_type == 'svd':
ui = self.vectorize_users(user_i)
uj = self.vectorize_users(user_j)
sim = 1 - cos_dist(ui, uj)
return sim
elif vector_type == 'chen':
num = 0
sum1 = 0
sum2 = 0
for it in self.items:
num += self.R[it - 1001][user_i - 1] * self.R[it -
1001][user_j - 1]
sum1 += np.power(self.R[it - 1001][user_i - 1], 2)
sum2 += np.power(self.R[it - 1001][user_j - 1], 2)
den = np.sqrt(sum1) * np.sqrt(sum2)
return num / den
def analogy(self, pos1, neg1, pos2, neg2):
N, D = self.W.shape
assert (N == self.vocab_size)
# don't actually use pos2 in calculation, just print what's expected
print("testing: %s - %s = %s - %s" % (pos1, neg1, pos2, neg2))
for w in (pos1, neg1, pos2, neg2):
if w not in self.word2idx:
print("Sorry, %s not in pre-trained word2idx" % w)
return
p1_w2v = self.wv[self.word2idx[pos1]]
n1_w2v = self.wv[self.word2idx[neg1]]
p2_w2v = self.wv[self.word2idx[pos2]]
n2_w2v = self.wv[self.word2idx[neg2]]
w2v = p1_w2v - n1_w2v + n2_w2v # (D,)
distances = pairwise_distances(w2v.reshape(1, D),
self.W,
metric='cosine').reshape(N)
idx = distances.argsort()[:10] # smaller distance, more similar
# pick one that's not p1, n1, or n2
best_idx = -1
idx_not_use = [self.word2idx[w] for w in (pos1, neg1, neg2)]
for i in idx:
if i not in idx_not_use:
best_idx = i
break
print("analogy results: %s - %s = %s - %s" %
(pos1, neg1, self.idx2word[best_idx], neg2))
print("closest top 10:")
for i in idx:
print(self.idx2word[i], distances[i])
print("cosine distance to %s:" % pos2, cos_dist(p2_w2v, w2v))
def analogy(pos1, neg1, pos2, neg2, word2index, index2word, W):
V, D = W.shape
# do not actually use pos2 in calculation just print what's excepted
print("testing: %s - %s = %s - %s" % (pos1, neg1, pos2, neg2))
for w in (pos1, neg1, pos2, neg2):
if w not in word2index:
print("Sorry, %s not in word2index" % w)
return
p1 = W[word2index[pos1]]
n1 = W[word2index[neg1]]
p2 = W[word2index[pos2]]
n2 = W[word2index[neg2]]
vec = p1 - n1 + n2
distances = pairwise_distances(vec.reshape(1, D), W,
metric='cosine').reshape(V)
index = distances.argsort()[:10]
# pick one that's not p1, n1 or n2
best_index = -1
keep_out = [word2index[w] for w in (pos1, neg1, neg2)]
print("keep out:", keep_out)
for i in index:
if i not in keep_out:
best_index = i
break
print("best index:", best_index)
print("got: %s - %s = %s - %s" %
(pos1, neg1, index2word[best_index], neg2))
print("closest 10:")
for i in index:
print(index2word[i], distances[i])
print("dist to %s:" % pos2, cos_dist(p2, vec))
print("\n")
("yc, sn + tn", yc, sn + tn),
("yc, sh + th", yc, sh + th),
("yv + yc, sn", yv + yc, sn),
("yv + yc, sh", yv + yc, sh),
("yv + yc, tn", yv + yc, tn),
("yv + yc, th", yv + yc, th),
("yv + yc, sn + sh", yv + yc, sn + sh),
("yv + yc, tn + th", yv + yc, tn + th),
("yv + yc, sn + tn", yv + yc, sn + tn),
("yv + yc, sh + th", yv + yc, sh + th),
]
scores = np.zeros((4, len(pairs)))
for i, (s, x, y) in enumerate(pairs):
scores[0][i] = norm_dist(x, y)
scores[2][i] = cos_dist(x, y)
scores[1] = scores[0] / max(scores[0])
scores[3] = scores[2] / max(scores[2])
print("Distances between normalized vectors")
for count, i in enumerate(np.argsort(scores[3]), 1):
print(f"{pairs[i][0]} ({count}/{len(pairs)})")
print("norm\t{0:.4f}\t{1:.4f}".format(scores[0][i], scores[1][i]))
print("cos\t{0:.4f}\t{1:.4f}\n".format(scores[2][i], scores[3][i]))
print(
"Ready! Go to http://projector.tensorflow.org/ and upload the files in the "
"`projector` directory to visualize the embeddings.\n")
pearson = pearsonr(scores[1], scores[3])
def herify(sarsd, dyn_vals, obs_vectorizer, FLAGS):
"""return new sarsd's that came from HER
NOTE: this requires that sarsd not be flattened yet. a bit hacky
"""
her_sarsd = []
legit_reached_goals = []
no_move_reach = []
moves = []
for ridx in range(len(sarsd['s'])):
all_s = sarsd['s'][ridx]
all_s_next = sarsd['s_next'][ridx]
phi_s_next = dyn_vals['phi_s_next'][ridx]
if FLAGS['aac']:
value_phi_s_next = dyn_vals['value_phi_s_next'][ridx]
# We are kind of keeping this in format of rollouts
for k in range(FLAGS['her_k']):
new_sarsd = {}
new_sarsd['d'] = np.zeros_like(sarsd['r'][ridx])
new_sarsd['r'] = np.zeros_like(sarsd['r'][ridx])
num_tidx = len(new_sarsd['r'])
new_sarsd['r'] += FLAGS['old_weight'] * sarsd['og_reward'][
ridx][:num_tidx]
new_s = obs_vectorizer.to_np_dict(copy.deepcopy(all_s))
new_s_next = obs_vectorizer.to_np_dict(copy.deepcopy(all_s_next))
next_obs = obs_vectorizer.to_np_dict(all_s_next)
next_array = next_obs['array']
if FLAGS['use_image'] and not FLAGS['use_embed']:
next_image = next_obs['image']
if FLAGS['use_embed']:
new_sarsd['phi_s_pi'] = sarsd['phi_s_pi'][ridx].copy()
new_sarsd['phi_s_next_pi'] = sarsd['phi_s_next_pi'][ridx].copy(
)
new_sarsd['phi_g_pi'] = np.full_like(new_sarsd['phi_s_pi'],
np.nan)
if FLAGS['aac']:
new_sarsd['phi_s_vf'] = sarsd['phi_s_vf'][ridx].copy()
new_sarsd['phi_s_next_vf'] = sarsd['phi_s_next_vf'][
ridx].copy()
new_sarsd['phi_g_vf'] = np.full_like(
new_sarsd['phi_s_vf'], np.nan)
for tidx in range(num_tidx):
# sample random new future
new_idx = np.random.randint(tidx, num_tidx)
new_goal_array = next_array[new_idx].copy()
new_s['goal_array'][tidx] = new_goal_array
new_s_next['goal_array'][tidx] = new_goal_array
# Replace observations
if FLAGS['use_embed']:
new_sarsd['phi_g_pi'][tidx] = phi_s_next[new_idx].copy()
if FLAGS['aac']:
new_sarsd['phi_g_vf'][tidx] = value_phi_s_next[
new_idx].copy()
if FLAGS['use_image'] and not FLAGS['use_embed']:
new_goal_image = next_image[new_idx].copy()
new_s['goal_image'][tidx] = new_goal_image
new_s_next['goal_image'][tidx] = new_goal_image
# Compute replaced rewards
moved = True
if FLAGS['penalize_stasis']:
# Penalize the agent if it does not move
curr_s, curr_s_next = new_s['array'][tidx], new_s_next[
'array'][tidx]
deltas = (curr_s_next - curr_s)
sum_deltas = np.sum(np.abs(deltas))
moved = sum_deltas >= FLAGS['stasis_threshold']
if moved:
new_sarsd['r'][tidx] += FLAGS['stasis_rew']
else:
new_sarsd['r'][tidx] += -FLAGS['stasis_rew']
moves.append(1 if moved else 0)
# Reward is sparse random new goal in future
if FLAGS['aac']:
# compute goal using value (GN) because it is probably a bit better
#dgoal = cos_dist(value_phi_s_next[tidx], value_phi_s_next[new_idx])
dgoal = cos_dist(
dyn_vals['value_reward_phi_s_next'][ridx][tidx],
dyn_vals['value_reward_phi_s_next'][ridx][new_idx])
else:
#goal = cos_dist(phi_s_next[tidx], phi_s_next[new_idx])
dgoal = cos_dist(
dyn_vals['reward_phi_s_next'][ridx][tidx],
dyn_vals['reward_phi_s_next'][ridx][new_idx])
if dgoal < FLAGS['goal_threshold']:
if moved:
new_sarsd['r'][tidx] += 1.0
legit_reached_goals.append(1)
no_move_reach.append(0)
else:
no_move_reach.append(1)
legit_reached_goals.append(0)
# DONE
new_sarsd['d'][tidx] = 1.0 # done if we reach goal
else:
new_sarsd['r'][tidx] += -1.0 # else -1
no_move_reach.append(0)
legit_reached_goals.append(0)
# combine new stuff
if not FLAGS['use_embed']:
new_sarsd['s'] = obs_vectorizer.to_vecs(
[ns for ns in zip(*new_s.values())])
new_sarsd['s_next'] = obs_vectorizer.to_vecs(
[ns_next for ns_next in zip(*new_s_next.values())])
new_sarsd['a'] = sarsd['a'][ridx]
her_sarsd.append(new_sarsd)
# TODO: add all new stuff to a sarsd or rollout
out_sarsd = {}
for key in her_sarsd[0].keys():
out_sarsd[key] = np.concatenate([hs[key] for hs in her_sarsd])
extra_info = {
'her_reach_frac': np.mean(legit_reached_goals),
'her_nomove_reach_frac': np.mean(no_move_reach),
'her_move_frac': np.mean(moves)
}
return out_sarsd, extra_info
def swap_rewards(rollouts, sarsd, dyn_vals, obs_vectorizer, FLAGS):
"""
Swap rewards in rollouts for rewards in new_rewards and replace information to make things consistent.
NOTE: this reduces the lengths of every rollout by 1
Also does some other reward post-processing
Args:
add_old (bool): True to combine the two rewards into a single signal
"""
new_rewards = [
np.zeros(rs.shape[0], np.float32) for rs in dyn_vals['phi_s']
]
assert len(sarsd['s']) == len(new_rewards)
new_sarsd = {
key: []
for key in list(sarsd.keys()) + ['og_reward', 'dgoal']
}
new_dyn_vals = {key: [] for key in dyn_vals.keys()}
new_rollouts = []
reached_goals = []
for ridx in range(len(sarsd['s'])):
og_rewards = np.array(sarsd['r'][ridx][:len(new_rewards[ridx])])
new_rewards[ridx] += FLAGS['old_weight'] * og_rewards
s_array, s_next_array, goal_array = obs_vectorizer.to_np_dict(
sarsd['s'][ridx])['array'], obs_vectorizer.to_np_dict(
sarsd['s_next'][ridx])['array'], obs_vectorizer.to_np_dict(
sarsd['s'][ridx])['goal_array']
if FLAGS['penalize_stasis']:
# Penalize the agent if it does not move
deltas = (s_next_array - s_array)
sum_deltas = np.sum(np.abs(deltas), axis=(-1, -2))
assert sum_deltas.shape == new_rewards[ridx].shape
stasis_mask = (sum_deltas < FLAGS['stasis_threshold'])
move_mask = np.logical_not(stasis_mask)
new_rewards[ridx][move_mask] += FLAGS['stasis_rew']
new_rewards[ridx][stasis_mask] += -FLAGS['stasis_rew']
if FLAGS['goal_conditioned'] and not FLAGS['only_stasis']:
if FLAGS['aac']:
# compute rewards using value (GN) based embedding.
# i think this would help the reward signal be a bit smoother than image
phi_s = dyn_vals['value_reward_phi_s'][ridx]
phi_s_next = dyn_vals['value_reward_phi_s_next'][ridx]
phi_g = dyn_vals['value_reward_phi_g'][ridx]
else:
phi_s = dyn_vals['reward_phi_s'][ridx]
phi_s_next = dyn_vals['reward_phi_s_next'][ridx]
phi_g = dyn_vals['reward_phi_g'][ridx]
dgoal = np.array(
[cos_dist(phi_g[i], phi_s_next[i]) for i in range(len(phi_g))])
new_sarsd['dgoal'].append(dgoal)
achieved_goal = dgoal < FLAGS['goal_threshold']
done_idx = np.argmax(achieved_goal)
# TODO: really fix this so that we have a consistent reward function. It should
# just be a function that takes in certain things so that we can reuse it.
# Everything before goal is -1 (including if we don't reach goal)
# at goal, reward = 0.0
if done_idx == 0:
if achieved_goal[0] > 0.0:
# This means get rid of rollouts if we are already at the goal
continue
else:
# This means we never reached it, so all steps get -1
new_rewards[ridx][:] += -1.0
reached_goals.append(0)
else:
# This means we reached goal at some point
# Shorten rollout to end at goal and give rewards out
new_rewards[ridx] = new_rewards[ridx][:done_idx + 1]
new_rewards[ridx][:done_idx] += -1.0
# Test here to only give good reward if they actually moved
# NOTE: this probably is not required since I think it is impossible to get here without HER, but at least this makes things consisten
if new_rewards[ridx][done_idx] > 0.0:
new_rewards[ridx][done_idx] += 1.0
reached_goals.append(1)
sarsd['r'][ridx] = new_rewards[ridx]
new_len = len(new_rewards[ridx])
# store new rewards in sarsd
sarsd['r'][ridx] = new_rewards[ridx]
# Adapt length of sarsd and dyn vals so they will match up downstream
for key in sarsd:
new_sarsd[key].append(sarsd[key][ridx][:new_len])
new_sarsd['og_reward'].append(og_rewards)
for key in new_dyn_vals:
new_dyn_vals[key].append(dyn_vals[key][ridx][:new_len])
# Fix the shape of rollouts
rollouts[ridx].rewards = new_rewards[ridx]
# correct the length of the other settings so anyrl.rollouts doesn't bark at us
rollouts[ridx].infos = rollouts[ridx].infos[:new_len]
rollouts[ridx].model_outs = rollouts[ridx].model_outs[:new_len]
rollouts[ridx].observations = rollouts[ridx].observations[:new_len]
# reset this because it does not make sense for endless envs
rollouts[ridx].prev_reward = 0.0
new_rollouts.append(rollouts[ridx])
return new_rollouts, new_sarsd, new_dyn_vals, {
'reached_frac': np.mean(reached_goals)
}
# if dist < min_dist:
# closest = i
# min_dist = dist
# set word embedding matrix
# W = (W + U) / 2
distances = pairwise_distances(vec.reshape(1, D), W,
metric='cosine').reshape(V)
idx = distances.argsort()[:10]
print("closest 10:")
for i in idx:
print(top_words[i], distances[i])
print("dist to queen:", cos_dist(W[word2idx['queen']], vec))
def analogy(pos1, neg1, pos2, neg2):
# don't actually use pos2 in calculation, just print what's expected
print("testing: %s - %s = %s - %s" % (pos1, neg1, pos2, neg2))
for w in (pos1, neg1, pos2, neg2):
if w not in word2idx:
print("Sorry, %s not in word2idx" % w)
return
p1 = W[word2idx[pos1]]
n1 = W[word2idx[neg1]]
p2 = W[word2idx[pos2]]
n2 = W[word2idx[neg2]]
# dist = cos_dist(W[i], vec)
# if dist < min_dist:
# closest = i
# min_dist = dist
# set word embedding matrix
# W = (W + U) / 2
distances = pairwise_distances(vec.reshape(1, D), W, metric='cosine').reshape(V)
idx = distances.argsort()[:10]
print("closest 10:")
for i in idx:
print(top_words[i], distances[i])
print("dist to queen:", cos_dist(W[word2idx['queen']], vec))
def analogy(pos1, neg1, pos2, neg2):
# don't actually use pos2 in calculation, just print what's expected
print("testing: %s - %s = %s - %s" % (pos1, neg1, pos2, neg2))
for w in (pos1, neg1, pos2, neg2):
if w not in word2idx:
print("Sorry, %s not in word2idx" % w)
return
p1 = W[word2idx[pos1]]
n1 = W[word2idx[neg1]]
p2 = W[word2idx[pos2]]
n2 = W[word2idx[neg2]]
def get_similarity(self, w1, w2):
if w1 not in self.wv or w2 not in self.wv: return -0.5
sim = 1.0 - cos_dist(self.wv[w1], self.wv[w2])
return sim
