def get_embeddings(mdl, fm, ofn, embedding_layer=2, fn=None, vob=None):
'''
extract embeddings
:param mdl : the model
:fm : feature manager
:param fn: file containing words that need to be represented as embeddings/ Word,Role separated by a space
:return: the new file containing the embeddings of the verbs
'''
if fn is not None:
vobs = set()
f = open(fn, "r")
for l in f.readlines():
tmps = l.split(" ")
for tmp in tmps:
if tmp != "":
tmpps = tmp.split(",")
if len(tmpps) == 2:
vobs.add((tmpps[0], tmpps[1]))
vobs = list(vobs)
else:
if vob is not None:
vobs = vob
vv = []
for v in vobs: # vobs must be a pair of word and label
if v[0] + "_" + v[1] in fm.input_key_map:
vv.append(v[0] + "_" + v[1])
vobs = vv
print(vobs)
X = [[v] for v in vobs]
Y = [["EOS"] for i in range(len(X))]
X = [[fm.input_key_map[x] + 1 for x in XX] for XX in X]
Y = [[fm.input_key_map[x] + 1 for x in XX] for XX in Y]
x, x_mask, y, y_mask = preprare_seq_seq_data(X, Y)
x, y, mask_x, mask_y, _, _, _, _ = mdl.standardize_data(
x, y, x_mask, y_mask, None, None, None, None)
rs = mdl.get_output_layer(embedding_layer, x, mask_x)
f = open(ofn, "w")
for i in range(len(vobs)):
w = vobs[i]
em = rs[0][i]
f.write(w + " ")
for e in em:
f.write(str(e))
f.write(" ")
f.write("\n")
f.close()
def get_verb_embeddings(mdl, fm, ofn, embedding_layer=3, fn=None, vob=None):
'''
extract the verb embeddings
:param mdl : the model
:fm : feature manager
:param fn: file containing the verbs, verbs are separated by a space
:ofn : output file
:return: the new file containing the embeddings of the verbs
'''
if fn is not None:
vobs = set()
f = open(fn, "r")
for l in f.readlines():
tmps = l.split(" ")
for tmp in tmps:
if tmp != "":
vobs.add(tmp)
vobs = list(vobs)
else:
if vob is not None:
vobs = vob
vv = []
for v in vobs:
if v + "_" + "PRED" in fm.input_key_map:
vv.append(v)
vobs = vv
X = [[v + "_" + "PRED"] for v in vobs]
Y = [["EOS"] for i in range(len(X))]
X = [[fm.input_key_map[x] + 1 for x in XX] for XX in X]
Y = [[fm.input_key_map[x] + 1 for x in XX] for XX in Y]
x, x_mask, y, y_mask = preprare_seq_seq_data(X, Y)
x, y, mask_x, mask_y, _, _, _, _ = mdl.standardize_data(
x, y, x_mask, y_mask, None, None, None, None)
rs = mdl.get_output_layer(embedding_layer, x, mask_x)
f = open(ofn, "w")
for i in range(len(vobs)):
w = vobs[i]
em = rs[0][i]
f.write(w + " ")
for e in em:
f.write(str(e))
f.write(" ")
f.write("\n")
f.close()
def get_scores_all(mdl, fm, X, X_new, num_select=10):
#f = open(output, "w")
X1 = []
# x = X[i], we add new values to the end of x
print(X_new)
for j in range(len(X_new)):
for k in range(len(X_new[j])):
xx = [xxx for xxx in X[j]]
xx.append(X_new[j][k][0] + "_" + X_new[j][k][1])
X1.append(xx)
print(X1)
X = [[fm.input_key_map[x] for x in XX] for XX in X1]
x, x_mask = preprare_seq_seq_data(X)
x, _, mask_x, _, _, _, _, _ = mdl.standardize_data(x, None, x_mask, None,
None, None, None, None)
score_pos = mdl.get_output_layer(-1, x, mask_x)
score_pos = score_pos.swapaxes(0, 1)
score_pos = score_pos[:, -1]
x = T.matrix("score")
sort_f = th.function([x], T.argsort(x))
sorted_values = sort_f(score_pos)
sorted_values = sorted_values
rs = []
rs_scores = []
my_scores = []
for i in range(sorted_values.shape[0]):
#f.write(to_string(X1[i]) + " ")
ss = []
for j in range(1, sorted_values.shape[1] + 1):
val = sorted_values[i][sorted_values.shape[1] - j]
#val_map = fm.fY.map_inversed[val-1]
score = score_pos[i][val]
#f.write(str(val) + ":" + str(score) + " ")
ss.append((val, score))
#f.write("\n")
my_scores.append((to_string(X1[i]), ss))
vals = []
c = 0
for t in range(sorted_values.shape[1] - 1, -1, -1):
if c == num_select:
break
v = sorted_values[i][t]
if fm.get_key(v) != "EOS":
vals.append(v)
c += 1
#vals = sorted_values[i][sorted_values.shape[1]-num_select:sorted_values.shape[1]]
val_maps = [fm.get_key(v).split("_")
for v in list(vals)] #if fm.fY.map_inversed[v-1]!="EOS" ]
scores = [score_pos[i][v]
for v in list(vals)] # if fm.fY.map_inversed[v-1]!="EOS"]
rs.append(val_maps)
rs_scores.append(scores)
print(rs)
print(rs_scores)
print(X1)
print(my_scores)
return rs, rs_scores, X1, my_scores
def get_scores_all(mdl, fm1, fm2,fm3, X1, X2, X_new1, X_new2, num_select = 10):
X11 = []
X21 = []
# x = X[i], we add new values to the end of x
for j in range(len(X_new1)):
for k in range(len(X_new1[j])):
xx =[ xxx for xxx in X1[j]]
xx.append(X_new1[j][k] )
X11.append(xx)
for j in range(len(X_new2)):
for k in range(len(X_new2[j])):
xx =[ xxx for xxx in X2[j]]
xx.append(X_new2[j][k] )
X21.append(xx)
X1 = [[fm1.input_key_map[x] for x in XX] for XX in X11 ]
X2 = [[fm2.input_key_map[x] for x in XX] for XX in X21 ]
x1,x_mask1= preprare_seq_seq_data(X1)
x1, _, mask_x1,_, _, _,_, _ = mdl.standardize_data(x1, None, x_mask1, None, None,None, None,None)
x2,x_mask2= preprare_seq_seq_data(X2)
x2, _, mask_x2,_, _, _,_, _ = mdl.standardize_data(x2, None, x_mask2, None, None,None, None,None)
score_pos = mdl.get_output_layer(-1, x1, x2, mask_x1)
score_pos=score_pos.swapaxes(0,1)
score_pos = score_pos[:,-1]
x = T.matrix("score")
sort_f = th.function([x], T.argsort(x))
sorted_values = sort_f(score_pos)
rs1 = []
rs2 = []
rs_scores = []
my_scores = []
for i in range(sorted_values.shape[0]):
#f.write(to_string(X1[i]) + " ")
ss=[]
for j in range(1,sorted_values.shape[1]):
val = sorted_values[i][sorted_values.shape[1]-j]
#val_map = fm.fY.map_inversed[val-1]
score = score_pos[i][val]
#f.write(str(val) + ":" + str(score) + " ")
ss.append((val,score))
#f.write("\n")
my_scores.append(("_", ss))
vals = []
c = 0
for t in range(sorted_values.shape[1]-1, -1, -1):
if c == num_select:
break
v = sorted_values[i][t]
if fm3.get_key(v)!="EOS_EOS" :
tm = fm3.get_key(v).split("_")
if tm[0] in fm1.input_key_map and tm[1] in fm2.input_key_map:
vals.append(v)
c+=1
#vals = sorted_values[i][sorted_values.shape[1]-num_select:sorted_values.shape[1]]
vals1=[]
vals2 = []
#val_maps = [fm1.fY.map_inversed[v-1].split("_") for v in list(vals) ]#if fm.fY.map_inversed[v-1]!="EOS" ]
scores = [score_pos[i][v] for v in list(vals)]# if fm.fY.map_inversed[v-1]!="EOS"]
for v in list(vals):
tm = fm3.get_key(v).split("_")
vals1.append(tm[0])
vals2.append(tm[1])
rs1.append(vals1)
rs2.append(vals2)
rs_scores.append(scores)
return (rs1,rs2), rs_scores, X11, X21, my_scores
def get_embeddings(mdl, fm1, fm2, ofn, embedding_layer=2, fn = None, vob = None):
'''
extract the verb embeddings
:param mdl : the model
:fm : feature manager
:param fn: file containing the verbs, verbs are separated by a space
:return: the new file containing the embeddings of the verbs
'''
if fn is not None:
vobs = set()
f = open(fn, "r")
for l in f.readlines():
tmps = l.split(" ")
for tmp in tmps:
if tmp != "":
tmps1=tmp.split(",")
if len(tmps1) ==2:
vobs.add((tmps1[0], tmps1[1]))
vobs = list(vobs)
else:
if vob is not None:
vobs = vob
vv = []
for v in vobs:
if v[0] in fm1.f.map.keys() and v[1] in fm2.f.map.keys():
vv.append(v)
vobs = vv
X1 = [ ["EOS", v[0]] for v in vobs]
X2 = [ ["EOS", v[1]] for v in vobs]
Y1 = [["EOS","EOS"] for i in range(len(X1))]
Y2 = [["EOS","EOS"] for i in range(len(X1))]
X1 = [[fm1.f.map[fm1.f.getFeatureValue(x)] +1 for x in XX] for XX in X1 ]
Y1 = [[fm1.fY.map[fm1.fY.getFeatureValue(x)] + 1 for x in XX] for XX in Y1]
X2 = [[fm2.f.map[fm2.f.getFeatureValue(x)] +1 for x in XX] for XX in X2 ]
Y2 = [[fm2.fY.map[fm2.fY.getFeatureValue(x)] + 1 for x in XX] for XX in Y2]
x1, mask_x, y1, mask_y =preprare_seq_seq_data (X1, Y1)
x2, mask_x, y2, mask_y = preprare_seq_seq_data(X2, Y2)
x1, y1, mask_x,mask_y, _, _,_, _ = mdl.standardize_data(x1, y1, mask_x, mask_y, None,None, None,None)
x2, y2, _,_, _, _,_, _ = mdl.standardize_data(x2, y2, None, None, None,None, None,None)
print (x1)
print (x2)
print (mask_x)
rs = mdl.get_output_layer(embedding_layer, x1, x2, mask_x)
print (rs.shape)
f = open (ofn, "w")
for i in range(len(vobs)):
w = vobs[i]
em = rs[0][i]
f.write(w + " ")
for e in em:
f.write(str(e))
f.write(" ")
f.write("\n")
f.close()
def get_verb_embeddings(mdl, map_x1, map_x2, map_y1, ofn, embedding_layer=2, fn = None, vob = None):
'''
extract the verb embeddings
:param mdl : the model
:fm : feature manager
:param fn: file containing the verbs, verbs are separated by a space
:return: the new file containing the embeddings of the verbs
'''
print (map_x1.input_key_map)
if fn is not None:
vobs = set()
f = open(fn, "r")
for l in f.readlines():
tmps = l.split(" ")
for tmp in tmps:
if tmp != "":
vobs.add(tmp)
vobs = list(vobs)
else:
if vob is not None:
vobs = vob
print (vobs)
vv = []
print (vobs)
for v in vobs:
if v in map_x1.input_key_map:
vv.append(v)
vobs = vv
print (vobs)
X1 = [ [ v ] for v in vobs]
X2 = [ ["PRED"] for v in vobs]
Y1 = [["EOS_EOS"] for i in range(len(X1))]
X1 = [[map_x1.input_key_map[x] for x in XX] for XX in X1 ]
X2 = [[map_x2.input_key_map[x] for x in XX] for XX in X2 ]
Y1 = [[map_y1.input_key_map[x] for x in XX] for XX in Y1]
x1, mask_x, y, mask_y = preprare_seq_seq_data(X1, Y1)
x2, mask_x = preprare_seq_seq_data(X2, None)
x1, y, mask_x, mask_y, _, _, _, _ = mdl.standardize_data(x1, y, mask_x, mask_y, None, None, None, None)
x2, _, _, _, _, _, _, _ = mdl.standardize_data(x2, None, None, None, None, None, None, None)
rs = mdl.get_output_layer(embedding_layer, x1, x2, mask_x)
print (rs.shape)
f = open (ofn, "w")
for i in range(len(vobs)):
w = vobs[i]
em = rs[0][i]
f.write(w + " ")
for e in em:
f.write(str(e))
f.write(" ")
f.write("\n")
f.close()