fix wt_clf bug

.gitignore

@@ -1,3 +1,4 @@
 
 *.pyc
 models/
+temp.py

src/classifier.py

@@ -225,9 +225,9 @@ def clf_interact(fld):
 		print('%.4f'%score)
 
 
-def clf_eval(path):
+def clf_eval(clf_fld, path):
     # path is a tsv, last col is hyp
-    clf = load_classifier(fld)
+    clf = load_classifier(clf_fld)
     sum_score = 0
     n = 0
     for line in open(path, encoding='utf-8'):

src/decode.py

@@ -192,9 +192,7 @@ def rank_nbest(hyps, logP, logP_center, master, inp, infer_args=dict(), base_ran
 		hyps_no_ie.append((' '+hyp+' ').replace(' i . e . ,',' ').replace(' i . e. ',' ').strip())
 	hyps = hyps_no_ie[:]
 
-	wts_classifier = []
-	for clf_name in master.clf_names:
-		wts_classifier.append(infer_args.get(clf_name, 0))
+	wt_clf = infer_args.get('wt_clf', 0) / len(master.classifiers)
 	wt_rep = infer_args.get('wt_rep', 0)
 	wt_len = infer_args.get('wt_len', 0)
 	wt_center = infer_args.get('wt_center', 0)
@@ -224,7 +222,7 @@ def rank_nbest(hyps, logP, logP_center, master, inp, infer_args=dict(), base_ran
 		clf_score_ = []
 		for k in range(len(master.classifiers)):
 			s = clf_score[k][i]
-			score += wts_classifier[k] * s
+			score += wt_clf * s
 			clf_score_.append(s)
 		pq.put((-score, hyp, (logP[i], logP_center[i], logP_base[i], rep, l) + tuple(clf_score_)))
 

src/infer_args.csv

@@ -1,17 +1,14 @@
 method,latent
 beam_width,100
-n_rand,100
-r_rand,2
+n_rand,10
+r_rand,3
 softmax_temperature,1
 min_logP,-1000.
-wt_center,1.
-wt_rep,1.
-wt_len,0.5
-wt_base,1.
-base10m_vs_holmes38k/ngram,1.5
-base10m_vs_holmes38k/neural,1.5
-base10m_vs_arxiv1m/ngram,1.5
-base10m_vs_arxiv1m/neural,1.5
+wt_center,0.
+wt_rep,0.
+wt_len,0.
+wt_base,0.
+wt_clf,1.
 crit_logP,-2.5
 crit_rep,-0.3
 rep_allow,0.9