Add tests for calibrator

genienlp/calibrate.py

@@ -196,6 +196,7 @@ def parse_argv(parser):
     parser.add_argument('--seed', type=int, default=123, help='Random seed to use for reproducibility')
     parser.add_argument('--save', type=str, help='The directory to save the calibrator model and plots after training')
     parser.add_argument('--plot', action='store_true', help='If True, will plot metrics and save them. Requires Matplotlib installation.')
+    parser.add_argument('--testing', action='store_true', help='If True, will change labels so that not all of them are equal. This is only used for testing purposes.')
 
 
 class ConfidenceEstimator():
@@ -492,17 +493,20 @@ def main(args):
                     ]:
         if name.startswith('raw'):
             estimator_class = RawConfidenceEstimator
-            mc_dropout = False
-            mc_dropout_num = 0
         else:
             estimator_class = TreeConfidenceEstimator
-            mc_dropout = train_confidences[0][0].mc_dropout
-            mc_dropout_num = train_confidences[0][0].mc_dropout_num
+        mc_dropout = train_confidences[0][0].mc_dropout
+        mc_dropout_num = train_confidences[0][0].mc_dropout_num
         estimator = estimator_class(name=name, featurizers=f, eval_metric=args.eval_metric, mc_dropout=mc_dropout, mc_dropout_num=mc_dropout_num)
         logger.info('name = %s', name)
         
         train_features, train_labels = estimator.convert_to_dataset(train_confidences, train=True)
         dev_features, dev_labels = estimator.convert_to_dataset(dev_confidences, train=False)
+        if args.testing:
+            if train_labels.all() or (~train_labels).all():
+                train_labels[0] = ~train_labels[0]
+            if dev_labels.all() or (~dev_labels).all():
+                dev_labels[0] = ~dev_labels[0]
         estimator.train_and_validate(train_features, train_labels, dev_features, dev_labels)
         precision, recall, pass_rate, accuracies, thresholds = estimator.evaluate(dev_features, dev_labels)
         if args.plot:

genienlp/server.py

@@ -118,7 +118,7 @@ class Server:
                 output = generate_with_model(self.model, [batch], self.numericalizer, task, self.args,
                                                   output_predictions_only=True,
                                                   confidence_estimator=self.confidence_estimator)
-                response = json.dumps(dict(id=request['id'], answer=output.predictions[0][0], score=output.confidence_scores[0]))
+                response = json.dumps(dict(id=request['id'], answer=output.predictions[0][0], score=float(output.confidence_scores[0])))
             else:
                 output = generate_with_model(self.model, [batch], self.numericalizer, task, self.args,
                                                   output_predictions_only=True)

tests/test.sh

@@ -63,6 +63,40 @@ do
     i=$((i+1))
 done
 
+# test calibration
+for hparams in \
+      "--model TransformerSeq2Seq --pretrained_model sshleifer/bart-tiny-random" ;
+do
+
+    # train
+    pipenv run python3 -m genienlp train --train_tasks almond --train_batch_tokens 100 --val_batch_size 100 --train_iterations 6 --preserve_case --save_every 2 --log_every 2 --val_every 2 --save $workdir/model_$i --data $SRCDIR/dataset/  $hparams --exist_ok --skip_cache --embeddings $embedding_dir --no_commit
+
+    # greedy prediction
+    pipenv run python3 -m genienlp predict --tasks almond --evaluate test --path $workdir/model_$i --overwrite --eval_dir $workdir/model_$i/eval_results/ --data $SRCDIR/dataset/ --embeddings $embedding_dir --skip_cache --save_confidence_features --confidence_feature_path $workdir/model_$i/confidences.pkl --mc_dropout --mc_dropout_num 10
+
+    # check if confidence file exists
+    if test ! -f $workdir/model_$i/confidences.pkl ; then
+        echo "File not found!"
+        exit
+    fi
+
+    # calibrate
+    pipenv run python3 -m genienlp calibrate --confidence_path $workdir/model_$i/confidences.pkl --save $workdir/model_$i --testing
+
+    # check if calibrator exists
+    if test ! -f $workdir/model_$i/calibrator.pkl ; then
+        echo "File not found!"
+        exit
+    fi
+
+    echo "Testing the server mode after calibration"
+    echo '{"id": "dummy_example_1", "context": "show me .", "question": "translate to thingtalk", "answer": "now => () => notify"}' | pipenv run python3 -m genienlp server --path $workdir/model_$i --stdin
+
+    rm -rf $workdir/model_$i $workdir/model_$i_exported
+
+    i=$((i+1))
+done
+
 # test almond_multilingual task
 for hparams in \
       "--model TransformerLSTM --pretrained_model bert-base-multilingual-cased --trainable_decoder_embeddings=50" \