fairseq/models/nat/levenshtein_transformer.py

Survived

Mutant 3060

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -33,7 +33,6 @@
 @register_model("levenshtein_transformer")
 class LevenshteinTransformerModel(FairseqNATModel):
 
-    @property
     def allow_length_beam(self):
         return False
 

Mutant 3061

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -37,7 +37,6 @@
     def allow_length_beam(self):
         return False
 
-    @staticmethod
     def add_args(parser):
         FairseqNATModel.add_args(parser)
         parser.add_argument(

Mutant 3062

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -67,7 +67,6 @@
             help='instead of argmax, use sampling to predict the tokens'
         )
 
-    @classmethod
     def build_decoder(cls, args, tgt_dict, embed_tokens):
         decoder = LevenshteinTransformerDecoder(args, tgt_dict, embed_tokens)
         if getattr(args, "apply_bert_init", False):

Mutant 3063

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -138,7 +138,7 @@
         }
 
     def forward_decoder(
-        self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs
+        self, decoder_out, encoder_out, eos_penalty=1.0, max_ratio=None, **kwargs
     ):
 
         output_tokens = decoder_out.output_tokens

Mutant 3064

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -270,7 +270,7 @@
 
 
 class LevenshteinTransformerDecoder(FairseqNATDecoder):
-    def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False):
+    def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=True):
         super().__init__(
             args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn
         )

Mutant 3065

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -365,7 +365,6 @@
 
         return x, {"attn": attn, "inner_states": inner_states}
 
-    @ensemble_decoder
     def forward_mask_ins(self, normalize, encoder_out, prev_output_tokens, **unused):
         features, extra = self.extract_features(
             prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[1], layers=self.layers_msk, **unused

Mutant 3066

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -376,7 +376,6 @@
             return F.log_softmax(decoder_out, -1), extra['attn']
         return decoder_out, extra['attn']
 
-    @ensemble_decoder
     def forward_word_ins(self, normalize, encoder_out, prev_output_tokens, **unused):
         features, extra = self.extract_features(
             prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[2], layers=self.layers, **unused

Mutant 3067

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -386,7 +386,6 @@
             return F.log_softmax(decoder_out, -1), extra['attn']
         return decoder_out, extra['attn']
 
-    @ensemble_decoder
     def forward_word_del(self, normalize, encoder_out, prev_output_tokens, **unused):
         features, extra = self.extract_features(
             prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[0], layers=self.layers_del, **unused

Mutant 3069

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -397,7 +397,7 @@
         return decoder_out, extra['attn']
 
 
-@register_model_architecture("levenshtein_transformer", "levenshtein_transformer")
+@register_model_architecture("levenshtein_transformer", "XXlevenshtein_transformerXX")
 def levenshtein_base_architecture(args):
     args.encoder_embed_path = getattr(args, "encoder_embed_path", None)
     args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)

Mutant 3070

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -396,8 +396,6 @@
             return F.log_softmax(decoder_out, -1), extra['attn']
         return decoder_out, extra['attn']
 
-
-@register_model_architecture("levenshtein_transformer", "levenshtein_transformer")
 def levenshtein_base_architecture(args):
     args.encoder_embed_path = getattr(args, "encoder_embed_path", None)
     args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)

Mutant 3072

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -444,7 +444,7 @@
 
 
 @register_model_architecture(
-    "levenshtein_transformer", "levenshtein_transformer_wmt_en_de"
+    "levenshtein_transformer", "XXlevenshtein_transformer_wmt_en_deXX"
 )
 def levenshtein_transformer_wmt_en_de(args):
     levenshtein_base_architecture(args)

Mutant 3074

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -452,7 +452,7 @@
 
 # similar parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017)
 @register_model_architecture(
-    "levenshtein_transformer", "levenshtein_transformer_vaswani_wmt_en_de_big"
+    "levenshtein_transformer", "XXlevenshtein_transformer_vaswani_wmt_en_de_bigXX"
 )
 def levenshtein_transformer_vaswani_wmt_en_de_big(args):
     args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024)

Mutant 3076

--- fairseq/models/nat/levenshtein_transformer.py
+++ fairseq/models/nat/levenshtein_transformer.py
@@ -468,7 +468,7 @@
 
 # default parameters used in tensor2tensor implementation
 @register_model_architecture(
-    "levenshtein_transformer", "levenshtein_transformer_wmt_en_de_big"
+    "levenshtein_transformer", "XXlevenshtein_transformer_wmt_en_de_bigXX"
 )
 def levenshtein_transformer_wmt_en_de_big_t2t(args):
     args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True)