Python options.eval_str_list方法代碼示例

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""
        # make sure all arguments are present in older models
        base_lm_architecture(args)

        if hasattr(args, 'max_target_positions'):
            args.tokens_per_sample = args.max_target_positions

        decoder = FConvDecoder(
            dictionary=task.target_dictionary,
            embed_dim=args.decoder_embed_dim,
            convolutions=eval(args.decoder_layers),
            out_embed_dim=args.decoder_embed_dim,
            attention=eval(args.decoder_attention),
            dropout=args.dropout,
            max_positions=args.tokens_per_sample,
            share_embed=False,
            positional_embeddings=False,
            adaptive_softmax_cutoff=(
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int)
                if args.criterion == 'adaptive_loss' else None
            ),
            normalization_constant=args.normalization_constant,
        )
        return FConvLanguageModel(decoder) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""
        # make sure all arguments are present in older models
        base_lm_architecture(args)

        if hasattr(args, 'max_target_positions') and not hasattr(args, 'tokens_per_sample'):
            args.tokens_per_sample = args.max_target_positions

        decoder = FConvDecoder(
            dictionary=task.target_dictionary,
            embed_dim=args.decoder_embed_dim,
            convolutions=eval(args.decoder_layers),
            out_embed_dim=args.decoder_embed_dim,
            attention=eval(args.decoder_attention),
            dropout=args.dropout,
            max_positions=args.tokens_per_sample,
            share_embed=False,
            positional_embeddings=False,
            adaptive_softmax_cutoff=(
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int)
                if args.criterion == 'adaptive_loss' else None
            ),
            adaptive_softmax_dropout=args.adaptive_softmax_dropout,
        )
        return FConvLanguageModel(decoder) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def add_ray_args(parser):
    """Add ray and fault-tolerance related parser arguments to the parser."""
    group = parser.add_argument_group("Ray related arguments")
    group.add_argument(
        "--ray-address",
        default="auto",
        type=str,
        help="address for ray initialization")
    group.add_argument(
        "--fix-batch-size",
        default=None,
        metavar="B1,B2,...,B_N",
        type=lambda uf: options.eval_str_list(uf, type=int),
        help="fix the actual batch size (max_sentences * update_freq "
        "* n_GPUs) to be the fixed input values by adjusting update_freq "
        "accroding to actual n_GPUs; the batch size is fixed to B_i for "
        "epoch i; all epochs >N are fixed to B_N")
    return group 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""
        # make sure all arguments are present in older models
        base_lm_architecture(args)

        if hasattr(args, 'max_target_positions'):
            args.tokens_per_sample = args.max_target_positions

        decoder = FConvDecoder(
            dictionary=task.target_dictionary,
            embed_dim=args.decoder_embed_dim,
            convolutions=eval(args.decoder_layers),
            out_embed_dim=args.decoder_embed_dim,
            attention=eval(args.decoder_attention),
            dropout=args.dropout,
            max_positions=args.tokens_per_sample,
            share_embed=False,
            positional_embeddings=False,
            adaptive_softmax_cutoff=(
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int)
                if args.criterion == 'adaptive_loss' else None
            ),
            adaptive_softmax_dropout=args.adaptive_softmax_dropout,
        )
        return FConvLanguageModel(decoder) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""

        # make sure all arguments are present in older models
        base_lm_architecture(args)

        if getattr(args, 'max_source_positions', None) is None:
            args.max_source_positions = args.tokens_per_sample
        if getattr(args, 'max_target_positions', None) is None:
            args.max_target_positions = args.tokens_per_sample

        if args.character_embeddings:
            embed_tokens = CharacterTokenEmbedder(task.dictionary, eval(args.character_filters),
                                                  args.character_embedding_dim,
                                                  args.decoder_embed_dim,
                                                  args.char_embedder_highway_layers,
                                                  )
        elif args.adaptive_input:
            embed_tokens = AdaptiveInput(len(task.dictionary), task.dictionary.pad(), args.decoder_input_dim,
                                         args.adaptive_input_factor, args.decoder_embed_dim,
                                         options.eval_str_list(args.adaptive_input_cutoff, type=int))
        else:
            embed_tokens = Embedding(len(task.dictionary), args.decoder_input_dim, task.dictionary.pad())

        if args.tie_adaptive_weights:
            assert args.adaptive_input
            assert args.adaptive_input_factor == args.adaptive_softmax_factor
            assert args.adaptive_softmax_cutoff == args.adaptive_input_cutoff, '{} != {}'.format(
                args.adaptive_softmax_cutoff, args.adaptive_input_cutoff)
            assert args.decoder_input_dim == args.decoder_output_dim

        decoder = LightConvDecoder(args, task.output_dictionary, embed_tokens, no_encoder_attn=True, final_norm=False)
        return LightConvLanguageModel(decoder) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""

        # make sure all arguments are present in older models
        base_lm_architecture(args)

        if args.decoder_layers_to_keep:
            args.decoder_layers = len(args.decoder_layers_to_keep.split(","))

        if getattr(args, 'max_target_positions', None) is None:
            args.max_target_positions = getattr(args, 'tokens_per_sample', DEFAULT_MAX_TARGET_POSITIONS)

        if args.character_embeddings:
            embed_tokens = CharacterTokenEmbedder(
                task.source_dictionary, eval(args.character_filters),
                args.character_embedding_dim, args.decoder_embed_dim,
                args.char_embedder_highway_layers,
            )
        elif args.adaptive_input:
            embed_tokens = AdaptiveInput(
                len(task.source_dictionary), task.source_dictionary.pad(), args.decoder_input_dim,
                args.adaptive_input_factor, args.decoder_embed_dim,
                options.eval_str_list(args.adaptive_input_cutoff, type=int),
                args.quant_noise_pq, args.quant_noise_pq_block_size,
            )
        else:
            embed_tokens = cls.build_embedding(args, task.source_dictionary, args.decoder_input_dim)

        if args.tie_adaptive_weights:
            assert args.adaptive_input
            assert args.adaptive_input_factor == args.adaptive_softmax_factor
            assert args.adaptive_softmax_cutoff == args.adaptive_input_cutoff, '{} != {}'.format(
                args.adaptive_softmax_cutoff, args.adaptive_input_cutoff)
            assert args.decoder_input_dim == args.decoder_output_dim

        decoder = TransformerDecoder(
            args, task.target_dictionary, embed_tokens, no_encoder_attn=True,
        )
        return cls(decoder) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False, left_pad=False):
        super().__init__(dictionary)
        self.dropout = args.dropout
        self.share_input_output_embed = args.share_decoder_input_output_embed
        self.fuse_dropout_add = args.fuse_dropout_add
        self.fuse_relu_dropout = args.fuse_relu_dropout

        embed_dim = embed_tokens.embedding_dim
        padding_idx = embed_tokens.padding_idx
        self.max_target_positions = args.max_target_positions

        self.embed_tokens = embed_tokens
        self.embed_scale = math.sqrt(embed_dim)
        self.embed_positions = PositionalEmbedding(
            args.max_target_positions, embed_dim, padding_idx,
            left_pad=left_pad,
            learned=args.decoder_learned_pos,
        ) if not args.no_token_positional_embeddings else None

        self.layers = nn.ModuleList([])
        self.layers.extend([
            TransformerDecoderLayer(args, no_encoder_attn)
            for _ in range(args.decoder_layers)
        ])
        transformer_print(key=mlperf_log.MODEL_HP_NUM_HIDDEN_LAYERS, value=args.decoder_layers)
        self.adaptive_softmax = None

        if args.adaptive_softmax_cutoff is not None:
            self.adaptive_softmax = AdaptiveSoftmax(
                len(dictionary), args.decoder_embed_dim,
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int),
                dropout=args.dropout
            )
        elif not self.share_input_output_embed:
            self.embed_out = nn.Parameter(torch.Tensor(len(dictionary), embed_dim))
            nn.init.normal_(self.embed_out, mean=0, std=embed_dim ** -0.5)
        self.normalize = args.decoder_normalize_before
        if self.normalize:
           self.layer_norm = FusedLayerNorm(embed_dim) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""

        # make sure all arguments are present in older models
        base_lm_architecture(args)

        if hasattr(args, 'no_tie_adaptive_proj') and args.no_tie_adaptive_proj is False:
            # backward compatibility
            args.tie_adaptive_proj = True

        if not hasattr(args, 'max_source_positions'):
            args.max_source_positions = args.tokens_per_sample
        if not hasattr(args, 'max_target_positions'):
            args.max_target_positions = args.tokens_per_sample

        if args.character_embeddings:
            embed_tokens = CharacterTokenEmbedder(
                task.dictionary, eval(args.character_filters),
                args.character_embedding_dim, args.decoder_embed_dim,
                args.char_embedder_highway_layers,
            )
        elif args.adaptive_input:
            embed_tokens = AdaptiveInput(
                len(task.dictionary), task.dictionary.pad(), args.decoder_input_dim,
                args.adaptive_input_factor, args.decoder_embed_dim,
                options.eval_str_list(args.adaptive_input_cutoff, type=int),
            )
        else:
            embed_tokens = Embedding(len(task.dictionary), args.decoder_input_dim, task.dictionary.pad())

        if args.tie_adaptive_weights:
            assert args.adaptive_input
            assert args.adaptive_input_factor == args.adaptive_softmax_factor
            assert args.adaptive_softmax_cutoff == args.adaptive_input_cutoff, '{} != {}'.format(
                args.adaptive_softmax_cutoff, args.adaptive_input_cutoff)
            assert args.decoder_input_dim == args.decoder_output_dim

        decoder = TransformerDecoder(
            args, task.output_dictionary, embed_tokens, no_encoder_attn=True, final_norm=False,
        )
        return TransformerLanguageModel(decoder) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""

        # make sure all arguments are present in older models
        base_lm_architecture(args)

        if not hasattr(args, 'max_source_positions'):
            args.max_source_positions = args.tokens_per_sample
        if not hasattr(args, 'max_target_positions'):
            args.max_target_positions = args.tokens_per_sample

        if args.character_embeddings:
            embed_tokens = CharacterTokenEmbedder(task.dictionary, eval(args.character_filters),
                                                  args.character_embedding_dim,
                                                  args.decoder_embed_dim,
                                                  args.char_embedder_highway_layers,
                                                  )
        elif args.adaptive_input:
            embed_tokens = AdaptiveInput(len(task.dictionary), task.dictionary.pad(), args.decoder_input_dim,
                                         args.adaptive_input_factor, args.decoder_embed_dim,
                                         options.eval_str_list(args.adaptive_input_cutoff, type=int))
        else:
            embed_tokens = Embedding(len(task.dictionary), args.decoder_input_dim, task.dictionary.pad())

        if args.tie_adaptive_weights:
            assert args.adaptive_input
            assert args.adaptive_input_factor == args.adaptive_softmax_factor
            assert args.adaptive_softmax_cutoff == args.adaptive_input_cutoff, '{} != {}'.format(
                args.adaptive_softmax_cutoff, args.adaptive_input_cutoff)
            assert args.decoder_input_dim == args.decoder_output_dim

        decoder = LightConvDecoder(args, task.output_dictionary, embed_tokens, no_encoder_attn=True, final_norm=False)
        return LightConvLanguageModel(decoder) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""

        # make sure all arguments are present in older models
        base_architecture(args)

        if getattr(args, 'max_target_positions', None) is not None:
            max_target_positions = args.max_target_positions
        else:
            max_target_positions = getattr(args, 'tokens_per_sample', DEFAULT_MAX_TARGET_POSITIONS)

        def load_pretrained_embedding_from_file(embed_path, dictionary, embed_dim):
            num_embeddings = len(dictionary)
            padding_idx = dictionary.pad()
            embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx)
            embed_dict = utils.parse_embedding(embed_path)
            utils.print_embed_overlap(embed_dict, dictionary)
            return utils.load_embedding(embed_dict, dictionary, embed_tokens)

        pretrained_decoder_embed = None
        if args.decoder_embed_path:
            pretrained_decoder_embed = load_pretrained_embedding_from_file(
                args.decoder_embed_path,
                task.target_dictionary,
                args.decoder_embed_dim
            )

        if args.share_decoder_input_output_embed:
            # double check all parameters combinations are valid
            if task.source_dictionary != task.target_dictionary:
                raise ValueError('--share-decoder-input-output-embeddings requires a joint dictionary')

            if args.decoder_embed_dim != args.decoder_out_embed_dim:
                raise ValueError(
                    '--share-decoder-input-output-embeddings requires '
                    '--decoder-embed-dim to match --decoder-out-embed-dim'
                    )

        decoder = LSTMDecoder(
            dictionary=task.dictionary,
            embed_dim=args.decoder_embed_dim,
            hidden_size=args.decoder_hidden_size,
            out_embed_dim=args.decoder_out_embed_dim,
            num_layers=args.decoder_layers,
            dropout_in=args.decoder_dropout_in,
            dropout_out=args.decoder_dropout_out,
            attention=False,  # decoder-only language model doesn't support attention
            encoder_output_units=0,
            pretrained_embed=pretrained_decoder_embed,
            share_input_output_embed=args.share_decoder_input_output_embed,
            adaptive_softmax_cutoff=(
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int)
                if args.criterion == 'adaptive_loss' else None
            ),
            max_target_positions=max_target_positions,
            residuals=args.residuals
        )

        return cls(decoder) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False, final_norm=True):
        super().__init__(dictionary)
        self.dropout = args.dropout
        self.share_input_output_embed = args.share_decoder_input_output_embed

        input_embed_dim = embed_tokens.embedding_dim
        embed_dim = args.decoder_embed_dim
        output_embed_dim = args.decoder_output_dim

        padding_idx = embed_tokens.padding_idx
        self.max_target_positions = args.max_target_positions

        self.embed_tokens = embed_tokens
        self.embed_scale = math.sqrt(embed_dim)  # todo: try with input_embed_dim

        self.project_in_dim = Linear(input_embed_dim, embed_dim, bias=False) if embed_dim != input_embed_dim else None

        self.embed_positions = PositionalEmbedding(
            args.max_target_positions, embed_dim, padding_idx,
            learned=args.decoder_learned_pos,
        ) if not args.no_token_positional_embeddings else None

        self.layers = nn.ModuleList([])
        self.layers.extend([
            LightConvDecoderLayer(args, no_encoder_attn, kernel_size=args.decoder_kernel_size_list[i])
            for i in range(args.decoder_layers)
        ])

        self.adaptive_softmax = None

        self.project_out_dim = Linear(embed_dim, output_embed_dim, bias=False) \
            if embed_dim != output_embed_dim and not args.tie_adaptive_weights else None

        if args.adaptive_softmax_cutoff is not None:
            self.adaptive_softmax = AdaptiveSoftmax(
                len(dictionary),
                output_embed_dim,
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int),
                dropout=args.adaptive_softmax_dropout,
                adaptive_inputs=embed_tokens if args.tie_adaptive_weights else None,
                factor=args.adaptive_softmax_factor,
                tie_proj=args.tie_adaptive_proj,
            )
        elif not self.share_input_output_embed:
            self.embed_out = nn.Parameter(torch.Tensor(len(dictionary), output_embed_dim))
            nn.init.normal_(self.embed_out, mean=0, std=output_embed_dim ** -0.5)
        self.register_buffer('version', torch.Tensor([2]))
        self.normalize = args.decoder_normalize_before and final_norm
        if self.normalize:
            self.layer_norm = LayerNorm(embed_dim) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False, left_pad=False, final_norm=True):
        super().__init__(dictionary)
        self.dropout = args.dropout
        self.share_input_output_embed = args.share_decoder_input_output_embed

        input_embed_dim = embed_tokens.embedding_dim
        embed_dim = args.decoder_embed_dim
        output_embed_dim = args.decoder_output_dim

        padding_idx = embed_tokens.padding_idx
        self.max_target_positions = args.max_target_positions

        self.embed_tokens = embed_tokens
        self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim

        self.project_in_dim = Linear(input_embed_dim, embed_dim, bias=False,
                                     uniform=False) if embed_dim != input_embed_dim else None

        self.embed_positions = PositionalEmbedding(
            args.max_target_positions, embed_dim, padding_idx,
            left_pad=left_pad,
            learned=args.decoder_learned_pos,
        ) if not args.no_token_positional_embeddings else None

        self.layers = nn.ModuleList([])
        self.layers.extend([
            TransformerDecoderLayer(args, no_encoder_attn)
            for _ in range(args.decoder_layers)
        ])

        self.adaptive_softmax = None

        self.project_out_dim = Linear(embed_dim, output_embed_dim,
                              bias=False, uniform=False) if embed_dim != output_embed_dim else None

        if args.adaptive_softmax_cutoff is not None:
            self.adaptive_softmax = AdaptiveSoftmax(
                len(dictionary), output_embed_dim,
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int),
                dropout=args.adaptive_softmax_dropout,
            )
        elif not self.share_input_output_embed:
            self.embed_out = nn.Parameter(torch.Tensor(len(dictionary), output_embed_dim))
            nn.init.normal_(self.embed_out, mean=0, std=output_embed_dim ** -0.5)
        self.register_buffer('version', torch.Tensor([2]))
        self.normalize = args.decoder_normalize_before and final_norm
        if self.normalize:
           self.layer_norm = LayerNorm(embed_dim) 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False, left_pad=False, final_norm=True):
        super().__init__(dictionary)
        self.dropout = args.dropout
        self.share_input_output_embed = args.share_decoder_input_output_embed

        input_embed_dim = embed_tokens.embedding_dim
        embed_dim = args.decoder_embed_dim
        output_embed_dim = args.decoder_output_dim

        padding_idx = embed_tokens.padding_idx
        self.max_target_positions = args.max_target_positions

        self.embed_tokens = embed_tokens
        self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim

        self.project_in_dim = Linear(input_embed_dim, embed_dim, bias=False,
                                     uniform=False) if embed_dim != input_embed_dim else None

        self.embed_positions = PositionalEmbedding(
            args.max_target_positions, embed_dim, padding_idx,
            left_pad=left_pad,
            learned=args.decoder_learned_pos,
        ) if not args.no_token_positional_embeddings else None

        self.layers = nn.ModuleList([])
        self.layers.extend([
            TransformerZeroDecoderLayer(args, no_encoder_attn)
            for _ in range(args.decoder_layers)
        ])

        self.adaptive_softmax = None

        self.project_out_dim = Linear(embed_dim, output_embed_dim,
                              bias=False, uniform=False) if embed_dim != output_embed_dim else None

        if args.adaptive_softmax_cutoff is not None:
            self.adaptive_softmax = AdaptiveSoftmax(
                len(dictionary), output_embed_dim,
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int),
                dropout=args.adaptive_softmax_dropout,
            )
        elif not self.share_input_output_embed:
            self.embed_out = nn.Parameter(torch.Tensor(len(dictionary), output_embed_dim))
            self.embed_out.data.fill_(0)
            # nn.init.normal_(self.embed_out, mean=0, std=output_embed_dim ** -0.5)
        self.register_buffer('version', torch.Tensor([2]))
        self.normalize = args.decoder_normalize_before and final_norm
        # if self.normalize:
        #    self.layer_norm = LayerNorm(embed_dim)
        self.shift0 = LayerShift()
        # self.scale0 = LayerScale() 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False, left_pad=False):
        super().__init__(dictionary)
        self.dropout = args.dropout
        self.share_input_output_embed = args.share_decoder_input_output_embed

        input_embed_dim = embed_tokens.embedding_dim
        embed_dim = args.decoder_embed_dim
        output_embed_dim = args.decoder_output_dim

        padding_idx = embed_tokens.padding_idx
        self.max_target_positions = args.max_target_positions

        self.embed_tokens = embed_tokens
        self.embed_scale = math.sqrt(embed_dim)  # todo: try with input_embed_dim

        self.project_in_dim = Linear(input_embed_dim, embed_dim, bias=False) if embed_dim != input_embed_dim else None

        self.embed_positions = PositionalEmbedding(
            args.max_target_positions, embed_dim, padding_idx,
            left_pad=left_pad,
            learned=args.decoder_learned_pos,
        ) if not args.no_token_positional_embeddings else None

        self.layers = nn.ModuleList([])
        self.layers.extend([
            TransformerDecoderLayer(args, no_encoder_attn)
            for _ in range(args.decoder_layers)
        ])

        self.adaptive_softmax = None

        self.project_out_dim = Linear(embed_dim, output_embed_dim, bias=False) \
            if embed_dim != output_embed_dim and not args.tie_adaptive_weights else None

        if args.adaptive_softmax_cutoff is not None:
            self.adaptive_softmax = AdaptiveSoftmax(
                len(dictionary),
                output_embed_dim,
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int),
                dropout=args.adaptive_softmax_dropout,
                adaptive_inputs=embed_tokens if args.tie_adaptive_weights else None,
                factor=args.adaptive_softmax_factor,
                tie_proj=args.tie_adaptive_proj,
            )
        elif not self.share_input_output_embed:
            self.embed_out = nn.Parameter(torch.Tensor(len(dictionary), output_embed_dim))
            nn.init.normal_(self.embed_out, mean=0, std=output_embed_dim ** -0.5)
        self.register_buffer('version', torch.Tensor([2]))
        if args.decoder_normalize_before and not getattr(args, 'no_decoder_final_norm', False):
            self.layer_norm = LayerNorm(embed_dim)
        else:
            self.layer_norm = None 

# 需要導入模塊: from fairseq import options [as 別名]
# 或者: from fairseq.options import eval_str_list [as 別名]
def build_model(cls, args, task):
        """Build a new model instance."""

        # make sure all arguments are present in older models
        base_architecture(args)

        if getattr(args, 'max_target_positions', None) is not None:
            max_target_positions = args.max_target_positions
        else:
            max_target_positions = getattr(args, 'tokens_per_sample', DEFAULT_MAX_TARGET_POSITIONS)

        def load_pretrained_embedding_from_file(embed_path, dictionary, embed_dim):
            num_embeddings = len(dictionary)
            padding_idx = dictionary.pad()
            embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx)
            embed_dict = utils.parse_embedding(embed_path)
            utils.print_embed_overlap(embed_dict, dictionary)
            return utils.load_embedding(embed_dict, dictionary, embed_tokens)

        pretrained_decoder_embed = None
        if args.decoder_embed_path:
            pretrained_decoder_embed = load_pretrained_embedding_from_file(
                args.decoder_embed_path,
                task.target_dictionary,
                args.decoder_embed_dim
            )

        if args.share_decoder_input_output_embed:
            # double check all parameters combinations are valid
            if task.source_dictionary != task.target_dictionary:
                raise ValueError('--share-decoder-input-output-embeddings requires a joint dictionary')

            if args.decoder_embed_dim != args.decoder_out_embed_dim:
                raise ValueError(
                    '--share-decoder-input-output-embeddings requires '
                    '--decoder-embed-dim to match --decoder-out-embed-dim'
                    )

        decoder = LSTMDecoder(
            dictionary=task.dictionary,
            embed_dim=args.decoder_embed_dim,
            hidden_size=args.decoder_hidden_size,
            out_embed_dim=args.decoder_out_embed_dim,
            num_layers=args.decoder_layers,
            dropout_in=args.decoder_dropout_in,
            dropout_out=args.decoder_dropout_out,
            attention=options.eval_bool(args.decoder_attention),
            encoder_output_units=0,
            pretrained_embed=pretrained_decoder_embed,
            share_input_output_embed=args.share_decoder_input_output_embed,
            adaptive_softmax_cutoff=(
                options.eval_str_list(args.adaptive_softmax_cutoff, type=int)
                if args.criterion == 'adaptive_loss' else None
            ),
            max_target_positions=max_target_positions
        )

        return cls(decoder)