Language model embeddings. Arguments: hidden_size: hidden size vocab_size: vocabulary size max_sequence_length: maximum size of sequence. This is used for positional embedding embedding_dropout_prob: dropout probability for embeddings
| 98 | |
| 99 | |
| 100 | class Embedding(MegatronModule): |
| 101 | """Language model embeddings. |
| 102 | |
| 103 | Arguments: |
| 104 | hidden_size: hidden size |
| 105 | vocab_size: vocabulary size |
| 106 | max_sequence_length: maximum size of sequence. This |
| 107 | is used for positional embedding |
| 108 | embedding_dropout_prob: dropout probability for embeddings |
| 109 | init_method: weight initialization method |
| 110 | num_tokentypes: size of the token-type embeddings. 0 value |
| 111 | will ignore this embedding |
| 112 | """ |
| 113 | |
| 114 | def __init__( |
| 115 | self, |
| 116 | hidden_size, |
| 117 | vocab_size, |
| 118 | max_sequence_length, |
| 119 | embedding_dropout_prob, |
| 120 | init_method, |
| 121 | num_tokentypes=0, |
| 122 | ): |
| 123 | super(Embedding, self).__init__() |
| 124 | |
| 125 | args = get_args() |
| 126 | |
| 127 | self.hidden_size = hidden_size |
| 128 | self.init_method = init_method |
| 129 | self.num_tokentypes = num_tokentypes |
| 130 | self.max_sequence_length = max_sequence_length |
| 131 | |
| 132 | # Word embeddings (parallel). |
| 133 | self.word_embeddings = mpu.VocabParallelEmbedding( |
| 134 | vocab_size, self.hidden_size, init_method=self.init_method) |
| 135 | self._word_embeddings_key = 'word_embeddings' |
| 136 | |
| 137 | self.vocab_size = vocab_size |
| 138 | |
| 139 | # Position embedding (serial). |
| 140 | self.position_embeddings = torch.nn.Embedding( |
| 141 | max_sequence_length, self.hidden_size) |
| 142 | self.position_embeddings = self.position_embeddings.half() |
| 143 | self._position_embeddings_key = 'position_embeddings' |
| 144 | |
| 145 | # Initialize the position embeddings. |
| 146 | self.init_method(self.position_embeddings.weight) |
| 147 | |
| 148 | # Token type embedding. |
| 149 | # Add this as an optional field that can be added through |
| 150 | # method call so we can load a pretrain model without |
| 151 | # token types and add them as needed. |
| 152 | self._tokentype_embeddings_key = 'tokentype_embeddings' |
| 153 | if self.num_tokentypes > 0: |
| 154 | self.tokentype_embeddings = torch.nn.Embedding(self.num_tokentypes, |
| 155 | self.hidden_size) |
| 156 | # Initialize the token-type embeddings. |
| 157 | self.init_method(self.tokentype_embeddings.weight) |