diff options
Diffstat (limited to 'training/conf/network')
| -rw-r--r-- | training/conf/network/conv_transformer.yaml | 2 | ||||
| -rw-r--r-- | training/conf/network/decoder/transformer_decoder.yaml | 4 |
2 files changed, 3 insertions, 3 deletions
diff --git a/training/conf/network/conv_transformer.yaml b/training/conf/network/conv_transformer.yaml index 7d57a2d..a97157d 100644 --- a/training/conf/network/conv_transformer.yaml +++ b/training/conf/network/conv_transformer.yaml @@ -4,7 +4,7 @@ defaults: _target_: text_recognizer.networks.conv_transformer.ConvTransformer input_dims: [1, 576, 640] -hidden_dim: 256 +hidden_dim: 96 dropout_rate: 0.2 max_output_len: 451 num_classes: 1006 diff --git a/training/conf/network/decoder/transformer_decoder.yaml b/training/conf/network/decoder/transformer_decoder.yaml index 3122de1..90b9d8a 100644 --- a/training/conf/network/decoder/transformer_decoder.yaml +++ b/training/conf/network/decoder/transformer_decoder.yaml @@ -2,12 +2,12 @@ defaults: - rotary_emb: null _target_: text_recognizer.networks.transformer.Decoder -dim: 256 +dim: 96 depth: 2 num_heads: 8 attn_fn: text_recognizer.networks.transformer.attention.Attention attn_kwargs: - dim_head: 64 + dim_head: 16 dropout_rate: 0.2 norm_fn: torch.nn.LayerNorm ff_fn: text_recognizer.networks.transformer.mlp.FeedForward |