diff options
| author | Gustaf Rydholm <gustaf.rydholm@gmail.com> | 2021-08-04 05:03:51 +0200 |
|---|---|---|
| committer | Gustaf Rydholm <gustaf.rydholm@gmail.com> | 2021-08-04 05:03:51 +0200 |
| commit | d3afa310f77f47553586eeee58e3d3345a754e2c (patch) | |
| tree | 08b7de1daf2550852d0a1e4d4d75202f14bb03d4 /training/conf/network | |
| parent | 65d5f6c694e73792e40ed693a1381a792da8d277 (diff) | |
New VQVAE
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 | ||||
| -rw-r--r-- | training/conf/network/vqvae.yaml | 21 |
3 files changed, 11 insertions, 16 deletions
diff --git a/training/conf/network/conv_transformer.yaml b/training/conf/network/conv_transformer.yaml index f76e892..d3a3b0f 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: 96 +hidden_dim: 128 dropout_rate: 0.2 num_classes: 1006 pad_index: 1002 diff --git a/training/conf/network/decoder/transformer_decoder.yaml b/training/conf/network/decoder/transformer_decoder.yaml index eb80f64..c326c04 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: 96 +dim: 128 depth: 2 num_heads: 8 attn_fn: text_recognizer.networks.transformer.attention.Attention attn_kwargs: - dim_head: 16 + dim_head: 64 dropout_rate: 0.2 norm_fn: torch.nn.LayerNorm ff_fn: text_recognizer.networks.transformer.mlp.FeedForward diff --git a/training/conf/network/vqvae.yaml b/training/conf/network/vqvae.yaml index 22eebf8..5a5c066 100644 --- a/training/conf/network/vqvae.yaml +++ b/training/conf/network/vqvae.yaml @@ -1,13 +1,8 @@ -type: VQVAE -args: - in_channels: 1 - channels: [64, 96] - kernel_sizes: [4, 4] - strides: [2, 2] - num_residual_layers: 2 - embedding_dim: 64 - num_embeddings: 256 - upsampling: null - beta: 0.25 - activation: leaky_relu - dropout_rate: 0.2 +_target_: text_recognizer.networks.vqvae.VQVAE +in_channels: 1 +res_channels: 32 +num_residual_layers: 2 +embedding_dim: 64 +num_embeddings: 512 +decay: 0.99 +activation: mish |