Remove VQVAE stuff, did not work...

16 files changed, 1 insertions, 430 deletions

diff --git a/training/conf/callbacks/vae.yaml b/training/conf/callbacks/vae.yaml
deleted file mode 100644
index 52adf69..0000000
--- a/training/conf/callbacks/vae.yaml
+++ /dev/null
@@ -1,3 +0,0 @@
-defaults:
-  - default
-  - wandb/reconstructions
diff --git a/training/conf/callbacks/wandb/reconstructions.yaml b/training/conf/callbacks/wandb/reconstructions.yaml
deleted file mode 100644
index 92f2d12..0000000
--- a/training/conf/callbacks/wandb/reconstructions.yaml
+++ /dev/null
@@ -1,4 +0,0 @@
-log_image_reconstruction:
-  _target_: callbacks.wandb_callbacks.LogReconstuctedImages
-  num_samples: 8
-  use_sigmoid: true
diff --git a/training/conf/criterion/mae.yaml b/training/conf/criterion/mae.yaml
deleted file mode 100644
index cb07467..0000000
--- a/training/conf/criterion/mae.yaml
+++ /dev/null
@@ -1,2 +0,0 @@
-_target_: torch.nn.L1Loss
-reduction: mean
diff --git a/training/conf/criterion/mse.yaml b/training/conf/criterion/mse.yaml
deleted file mode 100644
index ffd1403..0000000
--- a/training/conf/criterion/mse.yaml
+++ /dev/null
@@ -1,2 +0,0 @@
-_target_: torch.nn.MSELoss
-reduction: mean
diff --git a/training/conf/criterion/vqgan_loss.yaml b/training/conf/criterion/vqgan_loss.yaml
deleted file mode 100644
index 34a67ae..0000000
--- a/training/conf/criterion/vqgan_loss.yaml
+++ /dev/null
@@ -1,11 +0,0 @@
-_target_: text_recognizer.criterion.vqgan_loss.VQGANLoss
-reconstruction_loss:
-  _target_: torch.nn.MSELoss
-  reduction: mean
-discriminator:
-  _target_: text_recognizer.criterions.n_layer_discriminator.NLayerDiscriminator
-  in_channels: 1
-  num_channels: 32
-  num_layers: 3
-vq_loss_weight: 1.0
-discriminator_weight: 1.0
diff --git a/training/conf/datamodule/transform/paragraphs.yaml b/training/conf/datamodule/transform/paragraphs.yaml
index c1d2e4f..1b0ab04 100644
--- a/training/conf/datamodule/transform/paragraphs.yaml
+++ b/training/conf/datamodule/transform/paragraphs.yaml
@@ -18,7 +18,7 @@ random_affine:
 
 random_perspective:
   _target_: torchvision.transforms.RandomPerspective
-  distortion_scale: 0.07
+  distortion_scale: 0.1
   p: 0.5
   fill: 0
 
diff --git a/training/conf/experiment/vq_transformer_lines.yaml b/training/conf/experiment/vq_transformer_lines.yaml
deleted file mode 100644
index bbe1178..0000000
--- a/training/conf/experiment/vq_transformer_lines.yaml
+++ /dev/null
@@ -1,149 +0,0 @@
-# @package _global_
-
-defaults:
-  - override /mapping: null
-  - override /criterion: cross_entropy
-  - override /callbacks: htr
-  - override /datamodule: iam_lines
-  - override /network: null
-  - override /model: null
-  - override /lr_schedulers: null
-  - override /optimizers: null
-
-epochs: &epochs 512
-ignore_index: &ignore_index 3
-num_classes: &num_classes 57
-max_output_len: &max_output_len 89
-summary: [[1, 1, 56, 1024], [1, 89]]
-
-criterion:
-  ignore_index: *ignore_index
-    
-mapping: &mapping
-  mapping:
-    _target_: text_recognizer.data.mappings.emnist.EmnistMapping
-
-callbacks:
-  stochastic_weight_averaging:
-    _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
-    swa_epoch_start: 0.75
-    swa_lrs: 1.0e-5
-    annealing_epochs: 10
-    annealing_strategy: cos
-    device: null
-
-optimizers:
-  madgrad:
-    _target_: madgrad.MADGRAD
-    lr: 3.0e-4
-    momentum: 0.9
-    weight_decay: 0
-    eps: 1.0e-6
-    parameters: network
-
-lr_schedulers:
-  network:
-    _target_: torch.optim.lr_scheduler.CosineAnnealingLR
-    T_max: *epochs
-    eta_min: 1.0e-5
-    last_epoch: -1
-    interval: epoch
-    monitor: val/loss
-
-datamodule:
-  batch_size: 16
-  num_workers: 12
-  train_fraction: 0.9
-  pin_memory: true
-  << : *mapping
-
-rotary_embedding: &rotary_embedding
-  rotary_embedding: 
-    _target_: text_recognizer.networks.transformer.embeddings.rotary.RotaryEmbedding
-    dim: 64
-
-attn: &attn
-  dim: &hidden_dim 512
-  num_heads: 4
-  dim_head: 64
-  dropout_rate: &dropout_rate 0.4
-
-network:
-  _target_: text_recognizer.networks.vq_transformer.VqTransformer
-  input_dims: [1, 56, 1024]
-  hidden_dim: *hidden_dim
-  num_classes: *num_classes
-  pad_index: *ignore_index
-  encoder:
-    _target_: text_recognizer.networks.encoders.efficientnet.EfficientNet
-    arch: b1
-    stochastic_dropout_rate: 0.2
-    bn_momentum: 0.99
-    bn_eps: 1.0e-3
-  decoder:
-    depth: 6
-    _target_: text_recognizer.networks.transformer.layers.Decoder
-    self_attn:
-      _target_: text_recognizer.networks.transformer.attention.Attention
-      << : *attn
-      causal: true
-      << : *rotary_embedding
-    cross_attn:
-      _target_: text_recognizer.networks.transformer.attention.Attention
-      << : *attn
-      causal: false
-    norm:
-      _target_: text_recognizer.networks.transformer.norm.ScaleNorm
-      normalized_shape: *hidden_dim
-    ff: 
-      _target_: text_recognizer.networks.transformer.mlp.FeedForward
-      dim: *hidden_dim
-      dim_out: null
-      expansion_factor: 4
-      glu: true
-      dropout_rate: *dropout_rate
-    pre_norm: true
-  pixel_pos_embedding:
-    _target_: text_recognizer.networks.transformer.embeddings.axial.AxialPositionalEmbedding
-    dim: *hidden_dim
-    shape: [1, 32]
-  quantizer:
-    _target_: text_recognizer.networks.quantizer.quantizer.VectorQuantizer
-    input_dim: 512
-    codebook:
-      _target_: text_recognizer.networks.quantizer.codebook.CosineSimilarityCodebook
-      dim: 16
-      codebook_size: 4096
-      kmeans_init: true
-      kmeans_iters: 10
-      decay: 0.8
-      eps: 1.0e-5
-      threshold_dead: 2
-    commitment: 1.0
-
-model:
-  _target_: text_recognizer.models.vq_transformer.VqTransformerLitModel
-  << : *mapping
-  max_output_len: *max_output_len
-  start_token: <s>
-  end_token: <e>
-  pad_token: <p>
-
-trainer:
-  _target_: pytorch_lightning.Trainer
-  stochastic_weight_avg: true
-  auto_scale_batch_size: binsearch
-  auto_lr_find: false
-  gradient_clip_val: 0.5
-  fast_dev_run: false
-  gpus: 1
-  precision: 16
-  max_epochs: *epochs
-  terminate_on_nan: true
-  weights_summary: null
-  limit_train_batches: 1.0 
-  limit_val_batches: 1.0
-  limit_test_batches: 1.0
-  resume_from_checkpoint: null
-  accumulate_grad_batches: 1
-  overfit_batches: 0
diff --git a/training/conf/experiment/vqgan.yaml b/training/conf/experiment/vqgan.yaml
deleted file mode 100644
index 726757f..0000000
--- a/training/conf/experiment/vqgan.yaml
+++ /dev/null
@@ -1,98 +0,0 @@
-# @package _global_
-
-defaults:
-  - override /network: vqvae
-  - override /criterion: null
-  - override /model: lit_vqgan
-  - override /callbacks: vae
-  - override /optimizers: null
-  - override /lr_schedulers: null
-
-epochs: &epochs 100
-ignore_index: &ignore_index 3
-num_classes: &num_classes 58
-max_output_len: &max_output_len 682
-summary: [[1, 1, 576, 640]]
-
-criterion:
-  _target_: text_recognizer.criterion.vqgan_loss.VQGANLoss
-  reconstruction_loss:
-    _target_: torch.nn.BCEWithLogitsLoss
-    reduction: mean
-  discriminator:
-    _target_: text_recognizer.criterion.n_layer_discriminator.NLayerDiscriminator
-    in_channels: 1
-    num_channels: 64
-    num_layers: 3
-  commitment_weight: 0.25
-  discriminator_weight: 0.8
-  discriminator_factor: 1.0
-  discriminator_iter_start: 8.0e4
-
-mapping: &mapping
-  mapping:
-    _target_: text_recognizer.data.mappings.emnist.EmnistMapping
-    extra_symbols: [ "\n" ]
-
-datamodule:
-  _target_: text_recognizer.data.iam_extended_paragraphs.IAMExtendedParagraphs
-  batch_size: 4
-  num_workers: 12
-  train_fraction: 0.9
-  pin_memory: true
-  << : *mapping
-
-lr_schedulers:
-  network:
-    _target_: torch.optim.lr_scheduler.CosineAnnealingLR
-    T_max: *epochs
-    eta_min: 1.0e-5
-    last_epoch: -1
-    interval: epoch
-    monitor: val/loss
- 
-  discriminator:
-    _target_: torch.optim.lr_scheduler.CosineAnnealingLR
-    T_max: *epochs
-    eta_min: 1.0e-5
-    last_epoch: -1
-    interval: epoch
-    monitor: val/loss
-
-optimizers:
-  generator:
-    _target_: madgrad.MADGRAD
-    lr: 1.0e-4
-    momentum: 0.5
-    weight_decay: 0
-    eps: 1.0e-7
-
-    parameters: network
-
-  discriminator:
-    _target_: madgrad.MADGRAD
-    lr: 4.5e-6
-    momentum: 0.5
-    weight_decay: 0
-    eps: 1.0e-6
-
-    parameters: loss_fn.discriminator
-
-trainer:
-  _target_: pytorch_lightning.Trainer
-  stochastic_weight_avg: false
-  auto_scale_batch_size: binsearch
-  auto_lr_find: false
-  gradient_clip_val: 0
-  fast_dev_run: false
-  gpus: 1
-  precision: 16
-  max_epochs: *epochs
-  terminate_on_nan: true
-  weights_summary: null
-  limit_train_batches: 1.0 
-  limit_val_batches: 1.0
-  limit_test_batches: 1.0
-  resume_from_checkpoint: null
-  accumulate_grad_batches: 2
-  overfit_batches: 0
diff --git a/training/conf/experiment/vqgan_htr_char.yaml b/training/conf/experiment/vqgan_htr_char.yaml
deleted file mode 100644
index af3fa40..0000000
--- a/training/conf/experiment/vqgan_htr_char.yaml
+++ /dev/null
@@ -1,59 +0,0 @@
-defaults:
-  - override /mapping: null
-  - override /network: null
-  - override /model: null
-
-mapping:
-  _target_: text_recognizer.data.emnist_mapping.EmnistMapping
-  extra_symbols: [ "\n" ]
-
-datamodule:
-  word_pieces: false
-  batch_size: 8
-  augment: false
-
-criterion:
-  ignore_index: 3
-
-network:
-  _target_: text_recognizer.networks.vq_transformer.VqTransformer
-  input_dims: [1, 576, 640]
-  encoder_dim: 32
-  hidden_dim: 256
-  dropout_rate: 0.1
-  num_classes: 58
-  pad_index: 3
-  no_grad: true
-  decoder:
-    _target_: text_recognizer.networks.transformer.Decoder
-    dim: 256
-    depth: 2
-    num_heads: 8
-    attn_fn: text_recognizer.networks.transformer.attention.Attention
-    attn_kwargs:
-      dim_head: 32
-      dropout_rate: 0.2
-    norm_fn: torch.nn.LayerNorm
-    ff_fn: text_recognizer.networks.transformer.mlp.FeedForward
-    ff_kwargs:
-      dim_out: null
-      expansion_factor: 4
-      glu: true
-      dropout_rate: 0.2
-    cross_attend: true
-    pre_norm: true
-    rotary_emb: null
-  pretrained_encoder_path: "training/logs/runs/2021-09-25/23-07-28"
-
-model:
-  _target_: text_recognizer.models.vq_transformer.VqTransformerLitModel
-  start_token: <s>
-  end_token: <e>
-  pad_token: <p>
-  max_output_len: 682  # 451
-  alpha: 1.0
-
-trainer:
-  max_epochs: 64
-  limit_train_batches: 0.1
-  limit_val_batches: 0.1
diff --git a/training/conf/experiment/vqvae.yaml b/training/conf/experiment/vqvae.yaml
deleted file mode 100644
index d069aef..0000000
--- a/training/conf/experiment/vqvae.yaml
+++ /dev/null
@@ -1,51 +0,0 @@
-defaults:
-  - override /network: vqvae
-  - override /criterion: mse
-  - override /model: lit_vqvae
-  - override /callbacks: wandb_vae
-  - override /optimizers: null
-  # - override /lr_schedulers: 
-      # - cosine_annealing
-
-# lr_schedulers: null
-#   network:
-#     _target_: torch.optim.lr_scheduler.OneCycleLR
-#     max_lr: 1.0e-2
-#     total_steps: null
-#     epochs: 100
-#     steps_per_epoch: 200
-#     pct_start: 0.1
-#     anneal_strategy: cos
-#     cycle_momentum: true
-#     base_momentum: 0.85
-#     max_momentum: 0.95
-#     div_factor: 25
-#     final_div_factor: 1.0e4
-#     three_phase: true
-#     last_epoch: -1
-#     verbose: false
-
-#     # Non-class arguments
-#     interval: step
-#     monitor: val/loss
-
-optimizers:
-  network:
-    _target_: madgrad.MADGRAD
-    lr: 1.0e-4
-    momentum: 0.9
-    weight_decay: 0
-    eps: 1.0e-7
-
-    parameters: network
-
-trainer:
-  max_epochs: 128
-  limit_train_batches: 0.1
-  limit_val_batches: 0.1
-
-datamodule:
-  batch_size: 8
-  # resize: [288, 320]
-
-summary: null
diff --git a/training/conf/model/lit_vqgan.yaml b/training/conf/model/lit_vqgan.yaml
deleted file mode 100644
index 9ee1046..0000000
--- a/training/conf/model/lit_vqgan.yaml
+++ /dev/null
@@ -1 +0,0 @@
-_target_: text_recognizer.models.vqgan.VQGANLitModel
diff --git a/training/conf/model/lit_vqvae.yaml b/training/conf/model/lit_vqvae.yaml
deleted file mode 100644
index 6dc44d7..0000000
--- a/training/conf/model/lit_vqvae.yaml
+++ /dev/null
@@ -1 +0,0 @@
-_target_: text_recognizer.models.vqvae.VQVAELitModel
diff --git a/training/conf/network/decoder/vae_decoder.yaml b/training/conf/network/decoder/vae_decoder.yaml
deleted file mode 100644
index aed5733..0000000
--- a/training/conf/network/decoder/vae_decoder.yaml
+++ /dev/null
@@ -1,9 +0,0 @@
-_target_: text_recognizer.networks.vqvae.decoder.Decoder
-out_channels: 1 
-hidden_dim: 32
-channels_multipliers: [4, 2, 1]
-dropout_rate: 0.0
-activation: mish
-use_norm: true
-num_residuals: 4
-residual_channels: 32
diff --git a/training/conf/network/encoder/vae_encoder.yaml b/training/conf/network/encoder/vae_encoder.yaml
deleted file mode 100644
index 5d39bf7..0000000
--- a/training/conf/network/encoder/vae_encoder.yaml
+++ /dev/null
@@ -1,9 +0,0 @@
-_target_: text_recognizer.networks.vqvae.encoder.Encoder
-in_channels: 1 
-hidden_dim: 32
-channels_multipliers: [1, 2, 4]
-dropout_rate: 0.0
-activation: mish
-use_norm: true
-num_residuals: 4
-residual_channels: 32
diff --git a/training/conf/network/quantizer.yaml b/training/conf/network/quantizer.yaml
deleted file mode 100644
index 827a247..0000000
--- a/training/conf/network/quantizer.yaml
+++ /dev/null
@@ -1,12 +0,0 @@
-_target_: text_recognizer.networks.quantizer.quantizer.VectorQuantizer
-input_dim: 192
-codebook:
-  _target_: text_recognizer.networks.quantizer.codebook.CosineSimilarityCodebook
-  dim: 16
-  codebook_size: 2048
-  kmeans_init: true
-  kmeans_iters: 10
-  decay: 0.8
-  eps: 1.0e-5
-  threshold_dead: 2
-commitment: 1.0
diff --git a/training/conf/network/vqvae.yaml b/training/conf/network/vqvae.yaml
deleted file mode 100644
index 22f786f..0000000
--- a/training/conf/network/vqvae.yaml
+++ /dev/null
@@ -1,18 +0,0 @@
-defaults:
-  - encoder: vae_encoder
-  - decoder: vae_decoder
-
-_target_: text_recognizer.networks.vqvae.vqvae.VQVAE
-quantizer:
-  _target_: text_recognizer.networks.quantizer.quantizer.VectorQuantizer
-  input_dim: 128
-  codebook:
-    _target_: text_recognizer.networks.quantizer.codebook.CosineSimilarityCodebook
-    dim: 8
-    codebook_size: 512
-    kmeans_init: true
-    kmeans_iters: 10
-    decay: 0.8
-    eps: 1.0e-5
-    threshold_dead: 2
-  commitment: 1.0