summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--training/conf/experiment/conv_transformer_lines.yaml26
-rw-r--r--training/conf/experiment/conv_transformer_paragraphs.yaml51
-rw-r--r--training/conf/experiment/vqgan.yaml89
-rw-r--r--training/conf/network/vqvae.yaml17
4 files changed, 92 insertions, 91 deletions
diff --git a/training/conf/experiment/conv_transformer_lines.yaml b/training/conf/experiment/conv_transformer_lines.yaml
index 11646ca..20e369e 100644
--- a/training/conf/experiment/conv_transformer_lines.yaml
+++ b/training/conf/experiment/conv_transformer_lines.yaml
@@ -10,7 +10,7 @@ defaults:
- override /lr_schedulers: null
- override /optimizers: null
-epochs: &epochs 620
+epochs: &epochs 300
ignore_index: &ignore_index 3
num_classes: &num_classes 57
max_output_len: &max_output_len 89
@@ -27,7 +27,7 @@ callbacks:
stochastic_weight_averaging:
_target_: pytorch_lightning.callbacks.StochasticWeightAveraging
swa_epoch_start: 0.75
- swa_lrs: 1.0e-5
+ swa_lrs: 1.0e-4
annealing_epochs: 10
annealing_strategy: cos
device: null
@@ -43,15 +43,15 @@ optimizers:
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
+ _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+ T_max: *epochs
+ eta_min: 1.0e-4
+ last_epoch: -1
+ interval: epoch
+ monitor: val/loss
datamodule:
- batch_size: 32
+ batch_size: 16
num_workers: 12
train_fraction: 0.9
pin_memory: true
@@ -64,7 +64,7 @@ rotary_embedding: &rotary_embedding
attn: &attn
dim: &hidden_dim 256
- num_heads: 4
+ num_heads: 6
dim_head: 64
dropout_rate: &dropout_rate 0.5
@@ -76,12 +76,12 @@ network:
pad_index: *ignore_index
encoder:
_target_: text_recognizer.networks.encoders.efficientnet.EfficientNet
- arch: b3
+ arch: b0
stochastic_dropout_rate: 0.2
bn_momentum: 0.99
bn_eps: 1.0e-3
decoder:
- depth: 6
+ depth: 3
_target_: text_recognizer.networks.transformer.layers.Decoder
self_attn:
_target_: text_recognizer.networks.transformer.attention.Attention
@@ -106,7 +106,7 @@ network:
pixel_pos_embedding:
_target_: text_recognizer.networks.transformer.embeddings.axial.AxialPositionalEmbedding
dim: *hidden_dim
- shape: [1, 32]
+ shape: [3, 64]
model:
_target_: text_recognizer.models.transformer.TransformerLitModel
diff --git a/training/conf/experiment/conv_transformer_paragraphs.yaml b/training/conf/experiment/conv_transformer_paragraphs.yaml
index 00ad389..d2916e1 100644
--- a/training/conf/experiment/conv_transformer_paragraphs.yaml
+++ b/training/conf/experiment/conv_transformer_paragraphs.yaml
@@ -10,8 +10,7 @@ defaults:
- override /lr_schedulers: null
- override /optimizers: null
-
-epochs: &epochs 720
+epochs: &epochs 512
ignore_index: &ignore_index 3
num_classes: &num_classes 58
max_output_len: &max_output_len 682
@@ -29,7 +28,7 @@ callbacks:
stochastic_weight_averaging:
_target_: pytorch_lightning.callbacks.StochasticWeightAveraging
swa_epoch_start: 0.75
- swa_lrs: 1.0e-5
+ swa_lrs: 3.0e-5
annealing_epochs: 10
annealing_strategy: cos
device: null
@@ -37,7 +36,7 @@ callbacks:
optimizers:
madgrad:
_target_: madgrad.MADGRAD
- lr: 1.0e-4
+ lr: 3.0e-4
momentum: 0.9
weight_decay: 5.0e-6
eps: 1.0e-6
@@ -45,27 +44,16 @@ optimizers:
lr_schedulers:
network:
- _target_: torch.optim.lr_scheduler.OneCycleLR
- max_lr: 1.0e-4
- total_steps: null
- epochs: *epochs
- steps_per_epoch: 1264
- pct_start: 0.01
- anneal_strategy: cos
- cycle_momentum: true
- base_momentum: 0.85
- max_momentum: 0.95
- div_factor: 25
- final_div_factor: 1.0e2
- three_phase: false
+ _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+ T_max: *epochs
+ eta_min: 1.0e-5
last_epoch: -1
- verbose: false
- interval: step
+ interval: epoch
monitor: val/loss
datamodule:
_target_: text_recognizer.data.iam_extended_paragraphs.IAMExtendedParagraphs
- batch_size: 6
+ batch_size: 4
num_workers: 12
train_fraction: 0.8
pin_memory: true
@@ -77,27 +65,25 @@ rotary_embedding: &rotary_embedding
dim: 64
attn: &attn
- dim: 192
+ dim: &hidden_dim 192
num_heads: 4
dim_head: 64
- dropout_rate: 0.05
+ dropout_rate: &dropout_rate 0.5
network:
_target_: text_recognizer.networks.conv_transformer.ConvTransformer
input_dims: [1, 576, 640]
- hidden_dim: &hidden_dim 192
+ hidden_dim: *hidden_dim
num_classes: *num_classes
pad_index: *ignore_index
encoder:
_target_: text_recognizer.networks.encoders.efficientnet.EfficientNet
- arch: b0
- out_channels: 1280
+ arch: b1
stochastic_dropout_rate: 0.2
bn_momentum: 0.99
bn_eps: 1.0e-3
decoder:
- depth: 3
- local_depth: 2
+ depth: 6
_target_: text_recognizer.networks.transformer.layers.Decoder
self_attn:
_target_: text_recognizer.networks.transformer.attention.Attention
@@ -108,13 +94,6 @@ network:
_target_: text_recognizer.networks.transformer.attention.Attention
<< : *attn
causal: false
- local_self_attn:
- _target_: text_recognizer.networks.transformer.local_attention.LocalAttention
- << : *attn
- window_size: 31
- look_back: 1
- autopad: true
- << : *rotary_embedding
norm:
_target_: text_recognizer.networks.transformer.norm.ScaleNorm
normalized_shape: *hidden_dim
@@ -124,7 +103,7 @@ network:
dim_out: null
expansion_factor: 4
glu: true
- dropout_rate: 0.05
+ dropout_rate: *dropout_rate
pre_norm: true
pixel_pos_embedding:
_target_: text_recognizer.networks.transformer.embeddings.axial.AxialPositionalEmbedding
@@ -155,5 +134,5 @@ trainer:
limit_val_batches: 1.0
limit_test_batches: 1.0
resume_from_checkpoint: null
- accumulate_grad_batches: 4
+ accumulate_grad_batches: 2
overfit_batches: 0
diff --git a/training/conf/experiment/vqgan.yaml b/training/conf/experiment/vqgan.yaml
index 98f3346..726757f 100644
--- a/training/conf/experiment/vqgan.yaml
+++ b/training/conf/experiment/vqgan.yaml
@@ -1,18 +1,26 @@
+# @package _global_
+
defaults:
- override /network: vqvae
- override /criterion: null
- override /model: lit_vqgan
- - override /callbacks: wandb_vae
+ - 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.criterions.vqgan_loss.VQGANLoss
+ _target_: text_recognizer.criterion.vqgan_loss.VQGANLoss
reconstruction_loss:
_target_: torch.nn.BCEWithLogitsLoss
reduction: mean
discriminator:
- _target_: text_recognizer.criterions.n_layer_discriminator.NLayerDiscriminator
+ _target_: text_recognizer.criterion.n_layer_discriminator.NLayerDiscriminator
in_channels: 1
num_channels: 64
num_layers: 3
@@ -21,39 +29,35 @@ criterion:
discriminator_factor: 1.0
discriminator_iter_start: 8.0e4
+mapping: &mapping
+ mapping:
+ _target_: text_recognizer.data.mappings.emnist.EmnistMapping
+ extra_symbols: [ "\n" ]
+
datamodule:
- batch_size: 12
- # resize: [288, 320]
- augment: false
+ _target_: text_recognizer.data.iam_extended_paragraphs.IAMExtendedParagraphs
+ batch_size: 4
+ num_workers: 12
+ train_fraction: 0.9
+ pin_memory: true
+ << : *mapping
lr_schedulers:
- generator:
- _target_: torch.optim.lr_scheduler.OneCycleLR
- max_lr: 3.0e-4
- total_steps: null
- epochs: 64
- steps_per_epoch: 1685
- pct_start: 0.3
- anneal_strategy: cos
- cycle_momentum: true
- base_momentum: 0.85
- max_momentum: 0.95
- div_factor: 25.0
- final_div_factor: 10000.0
- three_phase: true
+ network:
+ _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+ T_max: *epochs
+ eta_min: 1.0e-5
last_epoch: -1
- verbose: false
- interval: step
+ interval: epoch
monitor: val/loss
-# discriminator:
-# _target_: torch.optim.lr_scheduler.CosineAnnealingLR
-# T_max: 64
-# eta_min: 0.0
-# 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:
@@ -75,11 +79,20 @@ optimizers:
parameters: loss_fn.discriminator
trainer:
- max_epochs: 64
- # limit_train_batches: 0.1
- # limit_val_batches: 0.1
- # gradient_clip_val: 100
-
-# tune: false
-# train: true
-# test: false
+ _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/network/vqvae.yaml b/training/conf/network/vqvae.yaml
index 8210f04..22f786f 100644
--- a/training/conf/network/vqvae.yaml
+++ b/training/conf/network/vqvae.yaml
@@ -3,7 +3,16 @@ defaults:
- decoder: vae_decoder
_target_: text_recognizer.networks.vqvae.vqvae.VQVAE
-hidden_dim: 128
-embedding_dim: 32
-num_embeddings: 8192
-decay: 0.99
+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
generated by cgit v1.2.3-70-g09d2 (git 2.46.0) at 2024-11-10 06:01:43 +0000