From 441b7484348953deb7c94150675d54583ef5a81a Mon Sep 17 00:00:00 2001
From: Gustaf Rydholm <gustaf.rydholm@gmail.com>
Date: Tue, 10 Aug 2021 23:25:13 +0200
Subject: Update to config and logging in VQGAN

---
 training/conf/experiment/vqgan.yaml               | 17 +++++++++++++++--
 training/conf/experiment/vqvae.yaml               |  8 --------
 training/conf/lr_scheduler/cosine_annealing.yaml  |  8 --------
 training/conf/lr_scheduler/one_cycle.yaml         | 20 --------------------
 training/conf/lr_schedulers/cosine_annealing.yaml |  8 ++++++++
 training/conf/lr_schedulers/one_cycle.yaml        | 20 ++++++++++++++++++++
 training/conf/optimizer/madgrad.yaml              |  8 --------
 training/conf/optimizers/madgrad.yaml             |  8 ++++++++
 8 files changed, 51 insertions(+), 46 deletions(-)
 delete mode 100644 training/conf/lr_scheduler/cosine_annealing.yaml
 delete mode 100644 training/conf/lr_scheduler/one_cycle.yaml
 create mode 100644 training/conf/lr_schedulers/cosine_annealing.yaml
 create mode 100644 training/conf/lr_schedulers/one_cycle.yaml
 delete mode 100644 training/conf/optimizer/madgrad.yaml
 create mode 100644 training/conf/optimizers/madgrad.yaml

(limited to 'training')

diff --git a/training/conf/experiment/vqgan.yaml b/training/conf/experiment/vqgan.yaml
index 570e7f9..554ec9e 100644
--- a/training/conf/experiment/vqgan.yaml
+++ b/training/conf/experiment/vqgan.yaml
@@ -8,6 +8,19 @@ defaults:
   - override /optimizers: null
   - override /lr_schedulers: null
 
+criterion:
+  _target_: text_recognizer.criterions.vqgan_loss.VQGANLoss
+  reconstruction_loss:
+    _target_: torch.nn.L1Loss
+    reduction: mean
+  discriminator:
+    _target_: text_recognizer.criterions.n_layer_discriminator.NLayerDiscriminator
+    in_channels: 1
+    num_channels: 32
+    num_layers: 3
+  vq_loss_weight: 0.8
+  discriminator_weight: 0.6
+
 datamodule:
   batch_size: 8
 
@@ -33,7 +46,7 @@ lr_schedulers:
 optimizers:
   generator:
     _target_: madgrad.MADGRAD
-    lr: 2.0e-5
+    lr: 4.5e-6
     momentum: 0.5
     weight_decay: 0
     eps: 1.0e-6
@@ -42,7 +55,7 @@ optimizers:
 
   discriminator:
     _target_: madgrad.MADGRAD
-    lr: 2.0e-5
+    lr: 4.5e-6
     momentum: 0.5
     weight_decay: 0
     eps: 1.0e-6
diff --git a/training/conf/experiment/vqvae.yaml b/training/conf/experiment/vqvae.yaml
index 397a039..8dbb257 100644
--- a/training/conf/experiment/vqvae.yaml
+++ b/training/conf/experiment/vqvae.yaml
@@ -10,16 +10,8 @@ defaults:
 
 trainer:
   max_epochs: 256
-  # gradient_clip_val: 0.25
 
 datamodule:
   batch_size: 8
 
-# lr_scheduler:
-  # epochs: 64
-  # steps_per_epoch: 1245
-
-# optimizer:
-  # lr: 1.0e-3
-
 summary: null
diff --git a/training/conf/lr_scheduler/cosine_annealing.yaml b/training/conf/lr_scheduler/cosine_annealing.yaml
deleted file mode 100644
index c53ee3a..0000000
--- a/training/conf/lr_scheduler/cosine_annealing.yaml
+++ /dev/null
@@ -1,8 +0,0 @@
-cosine_annealing:
-  _target_: torch.optim.lr_scheduler.CosineAnnealingLR
-  T_max: 256
-  eta_min: 0.0
-  last_epoch: -1
-
-  interval: epoch
-  monitor: val/loss
diff --git a/training/conf/lr_scheduler/one_cycle.yaml b/training/conf/lr_scheduler/one_cycle.yaml
deleted file mode 100644
index c60577a..0000000
--- a/training/conf/lr_scheduler/one_cycle.yaml
+++ /dev/null
@@ -1,20 +0,0 @@
-onc_cycle:
-  _target_: torch.optim.lr_scheduler.OneCycleLR
-  max_lr: 1.0e-3
-  total_steps: null
-  epochs: 512
-  steps_per_epoch: 4992
-  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
-  last_epoch: -1
-  verbose: false
-
-  # Non-class arguments
-  interval: step
-  monitor: val/loss
diff --git a/training/conf/lr_schedulers/cosine_annealing.yaml b/training/conf/lr_schedulers/cosine_annealing.yaml
new file mode 100644
index 0000000..c53ee3a
--- /dev/null
+++ b/training/conf/lr_schedulers/cosine_annealing.yaml
@@ -0,0 +1,8 @@
+cosine_annealing:
+  _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+  T_max: 256
+  eta_min: 0.0
+  last_epoch: -1
+
+  interval: epoch
+  monitor: val/loss
diff --git a/training/conf/lr_schedulers/one_cycle.yaml b/training/conf/lr_schedulers/one_cycle.yaml
new file mode 100644
index 0000000..c60577a
--- /dev/null
+++ b/training/conf/lr_schedulers/one_cycle.yaml
@@ -0,0 +1,20 @@
+onc_cycle:
+  _target_: torch.optim.lr_scheduler.OneCycleLR
+  max_lr: 1.0e-3
+  total_steps: null
+  epochs: 512
+  steps_per_epoch: 4992
+  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
+  last_epoch: -1
+  verbose: false
+
+  # Non-class arguments
+  interval: step
+  monitor: val/loss
diff --git a/training/conf/optimizer/madgrad.yaml b/training/conf/optimizer/madgrad.yaml
deleted file mode 100644
index a6c059d..0000000
--- a/training/conf/optimizer/madgrad.yaml
+++ /dev/null
@@ -1,8 +0,0 @@
-madgrad:
-  _target_: madgrad.MADGRAD
-  lr: 1.0e-3
-  momentum: 0.9
-  weight_decay: 0
-  eps: 1.0e-6
-
-  parameters: network
diff --git a/training/conf/optimizers/madgrad.yaml b/training/conf/optimizers/madgrad.yaml
new file mode 100644
index 0000000..a6c059d
--- /dev/null
+++ b/training/conf/optimizers/madgrad.yaml
@@ -0,0 +1,8 @@
+madgrad:
+  _target_: madgrad.MADGRAD
+  lr: 1.0e-3
+  momentum: 0.9
+  weight_decay: 0
+  eps: 1.0e-6
+
+  parameters: network
-- 
cgit v1.2.3-70-g09d2