Inital commit for refactoring to lightning

3 files changed, 233 insertions, 0 deletions

diff --git a/training/experiments/default_config_emnist.yml b/training/experiments/default_config_emnist.yml
new file mode 100644
index 0000000..bf2ed0a
--- /dev/null
+++ b/training/experiments/default_config_emnist.yml
@@ -0,0 +1,70 @@
+dataset: EmnistDataset
+dataset_args:
+  sample_to_balance: true
+  subsample_fraction: 0.33
+  transform: null
+  target_transform: null
+  seed: 4711
+
+data_loader_args:
+  splits: [train, val]
+  shuffle: true
+  num_workers: 8
+  cuda: true
+
+model: CharacterModel
+metrics: [accuracy]
+
+network_args:
+  in_channels: 1
+  num_classes: 80
+  depths: [2]
+  block_sizes: [256]
+
+train_args:
+  batch_size: 256
+  epochs: 5
+
+criterion: CrossEntropyLoss
+criterion_args:
+  weight: null
+  ignore_index: -100
+  reduction: mean
+
+optimizer: AdamW
+optimizer_args:
+  lr: 1.e-03
+  betas: [0.9, 0.999]
+  eps: 1.e-08
+  # weight_decay: 5.e-4
+  amsgrad: false
+
+lr_scheduler: OneCycleLR
+lr_scheduler_args:
+  max_lr: 1.e-03
+  epochs: 5
+  anneal_strategy: linear
+
+
+callbacks: [Checkpoint, ProgressBar, EarlyStopping, WandbCallback, WandbImageLogger, OneCycleLR]
+callback_args:
+  Checkpoint:
+    monitor: val_accuracy
+  ProgressBar:
+    epochs: 5
+    log_batch_frequency: 100
+  EarlyStopping:
+    monitor: val_loss
+    min_delta: 0.0
+    patience: 3
+    mode: min
+  WandbCallback:
+    log_batch_frequency: 10
+  WandbImageLogger:
+    num_examples: 4
+  OneCycleLR:
+    null
+verbosity: 1 # 0, 1, 2
+resume_experiment: null
+train: true
+validation_metric: val_accuracy
diff --git a/training/experiments/embedding_experiment.yml b/training/experiments/embedding_experiment.yml
new file mode 100644
index 0000000..1e5f941
--- /dev/null
+++ b/training/experiments/embedding_experiment.yml
@@ -0,0 +1,64 @@
+experiment_group: Embedding Experiments
+experiments:
+    - train_args:
+        transformer_model: false
+        batch_size: &batch_size 256
+        max_epochs: &max_epochs 32
+        input_shape: [[1, 28, 28]]
+      dataset:
+        type: EmnistDataset
+        args:
+          sample_to_balance: true
+          subsample_fraction: null
+          transform: null
+          target_transform: null
+          seed: 4711
+        train_args:
+          num_workers: 8
+          train_fraction: 0.85
+          batch_size: *batch_size
+      model: CharacterModel
+      metrics: []
+      network:
+        type: DenseNet
+        args:
+          growth_rate: 4
+          block_config: [4, 4]
+          in_channels: 1
+          base_channels: 24
+          num_classes: 128
+          bn_size: 4
+          dropout_rate: 0.1
+          classifier: true
+          activation: elu
+      criterion:
+        type: EmbeddingLoss
+        args:
+          margin: 0.2
+          type_of_triplets: semihard
+      optimizer:
+        type: AdamW
+        args:
+          lr: 1.e-02
+          betas: [0.9, 0.999]
+          eps: 1.e-08
+          weight_decay: 5.e-4
+          amsgrad: false
+      lr_scheduler:
+        type: CosineAnnealingLR
+        args:
+          T_max: *max_epochs
+      callbacks: [Checkpoint, ProgressBar, WandbCallback]
+      callback_args:
+        Checkpoint:
+          monitor: val_loss
+          mode: min
+        ProgressBar:
+          epochs: *max_epochs
+        WandbCallback:
+          log_batch_frequency: 10
+      verbosity: 1 # 0, 1, 2
+      resume_experiment: null
+      train: true
+      test: true
+      test_metric: mean_average_precision_at_r
diff --git a/training/experiments/sample_experiment.yml b/training/experiments/sample_experiment.yml
new file mode 100644
index 0000000..8f94475
--- /dev/null
+++ b/training/experiments/sample_experiment.yml
@@ -0,0 +1,99 @@
+experiment_group: Sample Experiments
+experiments:
+    - train_args:
+        batch_size: 256
+        max_epochs: &max_epochs 32
+      dataset:
+        type: EmnistDataset
+        args:
+          sample_to_balance: true
+          subsample_fraction: null
+          transform: null
+          target_transform: null
+          seed: 4711
+        train_args:
+          num_workers: 6
+          train_fraction: 0.8
+
+      model: CharacterModel
+      metrics: [accuracy]
+      # network: MLP
+      # network_args:
+      #   input_size: 784
+      #   hidden_size: 512
+      #   output_size: 80
+      #   num_layers: 5
+      #   dropout_rate: 0.2
+      #   activation_fn: SELU
+      network:
+        type: ResidualNetwork
+        args:
+          in_channels: 1
+          num_classes: 80
+          depths: [2, 2]
+          block_sizes: [64, 64]
+          activation: leaky_relu
+      # network:
+      #   type: WideResidualNetwork
+      #   args:
+      #     in_channels: 1
+      #     num_classes: 80
+      #     depth: 10
+      #     num_layers: 3
+      #     width_factor: 4
+      #     dropout_rate: 0.2
+      #     activation: SELU
+      # network: LeNet
+      # network_args:
+      #   output_size: 62
+      #   activation_fn: GELU
+      criterion:
+        type: CrossEntropyLoss
+        args:
+          weight: null
+          ignore_index: -100
+          reduction: mean
+      optimizer:
+        type: AdamW
+        args:
+          lr: 1.e-02
+          betas: [0.9, 0.999]
+          eps: 1.e-08
+          # weight_decay: 5.e-4
+          amsgrad: false
+      # lr_scheduler:
+      #   type: OneCycleLR
+      #   args:
+      #     max_lr: 1.e-03
+      #     epochs: *max_epochs
+      #     anneal_strategy: linear
+      lr_scheduler:
+        type: CosineAnnealingLR
+        args:
+          T_max: *max_epochs
+          interval: epoch
+      swa_args:
+        start: 2
+        lr: 5.e-2
+      callbacks: [Checkpoint, ProgressBar, WandbCallback, WandbImageLogger, EarlyStopping]
+      callback_args:
+        Checkpoint:
+          monitor: val_accuracy
+        ProgressBar:
+          epochs: null
+          log_batch_frequency: 100
+        EarlyStopping:
+          monitor: val_loss
+          min_delta: 0.0
+          patience: 5
+          mode: min
+        WandbCallback:
+          log_batch_frequency: 10
+        WandbImageLogger:
+          num_examples: 4
+          use_transpose: true
+      verbosity: 0 # 0, 1, 2
+      resume_experiment: null
+      train: true
+      test: true
+      test_metric: test_accuracy