diff options
Diffstat (limited to 'src/training/experiments')
| -rw-r--r-- | src/training/experiments/sample_experiment.yml | 37 |
1 files changed, 24 insertions, 13 deletions
diff --git a/src/training/experiments/sample_experiment.yml b/src/training/experiments/sample_experiment.yml index 355305c..bae02ac 100644 --- a/src/training/experiments/sample_experiment.yml +++ b/src/training/experiments/sample_experiment.yml @@ -9,25 +9,32 @@ experiments: seed: 4711 data_loader_args: splits: [train, val] - batch_size: 256 shuffle: true num_workers: 8 cuda: true model: CharacterModel metrics: [accuracy] - network: MLP + # network: MLP + # network_args: + # input_size: 784 + # hidden_size: 512 + # output_size: 80 + # num_layers: 3 + # dropout_rate: 0 + # activation_fn: SELU + network: ResidualNetwork network_args: - input_size: 784 - output_size: 62 - num_layers: 3 - activation_fn: GELU + in_channels: 1 + num_classes: 80 + depths: [1, 1] + block_sizes: [128, 256] # network: LeNet # network_args: # output_size: 62 # activation_fn: GELU train_args: batch_size: 256 - epochs: 16 + epochs: 32 criterion: CrossEntropyLoss criterion_args: weight: null @@ -43,20 +50,24 @@ experiments: # centered: false optimizer: AdamW optimizer_args: - lr: 1.e-2 + lr: 1.e-03 betas: [0.9, 0.999] eps: 1.e-08 - weight_decay: 0 + # weight_decay: 5.e-4 amsgrad: false # lr_scheduler: null lr_scheduler: OneCycleLR lr_scheduler_args: - max_lr: 1.e-3 - epochs: 16 - callbacks: [Checkpoint, EarlyStopping, WandbCallback, WandbImageLogger, OneCycleLR] + max_lr: 1.e-03 + epochs: 32 + anneal_strategy: linear + callbacks: [Checkpoint, ProgressBar, EarlyStopping, WandbCallback, WandbImageLogger, OneCycleLR] callback_args: Checkpoint: monitor: val_accuracy + ProgressBar: + epochs: 32 + log_batch_frequency: 100 EarlyStopping: monitor: val_loss min_delta: 0.0 @@ -68,5 +79,5 @@ experiments: num_examples: 4 OneCycleLR: null - verbosity: 2 # 0, 1, 2 + verbosity: 1 # 0, 1, 2 resume_experiment: null |