summaryrefslogtreecommitdiff
path: root/src/training/experiments
diff options
context:
space:
mode:
authoraktersnurra <gustaf.rydholm@gmail.com>2020-10-22 22:45:58 +0200
committeraktersnurra <gustaf.rydholm@gmail.com>2020-10-22 22:45:58 +0200
commit4d7713746eb936832e84852e90292936b933e87d (patch)
tree2b2519d1d2ce53d4e1390590f52018d55dadbc7c /src/training/experiments
parent1b3b8073a19f939d18a0bb85247eb0d99284f7cc (diff)
Transfomer added, many other changes.
Diffstat (limited to 'src/training/experiments')
-rw-r--r--src/training/experiments/embedding_experiment.yml22
-rw-r--r--src/training/experiments/line_ctc_experiment.yml92
2 files changed, 14 insertions, 100 deletions
diff --git a/src/training/experiments/embedding_experiment.yml b/src/training/experiments/embedding_experiment.yml
index e674c26..1e5f941 100644
--- a/src/training/experiments/embedding_experiment.yml
+++ b/src/training/experiments/embedding_experiment.yml
@@ -1,8 +1,10 @@
experiment_group: Embedding Experiments
experiments:
- train_args:
- batch_size: 256
- max_epochs: &max_epochs 8
+ transformer_model: false
+ batch_size: &batch_size 256
+ max_epochs: &max_epochs 32
+ input_shape: [[1, 28, 28]]
dataset:
type: EmnistDataset
args:
@@ -14,17 +16,21 @@ experiments:
train_args:
num_workers: 8
train_fraction: 0.85
+ batch_size: *batch_size
model: CharacterModel
metrics: []
network:
- type: ResidualNetwork
+ type: DenseNet
args:
+ growth_rate: 4
+ block_config: [4, 4]
in_channels: 1
- num_classes: 64 # Embedding
- depths: [2,2]
- block_sizes: [32, 64]
- activation: selu
- stn: false
+ base_channels: 24
+ num_classes: 128
+ bn_size: 4
+ dropout_rate: 0.1
+ classifier: true
+ activation: elu
criterion:
type: EmbeddingLoss
args:
diff --git a/src/training/experiments/line_ctc_experiment.yml b/src/training/experiments/line_ctc_experiment.yml
deleted file mode 100644
index ef97527..0000000
--- a/src/training/experiments/line_ctc_experiment.yml
+++ /dev/null
@@ -1,92 +0,0 @@
-experiment_group: Lines Experiments
-experiments:
- - train_args:
- batch_size: 64
- max_epochs: &max_epochs 64
- dataset:
- type: IamLinesDataset
- args:
- subsample_fraction: null
- transform: null
- target_transform: null
- train_args:
- num_workers: 8
- train_fraction: 0.85
- model: LineCTCModel
- metrics: [cer, wer]
- network:
- type: LineRecurrentNetwork
- args:
- # backbone: ResidualNetwork
- # backbone_args:
- # in_channels: 1
- # num_classes: 64 # Embedding
- # depths: [2,2]
- # block_sizes: [32, 64]
- # activation: selu
- # stn: false
- backbone: ResidualNetwork
- backbone_args:
- pretrained: training/experiments/CharacterModel_EmnistDataset_ResidualNetwork/0920_025816/model/best.pt
- freeze: false
- flatten: false
- input_size: 64
- hidden_size: 64
- bidirectional: true
- num_layers: 2
- num_classes: 80
- patch_size: [28, 18]
- stride: [1, 4]
- criterion:
- type: CTCLoss
- args:
- blank: 79
- 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-02
- epochs: *max_epochs
- anneal_strategy: cos
- pct_start: 0.475
- cycle_momentum: true
- base_momentum: 0.85
- max_momentum: 0.9
- div_factor: 10
- final_div_factor: 10000
- interval: step
- # lr_scheduler:
- # type: CosineAnnealingLR
- # args:
- # T_max: *max_epochs
- swa_args:
- start: 48
- lr: 5.e-2
- callbacks: [Checkpoint, ProgressBar, WandbCallback, WandbImageLogger, EarlyStopping]
- callback_args:
- Checkpoint:
- monitor: val_loss
- mode: min
- ProgressBar:
- epochs: *max_epochs
- EarlyStopping:
- monitor: val_loss
- min_delta: 0.0
- patience: 10
- mode: min
- WandbCallback:
- log_batch_frequency: 10
- WandbImageLogger:
- num_examples: 6
- verbosity: 1 # 0, 1, 2
- resume_experiment: null
- train: true
- test: true
- test_metric: test_cer