config.yml.example

### WANDB METADATA

# this will be used to associate this run with a wandb project
WANDB_PROJECT         : "iNat CV"
WANDB_LOG_FREQ        : 100

#### EXPERIMENT PARAMETERS
TENSORBOARD_LOG_DIR   : "/data-ssd/alex/experiments/slim_export/1608157217/log_dir"
CHECKPOINT_DIR        : "/data-ssd/alex/experiments/slim_export/1608157217/checkpoints/ckpt"
FINAL_SAVE_DIR        : "/data-ssd/alex/experiments/slim_export/1608157217/final_model/final"
BACKUP_DIR            : "/data-ssd/alex/experiments/slim_export/1608157217/backup"


#### DATASET PARAMETERS

TRAINING_DATA         : "/data-ssd/alex/datasets/slim_export_20201213/train_cleaned.json"
VAL_DATA              : "/data-ssd/alex/datasets/slim_export_20201213/val_cleaned.json"
TEST_DATA             : "/data-ssd/alex/datasets/slim_export_20201213/test_cleaned.json"
NUM_CLASSES           : 38148


#### TRAINING PARAMETERS

# training policy - only use mixed precision=true if you
# have a recent NVIDIA GPU that supports CUDA 7.0 or later
TRAIN_MIXED_PRECISION : True

# size of batch, per gpu
BATCH_SIZE            : 256

# number of training epochs
NUM_EPOCHS            : 80

# initial learning rate for the model
INITIAL_LEARNING_RATE : 0.05
LR_DECAY_FACTOR       : 0.94
EPOCHS_PER_LR_DECAY   : 4


#### MODEL PARAMETERS

activation          : "softmax" # or null (python None) for logits

# neural network architecture
MODEL_NAME            : "xception"

# size of input
IMAGE_SIZE            : [299,299]

# dropout percentage for layer between pool & logits
DROPOUT_PCT           : 0.5

# optiimzer
OPTIMIZER_NAME        : "rmsprop"
RMSPROP_RHO           : 0.9
RMSPROP_MOMENTUM      : 0.9
RMSPROP_EPSILON       : 1.0