forked from chrisgoringe/Custom-classifier-training
-
Notifications
You must be signed in to change notification settings - Fork 0
/
arguments-example.txt
122 lines (112 loc) · 4.77 KB
/
arguments-example.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
# One argument per line; white space either side of '=' is removed
# lines starting '#' are ignored, as are blank lines
#
# Main arguments.
#
--directory = my_image_directory
# Where the images are saved. Subfolders are included. .png, .jpg and .jpeg are considered images
#--model = model.safetensors
#--modelpath
# File to save the model in when trained. --model is relative to --directory, or --modelpath is full path.
# If both are set, modelpath is used.
#--scores = saved_scores.csv
# Where to find the scores for the images. Can be a .json file or a .csv file
#--savefile = saved_scores.csv
# Where to save final scores as a .csv file which includes the score predicted by the best model, and the train/eval split
#--name = my_training_run
# Prefix for study name in optuna. Studies are saved in db.sqlite
#--server={on,daemon,off}
# Start an optuna dashboard server. Use daemon to start in daemon thread (terminates with training), or off for no server
#--database = sqlite:///db.sqlite
# Storage for optuna. Set --database= for no database (implies --server=off)
#
# Defining the model architecture:
#
#--model_seed=MODEL_SEED
# Seed for initialising model (default none)
#--min_first_layer_size=MIN_FIRST_LAYER_SIZE
# Minimum number of features in first hidden layer (default 10)
#--max_first_layer_size=MAX_FIRST_LAYER_SIZE
# Maximum number of features in first hidden layer (default 1000)
#--min_second_layer_size=MIN_SECOND_LAYER_SIZE
# Minimum number of features in second hidden layer (default 10)
#--max_second_layer_size=MAX_SECOND_LAYER_SIZE
# Maximum number of features in second hidden layer (default 1000)
#
# Feature extraction:
#
#--feature_extractor_model=FEATURE_EXTRACTOR_MODEL
# Model to use for feature extraction
#--hidden_states_used=HIDDEN_STATES_USED
# Comma separated list of the hidden states to extract features from (0 is output layer, 1 is last hidden layer etc.). Model default used if not specified.
#--hidden_states_mode={"default", "join", "average", "weight"}
# Combine multiple layers from feature extractor by join, average, or weight. Model default used if not specified.
#--fp16_features
# Store features in fp16. Significantly reduces VRAM usage
#
# Training constants:
#
#--loss_model={mse,ab,nll,wmse}
# Loss model (default mse) (mse=mean square error, ab=ab ranking, nll=negative log likelihood, wmse=weighted mse)
#--set_for_scoring={eval,full,train}
# Image set to be used for scoring a model when trained (default eval)
#--metric_for_scoring={mse,ab,nll,wmse,spearman,pearson,accuracy}
# Metric to be used for scoring a model when trained (default is the loss_model)
#--calculate_ab
#--calculate_mse
#--calculate_wmse
#--calculate_spearman
#--calculate_pearson
#--calculate_accuracy
# Calculate the specified metric. Metrics required for loss or for scoring are calculated as required even if not specified
#--accuracy_divider=ACCURACY_DIVIDER
# Divider between 'high' and 'low' for accuracy. If not specified the median score is used.
#--normalise_weights
# If using weights, normalise them to a mean value of 1.0
#--fraction_for_eval=FRACTION_FOR_EVAL
# fraction of images to be reserved for eval (default 0.25)
#--eval_pick_seed=EVAL_PICK_SEED
# Seed for random numbers when choosing eval images (default 42)
#--ignore_existing_split
# Default behaviour if the scorefile includes splits is to use them. Use this to discard existing split and generate a new eval set
#
# Metaparameters:
#
#--name=NAME
# Name prefix for Optuna
#--trials=TRIALS
# Number of metaparameter trials
#--sampler={CmaEs,random,QMC}
# Metaparameter search algorithm
#--min_train_epochs=MIN_TRAIN_EPOCHS
# (default 5)
#--max_train_epochs=MAX_TRAIN_EPOCHS
# (default 50)
#--min_warmup_ratio=MIN_WARMUP_RATIO
# (default 0.0)
#--max_warmup_ratio=MAX_WARMUP_RATIO
# (default 0.2)
#--min_log_lr=MIN_LOG_LR
# (default -4.5)
#--max_log_lr=MAX_LOG_LR
# (default -2.5)
#--min_log_weight_lr=MIN_LOG_WEIGHT_LR
# (default -4.5)
#--max_log_weight_lr=MAX_LOG_WEIGHT_LR
# (default -2.5)
#--min_batch_size=MIN_BATCH_SIZE
# (default 1)
#--max_batch_size=MAX_BATCH_SIZE
# (default 100)
#--min_dropout=MIN_DROPOUT
# Minimum dropout between two hidden layers (default 0.0)
#--max_dropout=MAX_DROPOUT
# Maximum dropout between two hidden layers (default 0.8)
#--min_input_dropout=MIN_INPUT_DROPOUT
# Minimum dropout between features and first hidden layer (default 0.0)
#--max_input_dropout=MAX_INPUT_DROPOUT
# Maximum dropout between features and first hidden layer (default 0.8)
#--min_output_dropout=MIN_OUTPUT_DROPOUT
# Minimum dropout before final projection (default 0.0)
#--max_output_dropout=MAX_OUTPUT_DROPOUT
# Maximum dropout before final projection (default 0.0)