Parameter reference for all architectures

Defaults and numeric ranges below are read from the current UI; explanations were checked against the application's adapters. After changing presets, devices or advanced YAML, the experiment's effective configuration determines the actual training values.

Use configuration placeholders are not the engine's training values; an empty path means no selection yet. Modern workers and classic Windows workers use different implementations.

Modern object detection

Parameter UI default / passed value Numeric input range Actual meaning
imgsz 640 See description (text, option or path) Single input size for modern architectures. Classification crops according to its preprocessing; detection, segmentation and oriented detection resize and pad according to the engine rules. This does not set arbitrary height and width.
epochs 1000 1～10000 Maximum training epochs; modern architectures may stop earlier through early stopping or manual cancellation.
patience 100 0～1000 Early stopping patience for modern architectures: consecutive epochs without validation metric improvement. Set 0 to disable early stopping.
workers 8 0～64 Modern architectures use data loader processes; on Windows, classic architectures use threads for parallel image reading and preprocessing when above 0, or the original loading method at 0.
batch 16 1～512 Images processed per batch; for anomaly detection, this is the feature extraction batch size. Reduce it if GPU memory is insufficient.
device Not selected / use configuration See description (text, option or path) Select the compute device. Modern auto lets the engine decide; gpu uses device 0. Classic and anomaly GPU tasks use the selected card and do not silently fall back to CPU on failure.
seed 42 0～99999 Random seed; identical seeds do not guarantee identical results across devices or environments.
amp true See description (text, option or path) Mixed precision. Modern UI: True/False. Classic UI: use configuration/off/on; explicitly enabling it requires GPU. Classic AMP is disabled on CPU.
optimizer "auto" See description (text, option or path) Select an optimizer. Modern auto overrides the manual initial learning rate and momentum and sets the warmup bias learning rate to 0; classic config uses the YAML settings.
momentum 0.937 0.0～1.0 Momentum for SGD-type optimizers; usually beta1 for Adam-type optimizers. Modern auto may override it.
weight_decay 0.0005 0.0～0.1 Weight decay coefficient. For classic architectures, the default negative value means use the configuration; negative decay is not passed to the engine.
warmup_epochs 3.0 0.0～50.0 Warmup epochs. Modern architectures accept fractions; classic architectures allow adjustment only with the cosine policy and require a value smaller than total epochs.
warmup_momentum 0.8 0.0～1.0 Initial momentum during modern detection/segmentation/oriented detection warmup.
warmup_bias_lr 0.1 0.0～1.0 Initial bias learning rate during modern detection/segmentation/oriented detection warmup; the auto optimizer sets it to 0.
degrees 0.0 0.0～180.0 Maximum absolute random rotation angle during training; 0 disables it. Classification uses the software's explicit random rotation adapter.
flipud 0.0 0.0～1.0 Probability of vertical flipping during training.
fliplr 0.5 0.0～1.0 Probability of horizontal flipping during training.
translate 0.1 0.0～1.0 Training translation magnitude relative to image size, not a pixel count.
scale 0.0 0.0～1.0 Random scaling magnitude during training, not the target input size or inference magnification.
mosaic 0.0 0.0～1.0 Mosaic probability during training; the engine's end-of-training disabling rules still apply.
box 7.5 0.0～20.0 Box localization loss weight, not a detection threshold.
cls 0.5 0.0～20.0 Object class loss weight, not the number of classes.
dfl 1.5 0.0～20.0 Distribution Focal Loss (DFL) weight; whether the loss exists depends on the model architecture.
copy_paste 0.1 0.0～1.0 Detection probabilistically copies an object within the same image to a non-overlapping location; segmentation uses the native mask copy-paste fraction. Skipped when no safe position exists; unavailable for OBB.
erasing 0.4 0.0～1.0 Detection and segmentation erase only background outside object boxes, preserving objects and annotations; OBB has no such control.
model_path "ats-m.pt" See description (text, option or path) Select preset weights matching the task. The software locates the actual files in its bundled directories.
dataset_path Not selected / use configuration See description (text, option or path) Select a dataset for the current task; see the corresponding architecture chapter for formats and training/validation rules.

Modern image classification

Parameter UI default / passed value Numeric input range Actual meaning
imgsz 960 See description (text, option or path) Single input size for modern architectures. Classification crops according to its preprocessing; detection, segmentation and oriented detection resize and pad according to the engine rules. This does not set arbitrary height and width.
epochs 500 1～10000 Maximum training epochs; modern architectures may stop earlier through early stopping or manual cancellation.
patience 100 0～1000 Early stopping patience for modern architectures: consecutive epochs without validation metric improvement. Set 0 to disable early stopping.
workers 4 0～64 Modern architectures use data loader processes; on Windows, classic architectures use threads for parallel image reading and preprocessing when above 0, or the original loading method at 0.
batch 8 1～512 Images processed per batch; for anomaly detection, this is the feature extraction batch size. Reduce it if GPU memory is insufficient.
device Not selected / use configuration See description (text, option or path) Select the compute device. Modern auto lets the engine decide; gpu uses device 0. Classic and anomaly GPU tasks use the selected card and do not silently fall back to CPU on failure.
seed 42 0～99999 Random seed; identical seeds do not guarantee identical results across devices or environments.
amp true See description (text, option or path) Mixed precision. Modern UI: True/False. Classic UI: use configuration/off/on; explicitly enabling it requires GPU. Classic AMP is disabled on CPU.
optimizer "AdamW" See description (text, option or path) Select an optimizer. Modern auto overrides the manual initial learning rate and momentum and sets the warmup bias learning rate to 0; classic config uses the YAML settings.
lr0 0.0005 1e-05～0.1 Initial learning rate for modern classification; optimizer=auto lets the engine choose it again. Detection, segmentation and oriented detection pages do not have this field.
lrf 0.01 0.0001～1.0 Final learning rate multiplier for modern classification, relative to the initial rate; not an independent absolute learning rate.
momentum 0.9 0.0～1.0 Momentum for SGD-type optimizers; usually beta1 for Adam-type optimizers. Modern auto may override it.
weight_decay 0.05 0.0～0.5 Weight decay coefficient. For classic architectures, the default negative value means use the configuration; negative decay is not passed to the engine.
warmup_epochs 5.0 0.0～50.0 Warmup epochs. Modern architectures accept fractions; classic architectures allow adjustment only with the cosine policy and require a value smaller than total epochs.
degrees 180.0 0.0～180.0 Maximum absolute random rotation angle during training; 0 disables it. Classification uses the software's explicit random rotation adapter.
flipud 0.5 0.0～1.0 Probability of vertical flipping during training.
fliplr 0.5 0.0～1.0 Probability of horizontal flipping during training.
crop_fraction 1.0 0.1～1.0 Short-edge retention fraction for modern classification. Below 1, the short edge is first enlarged to imgsz/fraction, then center-cropped. Training, validation and software inference share this central-region rule; training still adds random augmentation.
model_path "ats-m-cls.pt" See description (text, option or path) Select preset weights matching the task. The software locates the actual files in its bundled directories.
dataset_path Not selected / use configuration See description (text, option or path) Select a dataset for the current task; see the corresponding architecture chapter for formats and training/validation rules.

Modern instance segmentation

Parameter UI default / passed value Numeric input range Actual meaning
imgsz 640 See description (text, option or path) Single input size for modern architectures. Classification crops according to its preprocessing; detection, segmentation and oriented detection resize and pad according to the engine rules. This does not set arbitrary height and width.
epochs 1000 1～10000 Maximum training epochs; modern architectures may stop earlier through early stopping or manual cancellation.
patience 100 0～1000 Early stopping patience for modern architectures: consecutive epochs without validation metric improvement. Set 0 to disable early stopping.
workers 8 0～64 Modern architectures use data loader processes; on Windows, classic architectures use threads for parallel image reading and preprocessing when above 0, or the original loading method at 0.
batch 16 1～512 Images processed per batch; for anomaly detection, this is the feature extraction batch size. Reduce it if GPU memory is insufficient.
device Not selected / use configuration See description (text, option or path) Select the compute device. Modern auto lets the engine decide; gpu uses device 0. Classic and anomaly GPU tasks use the selected card and do not silently fall back to CPU on failure.
seed 42 0～99999 Random seed; identical seeds do not guarantee identical results across devices or environments.
amp true See description (text, option or path) Mixed precision. Modern UI: True/False. Classic UI: use configuration/off/on; explicitly enabling it requires GPU. Classic AMP is disabled on CPU.
optimizer "auto" See description (text, option or path) Select an optimizer. Modern auto overrides the manual initial learning rate and momentum and sets the warmup bias learning rate to 0; classic config uses the YAML settings.
momentum 0.937 0.0～1.0 Momentum for SGD-type optimizers; usually beta1 for Adam-type optimizers. Modern auto may override it.
weight_decay 0.0005 0.0～0.1 Weight decay coefficient. For classic architectures, the default negative value means use the configuration; negative decay is not passed to the engine.
warmup_epochs 3.0 0.0～50.0 Warmup epochs. Modern architectures accept fractions; classic architectures allow adjustment only with the cosine policy and require a value smaller than total epochs.
warmup_momentum 0.8 0.0～1.0 Initial momentum during modern detection/segmentation/oriented detection warmup.
warmup_bias_lr 0.1 0.0～1.0 Initial bias learning rate during modern detection/segmentation/oriented detection warmup; the auto optimizer sets it to 0.
degrees 0.0 0.0～180.0 Maximum absolute random rotation angle during training; 0 disables it. Classification uses the software's explicit random rotation adapter.
flipud 0.0 0.0～1.0 Probability of vertical flipping during training.
fliplr 0.5 0.0～1.0 Probability of horizontal flipping during training.
translate 0.1 0.0～1.0 Training translation magnitude relative to image size, not a pixel count.
scale 0.0 0.0～1.0 Random scaling magnitude during training, not the target input size or inference magnification.
mosaic 0.0 0.0～1.0 Mosaic probability during training; the engine's end-of-training disabling rules still apply.
copy_paste 0.1 0.0～1.0 Detection probabilistically copies an object within the same image to a non-overlapping location; segmentation uses the native mask copy-paste fraction. Skipped when no safe position exists; unavailable for OBB.
erasing 0.4 0.0～1.0 Detection and segmentation erase only background outside object boxes, preserving objects and annotations; OBB has no such control.
box 7.5 0.0～20.0 Box localization loss weight, not a detection threshold.
cls 0.5 0.0～20.0 Object class loss weight, not the number of classes.
dfl 1.5 0.0～20.0 Distribution Focal Loss (DFL) weight; whether the loss exists depends on the model architecture.
model_path "ats-m-seg.pt" See description (text, option or path) Select preset weights matching the task. The software locates the actual files in its bundled directories.
dataset_path Not selected / use configuration See description (text, option or path) Select a dataset for the current task; see the corresponding architecture chapter for formats and training/validation rules.

Modern oriented object detection

Parameter UI default / passed value Numeric input range Actual meaning
imgsz 640 See description (text, option or path) Single input size for modern architectures. Classification crops according to its preprocessing; detection, segmentation and oriented detection resize and pad according to the engine rules. This does not set arbitrary height and width.
epochs 1000 1～10000 Maximum training epochs; modern architectures may stop earlier through early stopping or manual cancellation.
patience 100 0～1000 Early stopping patience for modern architectures: consecutive epochs without validation metric improvement. Set 0 to disable early stopping.
workers 0 0～64 Modern architectures use data loader processes; on Windows, classic architectures use threads for parallel image reading and preprocessing when above 0, or the original loading method at 0.
batch 16 1～512 Images processed per batch; for anomaly detection, this is the feature extraction batch size. Reduce it if GPU memory is insufficient.
device Not selected / use configuration See description (text, option or path) Select the compute device. Modern auto lets the engine decide; gpu uses device 0. Classic and anomaly GPU tasks use the selected card and do not silently fall back to CPU on failure.
seed 42 0～99999 Random seed; identical seeds do not guarantee identical results across devices or environments.
amp true See description (text, option or path) Mixed precision. Modern UI: True/False. Classic UI: use configuration/off/on; explicitly enabling it requires GPU. Classic AMP is disabled on CPU.
optimizer "auto" See description (text, option or path) Select an optimizer. Modern auto overrides the manual initial learning rate and momentum and sets the warmup bias learning rate to 0; classic config uses the YAML settings.
momentum 0.937 0.0～1.0 Momentum for SGD-type optimizers; usually beta1 for Adam-type optimizers. Modern auto may override it.
weight_decay 0.0005 0.0～0.1 Weight decay coefficient. For classic architectures, the default negative value means use the configuration; negative decay is not passed to the engine.
warmup_epochs 3.0 0.0～50.0 Warmup epochs. Modern architectures accept fractions; classic architectures allow adjustment only with the cosine policy and require a value smaller than total epochs.
warmup_momentum 0.8 0.0～1.0 Initial momentum during modern detection/segmentation/oriented detection warmup.
warmup_bias_lr 0.1 0.0～1.0 Initial bias learning rate during modern detection/segmentation/oriented detection warmup; the auto optimizer sets it to 0.
degrees 0.0 0.0～180.0 Maximum absolute random rotation angle during training; 0 disables it. Classification uses the software's explicit random rotation adapter.
flipud 0.0 0.0～1.0 Probability of vertical flipping during training.
fliplr 0.5 0.0～1.0 Probability of horizontal flipping during training.
translate 0.1 0.0～1.0 Training translation magnitude relative to image size, not a pixel count.
scale 0.0 0.0～1.0 Random scaling magnitude during training, not the target input size or inference magnification.
mosaic 0.0 0.0～1.0 Mosaic probability during training; the engine's end-of-training disabling rules still apply.
box 7.5 0.0～20.0 Box localization loss weight, not a detection threshold.
cls 0.5 0.0～20.0 Object class loss weight, not the number of classes.
dfl 1.5 0.0～20.0 Distribution Focal Loss (DFL) weight; whether the loss exists depends on the model architecture.
angle 1.0 0.0～20.0 Oriented-box angle loss weight, not the augmentation rotation angle; whether it is used depends on the OBB model architecture.
model_path "ats-m-obb.pt" See description (text, option or path) Select preset weights matching the task. The software locates the actual files in its bundled directories.
dataset_path Not selected / use configuration See description (text, option or path) Select a dataset for the current task; see the corresponding architecture chapter for formats and training/validation rules.

Classic object detection 2.7.0

Parameter UI default / passed value Numeric input range Actual meaning
python "<software directory>\\Paddle\\runtime-cpu\\python.exe" See description (text, option or path) Interpreter for the independent classic runtime; automatically located from the CPU/GPU selection, or manually set to a compatible environment.
config "<software directory>\\Paddle\\PaddleDetection-2.7.0\\configs\\ppyoloe\\ppyoloe_plus_crn_s_80e_coco.yml" See description (text, option or path) Reference YAML: s/m/l/x/sod are classic detection presets; classification uses the actual preset configuration names.
output "<software directory>\\runs\\classic_det" See description (text, option or path) Classic training output root; each run creates an independent timestamped experiment subdirectory.
epochs 100 1～100000 Maximum training epochs; modern architectures may stop earlier through early stopping or manual cancellation.
batch 4 1～4096 Images processed per batch; for anomaly detection, this is the feature extraction batch size. Reduce it if GPU memory is insufficient.
seed 42 0～99999 Random seed; identical seeds do not guarantee identical results across devices or environments.
device "cpu" See description (text, option or path) Select the compute device. Modern auto lets the engine decide; gpu uses device 0. Classic and anomaly GPU tasks use the selected card and do not silently fall back to CPU on failure.
weight_mode "config" See description (text, option or path) config follows pretrained settings in the configuration; local uses local .pdparams; random initializes randomly. config may download weights.
weights Not selected / use configuration See description (text, option or path) Local .pdparams used only when local pretrained weights are selected.
dataset.mode "coco" See description (text, option or path) Classic detection defaults to coco, classification to classification; config uses the dataset settings in YAML.
dataset.root Not selected / use configuration See description (text, option or path) Classic dataset root. Select one folder; recognition rules are in the dataset chapter.
dataset.evaluate_on_train false See description (text, option or path) Disabled by default. When enabled, all training images and annotations replace the separate validation set for evaluation; metrics reflect only the fit to the training set.
training_options.image_size_mode "config" See description (text, option or path) Classic detection: config keeps configured sizes; fixed unifies training/validation/export sizes; eval changes validation/test/export sizes only.
training_options.image_height 640 96～8192 Custom classic detection height, 96–8192 and a multiple of 32; inactive in config size mode.
training_options.image_width 640 96～8192 Custom classic detection width, 96–8192 and a multiple of 32; inactive in config size mode.
training_options.workers 0 0～32 Modern architectures use data loader processes; on Windows, classic architectures use threads for parallel image reading and preprocessing when above 0, or the original loading method at 0.
training_options.learning_rate Not selected / use configuration 0.0～100.0 Initial classic learning rate; 0 in the UI means use the configuration. An empty value is passed when no override applies.
training_options.weight_decay Not selected / use configuration -1.0～10.0 Weight decay coefficient. For classic architectures, the default negative value means use the configuration; negative decay is not passed to the engine.
training_options.optimizer "config" See description (text, option or path) Select an optimizer. Modern auto overrides the manual initial learning rate and momentum and sets the warmup bias learning rate to 0; classic config uses the YAML settings.
training_options.scheduler "config" See description (text, option or path) Classic learning rate policy. config retains YAML milestones and total schedule duration; constant fixes the rate; cosine rebuilds a cosine schedule using this run's total epochs.
training_options.warmup_epochs 0 0～10000 Warmup epochs. Modern architectures accept fractions; classic architectures allow adjustment only with the cosine policy and require a value smaller than total epochs.
training_options.amp "config" See description (text, option or path) Mixed precision. Modern UI: True/False. Classic UI: use configuration/off/on; explicitly enabling it requires GPU. Classic AMP is disabled on CPU.
training_options.augmentation "config" See description (text, option or path) For classic architectures, config retains model augmentation; off removes random augmentation and uses validation preprocessing, without removing normalization.
gpu_id 0 See description (text, option or path) Physical GPU index selected for classic/anomaly detection; mapped to device 0 inside the independent process, using one GPU per run.

Classic image classification 2.5.0

Parameter UI default / passed value Numeric input range Actual meaning
python "<software directory>\\Paddle\\runtime-cpu\\python.exe" See description (text, option or path) Interpreter for the independent classic runtime; automatically located from the CPU/GPU selection, or manually set to a compatible environment.
config "<software directory>\\Paddle\\PaddleClas-2.5.0\\ppcls\\configs\\ImageNet\\ResNet\\ResNet50.yaml" See description (text, option or path) Reference YAML: s/m/l/x/sod are classic detection presets; classification uses the actual preset configuration names.
output "<software directory>\\runs\\classic_cls" See description (text, option or path) Classic training output root; each run creates an independent timestamped experiment subdirectory.
epochs 100 1～100000 Maximum training epochs; modern architectures may stop earlier through early stopping or manual cancellation.
batch 4 1～4096 Images processed per batch; for anomaly detection, this is the feature extraction batch size. Reduce it if GPU memory is insufficient.
seed 42 0～99999 Random seed; identical seeds do not guarantee identical results across devices or environments.
device "cpu" See description (text, option or path) Select the compute device. Modern auto lets the engine decide; gpu uses device 0. Classic and anomaly GPU tasks use the selected card and do not silently fall back to CPU on failure.
weight_mode "config" See description (text, option or path) config follows pretrained settings in the configuration; local uses local .pdparams; random initializes randomly. config may download weights.
weights Not selected / use configuration See description (text, option or path) Local .pdparams used only when local pretrained weights are selected.
dataset.mode "classification" See description (text, option or path) Classic detection defaults to coco, classification to classification; config uses the dataset settings in YAML.
dataset.root Not selected / use configuration See description (text, option or path) Classic dataset root. Select one folder; recognition rules are in the dataset chapter.
training_options.image_size 0 0～8192 For classic classification, square side length; 0 uses the configuration. Anomaly detection resizes the entire image directly to this square size, which must be 64–1024 and a multiple of 32.
training_options.workers 0 0～32 Modern architectures use data loader processes; on Windows, classic architectures use threads for parallel image reading and preprocessing when above 0, or the original loading method at 0.
training_options.learning_rate Not selected / use configuration 0.0～100.0 Initial classic learning rate; 0 in the UI means use the configuration. An empty value is passed when no override applies.
training_options.weight_decay Not selected / use configuration -1.0～10.0 Weight decay coefficient. For classic architectures, the default negative value means use the configuration; negative decay is not passed to the engine.
training_options.optimizer "config" See description (text, option or path) Select an optimizer. Modern auto overrides the manual initial learning rate and momentum and sets the warmup bias learning rate to 0; classic config uses the YAML settings.
training_options.scheduler "config" See description (text, option or path) Classic learning rate policy. config retains YAML milestones and total schedule duration; constant fixes the rate; cosine rebuilds a cosine schedule using this run's total epochs.
training_options.warmup_epochs 0 0～10000 Warmup epochs. Modern architectures accept fractions; classic architectures allow adjustment only with the cosine policy and require a value smaller than total epochs.
training_options.amp "config" See description (text, option or path) Mixed precision. Modern UI: True/False. Classic UI: use configuration/off/on; explicitly enabling it requires GPU. Classic AMP is disabled on CPU.
training_options.augmentation "config" See description (text, option or path) For classic architectures, config retains model augmentation; off removes random augmentation and uses validation preprocessing, without removing normalization.
gpu_id 0 See description (text, option or path) Physical GPU index selected for classic/anomaly detection; mapped to device 0 inside the independent process, using one GPU per run.

Anomaly detection

Parameter UI default / passed value Numeric input range Actual meaning
train_dir Not selected / use configuration See description (text, option or path) OK training image directory for anomaly detection; include normal images only.
calibration_dir Not selected / use configuration See description (text, option or path) Independent OK calibration directory for anomaly detection. If blank, approximately 20% is held out after content-based deduplication; calibration images do not enter the memory bank.
backbone "resnet18" See description (text, option or path) Pretrained anomaly feature backbone: resnet18 or wide_resnet50_2; backbone parameters are frozen.
image_size 256 64～1024 For classic classification, square side length; 0 uses the configuration. Anomaly detection resizes the entire image directly to this square size, which must be 64–1024 and a multiple of 32.
batch 2 1～16 Images processed per batch; for anomaly detection, this is the feature extraction batch size. Reduce it if GPU memory is insufficient.
coreset_ratio 0.01 0.001～1.0 Fraction of candidate features retained in the anomaly detection coreset; also constrained by the coreset feature limit.
max_patches 50000 100～250000 Maximum number of anomaly detection candidate features; bounded sampling applies above this limit. This is not the number of training images.
max_bank 5000 1～10000 Maximum final coreset feature count for anomaly detection, not an image or class count.
model Not selected / use configuration See description (text, option or path) Anomaly inference uses the model.patchcore.pt generated during training.
source Not selected / use configuration See description (text, option or path) Image or image folder for anomaly inference; the output directory must not be inside the input directory.
threshold_multiplier 1.0 0.1～10.0 Anomaly inference threshold multiplier: below 1 is more sensitive; above 1 is more permissive. Anomaly scores are not probabilities.
output_dir "<software directory>\\runs\\patchcore" See description (text, option or path) Anomaly detection output root; training and inference create separate experiment subdirectories.
device "cpu" See description (text, option or path) Select the compute device. Modern auto lets the engine decide; gpu uses device 0. Classic and anomaly GPU tasks use the selected card and do not silently fall back to CPU on failure.
gpu_id 0 See description (text, option or path) Physical GPU index selected for classic/anomaly detection; mapped to device 0 inside the independent process, using one GPU per run.