EfficientDet (TF2)
With EfficientDet, the following tasks are supported:
dataset_convert
train
evaluate
prune
inference
export
These tasks may be invoked from the TAO Launcher by following the below convention from command line:
tao model efficientdet_tf2 <sub_task> <args_per_subtask>
Where args_per_subtask
are the command line arguments required for a given subtask. Each
of these sub-tasks are explained in detail below.
EfficientDet expects directories of images for training or validation and annotation JSON files in COCO format. See the Data Annotation Format page for more information about the data format for EfficientDet.
The raw image data and the corresponding annotation file need to be converted to TFRecords
before training and evaluation. The dataset_convert
tool helps to achieve seamless
conversion while providing insight on potential issues in an annotation file.
The following sections detail how to use dataset_convert
.
Sample Usage of the Dataset Converter Tool
The dataset_convert
tool is described below:
tao model efficientdet_tf2 dataset-convert [-h] -e <conversion spec file>
Below is a sample for the data conversion spec file. The format of the spec file is YAML, with configuration parameters under dataset_convert.
dataset_convert:
image_dir: '/workspace/tao-experiments/data/raw-data/train2017/'
annotations_file: '/workspace/tao-experiments/data/raw-data/annotations/instances_train2017.json'
output_dir: '/workspace/tao-experiments/data'
tag: 'train'
num_shards: 256
include_masks: True
The details of each parameter are summarized in the table below:
Field | Description | Data Type and Constraints | Recommended/Typical Value |
image_dir | The path to the directory where raw images are stored | String | – |
annotations_file | The path to the annotation JSON file | String | – |
output_dir | The output directory where TFRecords are saved | String | – |
tag | The number of shards for the converted TFRecords | Integer | 256 |
num_shards | The path to a TAO pruned model for re-training, if any | String | – |
include_mask | Whether to include segmentation groundtruth during conversion | Boolean | False |
A log file named <tag>_warnings.json
will be generated in the output_dir
if
the bounding box of an object is out of bounds with respect to the image frame or if an object mask
is out of bounds with respect to its bounding box. The log file records the image_id
that has
problematic object IDs. For example, {"200365": {"box": [918], "mask": []}
means the bounding
box of object 918
is out of bounds in image 200365
.
The following example shows how to use the command:
tao model efficientdet_tf2 dataset_convert -i /path/to/convert.yaml
Below is a sample for the EfficientDet spec file. It has 7 major components:
dataset
, model
, train
, evaluate
, inference
,
prune
and export
config as well as
the encryption key (encryption_key
) and the results directory (results_dir
).
dataset:
loader:
prefetch_size: 4
shuffle_file: False
shuffle_buffer: 10000
cycle_length: 32
block_length: 16
max_instances_per_image: 100
skip_crowd_during_training: True
num_classes: 91
train_tfrecords:
- '/datasets/coco/train-*'
val_tfrecords:
- '/datasets/coco/val-*'
val_json_file: '/datasets/coco/annotations/instances_val2017.json'
augmentation:
rand_hflip: True
random_crop_min_scale: 0.1
random_crop_max_scale: 2
auto_color_distortion: False
auto_translate_xy: False
train:
optimizer:
name: 'sgd'
momentum: 0.9
lr_schedule:
name: 'cosine'
warmup_epoch: 5
warmup_init: 0.0001
learning_rate: 0.2
amp: True
checkpoint: "/weights/efficientnet-b0_500.tlt"
num_examples_per_epoch: 100
moving_average_decay: 0.999
batch_size: 20
checkpoint_interval: 5
l2_weight_decay: 0.00004
l1_weight_decay: 0.0
clip_gradients_norm: 10.0
image_preview: True
qat: False
random_seed: 42
pruned_model_path: ''
num_epochs: 200
model:
name: 'efficientdet-d0'
aspect_ratios: '[(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)]'
anchor_scale: 4
min_level: 3
max_level: 7
num_scales: 3
freeze_bn: False
freeze_blocks: []
input_width: 512
input_height: 512
evaluate:
batch_size: 8
num_samples: 5000
max_detections_per_image: 100
checkpoint: ''
export:
batch_size: 8
dynamic_batch_size: True
min_score_thresh: 0.4
checkpoint: ""
onnx_file: ""
inference:
checkpoint: ""
image_dir: ""
dump_label: False
batch_size: 1
prune:
checkpoint: ""
normalizer: 'max'
output_path: ""
equalization_criterion: 'union'
granularity: 8
threshold: 0.5
min_num_filters: 16
excluded_layers: []
encryption_key: 'nvidia_tlt'
results_dir: '/workspace/results_dir'
num_gpus: 1
gpu_ids: [0]
The format of the spec file is YAML. The top level structure of the spec file is summarized in the table below:
Field | Description |
dataset | Configuration related to data sources and dataloader |
model | Configuration related to model construction |
train | Configuration related to the training process |
evaluate | Configuration related to the standalone evaluation process |
prune | Configuration for pruning a trained model |
inference | Configuration for running model inference |
export | Configuration for exporting a trained model |
encryption_key | Global encryption key |
results_dir | Directory where experiment results and status logging are saved |
Training Config
The training configuration(train
) defines the parameters needed for training,
evaluation. Details are summarized in the table below.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
batch_size | The batch size for each GPU, so the effective batch size is batch_size_per_gpu * num_gpus. | Unsigned int, positive | 16 |
num_epochs | The number of epochs to train the network | Unsigned int, positive | 300 |
num_examples_per _epoch | Total number of images in the training set | Unsigned int, positive | – |
checkpoint | The path to the pretrained model, if any | String | – |
pruned_model_path | The path to a TAO pruned model for re-training, if any | String | – |
checkpoint_interval | The number of training epochs that should run per model checkpoint/validation | Unsigned int, positive | 10 |
amp | Whether to use mixed precision training | Boolean | – |
moving_average_decay | Moving average decay | Float | 0.9999 |
l2_weight_decay | L2 weight decay | Float | – |
l1_weight_decay | L1 weight decay | Float | – |
random_seed | Random seed | Unsigned int, positive | 42 |
clip_gradients_norm | Clip gradients by the norm value | Float | 5 |
qat | Enabled quantization aware training | Boolean | False |
optimizer | Optimizer configuration | ||
lr_schedule | Learning rate scheduler configuration |
The optimizer configuration(train.optimizer
) specifies the type and parameters of an optimizer.
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| Field | Description | Data Type and Constraints | Recommended/Typical Value |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| name | Optimizer name (only sgd
is supported) | String | ‘sgd’ |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| momentum | Momentum | float | 0.9 |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
The learning rate scheduler configuration(train.lr_schedule
) specifies the type and parameters of a learning rate scheduler.
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| Field | Description | Data Type and Constraints | Recommended/Typical Value |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| name | The name of the learning rate scheduler. Available options are cosine
and soft_anneal
| String | ‘cosine’ |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| warmup_epoch | The number of warmup epochs in the learning rate schedule | Unsigned int, positive | – |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| warmup_init | The initial learning rate in the warmup period | Float | – |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| learning_rate | The maximum learning rate | Float | – |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
| annealing_epoch | Start annealing to warmup_init at this point | Unsigned int, positive | – |
+———————+——————————————————————————————————-+——————————-+————————————————————————————–+
Evaluation Config
The evaluation configuration (evaluate
) defines the parameters needed for the evaluation
either during training or standalone evaluation. Details are summarized in the table below.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
checkpoint | The path to the .tlt model to be evaluated | String | – |
max_detections_per_image | The maximum number of detections to visualize | Unsigned int, positive | 100 |
batch_size | The batch size for each GPU, so the effective batch size is batch_size_per_gpu * num_gpus | Unsigned int, positive | 16 |
num_samples | The number of samples for evaluation | Unsigned int | – |
label_map | YAML file that stores index to label name mapping. (Optional) If set, per class AP metric will be calculated | String | – |
start_eval_epoch | Evaluation will not start until this epoch (Default: 1) | Unsigned int | – |
results_dir | The directory where the evaluation result is stored (Optional) | Unsigned int | – |
Inference Config
The inference configuration (inference
) defines the parameters needed for the
standalone inference with the trained .tlt
model. Details are summarized in the table below.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
checkpoint | The path to the .tlt model to run inference with | String | – |
image_dir | The path to the image directory | String | – |
output_dir | The path to the output directory where annotated images will be saved | String | – |
dump_label | Whether to dump label files in KITTI format | Boolean | – |
batch_size | Batch size to run inference with | Unsigned int | – |
min_score_thresh | Minimum confidence threshold to render the predicted bounding boxes | String | – |
label_map | YAML file that stores index to label name mapping (Optional) If set, annotated images will have class labels associated with bounding boxes | String | – |
max_boxes_to_draw | The maximum number of bounding boxes that will be rendered in the annotated images | String | – |
results_dir | The directory where the inference result is stored (Optional) | Unsigned int | – |
Dataset Config
The dataset configuration (dataset
) specifies the input data source and format. This is
used for training, evaluation. A detailed description is summarized in the table
below.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
train_tfrecords | The TFRecord path for training | String | – |
val_tfrecords | The TFRecord path for validation | String | – |
val_json_file | The annotation file path for validation | String | – |
num_classes | The number of classes. If there are N categories in the annotation, num_classes should be N+1 (background class) | Unsigned int | – |
max_instances_per_image | The maximum number of object instances to parse (default: 100) | Unsigned int | 100 |
skip_crowd_during_training | Specifies whether to skip crowd during training | Boolean | True |
loader | Data loader configuration | ||
augmentation | Data augmentation configuration |
The dataloader configuration (dataset.loader
) specifies how batches of data are fed into the model.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
prefetch_size | The image dimension in “WxH” format, where W and H indicates the dimension of the resized and padded input. | String | “512x512” |
shuffle_file | The TFRecord path for training | String | – |
shuffle_buffer | The image dimension in “WxH” format, where W and H indicates the dimension of the resized and padded input. | String | “512x512” |
cycle_length | The TFRecord path for training | String | – |
block_length | The TFRecord path for training | String | – |
The dataset.augmentation
configuration specifies the image augmentation methods used after preprocessing.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
rand_hflip | A flag specifying whether to perform random horizontal flip | Boolean | – |
random_crop_min_scale | The minimum scale of RandomCrop augmentation (default: 0.1) | Float | 0.1 |
random_crop_max_scale | The maximum scale of RandomCrop augmentation (default: 2.0) | Float | 2.0 |
auto_color_distortion | A flag to enable automatic color augmentation | Boolean | False |
auto_translate_xy | A flag to enable automatic image translation on the X/Y axis | Boolean | False |
Model Config
The model configuration (model
) specifies the model structure. A detailed description is summarized in the table below.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
model_name | EfficientDet model name | string | “efficientdet_d0” |
min_level | The minimum level of the output feature pyramid | Unsigned int | 3 (only 3 is supported) |
max_level | The maximum level of the output feature pyramid | Unsigned int | 7 (only 7 is supported) |
num_scales | The number of anchor octave scales on each pyramid level (e.g. if set to 3, the anchor scales are [2^0, 2^(1/3), 2^(2/3)]) | Unsigned int | 3 |
max_instances_per_image | The maximum number of object instances to parse (default: 100) | Unsigned int | 100 |
aspect_ratios | A list of tuples representing the aspect ratios of anchors on each pyramid level | string | “[(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)]” |
anchor_scale | Scale of the base-anchor size to the feature-pyramid stride | Unsigned int | 4 |
input_width | Input width | Unsigned int | 512 |
input_height | Input height | Unsigned int | 512 |
Pruning Config
The prune
configuration defines the pruning process for a trained model. A detailed description is summarized in the table
below.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
normalizer | Normalization method. Specify max to normalize by dividing each
norm by the maximum norm within a layer or L2 to normalize
by dividing by the L2 norm of the vector comprising all kernel norms |
String | max |
equalization_criterion | The criteria to equalize the stats of inputs to an element-wise op
layer or depth-wise conv layer. Options are arithmetic_mean
geometric_mean ,``union``, and intersection . |
String | union |
granularity | The number of filters to remove at a time | Integer | 8 |
threshold | Pruning threshold | Float | – |
min_num_filters | The minimum number of filters to keep per layer. Default: 16 | Integer | 16 |
excluded_layers | A list of layers to be excluded from pruning | List | – |
checkpoint | The path to the .tlt model file to be pruned |
String | – |
Export Config
The export
configuration contains the parameters for exporting a .tlt
model to an .onnx
model, which can be used for deployment.
Field | Description | Data Type and Constraints | Recommended/Typical Value |
batch_size | The maximum batch size of the .onnx model
if dynamic_batch_size is set to False |
Boolean | – |
dynamic_batch_size | A flag specifying whether to use dynamic batch size in
the exported .onnx model |
Boolean | True |
checkpoint | The path to the .tlt model file to be exported |
String | – |
onnx_file | The path to save the exported .onnx model |
String | False |
min_score_thresh | The confidence threshold in the NMS layer (default: 0.01) | float | – |
Train the EfficientDet model using this command:
tao model efficientdet_tf2 train [-h] -e <experiment_spec>
[results_dir=<global_results_dir>]
[model.<model_option>=<model_option_value>]
[dataset.<dataset_option>=<dataset_option_value>]
[train.<train_option>=<train_option_value>]
[num_gpus=<num GPUs>]
[gpu_ids=<gpu_index>]
Required Arguments
-e, --experiment_spec
: The experiment specification file to set up the training experiment.
Optional Arguments
model.<model_option>
: The model options.dataset.<dataset_option>
: The dataset options.train.<train_option>
: The train options.num_gpus
: The number of GPUs to be used for training in a multi-GPU scenario. The default value is 1.gpu_ids
: The indices of the GPUs to use for training. This argument can be used when the machine has multiple GPUs installed.-h, --help
: Show this help message and exit.
Input Requirement
Input size: C * W * H (where C = 1 or 3, W >= 128, H >= 128; W, H are multiples of 32)
Image format: JPG
Label format: COCO detection
Sample Usage
Here’s an example of the train
command:
tao model efficientdet_tf2 train -e /path/to/spec.yaml num_gpus=2
To run evaluation with an EfficientDet model, use this command:
tao model efficientdet_tf2 evaluate [-h] -e <experiment_spec>
evaluate.checkpoint=<model to be evaluated>
[evaluate.<evaluate_option>=<evaluate_option_value>]
[gpu_ids=<gpu_index>]
Required Arguments
-e, --experiment_spec
: The experiment spec file to set up the evaluation experiment. This should be the same as the training specification file.evaluate.checkpoint
: The.pth
model to evaluate.
Optional Arguments
evaluate.<evaluate_option>
: The evaluate options.gpu_ids
: The index of the GPU to use for evaluation. This argument can be used when the machine has multiple GPUs installed. Note that evaluation can only run on a single GPU.-h, --help
: Show this help message and exit.
Sample Usage
Here’s an example of using the evaluate
command:
tao model efficientdet_tf2 evaluate -e /path/to/spec.yaml
The inference tool for EfficientDet models can be used to visualize bboxes and generate frame-by- frame KITTI format labels on a directory of images.
tao model efficientdet_tf2 inference [-h] -e <experiment spec file>
inference.checkpoint=<model to be inferenced>
[inference.<inference_option>=<inference_option_value>]
[gpu_ids=<gpu_index>]
Required Arguments
-e, --experiment_spec
: The path to an experiment spec fileinference.checkpoint
: The.pth
model to inference.
Optional Arguments
inference.<inference_option>
: The inference options.gpu_ids
: The index of the GPU to run inference on. This argument can be used when the machine has multiple GPUs installed. Note that inference can only run on a single GPU.-h, --help
: Show this help message and exit
Sample Usage
Here’s an example of using the inference
command:
tao model efficientdet_tf2 inference -e /path/to/spec.yaml
Pruning removes parameters from the model to reduce the model size without compromising the
integrity of the model itself using the tao model efficientdet_tf2 prune
command.
The tao model efficientdet_tf2 prune
command includes these parameters:
tao model efficientdet_tf2 prune [-h] -e <experiment spec file>
prune.checkpoint=<model to be pruned>
[prune.<prune_option>=<prune_option_value>]
Required Arguments
-e, --experiment_spec
: The path to an experiment spec fileprune.checkpoint
: The.pth
model to prune.
Optional Arguments
prune.<prune_option>
: The prune options.gpu_ids
: The index of the GPU to run pruning on. This argument can be used when the machine has multiple GPUs installed. Note that inference can only run on a single GPU.-h, --help
: Show this help message and exit.
After pruning, the model needs to be retrained. See Re-training the Pruned Model for more details.
Due to the complexity of larger EfficientDet models, the pruning process will take significantly longer to finish. For example, pruning the EfficientDet-D5 model may take at least 25 minutes on a V100 server.
Using the Prune Command
Here’s an example of using the tao model efficientdet_tf2 prune
command:
tao model efficientdet_tf2 prune -e /path/to/spec.yaml
Once the model has been pruned, there might be a slight decrease in accuracy
because some previously useful weights may have been removed. To regain the accuracy,
we recommend that you retrain this pruned model over the same dataset. To do this, use
the tao model efficientdet_tf2 train
command as documented in Training the model,
with an updated spec file that points to the newly pruned model as the pretrained model file.
We recommend turning off the regularizer or reducing the weight decay in the training_config
for EfficientDet to recover the accuracy when retraining a pruned model. To do this, set the
regularizer type to NO_REG
as mentioned in the Training config
section. All the other parameters may be retained in the spec file from the previous training.
Exporting the model decouples the training process from deployment and allows conversion to TensorRT engines outside the TAO environment. TensorRT engines are specific to each hardware configuration and should be generated for each unique inference environment. The exported model may be used universally across training and deployment hardware.
Exporting the EfficientDet Model
Here’s an example of the command line arguments of the tao model efficientdet_tf2 export
command:
tao model efficientdet_tf2 export [-h] -e <path to experiment spec>
export.checkpoint=<model to export>
export.onnx_file=<onnx path>
[export.<export_option>=<export_option_value>]
[gpu_ids=<gpu_index>]
Required Arguments
-e, --experiment_spec
: The path to the spec fileexport.checkpoint
: The.pth
model to export.export.onnx_file
: The path where the.etlt
or.onnx
model is saved.
Optional Arguments
export.<export_option>
: The export options.gpu_ids
: The index of (discrete) GPUs used for exporting the model. You can specify the index of the GPU to run export if the machine has multiple GPUs installed. Note that export can only run on a single GPU.-h, --help
: Show this help message and exits.
Sample usage
Here’s a sample command to export an EfficientDet model in INT8 mode.
tao model efficientdet_tf2 export -e /path/to/spec.yaml
For TensorRT engine generation, validation, and int8 calibration, refer to the TAO Deploy documentation.
Refer to the Integrating an EfficientDet (TF1/TF2) Model page to learn more about deploying an EfficientDet TF2 model to Deepsteram.