Skip to content

Latest commit

 

History

History
50 lines (41 loc) · 3.12 KB

train.md

File metadata and controls

50 lines (41 loc) · 3.12 KB
Raw

Train your model on LOOK / JAAD / PIE

General parameters for training:

Parameter Description
epochs Number of epochs for training
learning_rate Learning rate for training
path Path to save the models
batch_size Batch size for training
pose Choice of poses to train with between [full, body, head]
loss Choice of loss function to use with between [BCE, Focal]
dropout Dropout rate
grad_map Enables gradient maps visualization after training. Applicable only for joints models. Choice between [yes, no].
device cuda device
optimizer Choice of optimizers between [adam, sgd]
eval_it Number of iterations for negative sampling on the validation set
multi_dataset Enables the mutli dataset training configuration. Choice between [yes, no].

Model parameters :

Parameter Description
type Choice of poses to train with between [joints, heads, heads+joints]
backbone Backbone model for heads model. Applicable only if [heads, heads+joints] selected above
fine_tune Enable finetuning. Applicable only if [heads, heads+joints] selected above
trained_on Applicable only if heads+joints selected. Must be the name of the dataset where we trained the heads model.

Dataset parameters :

This section has to be modified only if you train your model on a single-dataset configuration. For a multi dataset configuration the parameters in the Multi_dataset parameters will be considered.

Parameter Description
name Name of the dataset to train on, choice between [JAAD, LOOK, PIE]
path_data Path of the dataset folder of the selected dataset. Please refer to create_data/ folder
split Splitting strategy, applicable only if [JAAD] selected above. Choice between [scenes, instances]. Otherwise you can put anything, it will be ignored.
path_txt path to the ground truth txt files, this parameter shouldn't be modified if the dataset has been created correctly. Default: ./create_data

Multi-Dataset parameters:

Parameter Description
train_datasets Name of the datasets to train on separated by a comma wihtout a space. Examples: JAAD,PIE, LOOK,JAAD, LOOK,PIE,JAAD]
weighted Enable the weighted sampling while training. Choice between [yes, no]

LOOK-Dataset parameters:

Parameter Description
data Name of subset to train the model on. Choice between [all, Kitti, JRDB, Nuscenes]
trained_on Applicable only for evaluation