a Tqe @s4dZddlZddlZddlZddlmZddZdS)a[ # -*- coding: utf-8 -*- ----------------------------------------------------------------------------------- # Author: Nguyen Mau Dung # DoC: 2020.08.17 # email: nguyenmaudung93.kstn@gmail.com ----------------------------------------------------------------------------------- # Description: The configurations of the project will be defined here N)EasyDictcCstjdd}|jdtddd|jdtdd d d |jd td ddd |jdtdddd |jdtdddd |jdtddd|jdddd|jdtddd|jdtdd d|jd!td"d#d|jd$td%d&d'd |jd(td%d&d)d |jd*td+d&d,d |jd-td.d&d/d |jd0td1d&d2d |jd3td4d5d|jd6td7d8d9d |jd:td;dtd?d@dAd |jdBdCtdDdEdFd |jdGtdHdIdJd |jdKdLdMdNgdOdP|jdQdRtd&dSdT|jdUdRtd&dVdT|jdWdXtdYdZ|jd[d\td]dZ|jd^dtd_dZ|jd`ddad|jdbddcd|jdddded|jdftdddgd |jdhtd%didtt|}t |j rdjndk|_ t j |_dl|_dm|_dn|_d|_d%|_dl|_do|_dp|_d+|_d.|_dp|_d+|_|j|j|j|j|jdq|_d|_tj|j drds|_!tj|j dt|j"|_#tj|j du|j"|_$tj%|j#sft&|j#tj%|j$st&|j$|S)vNz The Implementation using PyTorch) descriptionz--seediz're-produce the results with seed random)typedefaulthelpz --saved_fnZ fpn_resnet_18ZFNz*The name using for saving logs, models,...)rrmetavarrz --root-dirz../PATHzThe ROOT working directoryz--archARCHz"The name of the model architecturez--pretrained_pathz%the path of the pretrained checkpointz --hflip_probg?z"The probability of horizontal flipz--no-val store_truez/If true, dont evaluate the model on the val set)actionrz --num_samplesz-Take a subset of the dataset to run and debugz --num_workersz"Number of threads for loading dataz --batch_sizezmini-batch size (default: 16), this is the totalbatch size of all GPUs on the current node when usingData Parallel or Distributed Data Parallelz --print_freq2Nzprint frequency (default: 50)z--tensorboard_freqz-frequency of saving tensorboard (default: 50)z--checkpoint_freqz,frequency of saving checkpoints (default: 5)z --start_epochzthe starting epochz --num_epochsi,znumber of total epochs to runz --lr_typecosinzFthe type of learning rate scheduler (cosin or multi_step or one_cycle)z--lrgMbP?LRzinitial learning ratez --minimum_lrgHz>ZMIN_LRz%minimum learning rate during trainingz --momentumg|?5^?Mmomentumz-wdz--weight_decaygZWDzweight decay (default: 0.)z--optimizer_typeadamZ OPTIMIZERz,the type of optimizer, it can be sgd or adamz--steps*znumber of burn in step)nargsrrz --world-sizez(number of nodes for distributed training)rrrrz--rankz"node rank for distributed trainingz --dist-urlztcp://127.0.0.1:29500z'url used to set up distributed training)rrrz--dist-backendncclzdistributed backendz --gpu_idxzGPU index to use.z --no_cudazIf true, cuda is not used.z--multiprocessing-distributedzUse multi-processing distributed training to launch N processes per node, which has N GPUs. This is the fastest way to use PyTorch for either single node or multi node data parallel trainingz --evaluatez%only evaluate the model, not trainingz --resume_pathz"the path of the resumed checkpointz--Kzthe number of top KcpucudaT)`r)r @)hm_cen cen_offset directionz_coordimdatasetZkitti checkpointslogs)'argparseArgumentParser add_argumentintstrfloatedictvars parse_argstorchdeviceZno_cudar device_countngpus_per_node pin_memory input_sizehm_sizeZ down_ratio max_objectsimagenet_pretrained head_conv num_classesZnum_center_offsetZnum_zZnum_dimZ num_directionheadsZnum_input_featuresospathjoinroot_dir dataset_dirsaved_fncheckpoints_dirlogs_dirisdirmakedirs)parserconfigsrLJ/home/opencvuniv/work/pranav/ADAS_2_LIDAR/SFA3D/sfa/config/train_config.pyparse_train_configss                              rN)__doc__r@r+r4Zeasydictrr1rNrLrLrLrMs