#!/bin/bash export TPU_IP_ADDRESS=10.104.22.146; #export TPU_IP_ADDRESS=10.95.66.34; export XRT_TPU_CONFIG="tpu_worker;0;$TPU_IP_ADDRESS:8470" export TPU_NAME="grpc://$TPU_IP_ADDRESS:8470" #export XRT_WORKERS="localservice:0;grpc://localhost:40934" #export XRT_DEVICE_MAP="CPU:0;/job:localservice/replica:0/task:0/device:XLA_CPU:0|GPU:0;/job:localservice/replica:0/task:0/device:XLA_GPU:0" #export PYTHONPATH=$SCRATCH/:${PYTHONPATH} #export PYTHONPATH=/gpfsscratch/rech/imi/usc19dv/lib/python3.7/site-packages:${PYTHONPATH} py=python3 #root_dir=$SCRATCH/data root_dir=data ####aistpp_60hz #data_dir=${root_dir}/scaled_features #hparams_file=aistpp_60hz/transflower_aistpp_expmap #hparams_file=aistpp_60hz/transglower_aistpp_expmap ####aistpp_20hz data_dir=${root_dir}/aistpp_20hz #exp=$1 exp=testing #exp=transglower_aistpp_expmap #exp=transglower_residual_aistpp_expmap #exp=transflower_residual_aistpp_expmap #exp=transflower_aistpp_expmap #exp=residualflower2_transflower_aistpp_expmap #exp=moglow_aistpp_expmap #hparams_file=aistpp_20hz/${exp} hparams_file=aistpp_20hz/mowgli_aistpp_expmap_testing ## Fix: needs vmapped version of transformer: #hparams_file=aistpp_20hz/residualflower2_moglow_aistpp_expmap ####dance_combined #data_dir=${root_dir}/dance_combined #exp=$1 #exp=transflower_expmap #exp=moglow_expmap #hparams_file=dance_combined/${exp} #exp=${exp}_future3_actnorm #exp=${exp}_future3 #exp=${exp}_future3 echo $exp $py training/train.py --data_dir=${data_dir} --max_epochs=1000\ --model=mowgli2 \ --do_validation \ --val_batch_size=32 \ --batch_size=32 \ --experiment_name=$exp\ --workers=$(nproc) \ --tpu_cores=8 \ --hparams_file=training/hparams/${hparams_file}.yaml \ #--continue_train \ #--load_weights_only \ #--stage2 \ #--prior_use_x_transformers \ #--output_lengths="3" \ #--max_prior_loss_weight=0.01 \ #--accelerator=ddp \ #--scales="[[16,0]]" \ # --use_rotary_pos_emb \ #--residual_scales="[[16,0]]" # --glow_norm_layer="actnorm" \ #--use_pos_emb_output \ #--gpus=2 \