test_charades.sh

#!/bin/bash

root=/ROOT_DIR 
dataset=charades
data_type=features # features / raw
backbone=clip # clip / c3d_lstm
method=joint
end_epoch=200
bs=16
# num_videos
# train/test: 5338/1334
# num_proposals
# train/val/test: 12408/3720
lr=1e-4
lr_drop_step=50
num_input_frames=64
num_input_sentences=4
# num_sentences
# (train) avg: 2.324 / max:10 / min: 1 
# (test) avg: 2.789 / max: 12 / min: 1
enc_layers=4 # base: 2 / large: 4 / huge: 6
dec_layers=4 # base: 2 / large: 4 / huge: 6
num_proposals=$((10*$num_input_sentences)) # FIXME: empirically gt*10 is good
set_cost_span=1
set_cost_giou=3
set_cost_query=2
pred_label=cos # att / sim / cos / pred
resume=/SAVE_DIR/best_model_\
${dataset}_${backbone}_${bs}b_${enc_layers}l_${num_input_frames}f_${num_proposals}q_\
${pred_label}_${set_cost_span}_${set_cost_giou}_${set_cost_query}.ckpt

# TEST
PYTHONPATH=$PYTHONPATH:. python test.py \
--root ${root} \
--dataset ${dataset} \
--data_type ${data_type} \
--backbone ${backbone} \
--method ${method} \
--num_input_frames ${num_input_frames} \
--num_input_sentences ${num_input_sentences} \
--enc_layers ${enc_layers} \
--dec_layers ${dec_layers} \
--num_proposals ${num_proposals} \
--set_cost_span ${set_cost_span} \
--set_cost_giou ${set_cost_giou} \
--set_cost_query ${set_cost_query} \
--pred_label ${pred_label} \
--resume ${resume}