-
Download the
tsv databases for train/val/test
andgeodesic dist. matrix
from the RICH website. Unzip the zip file and place them underdatasets
folder following the structure below.${REPO_DIR} |-- models |-- metro |-- datasets | |-- rich_for_bstro_tsv_db | | |-- train.img.tsv | | |-- train.hw.tsv | | |-- train.label.tsv | | |-- ... | | |-- test.img.tsv | | |-- test.hw.tsv | | |-- test.label.tsv | | |-- ... | | |-- val.img.tsv | | |-- val.hw.tsv | | |-- val.label.tsv | | |-- ... | |-- smpl_neutral_geodesic_dist.npy |-- predictions |-- README.md |-- ...
-
Run the following command
bash scripts/training.sh 32 # 32 is the batch size.
We trained BSTRO with a GPU with 32GB memory. Please adjust this number according to your GPUs.
One should see the following output:
... METRO INFO: Update config parameter num_hidden_layers: 12 -> 4 METRO INFO: Update config parameter hidden_size: 768 -> 1024 METRO INFO: Update config parameter num_attention_heads: 12 -> 4 METRO INFO: Init model from scratch. METRO INFO: Update config parameter num_hidden_layers: 12 -> 4 METRO INFO: Update config parameter hidden_size: 768 -> 256 METRO INFO: Update config parameter num_attention_heads: 12 -> 4 METRO INFO: Init model from scratch. METRO INFO: Update config parameter num_hidden_layers: 12 -> 4 METRO INFO: Update config parameter hidden_size: 768 -> 128 METRO INFO: Update config parameter num_attention_heads: 12 -> 4 ... => loading pretrained model models/hrnet/hrnetv2_w64_imagenet_pretrained.pth METRO INFO: => loading hrnet-v2-w64 model METRO INFO: Transformers total parameters: 102256130 METRO INFO: Backbone total parameters: 128059944 METRO INFO: Loading state dict from checkpoint models/metro_release/metro_3dpw_state_dict.bin METRO INFO: => initializing with metro weights from models/metro_release/metro_3dpw_state_dict.bin ... rich_for_bstro_tsv_db/train.yaml rich_for_bstro_tsv_db/val.yaml ...
The log file
log.txt
and intermediate training files can be found under${REPO_DIR}/output
.