Model Training

Train

Move to the sarnn folder and start training the model. The trained weights are saved in the log folder. If you want to perform validation without training the model, download the trained weights from here and save it in the log folder.

$ cd ../sarnn/
$ python3 ./bin/train.py
[INFO] Set tag = YEAR_DAY_TIME
================================
batch_size : 5
compile : False
device : 0
epoch : 10000
heatmap_size : 0.1
img_loss : 0.1
joint_loss : 1.0
k_dim : 5
log_dir : log/
lr : 0.0001
model : sarnn
n_worker : 8
optimizer : adam
pt_loss : 0.1
rec_dim : 50
stdev : 0.1
tag : YEAR_DAY_TIME
temperature : 0.0001
vmax : 1.0
vmin : 0.0
================================
 96%|███████████████████ | 9551/10000 [1:56:01<05:31,  1.35it/s, train_loss=0.00066, test_loss=0.00111]

---

Test

Specifying a weight file as the argument of test.py will save a gif animation of the predicted image, attention points, and predicted joint angles in the output folder.

$ python3 ./bin/test.py --filename ./log/YEAR_DAY_TIME/SARNN.pth
$ ls ./output/
$ SARNN_YEAR_DAY_TIME_0.gif

---

Visualization of internal representation using PCA

Specifying a weight file as the argument of test_pca_sarnn.py will save the internal representation of the RNN as a gif animation.

$ python3 ./bin/test_pca_sarnn.py ./log/YEAR_DAY_TIME/SARNN.pth
$ ls ./output/
$ PCA_SARNN_YEAR_DAY_TIME.gif