about training steps about multiview2novelview

[ehabhelmy82]

?

[ehabhelmy82]

Any reply is really appreciated

[shaohua0116]

Sorry for the delayed response.

1. It takes a long time to train the model. If I remember correctly, it takes roughly 2 or 3 days (1M iterations).
2. The input is just images.
3. I tried your command and it works just fine. Usually, this error msg means there is a data point is corrupted and cannot be read. Can you check your dataset file?
4. Please download the provided checkpoints and run the evaluation script.

[ehabhelmy82]

Thanks for your replay
One more question:
The published results in the paper calculated after how many number of iterations?
When the model stop training, after how many iterations?
Thanks

[shaohua0116]

For ShapeNet (both cars and chairs), the models were trained for roughly 300k iterations without the GAN loss and 200k more iterations with the GAN loss.

[ehabhelmy82]

I used the following training command :
python trainer.py --batch_size 8 --dataset chair --num_input 4

and the following testing command:

python evaler.py --dataset chair --data_id_list ./testing_tuple_lists/id_chair_random_elevation.txt --loss --checkpoint /data/ehab/Multiview2NovelviewMaster/train_dir/chair-default-bs_8_lr_flow_0.0001_pixel_5e-05_d_0.0001-num_input-4-20190325-150046/model-335001 --write_summary --summary_file log_chair335.txt

The recorded results in the report for only two views not 4 as follow:

Checkpoint: /data/ehab/Multiview2NovelviewMaster/train_dir/chair-default-bs_8_lr_flow_0.0001_pixel_5e-05_d_0.0001-num_input-4-20190325-150046/model-335001
Dataset: chair
Id list: ./testing_tuple_lists/id_chair_random_elevation.txt
[Final Avg Report] Total datapoint: 10000 from ./testing_tuple_lists/id_chair_random_elevation.txt
[Loss]
aggregate_improvement: 0.00000
aggregate_l1_loss: 0.34866
aggregate_report_loss_0: 0.52299
aggregate_report_loss_1: 0.52299
aggregate_report_ssim_0: 0.85165
aggregate_report_ssim_1: 0.85165
aggregate_total_loss: 0.34866
best_of_pixel_of_flow_report_loss: 0.52299
best_of_pixel_of_flow_report_ssim: 0.85165
flow_avg_report_loss_0: 0.34866
flow_avg_report_loss_1: 0.34866
flow_improvement: 0.00000
flow_l1_loss: 0.34866
flow_only_aggregate_improvement:-0.00000
flow_only_aggregate_l1_loss: 0.34866
flow_only_aggregate_report_loss_0: 0.52299
flow_only_aggregate_report_loss_1: 0.52299
flow_only_aggregate_report_ssim_0: 0.85165
flow_only_aggregate_report_ssim_1: 0.85165
flow_only_aggregate_total_loss: 0.34866
flow_total_loss: 0.34866
pixel_improvement: 0.00000
pixel_l1_loss: 0.34866
pixel_report_loss_0: 0.52299
pixel_report_loss_1: 0.52299
pixel_report_ssim_0: 0.85165
pixel_report_ssim_1: 0.85165
pixel_total_loss: 0.34866
[Time] (542.412 sec)

do you have any idea why report contains only results for two views only?
Regards

[shaohua0116]

By default, the evaler only feeds two source views to the model (which can be seen here). You need to specify --num_input 4 to evaluate the model using four source images.