[3.3] Vanilla UNet w/ DP w/ Attn

[gauravkuppa]

Description

Explain why we're running this and what we expect. We are running this experiment to see affects of attention on UNet; quantify these effects

Planned Start Date: 9/7/20

Depends on Previous Experiment? N

Train Command

python train.py \
--name vanilla_dp_unet_attn \
--model unet \
--batch 4 \
--person_inputs densepose agnostic \
--cloth_inputs cloth \
--val_check_interval 0.2 \
--self_attn \
--accumulated_batches 16

Report Results

To report a result, copy this into a comment below:

# Result Description
<!--- 
For Experiment Number, use "Major.minor.patch", e.g. 1.2.0.
Major.minor should match the [M.m] in the title. 
Patch describes a bug fix (change in the code or branch).
-->
**Experiment Number:** 1.2.0
**Branch:** `master`
**Timestamp:** MM/DD/YYYY 9pm PT
**Epochs:** 

# Architecture
**Model Layers:**
<!-- Paste the printed Model Layers -->

**Module Parameters:**
<!-- Paste the Params table -->

# Loss Graphs
<!--- Put detailed loss graphs here. Please include all graphs! -->

# Image Results
<!--- Put detailed image results here. Please include all images! Multiple screenshots is good. -->

# Comments, Observations, or Insights
<!--- Optional -->

• [x] Open GitHub Issue
• [x] Start training with tmux (tensorboard and training)
• [x] Upload scalars, train, and validation images to GitHub
• [ ] Upload checkpoints to Google Drive
• [ ] Generate test results from latest epoch
• [ ] Calculate metrics (PSNR, SSIM)
• [ ] Visualize metrics

[gauravkuppa]

Result Description

Experiment Number: 3.3.0 Branch: unet_training Timestamp: 09/09/2020 1pm PT Epochs: 10

Architecture

Model Layers:

Module Parameters:

Loss Graphs

Image Results

Comments, Observations, or Insights

Validation results are not conclusive. No diversity in validation examples.. kind of concerning. Attention seems to help, but includes some artifacts from masks (Row 3, Col 1) and (Row 2, Col 3), which Experiment 3.2 does not.

[andrewjong]

@gauravkuppa Mind if we follow the format under the "Report Results" heading? It's our checklist to make sure we cover everything.

[gauravkuppa]

Test Command

python test.py \
--model unet \
--datamode test \
--checkpoint /data_hdd/gaurav/wacv_unet_experiments/checkpoints/3.3/version_1/checkpoints/epoch\=9_global_step\=0_val_loss\=0.00.ckpt --person_inputs densepose agnostic \
--cloth_inputs cloth \
--name vanilla_unet_densepose_attn \
--self_attn \
--batch 2 \
--result /data_hdd/gaurav/wacv_unet_experiments/outputs/3.3/