Small tweaks, doing test run after refactor

Author: lukemelas

README.md

@@ -24,13 +24,16 @@ The pretrained weights for most GANs are downloaded automatically. For those tha
 
 There are also some standard dependencies:
  - PyTorch (tested on version 1.7.1, but should work on any version)
+ - [PyTorch Lightning](https://github.com/PyTorchLightning/pytorch-lightning)
  - [Hydra](https://github.com/facebookresearch/hydra) 1.1
- - [Albumentations](https://github.com/albumentations-team/albumentations), [Kornia](https://github.com/kornia/kornia), [Retry](https://github.com/invl/retry)
+ - [Albumentations](https://github.com/albumentations-team/albumentations)
+ - [Kornia](https://github.com/kornia/kornia)
+ - [Retry](https://github.com/invl/retry)
  - [Optional] [Weights and Biases](https://wandb.ai/)
 
  Install them with:
  ```bash
-pip install hydra-core==1.1.0dev5 albumentations tqdm retry kornia
+pip install hydra-core==1.1.0dev5 pytorch_lightning albumentations tqdm retry kornia
  ```
 
 
@@ -154,7 +157,7 @@ In the example commands below, we use BigBiGAN. You can easily switch out BigBiG
 ```bash
 PYTHONPATH=. python optimization/main.py data_gen/generator=bigbigan name=NAME
 ```
-The output will be saved in `outputs/optimization/fixed-BigBiGAN-NAME/DATE/`, with the final checkpoint in `latest.pth`.
+This should take less than 5 minutes to run. The output will be saved in `outputs/optimization/fixed-BigBiGAN-NAME/DATE/`, with the final checkpoint in `latest.pth`.
 
 **Segmentation with precomputed generations**
 
@@ -170,7 +173,7 @@ data_gen.save_size=1000000 \
 data_gen.kwargs.batch_size=1 \
 data_gen.kwargs.generation_batch_size=128
 ```
-This will generate 1 million image-label pairs and save them to `YOUR_OUTPUT_DIR/images`. Note that `YOUR_OUTPUT_DIR` should be an _absolute path_, not a relative one, because Hydra changes the working directory. You may also want to tune the `generation_batch_size` to maximize GPU utilization on your machine.
+This will generate 1 million image-label pairs and save them to `YOUR_OUTPUT_DIR/images`. Note that `YOUR_OUTPUT_DIR` should be an _absolute path_, not a relative one, because Hydra changes the working directory. You may also want to tune the `generation_batch_size` to maximize GPU utilization on your machine. It takes around 3-4 hours to generate 1 million images on a single V100 GPU.
 
 Once you have generated data, you can train a segmentation model:
 ```bash
@@ -179,6 +182,7 @@ name=NAME \
 data_gen=saved \
 data_gen.data.root="YOUR_OUTPUT_DIR_FROM_ABOVE"
 ```
+It takes around 3 hours on 1 GPU to complete 18000 iterations, by which point the model has converged (in fact you can probably get away with fewer steps, I would guess around ~5000). 
 
 **Segmentation with on-the-fly generations** 
 

src/config/segment.yaml

@@ -57,7 +57,7 @@ dataloader:
 trainer:
   # See https://pytorch-lightning.readthedocs.io/en/stable/trainer.html#trainer-flags
   gpus: 1
-  max_steps: 12000
+  max_steps: 18000
   accelerator: null  # "ddp_spawn"
   num_sanity_val_steps: 5 
   fast_dev_run: False

src/models/__init__.py

@@ -1 +1,2 @@
 from .unet_model import UNet
+from .latent_shift_model import MODELS

src/segmentation/generate_and_save.py

@@ -46,7 +46,7 @@ def run(cfg: DictConfig):
             for output_dict in output_batch:
                 img = tensor_to_image(output_dict['img'])
                 mask = tensor_to_mask(output_dict['mask'])
-                y = int(output_dict['y'])
+                y = int(output_dict['y']) if 'y' in output_dict else 0
                 stem = f'{i:08d}-seed_{cfg.seed}-class_{y:03d}'
                 img.save(save_dir / f'{stem}.jpg')
                 mask.save(save_dir / f'{stem}.png')

src/segmentation/main.py

@@ -8,8 +8,8 @@
 import logging
 import hydra
 
-from . import utils
-from . import metrics
+from segmentation import utils
+from segmentation import metrics
 from models import UNet
 from datasets import SegmentationDataset, create_gan_dataset, create_train_and_val_dataloaders
 
@@ -189,7 +189,7 @@ def main(cfg: DictConfig):
         ]
 
         # Logging
-        logger = pl.loggers.WandbLogger(name=cfg.name) if cfg.wanbd else True
+        logger = pl.loggers.WandbLogger(name=cfg.name) if cfg.wandb else True
 
         # Trainer
         trainer = pl.Trainer(logger=logger, callbacks=callbacks, **cfg.trainer)