Readme 30k fix (#65)

* fixes

* fix

* nota bene

* link to supmat places section

* python -> python3 fix

* eval2_gpu fix

README.md

@@ -162,6 +162,10 @@ Then download models for _perceptual loss_:
 
 
 ## Places
+
+⚠️ NB: FID/SSIM/LPIPS metric values for Places that we see in LaMa paper are computed on 30000 images that we produce in evaluation section below.
+For more details on evaluation data check [[Section 3. Dataset splits in Supplementary](https://ashukha.com/projects/lama_21/lama_supmat_2021.pdf#subsection.3.1)]  ⚠️
+
 On the host machine:
 
     # Download data from http://places2.csail.mit.edu/download.html
@@ -170,18 +174,20 @@ On the host machine:
     wget http://data.csail.mit.edu/places/places365/val_large.tar
     wget http://data.csail.mit.edu/places/places365/test_large.tar
 
-    # Unpack and etc.
+    # Unpack train/test/val data and create .yaml config for it
     bash fetch_data/places_standard_train_prepare.sh
     bash fetch_data/places_standard_test_val_prepare.sh
-    bash fetch_data/places_standard_evaluation_prepare_data.sh
 
     # Sample images for test and viz at the end of epoch
     bash fetch_data/places_standard_test_val_sample.sh
     bash fetch_data/places_standard_test_val_gen_masks.sh
 
     # Run training
-    # You can change bs with data.batch_size=10
-    python bin/train.py -cn lama-fourier location=places_standard
+    python3 bin/train.py -cn lama-fourier location=places_standard
+
+    # To evaluate trained model and report metrics as in our paper
+    # we need to sample previously unseen 30k images and generate masks for them
+    bash fetch_data/places_standard_evaluation_prepare_data.sh
 
     # Infer model on thick/thin/medium masks in 256 and 512 and run evaluation 
     # like this:
@@ -191,9 +197,10 @@ On the host machine:
     outdir=$(pwd)/inference/random_thick_512 model.checkpoint=last.ckpt
 
     python3 bin/evaluate_predicts.py \
-    $(pwd)/configs/eval_2gpu.yaml \
+    $(pwd)/configs/eval2_gpu.yaml \
     $(pwd)/places_standard_dataset/evaluation/random_thick_512/ \
-    $(pwd)/inference/random_thick_512 $(pwd)/inference/random_thick_512_metrics.csv
+    $(pwd)/inference/random_thick_512 \
+    $(pwd)/inference/random_thick_512_metrics.csv
 
 
 
@@ -216,7 +223,7 @@ On the host machine:
     bash fetch_data/celebahq_gen_masks.sh
 
     # Run training
-    python bin/train.py -cn lama-fourier-celeba data.batch_size=10
+    python3 bin/train.py -cn lama-fourier-celeba data.batch_size=10
 
     # Infer model on thick/thin/medium masks in 256 and run evaluation 
     # like this:
@@ -335,7 +342,7 @@ On the host machine:
 
 
     # Run training
-    python bin/train.py -cn lama-fourier location=my_dataset data.batch_size=10
+    python3 bin/train.py -cn lama-fourier location=my_dataset data.batch_size=10
 
     # Evaluation: LaMa training procedure picks best few models according to 
     # scores on my_dataset/val/ 
@@ -353,7 +360,7 @@ On the host machine:
 
     # metrics calculation:
     python3 bin/evaluate_predicts.py \
-    $(pwd)/configs/eval_2gpu.yaml \
+    $(pwd)/configs/eval2_gpu.yaml \
     $(pwd)/my_dataset/eval/random_<size>_512/ \
     $(pwd)/inference/my_dataset/random_<size>_512 \
     $(pwd)/inference/my_dataset/random_<size>_512_metrics.csv

fetch_data/eval_sampler.py

@@ -1,19 +1,18 @@
 import os
 import random
 
-
-val_files_path = os.path.abspath('.') + '/places_standard_dataset/original/val/'
+val_files_path           = os.path.abspath('.') + '/places_standard_dataset/original/val/'
+list_of_random_val_files = os.path.abspath('.') + '/places_standard_dataset/original/eval_random_files.txt'
 val_files      = [val_files_path + image for image in os.listdir(val_files_path)]
 
-print(f'found {len(val_files)} images in {val_files_path}')
+print(f'Sampling 30000 images out of {len(val_files)} images in {val_files_path}' + \
+      f'and put their paths to {list_of_random_val_files}')
 
-random.shuffle(val_files)
-val_files_random = val_files[0:2000]
+print('In our paper we evaluate trained models on these 30k sampled (mask,image) pairs in our paper (check Sup. mat.)')
 
-list_of_random_val_files = os.path.abspath('.') \
-+ '/places_standard_dataset/original/eval_random_files.txt'
+random.shuffle(val_files)
+val_files_random = val_files[0:30000]
 
-print(f'copying 2000 random images to {list_of_random_val_files}')
 with open(list_of_random_val_files, 'w') as fw:
     for filename in val_files_random:
         fw.write(filename+'\n')

fetch_data/places_standard_evaluation_prepare_data.sh

@@ -7,7 +7,7 @@ mkdir -p places_standard_dataset/evaluation/random_thick_256/
 mkdir -p places_standard_dataset/evaluation/random_thin_256/
 mkdir -p places_standard_dataset/evaluation/random_medium_256/
 
-# 1. sample 2000 new images
+# 1. sample 30000 new images
 OUT=$(python3 fetch_data/eval_sampler.py)
 echo ${OUT}
 

fetch_data/sampler.py

@@ -1,37 +1,40 @@
 import os
 import random
 
-test_files_path = os.path.abspath('.') + '/places_standard_dataset/original/test/'
-test_files = [test_files_path + image for image in os.listdir(test_files_path)]
-print(f'found {len(test_files)} images in {test_files_path}')
+test_files_path           = os.path.abspath('.') + '/places_standard_dataset/original/test/'
+list_of_random_test_files = os.path.abspath('.') + '/places_standard_dataset/original/test_random_files.txt'
 
-random.shuffle(test_files)
-test_files_random = test_files[0:2000]
-#print(test_files_random[0:10])
+test_files = [
+    test_files_path + image for image in os.listdir(test_files_path)
+]
 
-list_of_random_test_files = os.path.abspath('.') \
-+ '/places_standard_dataset/original/test_random_files.txt'
+print(f'Sampling 2000 images out of {len(test_files)} images in {test_files_path}' + \
+      f'and put their paths to {list_of_random_test_files}')
+print('Our training procedure will pick best checkpoints according to metrics, computed on these images.')
 
-print(f'copying 100 random images to {list_of_random_test_files}')
+random.shuffle(test_files)
+test_files_random = test_files[0:2000]
 with open(list_of_random_test_files, 'w') as fw:
     for filename in test_files_random:
         fw.write(filename+'\n')
 print('...done')
 
-# ----------------------------------------------------------------------------------
 
+# --------------------------------
 
-val_files_path = os.path.abspath('.') + '/places_standard_dataset/original/val/'
-val_files = [val_files_path + image for image in os.listdir(val_files_path)]
-print(f'found {len(val_files)} images in {val_files_path}')
+val_files_path           = os.path.abspath('.') + '/places_standard_dataset/original/val/'
+list_of_random_val_files = os.path.abspath('.') + '/places_standard_dataset/original/val_random_files.txt'
 
-random.shuffle(val_files)
-val_files_random = val_files[0:100]
+val_files = [
+    val_files_path + image for image in os.listdir(val_files_path)
+]
 
-list_of_random_val_files = os.path.abspath('.') \
-+ '/places_standard_dataset/original/val_random_files.txt'
+print(f'Sampling 100 images out of {len(val_files)} in {val_files_path} ' + \
+      f'and put their paths to {list_of_random_val_files}')
+print('We use these images for visual check up of evolution of inpainting algorithm epoch to epoch' )
 
-print(f'copying 100 random images to {list_of_random_val_files}')
+random.shuffle(val_files)
+val_files_random = val_files[0:100]
 with open(list_of_random_val_files, 'w') as fw:
     for filename in val_files_random:
         fw.write(filename+'\n')