Reorganized the repository for easy installation

YuvalNirkin · Nov 16, 2022 · 61746a4 · 61746a4
1 parent 242e11c
commit 61746a4
Show file tree

Hide file tree

Showing 68 changed files with 83 additions and 34 deletions.
diff --git a/.gitignore b/.gitignore
@@ -1,6 +1,6 @@
 .idea
 *.pyc
 __pycache__
-experiments/results
+fsgan/experiments/results
 weights
 download_fsgan_models.py
diff --git a/README.md b/README.md
@@ -24,20 +24,27 @@ every pair of faces can produce a high quality face swap.
 ## Requirements
 - High-end NVIDIA GPUs with at least 11GB of DRAM.
 - Either Linux or Windows. We recommend Linux for better performance.
-- CUDA Toolkit 10.1, CUDNN 7.5, and the latest NVIDIA driver.
-- Python 3.6+ and PyTorch 1.4.0+.
+- CUDA Toolkit 10.1+, CUDNN 7.5+, and the latest NVIDIA driver.
 
 ## Installation
-- [Ubuntu installation guide](https://github.com/YuvalNirkin/fsgan/wiki/Ubuntu-Installation-Guide)
-- [Windows installation guide](https://github.com/YuvalNirkin/fsgan/wiki/Windows-Installation-Guide)
+```Bash
+git clone https://github.com/YuvalNirkin/hyperseg
+cd hyperseg
+conda env create -f hyperseg_env.yml
+conda activate hyperseg
+pip install .    # Alternatively add the root directory of the repository to PYTHONPATH.
+```
 
 For accessing FSGAN's pretrained models and auxiliary data, please fill out
 [this form](https://docs.google.com/forms/d/e/1FAIpQLScyyNWoFvyaxxfyaPLnCIAxXgdxLEMwR9Sayjh3JpWseuYlOA/viewform?usp=sf_link).
 We will then send you a link to FSGAN's shared directory and download script.
+```Bash
+python download_fsgan_models.py   # From the repository root directory
+```
 
 ## Inference
 - [Face swapping guide](https://github.com/YuvalNirkin/fsgan/wiki/Face-Swapping-Inference)
-- [Face swapping Google Colab](inference/face_swapping.ipynb)
+- [Face swapping Google Colab](fsgan/inference/face_swapping.ipynb)
 - [Paper models guide](https://github.com/YuvalNirkin/fsgan/wiki/Paper-Models-Inference)
 
 ## Training

diff --git a/criterions/__init__.py → fsgan/__init__.py b/criterions/__init__.py → fsgan/__init__.py
diff --git a/criterions/gan_loss.py → fsgan/criterions/gan_loss.py b/criterions/gan_loss.py → fsgan/criterions/gan_loss.py
diff --git a/criterions/vgg_loss.py → fsgan/criterions/vgg_loss.py b/criterions/vgg_loss.py → fsgan/criterions/vgg_loss.py
diff --git a/datasets/appearance_map.py → fsgan/datasets/appearance_map.py b/datasets/appearance_map.py → fsgan/datasets/appearance_map.py
diff --git a/datasets/image_list_dataset.py → fsgan/datasets/image_list_dataset.py b/datasets/image_list_dataset.py → fsgan/datasets/image_list_dataset.py
diff --git a/datasets/image_seg_dataset.py → fsgan/datasets/image_seg_dataset.py b/datasets/image_seg_dataset.py → fsgan/datasets/image_seg_dataset.py
diff --git a/datasets/img_landmarks_transforms.py → fsgan/datasets/img_landmarks_transforms.py b/datasets/img_landmarks_transforms.py → fsgan/datasets/img_landmarks_transforms.py
diff --git a/datasets/img_lms_pose_transforms.py → fsgan/datasets/img_lms_pose_transforms.py b/datasets/img_lms_pose_transforms.py → fsgan/datasets/img_lms_pose_transforms.py
diff --git a/datasets/opencv_video_seq_dataset.py → fsgan/datasets/opencv_video_seq_dataset.py b/datasets/opencv_video_seq_dataset.py → fsgan/datasets/opencv_video_seq_dataset.py
diff --git a/datasets/seq_dataset.py → fsgan/datasets/seq_dataset.py b/datasets/seq_dataset.py → fsgan/datasets/seq_dataset.py
diff --git a/datasets/video_inference_dataset.py → fsgan/datasets/video_inference_dataset.py b/datasets/video_inference_dataset.py → fsgan/datasets/video_inference_dataset.py
diff --git a/...nactment/ijbc_msrunet_reenactment_attr.py → ...nactment/ijbc_msrunet_reenactment_attr.py b/...nactment/ijbc_msrunet_reenactment_attr.py → ...nactment/ijbc_msrunet_reenactment_attr.py
@@ -29,8 +29,8 @@
     batch_size = [24, 12]
     workers = 32
     pretrained = False
-    criterion_id = VGGLoss('../../weights/vggface2_vgg19_256_1_2_id.pth')
-    criterion_attr = VGGLoss('../../weights/celeba_vgg19_256_2_0_28_attr.pth')
+    criterion_id = VGGLoss('../../../weights/vggface2_vgg19_256_1_2_id.pth')
+    criterion_attr = VGGLoss('../../../weights/celeba_vgg19_256_2_0_28_attr.pth')
     criterion_gan = GANLoss(use_lsgan=True)
     generator = MultiScaleResUNet(in_nc=101, out_nc=(3, 3), flat_layers=(2, 2, 2, 2), ngf=128)
     discriminator = MultiscaleDiscriminator(use_sigmoid=True, num_D=2)

diff --git a/...t/ijbc_msrunet_reenactment_attr_no_seg.py → ...t/ijbc_msrunet_reenactment_attr_no_seg.py b/...t/ijbc_msrunet_reenactment_attr_no_seg.py → ...t/ijbc_msrunet_reenactment_attr_no_seg.py
@@ -29,8 +29,8 @@
     batch_size = [32, 16]
     workers = 32
     pretrained = False
-    criterion_id = VGGLoss('../../weights/vggface2_vgg19_256_1_2_id.pth')
-    criterion_attr = VGGLoss('../../weights/celeba_vgg19_256_2_0_28_attr.pth')
+    criterion_id = VGGLoss('../../../weights/vggface2_vgg19_256_1_2_id.pth')
+    criterion_attr = VGGLoss('../../../weights/celeba_vgg19_256_2_0_28_attr.pth')
     criterion_gan = GANLoss(use_lsgan=True)
     generator = MultiScaleResUNet(in_nc=101, out_nc=3, flat_layers=(2, 2, 2, 2), ngf=128)
     discriminator = MultiscaleDiscriminator(use_sigmoid=True, num_D=2)

diff --git a/...v_msrunet_reenactment_attr_no_seg_v2.1.py → ...v_msrunet_reenactment_attr_no_seg_v2.1.py b/...v_msrunet_reenactment_attr_no_seg_v2.1.py → ...v_msrunet_reenactment_attr_no_seg_v2.1.py
@@ -27,8 +27,8 @@
     batch_size = [48, 24]
     workers = 32
     pretrained = False
-    criterion_id = VGGLoss('../../weights/vggface2_vgg19_256_1_2_id.pth')
-    criterion_attr = VGGLoss('../../weights/celeba_vgg19_256_2_0_28_attr.pth')
+    criterion_id = VGGLoss('../../../weights/vggface2_vgg19_256_1_2_id.pth')
+    criterion_attr = VGGLoss('../../../weights/celeba_vgg19_256_2_0_28_attr.pth')
     criterion_gan = GANLoss(use_lsgan=True)
     generator = MultiScaleResUNet(in_nc=101, out_nc=3, flat_layers=(2, 2, 2, 2), ngf=128)
     discriminator = MultiscaleDiscriminator(use_sigmoid=True, num_D=2)

diff --git a/experiments/segmentation/celeba_unet.py → ...n/experiments/segmentation/celeba_unet.py b/experiments/segmentation/celeba_unet.py → ...n/experiments/segmentation/celeba_unet.py
diff --git a/...riments/swapping/ijbc_msrunet_blending.py → ...riments/swapping/ijbc_msrunet_blending.py b/...riments/swapping/ijbc_msrunet_blending.py → ...riments/swapping/ijbc_msrunet_blending.py
@@ -29,8 +29,8 @@
     batch_size = [32, 16]
     workers = 32
     pretrained = False
-    criterion_id = VGGLoss('../../weights/vggface2_vgg19_256_1_2_id.pth')
-    criterion_attr = VGGLoss('../../weights/celeba_vgg19_256_2_0_28_attr.pth')
+    criterion_id = VGGLoss('../../../weights/vggface2_vgg19_256_1_2_id.pth')
+    criterion_attr = VGGLoss('../../../weights/celeba_vgg19_256_2_0_28_attr.pth')
     criterion_gan = GANLoss(use_lsgan=True)
     generator = MultiScaleResUNet(in_nc=7, out_nc=3, flat_layers=(2, 2, 2, 2), ngf=128)
     discriminator = MultiscaleDiscriminator(use_sigmoid=True, num_D=2)

diff --git a/...ments/swapping/ijbc_msrunet_inpainting.py → ...ments/swapping/ijbc_msrunet_inpainting.py b/...ments/swapping/ijbc_msrunet_inpainting.py → ...ments/swapping/ijbc_msrunet_inpainting.py
@@ -29,8 +29,8 @@
     batch_size = [32, 16]
     workers = 32
     pretrained = False
-    criterion_id = VGGLoss('../../weights/vggface2_vgg19_256_1_2_id.pth')
-    criterion_attr = VGGLoss('../../weights/celeba_vgg19_256_2_0_28_attr.pth')
+    criterion_id = VGGLoss('../../../weights/vggface2_vgg19_256_1_2_id.pth')
+    criterion_attr = VGGLoss('../../../weights/celeba_vgg19_256_2_0_28_attr.pth')
     criterion_gan = GANLoss(use_lsgan=True)
     generator = MultiScaleResUNet(in_nc=4, out_nc=3, flat_layers=(2, 2, 2, 2), ngf=128)
     discriminator = MultiscaleDiscriminator(use_sigmoid=True, num_D=2)

diff --git a/inference/face_swapping.ipynb → fsgan/inference/face_swapping.ipynb b/inference/face_swapping.ipynb → fsgan/inference/face_swapping.ipynb
diff --git a/inference/reenact.py → fsgan/inference/reenact.py b/inference/reenact.py → fsgan/inference/reenact.py
@@ -312,7 +312,7 @@ def __call__(self, source_path, target_path, output_path=None, select_source='lo
 
 class FaceReenactmentRenderer(VideoRenderer):
     def __init__(self, display=False, verbose=0, output_crop=False, resolution=256, crop_scale=1.2,
-                 encoder_codec='avc1', separate_process=False):
+                 encoder_codec='mp4v', separate_process=False):
         self._appearance_map = None
         self._fig = None
         self._figsize = (24, 16)

diff --git a/inference/swap.py → fsgan/inference/swap.py b/inference/swap.py → fsgan/inference/swap.py
@@ -51,12 +51,12 @@
 parser.add_argument('-b', '--batch_size', default=8, type=int, metavar='N',
                     help='mini-batch size')
 parser.add_argument('-rm', '--reenactment_model', metavar='PATH',
-                    default='../weights/nfv_msrunet_256_1_2_reenactment_v2.1.pth', help='reenactment model')
-parser.add_argument('-cm', '--completion_model', default='../weights/ijbc_msrunet_256_1_2_inpainting_v2.pth',
+                    default='weights/nfv_msrunet_256_1_2_reenactment_v2.1.pth', help='reenactment model')
+parser.add_argument('-cm', '--completion_model', default='weights/ijbc_msrunet_256_1_2_inpainting_v2.pth',
                     metavar='PATH', help='completion model')
-parser.add_argument('-bm', '--blending_model', default='../weights/ijbc_msrunet_256_1_2_blending_v2.pth',
+parser.add_argument('-bm', '--blending_model', default='weights/ijbc_msrunet_256_1_2_blending_v2.pth',
                     metavar='PATH', help='blending model')
-parser.add_argument('-ci', '--criterion_id', default="vgg_loss.VGGLoss('../weights/vggface2_vgg19_256_1_2_id.pth')",
+parser.add_argument('-ci', '--criterion_id', default="vgg_loss.VGGLoss('weights/vggface2_vgg19_256_1_2_id.pth')",
                     metavar='OBJ', help='id criterion object')
 parser.add_argument('-mr', '--min_radius', default=2.0, type=float, metavar='F',
                     help='minimum distance between points in the appearance map')
@@ -72,7 +72,7 @@
                       help='number of finetune iterations')
 finetune.add_argument('-fl', '--finetune_lr', default=1e-4, type=float, metavar='F',
                       help='finetune learning rate')
-finetune.add_argument('-fb', '--finetune_batch_size', default=4, type=int, metavar='N',
+finetune.add_argument('-fb', '--finetune_batch_size', default=3, type=int, metavar='N',
                       help='finetune batch size')
 finetune.add_argument('-fw', '--finetune_workers', default=4, type=int, metavar='N',
                       help='finetune workers')
@@ -362,7 +362,7 @@ def __call__(self, source_path, target_path, output_path=None, select_source='lo
 
 class FaceSwappingRenderer(VideoRenderer):
     def __init__(self, display=False, verbose=0, output_crop=False, resolution=256, crop_scale=1.2,
-                 encoder_codec='avc1', separate_process=False):
+                 encoder_codec='mp4v', separate_process=False):
         self._appearance_map = None
         self._fig = None
         self._figsize = (24, 16)

diff --git a/models/classifier1d.py → fsgan/models/classifier1d.py b/models/classifier1d.py → fsgan/models/classifier1d.py
diff --git a/models/discriminators_pix2pix.py → fsgan/models/discriminators_pix2pix.py b/models/discriminators_pix2pix.py → fsgan/models/discriminators_pix2pix.py
diff --git a/models/hopenet.py → fsgan/models/hopenet.py b/models/hopenet.py → fsgan/models/hopenet.py
diff --git a/models/hrnet.py → fsgan/models/hrnet.py b/models/hrnet.py → fsgan/models/hrnet.py
diff --git a/models/msba.py → fsgan/models/msba.py b/models/msba.py → fsgan/models/msba.py
diff --git a/models/res_unet.py → fsgan/models/res_unet.py b/models/res_unet.py → fsgan/models/res_unet.py
diff --git a/models/res_unet_msba.py → fsgan/models/res_unet_msba.py b/models/res_unet_msba.py → fsgan/models/res_unet_msba.py
diff --git a/models/res_unet_split.py → fsgan/models/res_unet_split.py b/models/res_unet_split.py → fsgan/models/res_unet_split.py
diff --git a/models/simple_unet.py → fsgan/models/simple_unet.py b/models/simple_unet.py → fsgan/models/simple_unet.py
diff --git a/models/simple_unet_02.py → fsgan/models/simple_unet_02.py b/models/simple_unet_02.py → fsgan/models/simple_unet_02.py
@@ -128,7 +128,8 @@ def __init__(self, in_size, out_size, is_deconv):
             self.up = nn.ConvTranspose2d(in_size, out_size, kernel_size=2, stride=2)
         else:
             self.up = nn.UpsamplingBilinear2d(scale_factor=2)
-        self.conv1d = nn.Conv1d(in_size, out_size, kernel_size=(1,1))
+        # self.conv1d = nn.Conv1d(in_size, out_size, kernel_size=(1,1))
+        self.conv1d = nn.Conv2d(in_size, out_size, kernel_size=(1, 1))
 
     def forward(self, inputs1, inputs2):
         outputs2 = self.up(inputs2)

diff --git a/models/vgg.py → fsgan/models/vgg.py b/models/vgg.py → fsgan/models/vgg.py
diff --git a/preprocess/clear_cache.py → fsgan/preprocess/clear_cache.py b/preprocess/clear_cache.py → fsgan/preprocess/clear_cache.py
diff --git a/preprocess/crop_image_sequences.py → fsgan/preprocess/crop_image_sequences.py b/preprocess/crop_image_sequences.py → fsgan/preprocess/crop_image_sequences.py
diff --git a/preprocess/crop_video_sequences.py → fsgan/preprocess/crop_video_sequences.py b/preprocess/crop_video_sequences.py → fsgan/preprocess/crop_video_sequences.py
@@ -9,7 +9,7 @@
 
 
 def main(input_path, output_dir=None, cache_path=None, seq_postfix='_dsfd_seq.pkl', resolution=256, crop_scale=2.0,
-         select='all', disable_tqdm=False, encoder_codec='avc1'):
+         select='all', disable_tqdm=False, encoder_codec='mp4v'):
     cache_path = os.path.splitext(input_path)[0] + seq_postfix if cache_path is None else cache_path
     if output_dir is None:
         output_dir = os.path.splitext(input_path)[0]
@@ -140,7 +140,7 @@ def main(input_path, output_dir=None, cache_path=None, seq_postfix='_dsfd_seq.pk
                         help='selection method [all|longest]')
     parser.add_argument('-dt', '--disable_tqdm', dest='disable_tqdm', action='store_true',
                           help='if specified disables tqdm progress bar')
-    parser.add_argument('-ec', '--encoder_codec', default='avc1', metavar='STR',
+    parser.add_argument('-ec', '--encoder_codec', default='mp4v', metavar='STR',
                         help='encoder codec code')
     args = parser.parse_args()
     main(args.input, args.output, args.cache, args.seq_postfix, args.resolution, args.crop_scale, args.select,

diff --git a/preprocess/crop_video_sequences_batch.py → .../preprocess/crop_video_sequences_batch.py b/preprocess/crop_video_sequences_batch.py → .../preprocess/crop_video_sequences_batch.py
diff --git a/preprocess/detections2sequences_1euro.py → .../preprocess/detections2sequences_1euro.py b/preprocess/detections2sequences_1euro.py → .../preprocess/detections2sequences_1euro.py
diff --git a/preprocess/detections2sequences_center.py → ...preprocess/detections2sequences_center.py b/preprocess/detections2sequences_center.py → ...preprocess/detections2sequences_center.py
diff --git a/preprocess/euler_sequences.py → fsgan/preprocess/euler_sequences.py b/preprocess/euler_sequences.py → fsgan/preprocess/euler_sequences.py
diff --git a/preprocess/preprocess_video.py → fsgan/preprocess/preprocess_video.py b/preprocess/preprocess_video.py → fsgan/preprocess/preprocess_video.py
@@ -47,13 +47,13 @@
                      help='display the rendering')
 general.add_argument('-v', '--verbose', default=0, type=int, metavar='N',
                      help='verbose level')
-general.add_argument('-ec', '--encoder_codec', default='avc1', metavar='STR',
+general.add_argument('-ec', '--encoder_codec', default='mp4v', metavar='STR',
                      help='encoder codec code')
 
 detection = base_parser.add_argument_group('detection')
-detection.add_argument('-dm', '--detection_model', metavar='PATH', default='../weights/WIDERFace_DSFD_RES152.pth',
+detection.add_argument('-dm', '--detection_model', metavar='PATH', default='weights/WIDERFace_DSFD_RES152.pth',
                        help='path to face detection model')
-detection.add_argument('-db', '--det_batch_size', default=8, type=int, metavar='N',
+detection.add_argument('-db', '--det_batch_size', default=4, type=int, metavar='N',
                        help='detection batch size')
 detection.add_argument('-dp', '--det_postfix', default='_dsfd.pkl', metavar='POSTFIX',
                        help='detection file postfix')
@@ -77,7 +77,7 @@
                        help='write empty sequence lists to file')
 
 pose = base_parser.add_argument_group('pose')
-pose.add_argument('-pm', '--pose_model', default='../weights/hopenet_robust_alpha1.pth', metavar='PATH',
+pose.add_argument('-pm', '--pose_model', default='weights/hopenet_robust_alpha1.pth', metavar='PATH',
                        help='path to face pose model file')
 pose.add_argument('-pb', '--pose_batch_size', default=128, type=int, metavar='N',
                        help='pose batch size')
@@ -91,7 +91,7 @@
                   help='poses temporal smoothing kernel size')
 
 landmarks = base_parser.add_argument_group('landmarks')
-landmarks.add_argument('-lm', '--lms_model', default='../weights/hr18_wflw_landmarks.pth', metavar='PATH',
+landmarks.add_argument('-lm', '--lms_model', default='weights/hr18_wflw_landmarks.pth', metavar='PATH',
                        help='landmarks model')
 landmarks.add_argument('-lb', '--lms_batch_size', default=64, type=int, metavar='N',
                        help='landmarks batch size')
@@ -103,9 +103,9 @@
                        help='landmarks temporal smoothing kernel size')
 
 segmentation = base_parser.add_argument_group('segmentation')
-segmentation.add_argument('-sm', '--seg_model', default='../weights/celeba_unet_256_1_2_segmentation_v2.pth',
+segmentation.add_argument('-sm', '--seg_model', default='weights/celeba_unet_256_1_2_segmentation_v2.pth',
                           metavar='PATH', help='segmentation model')
-segmentation.add_argument('-sb', '--seg_batch_size', default=32, type=int, metavar='N',
+segmentation.add_argument('-sb', '--seg_batch_size', default=16, type=int, metavar='N',
                           help='segmentation batch size')
 segmentation.add_argument('-sep', '--segmentation_postfix', default='_seg.pkl', metavar='POSTFIX',
                           help='segmentation file postfix')
@@ -240,6 +240,7 @@ def process_pose(self, input_path, output_dir, seq_file_path):
 
             if os.path.isfile(curr_pose_path):
                 continue
+            assert os.path.isfile(curr_vid_path), f'The cropped sequence video file is missing: {curr_vid_path}'
             print('=> Computing face poses for video: "%s"...' % curr_vid_name)
 
             # Initialize input video

diff --git a/preprocess/produce_train_val.py → fsgan/preprocess/produce_train_val.py b/preprocess/produce_train_val.py → fsgan/preprocess/produce_train_val.py
diff --git a/preprocess/render_sequences.py → fsgan/preprocess/render_sequences.py b/preprocess/render_sequences.py → fsgan/preprocess/render_sequences.py
diff --git a/preprocess/sequence_stats.py → fsgan/preprocess/sequence_stats.py b/preprocess/sequence_stats.py → fsgan/preprocess/sequence_stats.py
diff --git a/train_blending.py → fsgan/train_blending.py b/train_blending.py → fsgan/train_blending.py
diff --git a/train_inpainting.py → fsgan/train_inpainting.py b/train_inpainting.py → fsgan/train_inpainting.py
diff --git a/train_reenactment_attr.py → fsgan/train_reenactment_attr.py b/train_reenactment_attr.py → fsgan/train_reenactment_attr.py
diff --git a/train_reenactment_attr_no_seg.py → fsgan/train_reenactment_attr_no_seg.py b/train_reenactment_attr_no_seg.py → fsgan/train_reenactment_attr_no_seg.py
diff --git a/train_reenactment_attr_no_seg_v2_1.py → fsgan/train_reenactment_attr_no_seg_v2_1.py b/train_reenactment_attr_no_seg_v2_1.py → fsgan/train_reenactment_attr_no_seg_v2_1.py
diff --git a/train_segmentation.py → fsgan/train_segmentation.py b/train_segmentation.py → fsgan/train_segmentation.py
diff --git a/utils/batch.py → fsgan/utils/batch.py b/utils/batch.py → fsgan/utils/batch.py
diff --git a/utils/bbox_utils.py → fsgan/utils/bbox_utils.py b/utils/bbox_utils.py → fsgan/utils/bbox_utils.py
diff --git a/utils/blur.py → fsgan/utils/blur.py b/utils/blur.py → fsgan/utils/blur.py
diff --git a/utils/confusionmatrix.py → fsgan/utils/confusionmatrix.py b/utils/confusionmatrix.py → fsgan/utils/confusionmatrix.py
diff --git a/utils/img_utils.py → fsgan/utils/img_utils.py b/utils/img_utils.py → fsgan/utils/img_utils.py
diff --git a/utils/iou_metric.py → fsgan/utils/iou_metric.py b/utils/iou_metric.py → fsgan/utils/iou_metric.py
diff --git a/utils/landmarks_utils.py → fsgan/utils/landmarks_utils.py b/utils/landmarks_utils.py → fsgan/utils/landmarks_utils.py
diff --git a/utils/obj_factory.py → fsgan/utils/obj_factory.py b/utils/obj_factory.py → fsgan/utils/obj_factory.py
diff --git a/utils/one_euro_filter.py → fsgan/utils/one_euro_filter.py b/utils/one_euro_filter.py → fsgan/utils/one_euro_filter.py
diff --git a/utils/seg_utils.py → fsgan/utils/seg_utils.py b/utils/seg_utils.py → fsgan/utils/seg_utils.py
diff --git a/utils/set_checkpoint_arch.py → fsgan/utils/set_checkpoint_arch.py b/utils/set_checkpoint_arch.py → fsgan/utils/set_checkpoint_arch.py
diff --git a/utils/temporal_smoothing.py → fsgan/utils/temporal_smoothing.py b/utils/temporal_smoothing.py → fsgan/utils/temporal_smoothing.py
diff --git a/utils/tensorboard_logger.py → fsgan/utils/tensorboard_logger.py b/utils/tensorboard_logger.py → fsgan/utils/tensorboard_logger.py
diff --git a/utils/utils.py → fsgan/utils/utils.py b/utils/utils.py → fsgan/utils/utils.py
diff --git a/utils/video_renderer.py → fsgan/utils/video_renderer.py b/utils/video_renderer.py → fsgan/utils/video_renderer.py
@@ -24,7 +24,7 @@ class VideoRenderer(mp.Process):
         separate_process (bool): If True, the renderer will be run in a separate process
     """
     def __init__(self, display=False, verbose=0, verbose_size=None, output_crop=False, resolution=256, crop_scale=1.2,
-                 encoder_codec='avc1', separate_process=False):
+                 encoder_codec='mp4v', separate_process=False):
         super(VideoRenderer, self).__init__()
         self._display = display
         self._verbose = verbose

diff --git a/utils/video_utils.py → fsgan/utils/video_utils.py b/utils/video_utils.py → fsgan/utils/video_utils.py
diff --git a/fsgan_env.yml b/fsgan_env.yml
@@ -0,0 +1,26 @@
+name: fsgan
+channels:
+  - pytorch
+  - conda-forge
+  - anaconda
+dependencies:
+  - python=3.9.13
+  - pip=20.3.1
+  - cudatoolkit=11.6
+  - pytorch=1.12.1
+  - torchvision=0.13.1
+  - ffmpeg=4.4.1
+  - yacs=0.1.8
+  - pip:
+      - setuptools==58.2.0
+      - torch-summary==1.4.5
+      - opencv-contrib-python==4.5.4.60
+      - tensorflow==2.7.0
+      - tqdm==4.64.1
+      - matplotlib==3.6.2
+      - ffmpeg-python==0.2.0
+      - PyYAML==6.0
+      - pandas==1.5.1
+      - seaborn==0.12.1
+      - scipy==1.9.3
+      - git+https://github.com/YuvalNirkin/face_detection_dsfd.git
diff --git a/setup.py b/setup.py
@@ -0,0 +1,14 @@
+import setuptools
+
+setuptools.setup(
+    name="fsgan",
+    version="1.0.1",
+    author="Dr. Yuval Nirkin",
+    author_email="yuval.nirkin@gmail.com",
+    description="FSGAN: Subject Agnostic Face Swapping and Reenactment",
+    long_description_content_type="text/markdown",
+    package_data={'': ['license.txt']},
+    include_package_data=True,
+    packages=setuptools.find_packages(),
+    python_requires='>=3.6',
+)