fast-style-transfer-tutorial-pytorch

Simple Tutorials & Code Implementation of fast-style-transfer(Perceptual Losses for Real-Time Style Transfer and Super-Resolution, 2016 ECCV) using PyTorch. This code is based on pytorch example codes

Style Image from Battle Ground Game

Style Transfer Demo video (Left: original / Right: output)

For simplicity, i write codes in ipynb. So, you can easliy test my code.

Last update : 2019/03/05

Contributor

• hoya012

0. Requirements

python=3.5numpy
matplotlib
torch=1.0.0
torchvision
torchsummary
opencv-python

If you use google colab, you don't need to set up. Just run and run!!

1. Usage

You only run Fast-Style-Transfer-PyTorch.ipynb.

Or you can use Google Colab for free!! This is colab link.

After downloading ipynb, just upload to your google drive. and run!

2. Tutorial & Code implementation Blog Posting (Korean Only)

“Fast Style Transfer PyTorch Tutorial”

3. Dataset download

For simplicty, i use COCO 2017 validation set instead of COCO 2014 training set.

• COCO 2014 training: about 80000 images / 13GB

• COCO 2017 validation: about 5000 images / 1GB –> i will use training epoch multiplied by 16 times

You can download COCO 2017 validation dataset in this link

4. Link to google drive and upload files to google drive

If you use colab, you can simply link ipynb to google drive.

from google.colab import drive
drive.mount("/content/gdrive")

Upload COCO dataset & Style Image & Test Image or Videos to Your Google Drive.

You can use google drive location in ipynb like this codes.

style_image_location = "/content/gdrive/My Drive/Colab_Notebooks/data/vikendi.jpg"style_image_sample = Image.open(style_image_location, 'r')
display(style_image_sample)

5. Transfer learning, inference from checkpoint.

Since google colab only uses the GPU for 8 hours, we need to restart it from where it stopped.

To do this, the model can be saved as a checkpoint during training, and then the learning can be done. Also, you can also use trained checkpoints for inferencing.

transfer_learning = False # inference or training first --> False / Transfer learning --> Trueckpt_model_path = os.path.join(checkpoint_dir, "ckpt_epoch_63_batch_id_500.pth")if transfer_learning:
  checkpoint = torch.load(ckpt_model_path, map_location=device)
  transformer.load_state_dict(checkpoint['model_state_dict'])
  transformer.to(device)

6. Training phase

if running_option == "training":  if transfer_learning:
      transfer_learning_epoch = checkpoint['epoch'] 
  else:
      transfer_learning_epoch = 0

  for epoch in range(transfer_learning_epoch, num_epochs):
        transformer.train()
        agg_content_loss = 0.
        agg_style_loss = 0.
        count = 0

        for batch_id, (x, _) in enumerate(train_loader):
            n_batch = len(x)
            count += n_batch
            optimizer.zero_grad()

            x = x.to(device)
            y = transformer(x)

            y = normalize_batch(y)
            x = normalize_batch(x)

            features_y = vgg(y)
            features_x = vgg(x)

            content_loss = content_weight * mse_loss(features_y.relu2_2, features_x.relu2_2)

            style_loss = 0.            for ft_y, gm_s in zip(features_y, gram_style):
                gm_y = gram_matrix(ft_y)
                style_loss += mse_loss(gm_y, gm_s[:n_batch, :, :])
            style_loss *= style_weight

            total_loss = content_loss + style_loss
            total_loss.backward()
            optimizer.step()

            agg_content_loss += content_loss.item()
            agg_style_loss += style_loss.item()            if (batch_id + 1) % log_interval == 0:
                mesg = "{}tEpoch {}:t[{}/{}]tcontent: {:.6f}tstyle: {:.6f}ttotal: {:.6f}".format(
                    time.ctime(), epoch + 1, count, len(train_dataset),
                                  agg_content_loss / (batch_id + 1),
                                  agg_style_loss / (batch_id + 1),
                                  (agg_content_loss + agg_style_loss) / (batch_id + 1)
                )                print(mesg)            if checkpoint_dir is not None and (batch_id + 1) % checkpoint_interval == 0:
                transformer.eval().cpu()
                ckpt_model_filename = "ckpt_epoch_" + str(epoch) + "_batch_id_" + str(batch_id + 1) + ".pth"
                print(str(epoch), "th checkpoint is saved!")
                ckpt_model_path = os.path.join(checkpoint_dir, ckpt_model_filename)
                torch.save({                'epoch': epoch,                'model_state_dict': transformer.state_dict(),                'optimizer_state_dict': optimizer.state_dict(),                'loss': total_loss
                }, ckpt_model_path)

                transformer.to(device).train()

7. Test(Inference) Phase

I use video for demo. But you can use only single image. ( running_option == "test" ) The code below shows how to apply a style transfer with video as input and save the video as output.

If you download trained weight, you can test without any training!

if running_option == "test_video":    
    with torch.no_grad():
        style_model = TransformerNet()

        ckpt_model_path = os.path.join(checkpoint_dir, "ckpt_epoch_63_batch_id_500.pth")
        checkpoint = torch.load(ckpt_model_path, map_location=device)        # remove saved deprecated running_* keys in InstanceNorm from the checkpoint
        for k in list(checkpoint.keys()):            if re.search(r'ind+.running_(mean|var)$', k):                del checkpoint[k]

        style_model.load_state_dict(checkpoint['model_state_dict'])
        style_model.to(device)

        cap = cv2.VideoCapture("/content/gdrive/My Drive/Colab_Notebooks/data/mirama_demo.mp4")

        frame_cnt = 0
        
        fourcc = cv2.VideoWriter_fourcc(*'XVID') #cv2.VideoWriter_fourcc(*'MP42')
        out = cv2.VideoWriter('/content/gdrive/My Drive/Colab_Notebooks/data/mirama_demo_result.avi', fourcc, 60.0, (1920,1080))        while(cap.isOpened()):
            ret, frame = cap.read()            
            try:
              frame = frame[:,:,::-1] - np.zeros_like(frame)            except:              break
              
            print(frame_cnt, "th frame is loaded!")

            content_image = frame
            content_transform = transforms.Compose([
                transforms.ToTensor(),
                transforms.Lambda(lambda x: x.mul(255))
            ])
            content_image = content_transform(content_image)
            content_image = content_image.unsqueeze(0).to(device)

            output = style_model(content_image).cpu()            #save_image("/content/gdrive/My Drive/Colab_Notebooks/data/vikendi_video_result/" + str(frame_cnt) +".png", output[0])
            out.write(post_process_image(output[0]))
            frame_cnt += 1
            
        cap.release()
        out.release()
        cv2.destroyAllWindows()

Reference

• pytorch example code

• avi --> gif for demo

• google colaboratory