Tried hand on fastai wgan , and after 50 epoch it has now emerged perfectly as some weird abstract horror something.. Have a look at it .\nI dont yet know how to convert this output and save as images , So feel free to copy, fork and add that part for submission . But keep the kernel public . LoL.
\n\nhttps://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation
", "messages": [ { "id": "566903", "postDate": "07/02/2019 18:51:12", "content": "Tried hand on fastai wgan , and after 50 epoch it has now emerged perfectly as some weird abstract horror something.. Have a look at it .\nI dont yet know how to convert this output and save as images , So feel free to copy, fork and add that part for submission . But keep the kernel public . LoL.
\n\nhttps://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation
", "rawMarkdown": "Tried hand on fastai wgan , and after 50 epoch it has now emerged perfectly as some weird abstract horror something.. Have a look at it .\nI dont yet know how to convert this output and save as images , So feel free to copy, fork and add that part for submission . But keep the kernel public . LoL.\n\n[https://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation](https://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation)", "votes": null }, { "id": "567007", "postDate": "07/02/2019 23:17:26", "content": "That's a good start. The scenes look like nature landscapes. Now you just need the dogs to show up :-)
", "rawMarkdown": "That's a good start. The scenes look like nature landscapes. Now you just need the dogs to show up :-)", "votes": null }, { "id": "567046", "postDate": "07/03/2019 01:56:09", "content": "Lol Yes ...May be something else will show up :D
", "rawMarkdown": "Lol Yes ...May be something else will show up :D", "votes": null }, { "id": "567206", "postDate": "07/03/2019 07:09:04", "content": "I am experimenting with it too. I got some mediocre results the first time I ran it for 30 epoch with all default settings. There were some brown dogish-looking things on some of the generated images. And lots of green blobs. Though I haven't tried to save these images yet. \nI believe 'n_extra_layers=1' should be higher for this task. I am trying 3 to 6 extra layers for now. 3 was pretty stable for me so far, 6 didn't worked out.
", "rawMarkdown": "I am experimenting with it too. I got some mediocre results the first time I ran it for 30 epoch with all default settings. There were some brown dogish-looking things on some of the generated images. And lots of green blobs. Though I haven't tried to save these images yet. \nI believe 'n_extra_layers=1' should be higher for this task. I am trying 3 to 6 extra layers for now. 3 was pretty stable for me so far, 6 didn't worked out.", "votes": null }, { "id": "567240", "postDate": "07/03/2019 08:16:59", "content": "Thanks a ton for the insight . I am very new to this whole concept itself . So trying to learn as i go. I tried to increase the next layers , but it gave me some error in the dimensions . I need to seriously look into it now . I am getting varying result for different learning rate and different transformations as well . Best of luck to us !\nI tried to figure out now some option as to how to create and save the images . Looks like learner.get_prediction works here as well to get target images generated . Ofcourse we need to transpose it afterwords. However , I am not sure if I need to process the images further after the prediction . Have to look into that .
", "rawMarkdown": "Thanks a ton for the insight . I am very new to this whole concept itself . So trying to learn as i go. I tried to increase the next layers , but it gave me some error in the dimensions . I need to seriously look into it now . I am getting varying result for different learning rate and different transformations as well . Best of luck to us !\nI tried to figure out now some option as to how to create and save the images . Looks like learner.get_prediction works here as well to get target images generated . Ofcourse we need to transpose it afterwords. However , I am not sure if I need to process the images further after the prediction . Have to look into that .", "votes": null }, { "id": "567254", "postDate": "07/03/2019 08:43:09", "content": "You can call generator directly to get images. The code from official starter kernel works just fine, you just need to replace a line or two:
\n\n```python\nfor i_batch in range(0, n_images, im_batch_size):\n gen_z = torch.randn(im_batch_size, 100, 1, 1, device=device)\n gen_images = generator(gen_z)\n images = gen_images.to(\"cpu\").clone().detach()\n images = images.numpy().transpose(0, 2, 3, 1)
\n\nfor i_image in range(gen_images.size(0)):\n save_image(gen_images[i_image, :, :, :], os.path.join(OUT_DIR, f'image_{i_batch+i_image:05d}.png'))\n```
", "rawMarkdown": "You can call generator directly to get images. The code from official starter kernel works just fine, you just need to replace a line or two:\n\n```python\nfor i_batch in range(0, n_images, im_batch_size):\n gen_z = torch.randn(im_batch_size, 100, 1, 1, device=device)\n gen_images = generator(gen_z)\n images = gen_images.to(\"cpu\").clone().detach()\n images = images.numpy().transpose(0, 2, 3, 1)\n \n for i_image in range(gen_images.size(0)):\n save_image(gen_images[i_image, :, :, :], os.path.join(OUT_DIR, f'image_{i_batch+i_image:05d}.png'))\n```", "votes": null }, { "id": "567331", "postDate": "07/03/2019 11:03:14", "content": "I might be wrong about the number of extra layers. The basic_generator and basic_critic functions already know how many layers need to be added according to the size of the output image.
", "rawMarkdown": "I might be wrong about the number of extra layers. The basic_generator and basic_critic functions already know how many layers need to be added according to the size of the output image.", "votes": null }, { "id": "567425", "postDate": "07/03/2019 13:54:58", "content": "I thought so , it was giving error based on the noise and size that i selected .. So i thought there need to be tweaking and its not as simple .
", "rawMarkdown": "I thought so , it was giving error based on the noise and size that i selected .. So i thought there need to be tweaking and its not as simple .", "votes": null }, { "id": "568629", "postDate": "07/05/2019 08:42:41", "content": "I played a bit with the fast.ai WGAN implementation and decided to drop it. \nFrom what I understand, WGANs are kinda slow and I was able to achieve much better results using public scripts in less time. And they are easier to customize. But some parts of fast.ai library might be usefull though..
", "rawMarkdown": "I played a bit with the fast.ai WGAN implementation and decided to drop it. \nFrom what I understand, WGANs are kinda slow and I was able to achieve much better results using public scripts in less time. And they are easier to customize. But some parts of fast.ai library might be usefull though..", "votes": null }, { "id": "568718", "postDate": "07/05/2019 10:20:42", "content": "Well we both are on the same boat :) .I would take the scenic images and move towards DCGAN or so . May be torch implementation of WGAN-GP is another one in future I want to try .
", "rawMarkdown": "Well we both are on the same boat :) .I would take the scenic images and move towards DCGAN or so . May be torch implementation of WGAN-GP is another one in future I want to try .", "votes": null } ], "comments": [ { "id": 567007, "author_name": "cdeotte", "author_url": "", "post_date": "07/02/2019 23:17:26", "content": "That's a good start. The scenes look like nature landscapes. Now you just need the dogs to show up :-)
", "votes": null, "replies": [ { "id": 567046, "author_name": "phoenix9032", "author_url": "", "post_date": "07/03/2019 01:56:09", "content": "Lol Yes ...May be something else will show up :D
", "votes": null, "replies": [] } ] }, { "id": 567206, "author_name": "dvorobiev", "author_url": "", "post_date": "07/03/2019 07:09:04", "content": "I am experimenting with it too. I got some mediocre results the first time I ran it for 30 epoch with all default settings. There were some brown dogish-looking things on some of the generated images. And lots of green blobs. Though I haven't tried to save these images yet. \nI believe 'n_extra_layers=1' should be higher for this task. I am trying 3 to 6 extra layers for now. 3 was pretty stable for me so far, 6 didn't worked out.
", "votes": null, "replies": [ { "id": 567240, "author_name": "phoenix9032", "author_url": "", "post_date": "07/03/2019 08:16:59", "content": "Thanks a ton for the insight . I am very new to this whole concept itself . So trying to learn as i go. I tried to increase the next layers , but it gave me some error in the dimensions . I need to seriously look into it now . I am getting varying result for different learning rate and different transformations as well . Best of luck to us !\nI tried to figure out now some option as to how to create and save the images . Looks like learner.get_prediction works here as well to get target images generated . Ofcourse we need to transpose it afterwords. However , I am not sure if I need to process the images further after the prediction . Have to look into that .
", "votes": null, "replies": [] }, { "id": 567254, "author_name": "dvorobiev", "author_url": "", "post_date": "07/03/2019 08:43:09", "content": "You can call generator directly to get images. The code from official starter kernel works just fine, you just need to replace a line or two:
\n\n```python\nfor i_batch in range(0, n_images, im_batch_size):\n gen_z = torch.randn(im_batch_size, 100, 1, 1, device=device)\n gen_images = generator(gen_z)\n images = gen_images.to(\"cpu\").clone().detach()\n images = images.numpy().transpose(0, 2, 3, 1)
\n\nfor i_image in range(gen_images.size(0)):\n save_image(gen_images[i_image, :, :, :], os.path.join(OUT_DIR, f'image_{i_batch+i_image:05d}.png'))\n```
", "votes": null, "replies": [] }, { "id": 567331, "author_name": "dvorobiev", "author_url": "", "post_date": "07/03/2019 11:03:14", "content": "I might be wrong about the number of extra layers. The basic_generator and basic_critic functions already know how many layers need to be added according to the size of the output image.
", "votes": null, "replies": [] }, { "id": 567425, "author_name": "phoenix9032", "author_url": "", "post_date": "07/03/2019 13:54:58", "content": "I thought so , it was giving error based on the noise and size that i selected .. So i thought there need to be tweaking and its not as simple .
", "votes": null, "replies": [] }, { "id": 568629, "author_name": "dvorobiev", "author_url": "", "post_date": "07/05/2019 08:42:41", "content": "I played a bit with the fast.ai WGAN implementation and decided to drop it. \nFrom what I understand, WGANs are kinda slow and I was able to achieve much better results using public scripts in less time. And they are easier to customize. But some parts of fast.ai library might be usefull though..
", "votes": null, "replies": [] }, { "id": 568718, "author_name": "phoenix9032", "author_url": "", "post_date": "07/05/2019 10:20:42", "content": "Well we both are on the same boat :) .I would take the scenic images and move towards DCGAN or so . May be torch implementation of WGAN-GP is another one in future I want to try .
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "566903": "Tried hand on fastai wgan , and after 50 epoch it has now emerged perfectly as some weird abstract horror something.. Have a look at it .\nI dont yet know how to convert this output and save as images , So feel free to copy, fork and add that part for submission . But keep the kernel public . LoL.\n\n[https://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation](https://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation)", "567007": "That's a good start. The scenes look like nature landscapes. Now you just need the dogs to show up :-)", "567046": "Lol Yes ...May be something else will show up :D", "567206": "I am experimenting with it too. I got some mediocre results the first time I ran it for 30 epoch with all default settings. There were some brown dogish-looking things on some of the generated images. And lots of green blobs. Though I haven't tried to save these images yet. \nI believe 'n_extra_layers=1' should be higher for this task. I am trying 3 to 6 extra layers for now. 3 was pretty stable for me so far, 6 didn't worked out.", "567240": "Thanks a ton for the insight . I am very new to this whole concept itself . So trying to learn as i go. I tried to increase the next layers , but it gave me some error in the dimensions . I need to seriously look into it now . I am getting varying result for different learning rate and different transformations as well . Best of luck to us !\nI tried to figure out now some option as to how to create and save the images . Looks like learner.get_prediction works here as well to get target images generated . Ofcourse we need to transpose it afterwords. However , I am not sure if I need to process the images further after the prediction . Have to look into that .", "567254": "You can call generator directly to get images. The code from official starter kernel works just fine, you just need to replace a line or two:\n\n```python\nfor i_batch in range(0, n_images, im_batch_size):\n gen_z = torch.randn(im_batch_size, 100, 1, 1, device=device)\n gen_images = generator(gen_z)\n images = gen_images.to(\"cpu\").clone().detach()\n images = images.numpy().transpose(0, 2, 3, 1)\n \n for i_image in range(gen_images.size(0)):\n save_image(gen_images[i_image, :, :, :], os.path.join(OUT_DIR, f'image_{i_batch+i_image:05d}.png'))\n```", "567331": "I might be wrong about the number of extra layers. The basic_generator and basic_critic functions already know how many layers need to be added according to the size of the output image.", "567425": "I thought so , it was giving error based on the noise and size that i selected .. So i thought there need to be tweaking and its not as simple .", "568629": "I played a bit with the fast.ai WGAN implementation and decided to drop it. \nFrom what I understand, WGANs are kinda slow and I was able to achieve much better results using public scripts in less time. And they are easier to customize. But some parts of fast.ai library might be usefull though..", "568718": "Well we both are on the same boat :) .I would take the scenic images and move towards DCGAN or so . May be torch implementation of WGAN-GP is another one in future I want to try ." }, "source": "meta" }