def setup(opts):
checkpoint = opts['checkpoint'
use_gpu = True if torch.cuda.is_available() else False
# Load the model from the Pytorch Hub
model = torch.hub.load('facebookresearch/pytorch_GAN_zoo:hub',
'PGAN', model_name=checkpoint,
pretrained=True, useGPU=use_gpu)
return model
@runway.command('generate',
inputs={ 'z': runway.vector(length=512, sampling_std=0.5)},
outputs={ 'image': runway.image })
def generate(model, inputs):
# Generate ♾ infinite ♾ images
z = inputs['z']
latents = z.reshape((1, 559))
latents = torch.from_numpy(latents)
with torch.no_grad():
generated_image = model.test(latents.float())
generated_image = generated_image.clamp(min=-1, max=1)
generated_image = ((generated_image + 1.0) * 255 / 2.0)
# Now generated_image contains our generated image!
return generated_image[0].permute(1, 2, 0).numpy().astype(np.uint8)
if __name__ == '__main__':
runway.run(port=5232)
import runway
import numpy as np
from ISR.models import RDN
@runway.setup(options={'checkpoint': runway.file(extension='.hdf5')})
def setup(opts):
rdn = RDN(arch_params={'C':6, 'D':20, 'G':64, 'G0':64, 'x':2})
rdn.model.load_weights(opts['checkpoint'])
return rdn
@runway.command('upscale', inputs={'image': runway.image}, outputs={'upscaled': runway.image})
def upscale(rdn, inputs):
width, height = inputs['image'].size
if width >= 1000 or height >= 1000:
return rdn.predict(np.array(inputs['image']), by_patch_of_size=256)
else:
return rdn.predict(np.array(inputs['image']))
if __name__ == '__main__':
runway.run(port=4231)
sess = tf.Session(graph=g)
img_placeholder = tf.placeholder(tf.float32, shape=batch_shape, name='img_placeholder')
preds = transform.net(img_placeholder)
load_checkpoint(os.path.join(options['checkpoint_path'], 'fns.ckpt'), sess)
return sess
@runway.command('stylize',
inputs={'image': runway.image(channels=4), 'background_image': runway.image(channels=4)},
outputs={'output': runway.image(channels=4)})
def stylize(sess, inp):
img = inp['image']
background_image = inp['background_image'].resize(img.size)
alpha_mask = img.getchannel("A")
img = img.convert('RGB')
original_size = img.size
img = np.array(img.resize((640, 480)))
img = np.expand_dims(img, 0)
with g.as_default():
output = sess.run(preds, feed_dict={img_placeholder: img})
output = np.clip(output[0], 0, 255).astype(np.uint8)
output = Image.fromarray(output).resize(original_size)
output.putalpha(alpha_mask)
composite = Image.alpha_composite(background_image, output)
return dict(output=composite)
if __name__ == '__main__':
runway.run(model_options={'checkpoint_path': 'models/ckpt_hokusai_b20_e4_cw15'})
outputs={ 'image': image(description='Output image') })
def generate(model, args):
content_image = args['content_image'].convert('RGB')
style_image = args['style_image'].convert('RGB')
preserve_color = args['preserve_color']
alpha = args['alpha']
print('[GENERATE] Ran with preserve_color "{}". alpha "{}"'.format(preserve_color, alpha))
vgg = model['vgg']
decoder = model['decoder']
content_tf = model['content_tf']
style_tf = model['style_tf']
content = content_tf(content_image)
style = style_tf(style_image)
if preserve_color:
style = coral(style, content)
style = style.to(device).unsqueeze(0)
content = content.to(device).unsqueeze(0)
with torch.no_grad():
output = style_transfer(vgg, decoder, content, style, alpha)
ndarr = output[0].mul(255).clamp(0, 255).byte().permute(1, 2, 0).cpu().numpy()
return {
'image': Image.fromarray(ndarr)
}
if __name__ == '__main__':
runway.run(host='0.0.0.0', port=8888)
)
sess.run([tf.local_variables_initializer()])
init_fn(sess)
return sess
stylize_inputs = {
'content_image': runway.image,
'style_image': runway.image,
'interpolation_weight': runway.number(default=0.5, min=0, max=1, step=0.1)
}
@runway.command('stylize', inputs=stylize_inputs, outputs={'image': runway.image})
def stylize(sess, inputs):
content_img_np = np.array(inputs['content_image'])
style_image_np = np.array(inputs['style_image'])
interpolation_weight = inputs['interpolation_weight']
identity_params = sess.run(bottleneck_feat, feed_dict={style_img_ph: content_img_np})
style_params = sess.run(bottleneck_feat, feed_dict={style_img_ph: style_image_np})
stylized_image_res = sess.run(
stylized_images,
feed_dict={
bottleneck_feat: identity_params * (1 - interpolation_weight) + style_params * interpolation_weight,
content_img_ph: content_img_np
})
out = (stylized_image_res*255.0).astype(np.uint8)
return out[0]
if __name__ == '__main__':
runway.run(port=8540)
from PIL import Image
import cv2
sample_inputs = {
'source_imgs': runway.image(description='input image to be translated'),
'target': runway.image(description='input image to be translated'),
'amount': runway.number(min=0, max=16, default=0)
}
@runway.command(
'translate',
inputs=sample_inputs,
outputs={
'image':
runway.image(
description='output image containing the translated result')
})
def translate(learn, inputs):
listimg, h, w = imgint(inputs['source_imgs'], inputs['target'])
i = inputs['amount']
cvimg = (listimg[i][0] * 255).byte().cpu().numpy().transpose(1, 2,
0)[:h, :w]
img = cv2.cvtColor(cvimg, cv2.COLOR_BGR2RGB)
im_pil = Image.fromarray(img)
return im_pil
if __name__ == '__main__':
runway.run(port=8889)
'DCGAN',
pretrained=True,
useGPU=use_gpu)
return model
@runway.command('generate',
inputs={'z': runway.vector(length=64, sampling_std=0.5)},
outputs={'image': runway.image(width=64, height=64)})
def generate(model, inputs):
# Generate ♾ infinite ♾ images
z = inputs['z']
latents = z.reshape((1, 64))
latents = torch.from_numpy(latents)
# Generate one image
noise, _ = model.buildNoiseData(1)
with torch.no_grad():
generated_image = model.test(noise)
generated_image = generated_image.clamp(min=-1, max=1)
generated_image = ((generated_image + 1.0) * 255 / 2.0)
# Now generated_image contains our generated image! 🌞
# return generated_image[0].permute(1, 2, 0).numpy().astype(np.uint8)
return {
'image': generated_image[0].permute(1, 2,
0).cpu().numpy().astype(np.uint8)
}
if __name__ == '__main__':
runway.run(5133)
h, w = 480, 640
img_shape = (h, w, 3)
batch_shape = (1, ) + img_shape
g = tf.get_default_graph()
sess = tf.Session(graph=g)
img_placeholder = tf.placeholder(tf.float32,
shape=batch_shape,
name='img_placeholder')
preds = transform.net(img_placeholder)
load_checkpoint(os.path.join(options['checkpoint_path'], 'fns.ckpt'), sess)
return sess
@runway.command('stylize',
inputs={'image': runway.image},
outputs={'output': runway.image})
def stylize(sess, inp):
img = inp['image']
original_size = img.size
img = np.array(img.resize((640, 480)))
img = np.expand_dims(img, 0)
with g.as_default():
output = sess.run(preds, feed_dict={img_placeholder: img})
output = np.clip(output[0], 0, 255).astype(np.uint8)
output = Image.fromarray(output).resize(original_size)
return dict(output=output)
if __name__ == '__main__':
runway.run(model_options={'checkpoint_path': 'models/Cubist'})
}
command_outputs = {'output': runway.image}
@runway.command('convert', inputs=command_inputs, outputs=command_outputs)
def convert(model, inputs):
img = inputs['semantic_map']
original_size = img.size
img = img.resize((opt.load_size, opt.load_size))
params = get_params(opt, img.size)
transform_label = get_transform(opt,
params,
method=Image.NEAREST,
normalize=False)
label_tensor = transform_label(img) * 255.0
label_tensor[label_tensor == 255.0] = 0
data = {
'label': label_tensor.unsqueeze(0),
'instance': label_tensor.unsqueeze(0),
'image': None
}
generated = model(data, mode='inference')
output = util.tensor2im(generated[0])
output = Image.fromarray(output).resize(original_size)
return output
if __name__ == '__main__':
runway.run(port=5132)
def translate(models, inputs):
global aligned_target
global target_embedding
global last_target
fd = models['fd']
fp = models['fp']
idet = models['idet']
fv = models['fv']
generator = models['generator']
source = np.array(inputs['source'])
target = np.array(inputs['target'])
src, mask, aligned_im, (x0, y0, x1, y1), landmarks = utils.get_src_inputs(
source, fd, fp, idet)
if last_target is None or not np.array_equal(last_target, target):
aligned_target, target_embedding = utils.get_tar_inputs([target], fd,
fv)
last_target = target
out = generator.inference(src, mask, aligned_target, target_embedding)
result_face = np.squeeze(((out[0] + 1) * 255 / 2).astype(np.uint8))
result_img = utils.post_process_result(source, fd, result_face, aligned_im,
src, x0, y0, x1, y1, landmarks)
return result_img
if __name__ == "__main__":
# runway.run()
runway.run(model_options={
'encoder': './weights/encoder.h5',
'decoder': './weights/decoder.h5'
})
def setup():
cfg_from_file('cfg/coco_attn2.yml')
cfg.CUDA = torch.cuda.is_available()
wordtoix, ixtoword = word_index()
print('Loading Model...')
text_encoder, netG = models(len(wordtoix))
print('Models Loaded')
seed = 100
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if cfg.CUDA:
torch.cuda.manual_seed_all(seed)
return (wordtoix, ixtoword, text_encoder, netG)
@runway.command('generate',
inputs={'caption': runway.text},
outputs={'result': runway.image})
def generate_command(model, inp):
wordtoix, ixtoword, text_encoder, netG = model
caption = inp["caption"]
img = generate(caption, wordtoix, ixtoword, text_encoder, netG, False)
if img is None:
img = PIL.Image.new('RGB', (256, 256), color='black')
return dict(result=img)
if __name__ == "__main__":
runway.run(port=9000, debug=True)
global_feature = sess.run(model["global_f_ph"], feed_dict={model["img_ph"]: img})
h_np = np.zeros([batch_size, 1024])
c_np = np.zeros([batch_size, 1024])
while True:
action_np, h_np, c_np = sess.run((action, h, c), feed_dict={model["img_ph"]: img,
model["global_f_ph"]: global_feature,
model["h_ph"]: h_np,
model["c_ph"]: c_np})
ratios, terminals = command2action(action_np, ratios, terminals)
bbox = generate_bbox(origin_image, ratios)
if np.sum(terminals) == batch_size:
return bbox
img = crop_input(origin_image, bbox)
xmin, ymin, xmax, ymax = auto_cropping([ip_img - 0.5], model["sess"], model["action"], model["h"], model["c"])[0]
return {"output_image" : im[ymin:ymax, xmin:xmax]}
if __name__ == "__main__":
runway.run(model_options={"checkpoint" : "vfn_rl.pkl"})
def segment_humans(model, inputs):
frame = np.array(inputs["input_image"])
#image = frame[...,::-1]
h, w = frame.shape[0], frame.shape[1]
# Predict mask
X, pad_up, pad_left, h_new, w_new = utils.preprocessing(frame, expected_size=320, pad_value=0)
with torch.no_grad():
if torch.cuda.is_available():
mask = model(X.cuda())
mask = mask[..., pad_up: pad_up+h_new, pad_left: pad_left+w_new]
mask = F.interpolate(mask, size=(h,w), mode='bilinear', align_corners=True)
mask = F.softmax(mask, dim=1)
mask = mask[0,1,...].cpu().numpy()
else:
mask = model(X)
mask = mask[..., pad_up: pad_up+h_new, pad_left: pad_left+w_new]
mask = F.interpolate(mask, size=(h,w), mode='bilinear', align_corners=True)
mask = F.softmax(mask, dim=1)
mask = mask[0,1,...].numpy()
mask = 255*mask
mask = np.expand_dims(mask, axis=2)
image_alpha = np.concatenate((frame, mask), axis=2)
return image_alpha.astype(np.uint8)
if __name__ == "__main__":
runway.run(model_options={"backbone": "mobilenetv2", "checkpoint" : "../asdas/UNet_MobileNetV2.pth"})
print("\nPerforming object detection:")
input_imgs = Variable(input_img.type(Tensor))
# Get detections
with torch.no_grad():
detections = model["model"](input_imgs)
detections = non_max_suppression(detections, 0.8, 0.4)
bboxes = []
class_labels = []
scores = []
# Draw bounding boxes and labels of detections
if detections is not None:
# Rescale boxes to original image
detections = rescale_boxes(detections[0], 416)
unique_labels = detections[:, -1].cpu().numpy()
class_preds = [model["classes"][int(i)] for i in unique_labels]
op = detections.cpu().numpy()
for i in range(len(op)):
bboxes.append(op[i, 0:4])
scores.append(op[i, 5])
return dict(bboxes=bboxes, classes=class_preds, scores=scores)
if __name__ == "__main__":
runway.run(model_options={"checkpoint_dir": "./checkpoint"})
generator.set_dlatents(latent_vector)
img_array = generator.generate_images()[0]
img = PIL.Image.fromarray(img_array, 'RGB')
return img.resize((512, 512))
generate_inputs = {
'age': runway.number(min=-500, max=500, default=6, step=0.1),
}
generate_outputs = {
'image': runway.image(width=512, height=512),
}
@runway.command('generat3r', inputs=generate_inputs, outputs=generate_outputs)
def move_and_show(model, inputs):
latent_vector = (latent_vector_1 + latent_vector_2) * 2
# load direction
age_direction = np.load('ffhq_dataset/latent_directions/age.npy')
direction = age_direction
# model = generator
coeff = inputs['age']/5.0
new_latent_vector = latent_vector.copy()
new_latent_vector[:8] = (latent_vector + coeff*direction)[:8]
image = (generate_image(model, new_latent_vector))
#ax[i].set_title('Coeff: %0.1f' % coeff)
#plt.show()
return {'image': image}
if __name__ == '__main__':
runway.run(debug=True)
@runway.setup(options={'checkpoints_root': runway.file(is_directory=True)})
def setup(opts):
opt.name = opts['checkpoints_root'].split('/')[-1]
opt.checkpoints_dir = os.path.join(opts['checkpoints_root'], '..')
model = create_model(opt)
return model
@runway.command('generate',
inputs={'image': runway.image},
outputs={'image': runway.image})
def generate(model, inputs):
label = inputs['image']
params = get_params(opt, label.size)
transform_label = get_transform(opt,
params,
method=Image.NEAREST,
normalize=False)
label_tensor = transform_label(label)
label_tensor = label_tensor.unsqueeze(0)
generated = model.inference(label_tensor, None)
torch.cuda.synchronize()
im = util.tensor2im(generated.data[0])
return Image.fromarray(im)
if __name__ == '__main__':
runway.run(host='0.0.0.0',
port=8888,
model_options={'checkpoints_root': './checkpoints/Boston'})
inputs = {'noise_vector': vector(length=128, description='A random seed.')}
outputs = {'image': image(width=512, height=512)}
# The @runway.command() decorator is used to create interfaces to call functions
# remotely via an HTTP endpoint. This lets you send data to, or get data from,
# your model. Each command creates an HTTP route that the Runway app will use
# to communicate with your model (e.g. POST /generate). Multiple commands
# can be defined for the same model.
@runway.command('generate',
inputs=inputs,
outputs=outputs,
description='Generate an image.')
def generate(model, input_args):
# Functions wrapped by @runway.command() receive two arguments:
# 1. Whatever is returned by a function wrapped by @runway.setup(),
# usually a model.
# 2. The input arguments sent by the remote caller via HTTP. These values
# match the schema defined by inputs.
img = input_args['image']
return model.generate(img)
# The runway.run() function triggers a call to the function wrapped by
# @runway.setup() passing model_options as its single argument. It also
# creates an HTTP server that listens for and fulfills remote requests that
# trigger commands.
if __name__ == '__main__':
runway.run(host='0.0.0.0', port=9000, model_options={'model_size': 'big'})
"ear_r": [215, 175, 125],
"eye_g": [220, 180, 210],
"neck_l": [125, 125, 255]
}
to_tensor = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
@runway.command('parse',
inputs={'image': runway.image},
outputs={
'parsed_face':
runway.segmentation(label_to_id=label_to_id,
label_to_color=label_to_color)
})
def parse(model, inputs):
image = inputs['image'].resize((512, 512), Image.BILINEAR)
img = to_tensor(image)
img = torch.unsqueeze(img, 0)
img = img.cuda()
out = model(img)[0]
parsing = out.squeeze(0).cpu().detach().numpy().argmax(0)
return parsing.astype(np.uint8)
if __name__ == '__main__':
runway.run(model_options={'checkpoint': './79999_iter.pth'})
def process_tile(model, tile):
img = np.transpose(tile[:, :, [2, 1, 0]], (2, 0, 1))
img = torch.from_numpy(img).float()
img = img.unsqueeze(0)
img = img.to(device)
output = model(img).data.squeeze().float().cpu().clamp_(0, 1).numpy()
output = np.transpose(output[[2, 1, 0], :, :], (1, 2, 0))
output = resize(output, tile.shape)
return output
@runway.command(name='upscale',
inputs={'image': runway.image},
outputs={'upscaled': runway.image})
def upscale(model, inputs):
img = np.array(inputs['image'])
img = img * 1.0 / 255
tiler = ImageSlicer(img.shape,
tile_size=(512, 512),
tile_step=(256, 256),
weight='pyramid')
tiles = [process_tile(model, tile) for tile in tiler.split(img)]
output = tiler.merge(tiles)
output = (output * 255.0).round().astype('uint8')
return dict(upscaled=output)
if __name__ == '__main__':
runway.run(port=4323)
'mask': 1
},
label_to_color={
'background': [0, 0, 0],
'mask': [255, 255, 255]
})
}
@runway.command('inpaint',
inputs=command_inputs,
outputs={'inpainted': runway.image})
def inpaint(output_image, inputs):
image = inputs['image']
original_size = image.size
image = np.array(image.resize((256, 256)))
image = np.expand_dims(image, 0)
mask = np.array(inputs['mask'].resize((256, 256)))
mask = mask * 255
mask = np.stack((mask, ) * 3, axis=-1)
mask = np.expand_dims(mask, 0)
feed_dict = {input_image: np.concatenate([image, mask], axis=2)}
with g.as_default():
result = sess.run(output_image, feed_dict=feed_dict)
return Image.fromarray(result[0][:, :, ::-1]).resize(original_size)
if __name__ == '__main__':
runway.run(
port=5232,
model_options={'checkpoint_dir': './model_logs/release_places2_256'})
im = inputs['projectionImage']
if not os.path.exists('./projection'):
os.mkdir('./projection')
if not os.path.exists('./projection/imgs'):
os.mkdir ('./projection/imgs')
if not os.path.exists('./projection/records'):
os.mkdir('./projection/records')
if not os.path.exists('./projection/out'):
os.mkdir('./projection/out')
if os.path.isfile('./projection/imgs/project.png'):
os.remove('./projection/imgs/project.png')
for f in os.listdir('./projection/records/'):
if os.path.isfile(os.path.join('./projection/records/', f)):
os.remove (os.path.join('./projection/records/', f))
im.save('./projection/imgs/project.png')
dataset_tool.create_from_images("./projection/records/", "./projection/imgs/", True)
output = get_projected_real_images("records","./projection/",1,10, inputs['steps'], model)
#return the last item
return output[-1]
if __name__ == '__main__':
runway.run(model_options={ 'checkpoint': 'network-snapshot-001383.pkl' })
#runway.run(host='localhost', port=8888, debug=True, model_options={'checkpoint': './vox-cpk.pth.tar'})
}
@runway.command('inpaint',
inputs=command_inputs,
outputs={'output': runway.image})
def inpaint(model, inputs):
output, input_image_tf, input_mask_tf = model
image = inputs['image']
original_size = image.size
image = np.array(image.resize((256, 256)), dtype=np.float32)
mask = np.array(inputs['mask'].resize((256, 256)), dtype=np.float32)
mask = np.expand_dims(mask, -1)
result = sess.run(output,
feed_dict={
input_image_tf: np.expand_dims(image, 0),
input_mask_tf: np.expand_dims(mask, 0)
})
result = result[0][:, :, ::-1]
result = imresize(result, original_size[::-1])
mask = np.array(inputs['mask'].resize(original_size), dtype=np.float32)
mask = np.stack([mask, mask, mask], -1)
masked_result = mask * result
masked_result += (1 - mask) * np.array(inputs['image'])
return masked_result.astype(np.uint8)
if __name__ == "__main__":
runway.run(model_options={
'checkpoint_dir': 'checkpoints/places2_512x680_freeform'
})
result = learner.predict(im)
np_img = result[2].data.numpy()[1]
return np_img
@runway.command('mask',
inputs={
'image':
runway.image,
'threshold':
runway.number(default=0.5, min=0, max=1, step=0.001)
},
outputs={'image': runway.image(channels=4)})
def mask(learner, inputs):
inp = inputs['image'].convert('RGB')
original_size = inp.size
inp_resized = inp.resize((512, 512))
mask1 = inference(learner, inp_resized)
mask2 = np.fliplr(inference(learner, np.fliplr(inp_resized)))
mask = (mask1 + mask2) / 2
mask[mask > inputs['threshold']] = 255
mask = resize(mask, (original_size[1], original_size[0]),
anti_aliasing=False).astype(np.uint8)
masked = np.concatenate((np.array(inp), np.expand_dims(mask, -1)), axis=2)
return masked
if __name__ == '__main__':
runway.run(port=7843,
model_options={'checkpoint': './512x512_resnet34_2.pkl'})
sess = gpt2.start_tf_sess()
gpt2.load_gpt2(sess, run_name=run_name)
return None
# Every model needs to have at least one command. Every command allows to send
# inputs and process outputs. To see a complete list of supported inputs and
# outputs data types: https://sdk.runwayml.com/en/latest/data_types.html
@runway.command(name='generate',
inputs={'caption': text()},
outputs={'script': text()})
def generate(model, args):
print('[GENERATE] Ran with caption value "{}"'.format(args['caption']))
samples = gpt2.generate(sess,
nsamples=1,
return_as_list=True,
prefix=f"({args['caption']})",
include_prefix=True,
length=200,
temperature=0.7,
run_name=run_name)
return {'script': samples[0][len(args['caption']) + 2:]}
if __name__ == '__main__':
# run the model server using the default network interface and ports,
# displayed here for convenience
runway.run(host='0.0.0.0',
port=8000,
model_options={'run_name': 'first_345m_run'})