Add application file

app.py

@@ -0,0 +1,14 @@
+import gradio as gr
+from help_function import help_function
+from PIL import Image
+import numpy as np
+
+model_helper = help_function()
+
+def greet(numpy_image,text,float_value):
+    PIL_image = Image.fromarray(np.uint8(numpy_image)).convert('RGB')
+    image_edit = model_helper.image_from_text(text,PIL_image,float_value)
+    return image_edit
+
+iface = gr.Interface(fn=greet, inputs=["image", "text", gr.inputs.Slider(0.0, 1.0)], outputs="image")
+iface.launch()

help_function.py

@@ -0,0 +1,40 @@
+import torch
+import open_clip
+from torchvision import transforms
+from torchvision.transforms import ToPILImage
+
+class help_function:
+    def __init__(self):
+        self.clip_text_model = torch.jit.load('jit_models/clip_text_jit.pt', map_location=torch.device('cpu'))
+        self.decoder = torch.jit.load('jit_models/decoder_16w.pt', map_location=torch.device('cpu'))
+        self.mapper_clip = torch.jit.load('jit_models/mapper_clip_jit.pt', map_location=torch.device('cpu'))
+        self.mean_clip = torch.load('jit_models/mean_clip.pt')
+        self.mean_person = torch.load('jit_models/mean_person.pt')
+        self.encoder = torch.jit.load('jit_models/combined_encoder.pt', map_location=torch.device('cpu'))
+        self.tokenizer = open_clip.get_tokenizer('ViT-B-32')
+        self.transform = transforms.Compose([
+            transforms.Resize(224),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+        ])
+
+    def get_text_embedding(self, text):
+        text = self.clip_text_model(self.tokenizer(text))
+        return text
+
+    def get_image_inversion(self, image):
+        image = self.transform(image)
+        w_inversion = self.encoder(image.reshape(1,3,224,224)).reshape(1,16,512)
+        return w_inversion + self.mean_person
+
+    def get_text_delta(self,text_feachers):
+        w_delta = self.mapper_clip(text_feachers - self.mean_clip)
+        return w_delta
+    def image_from_text(self,text,image,power = 1.0):
+        w_inversion = self.get_image_inversion(image)
+        text_embedding = self.get_text_embedding(text)
+        w_delta = self.get_text_delta(text_embedding)
+
+        w_edit = w_inversion + w_delta * power
+        image_edit = self.decoder(w_edit)
+        return ToPILImage()((image_edit[0]+0.5)*0.5)

jit_models/mean_clip.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ee2ab527e32a6168ecb7cd2288d2374fc4ffd57d2287c5f1cf35e4ac677ac6a
+size 2801

jit_models/mean_person.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06f4b6b9913609b2a1ed3c63d198ce510f02dc9a0d70bdb037c073e76768df46
+size 33527

requirements.txt

@@ -0,0 +1,9 @@
+open-clip-torch==2.16.0
+orjson==3.8.10
+packaging==23.0
+pandas==2.0.0
+Pillow==9.5.0
+torch==2.0.0
+torchvision==0.15.1
+tqdm==4.65.0
+