Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -25,8 +25,121 @@ blip_model = blip_decoder(pretrained=blip_model_url, image_size=blip_image_eval_
|
|
25 |
blip_model.eval()
|
26 |
blip_model = blip_model.to(device)
|
27 |
|
28 |
-
def
|
29 |
-
|
|
|
|
|
|
|
|
|
30 |
|
31 |
-
|
32 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
blip_model.eval()
|
26 |
blip_model = blip_model.to(device)
|
27 |
|
28 |
+
def generate_caption(pil_image):
|
29 |
+
gpu_image = transforms.Compose([
|
30 |
+
transforms.Resize((blip_image_eval_size, blip_image_eval_size), interpolation=InterpolationMode.BICUBIC),
|
31 |
+
transforms.ToTensor(),
|
32 |
+
transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
|
33 |
+
])(pil_image).unsqueeze(0).to(device)
|
34 |
|
35 |
+
with torch.no_grad():
|
36 |
+
caption = blip_model.generate(gpu_image, sample=False, num_beams=3, max_length=20, min_length=5)
|
37 |
+
return caption[0]
|
38 |
+
|
39 |
+
def load_list(filename):
|
40 |
+
with open(filename, 'r', encoding='utf-8', errors='replace') as f:
|
41 |
+
items = [line.strip() for line in f.readlines()]
|
42 |
+
return items
|
43 |
+
|
44 |
+
def rank(model, image_features, text_array, top_count=1):
|
45 |
+
top_count = min(top_count, len(text_array))
|
46 |
+
text_tokens = clip.tokenize([text for text in text_array]).cuda()
|
47 |
+
with torch.no_grad():
|
48 |
+
text_features = model.encode_text(text_tokens).float()
|
49 |
+
text_features /= text_features.norm(dim=-1, keepdim=True)
|
50 |
+
|
51 |
+
similarity = torch.zeros((1, len(text_array))).to(device)
|
52 |
+
for i in range(image_features.shape[0]):
|
53 |
+
similarity += (100.0 * image_features[i].unsqueeze(0) @ text_features.T).softmax(dim=-1)
|
54 |
+
similarity /= image_features.shape[0]
|
55 |
+
|
56 |
+
top_probs, top_labels = similarity.cpu().topk(top_count, dim=-1)
|
57 |
+
return [(text_array[top_labels[0][i].numpy()], (top_probs[0][i].numpy()*100)) for i in range(top_count)]
|
58 |
+
|
59 |
+
def interrogate(cover):
|
60 |
+
image = Image.fromarray(cover)
|
61 |
+
#image = cover
|
62 |
+
models = models1
|
63 |
+
#caption = generate_caption(Image.fromarray(cover))
|
64 |
+
caption = generate_caption(image)
|
65 |
+
if len(models) == 0:
|
66 |
+
#print(f"\n\n{caption}")
|
67 |
+
return
|
68 |
+
|
69 |
+
table = []
|
70 |
+
bests = [[('',0)]]*5
|
71 |
+
for model_name in models:
|
72 |
+
#print(f"Interrogating with {model_name}...")
|
73 |
+
model, preprocess = clip.load(model_name)
|
74 |
+
model.cuda().eval()
|
75 |
+
|
76 |
+
images = preprocess(image).unsqueeze(0).cuda()
|
77 |
+
with torch.no_grad():
|
78 |
+
image_features = model.encode_image(images).float()
|
79 |
+
image_features /= image_features.norm(dim=-1, keepdim=True)
|
80 |
+
|
81 |
+
ranks = [
|
82 |
+
rank(model, image_features, mediums),
|
83 |
+
rank(model, image_features, ["by "+artist for artist in artists]),
|
84 |
+
rank(model, image_features, trending_list),
|
85 |
+
rank(model, image_features, movements),
|
86 |
+
rank(model, image_features, flavors, top_count=3)
|
87 |
+
]
|
88 |
+
|
89 |
+
for i in range(len(ranks)):
|
90 |
+
confidence_sum = 0
|
91 |
+
for ci in range(len(ranks[i])):
|
92 |
+
confidence_sum += ranks[i][ci][1]
|
93 |
+
if confidence_sum > sum(bests[i][t][1] for t in range(len(bests[i]))):
|
94 |
+
bests[i] = ranks[i]
|
95 |
+
|
96 |
+
row = [model_name]
|
97 |
+
for r in ranks:
|
98 |
+
row.append(', '.join([f"{x[0]} ({x[1]:0.1f}%)" for x in r]))
|
99 |
+
|
100 |
+
table.append(row)
|
101 |
+
|
102 |
+
del model
|
103 |
+
gc.collect()
|
104 |
+
#display(pd.DataFrame(table, columns=["Model", "Medium", "Artist", "Trending", "Movement", "Flavors"]))
|
105 |
+
|
106 |
+
flaves = ', '.join([f"{x[0]}" for x in bests[4]])
|
107 |
+
medium = bests[0][0][0]
|
108 |
+
if caption.startswith(medium):
|
109 |
+
return(f"{caption} {bests[1][0][0]}, {bests[2][0][0]}, {bests[3][0][0]}, {flaves}")
|
110 |
+
#print(f"{caption} {bests[3][0][0]}, {flaves}")
|
111 |
+
else:
|
112 |
+
return(f"{caption}, {medium} {bests[1][0][0]}, {bests[2][0][0]}, {bests[3][0][0]}, {flaves}")
|
113 |
+
#print(f"{caption} {bests[3][0][0]}, {flaves}")
|
114 |
+
|
115 |
+
data_path = "/clip-interrogator/data/"
|
116 |
+
|
117 |
+
artists = load_list(os.path.join(data_path, 'artists.txt'))
|
118 |
+
flavors = load_list(os.path.join(data_path, 'flavors.txt'))
|
119 |
+
mediums = load_list(os.path.join(data_path, 'mediums.txt'))
|
120 |
+
movements = load_list(os.path.join(data_path, 'movements.txt'))
|
121 |
+
|
122 |
+
sites = ['Artstation', 'behance', 'cg society', 'cgsociety', 'deviantart', 'dribble', 'flickr', 'instagram', 'pexels', 'pinterest', 'pixabay', 'pixiv', 'polycount', 'reddit', 'shutterstock', 'tumblr', 'unsplash', 'zbrush central']
|
123 |
+
trending_list = [site for site in sites]
|
124 |
+
trending_list.extend(["trending on "+site for site in sites])
|
125 |
+
trending_list.extend(["featured on "+site for site in sites])
|
126 |
+
trending_list.extend([site+" contest winner" for site in sites])
|
127 |
+
|
128 |
+
models1 = ['ViT-B/32']
|
129 |
+
|
130 |
+
width = 130
|
131 |
+
height = 180
|
132 |
+
|
133 |
+
cover = gr.inputs.Image(shape=(width, height), label='Upload cover image to classify')
|
134 |
+
label = gr.outputs.Label(label='Model prediction')
|
135 |
+
|
136 |
+
examples=["00064.jpg","00068.jpg", "00069.jpg"]
|
137 |
+
#gr.Interface(fn=interrogate,inputs=[gr.Image()],output_label,examples=examples).launch()
|
138 |
+
#gr.Interface(fn=interrogate,inputs=gr.Image(),outputs=gr.outputs.Label(),examples=examples).launch()
|
139 |
+
gr.Interface(fn=interrogate,inputs=cover,outputs=label,examples=examples).launch(share=True)
|
140 |
+
|
141 |
+
#def greet(name):
|
142 |
+
#return "hi " + name + "!!"
|
143 |
+
|
144 |
+
#iface = gr.Interface(fn=greet, inputs="text", outputs="text")
|
145 |
+
#iface.launch(share=True)
|