File size: 4,817 Bytes
f360117
 
 
 
 
 
 
 
 
 
 
 
 
94aebbe
f360117
94aebbe
f360117
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b6f6f81
 
 
f360117
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94aebbe
 
 
f360117
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163


DEVICE = 'cpu'

import gradio as gr
import numpy as np
from sklearn.svm import LinearSVC
from sklearn import preprocessing
import pandas as pd

import random
import time

import replicate
import torch
import pickle

prompt_list = [p for p in list(set(
                pd.read_csv('./twitter_prompts.csv').iloc[:, 1].tolist())) if type(p) == str]

calibrate_prompts = [
    "4k photo",
    'surrealist art',
    'a psychedelic, fractal view',
    'a beautiful collage',
    'an intricate portrait',
    'an impressionist painting',
    'abstract art',
    'an eldritch image',
    'a sketch',
    'a city full of darkness and graffiti',
    'a black & white photo',
    'a brilliant, timeless tarot card of the world',
    'a photo of a woman',
    '',
]

embs = []
ys = []

start_time = time.time()


glob_idx = 0

def next_image():
    global glob_idx
    glob_idx = glob_idx + 1
    with torch.no_grad():
        if len(calibrate_prompts) > 0:
            print('######### Calibrating with sample prompts #########')
            prompt = calibrate_prompts.pop(0)
            print(prompt)

            image, pooled_embeds = replicate.run(
            "rynmurdock/zahir:43177e0594f3bc2e3560170ff0ffb6d1cacdddda1be25fbcd4348ef02b0b7d0f",
            input={"prompt": prompt, 'im_emg': pickle.dumps(im_emb)}
            )

            embs.append(pooled_embeds)
            return image[0]
        else:
            print('######### Roaming #########')

            # sample only as many negatives as there are positives
            indices = range(len(ys))
            pos_indices = [i for i in indices if ys[i] == 1]
            neg_indices = [i for i in indices if ys[i] == 0]
            lower = min(len(pos_indices), len(neg_indices))
            neg_indices = random.sample(neg_indices, lower)
            pos_indices = random.sample(pos_indices, lower)

            cut_embs = [embs[i] for i in neg_indices] + [embs[i] for i in pos_indices]
            cut_ys = [ys[i] for i in neg_indices] + [ys[i] for i in pos_indices]

            feature_embs = torch.stack([e[0].detach().cpu() for e in cut_embs])
            scaler = preprocessing.StandardScaler().fit(feature_embs)
            feature_embs = scaler.transform(feature_embs)
            print(np.array(feature_embs).shape, np.array(ys).shape)

            lin_class = LinearSVC(max_iter=50000, dual='auto', class_weight='balanced').fit(np.array(feature_embs), np.array(cut_ys))
            lin_class.coef_ = torch.tensor(lin_class.coef_, dtype=torch.double)
            lin_class.coef_ = (lin_class.coef_.flatten() / (lin_class.coef_.flatten().norm())).unsqueeze(0)


            rng_prompt = random.choice(prompt_list)

            w = 1# if len(embs) % 2 == 0 else 0
            im_emb = w * lin_class.coef_.to(device=DEVICE, dtype=torch.float16)
            prompt= 'an image' if glob_idx % 2 == 0 else rng_prompt
            print(prompt)

            image, im_emb = replicate.run(
            "rynmurdock/zahir:43177e0594f3bc2e3560170ff0ffb6d1cacdddda1be25fbcd4348ef02b0b7d0f",
            input={"prompt": prompt, 'im_emg': pickle.dumps(im_emb)}
            )

            embs.append(im_emb)

            torch.save(lin_class.coef_, f'./{start_time}.pt')
            return image[0]









def start(_):
    return [
            gr.Button(value='Like', interactive=True), 
            gr.Button(value='Neither', interactive=True), 
            gr.Button(value='Dislike', interactive=True),
            gr.Button(value='Start', interactive=False),
            next_image()
            ]


def choose(choice):
    if choice == 'Like':
        choice = 1
    elif choice == 'Neither':
        _ = embs.pop(-1)
        return next_image()
    else:
        choice = 0
    ys.append(choice)
    return next_image()

css = "div#output-image {height: 768px !important; width: 768px !important; margin:auto;}"
with gr.Blocks(css=css) as demo:
    with gr.Row():
        html = gr.HTML('''<div style='text-align:center; font-size:32'>You will callibrate for several prompts and then roam.</ div>''')
    with gr.Row(elem_id='output-image'):
        img = gr.Image(interactive=False, elem_id='output-image',)
    with gr.Row(equal_height=True):
        b3 = gr.Button(value='Dislike', interactive=False,)
        b2 = gr.Button(value='Neither', interactive=False,)
        b1 = gr.Button(value='Like', interactive=False,)
        b1.click(
        choose, 
        [b1],
        [img]
        )
        b2.click(
        choose, 
        [b2],
        [img]
        )
        b3.click(
        choose, 
        [b3],
        [img]
        )
    with gr.Row():
        b4 = gr.Button(value='Start')
        b4.click(start,
                 [b4],
                 [b1, b2, b3, b4, img,])

demo.launch()  # Share your demo with just 1 extra parameter 🚀