app.py

from skimage import color
import numpy as np
import json

tokenizer_input_length = 77

import torch
device = 'cuda' if torch.cuda.is_available() else 'cpu'

def rgb_to_hex(rgb_array):
    return "{:02x}{:02x}{:02x}".format(*rgb_array)



def normalized_lab_to_rgb(lab_array):
    lab_array = np.array(lab_array, dtype=np.float32)
    lab_array = lab_array.copy()
    lab_array[0] *= 100.0
    lab_array[1] *= 127.0
    lab_array[2] *= 127.0
    
    if lab_array.ndim == 1:
        lab_array = lab_array.reshape(1, 3)

    rgb_array = color.lab2rgb(lab_array)
    rgb_array = (rgb_array * 255).astype(np.uint8)

    return tuple(rgb_array.squeeze())
    
    

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(repo_id="lasercatz/text2palette", filename="epoch_19.pth")





import torch.nn as nn
import torch.nn.functional as F
from transformers import CLIPTextModel, CLIPTokenizer


class AttentionPooling(nn.Module):
    def __init__(self, d_model):
        super().__init__()
        self.attn = nn.Linear(d_model, 1)

    def forward(self, x, mask=None):
        scores = self.attn(x).squeeze(-1)
        if mask is not None:
            scores = scores.masked_fill(mask == 0, float('-inf'))
        weights = F.softmax(scores, dim=-1).unsqueeze(-1)
        return torch.sum(x * weights, dim=1)


class SequencePriorNet(nn.Module):
    def __init__(self, d_model, d_z, n_heads=4):
        super().__init__()
        self.attn = nn.MultiheadAttention(d_model, n_heads, batch_first=True)
        self.pool = AttentionPooling(d_model)
        self.fc = nn.Linear(d_model, d_z * 2)
        self.norm = nn.LayerNorm(d_model)
        self.dropout = nn.Dropout(0.3)

    def forward(self, text_feats, attention_mask):
        attn_output, _ = self.attn(
            text_feats, text_feats, text_feats, key_padding_mask=~attention_mask.bool())
        x = self.norm(attn_output + text_feats)
        x = self.dropout(x)
        x = self.pool(x, attention_mask)
        x = self.fc(x)
        return x


class Text2PaletteModel(nn.Module):
    def __init__(self, d_model=768, d_z=256, max_seq_len=64,

                 n_layers=8, n_heads=8, dim_ff=3072):
        super().__init__()
        self.d_model = d_model
        self.max_seq_len = max_seq_len

        self.tokenizer = CLIPTokenizer.from_pretrained(
            'openai/clip-vit-base-patch32')
        self.clip_text = CLIPTextModel.from_pretrained(
            'openai/clip-vit-base-patch32')
        
        self.tokenizer_input_length = tokenizer_input_length


        self.text_proj = nn.Sequential(
            nn.Linear(512, d_model*2),
            nn.GELU(),
            nn.LayerNorm(d_model*2),
            nn.Dropout(0.3),
            nn.Linear(d_model*2, d_model)
        )

        self.color_embed = nn.Sequential(
            nn.Linear(3, d_model),
            nn.LayerNorm(d_model),
            nn.GELU(),
            nn.Dropout(0.3)
        )

        self.cross_attn = nn.MultiheadAttention(d_model, 8, batch_first=True)

        self.position_embed = nn.Embedding(max_seq_len, d_model)
        self.start_embed = nn.Parameter(torch.randn(1, d_model))

        self.palette_encoder = nn.TransformerEncoder(
            nn.TransformerEncoderLayer(
                d_model, n_heads, dim_ff, batch_first=True),
            n_layers
        )

        self.z_proj = nn.Sequential(
            nn.Linear(d_model*2, d_z),
            nn.LayerNorm(d_z),
            nn.GELU()
        )
        self.z_expand = nn.Linear(d_z, d_model)
        self.z_mu = nn.Linear(d_z, d_z)
        self.z_logvar = nn.Linear(d_z, d_z)

        self.decoder = nn.TransformerDecoder(
            nn.TransformerDecoderLayer(
                d_model, n_heads, dim_ff, batch_first=True),
            n_layers
        )

        self.out_mu_L = nn.Sequential(
            nn.Linear(d_model, 1),
            nn.Sigmoid()
        )
        self.out_mu_ab = nn.Sequential(
            nn.Linear(d_model, 2),
            nn.Tanh()
        )
        self.out_logvar = nn.Linear(d_model, 3)

        self.prior_net = SequencePriorNet(d_model, d_z, n_heads=4)

        self.text_pool = AttentionPooling(d_model)
        self.palette_pool = AttentionPooling(d_model)

    def reparameterize(self, mu, logvar):

        if self.training:
            std = torch.exp(0.5 * logvar)
            eps = torch.randn_like(std)
            return mu + eps * std
        else:
            return mu


    @torch.no_grad()
    def generate(self, text, palette_size, temp=1.0):
        self.eval()
        tokenized = self.tokenizer(text, return_tensors='pt', padding=True, truncation=True,
                                   max_length=self.tokenizer_input_length).to(next(self.parameters()).device)

        text_feats = self.clip_text(**tokenized).last_hidden_state
        text_feats = self.text_proj(text_feats)

        # Sample from prior
        prior_params = self.prior_net(text_feats, tokenized['attention_mask'])
        prior_mu, prior_logvar = prior_params.chunk(2, -1)
        z = prior_mu + torch.exp(0.5 * prior_logvar) * \
            torch.randn_like(prior_mu) * temp
        z_expanded = self.z_expand(z).unsqueeze(1)

        memory = torch.cat([z_expanded, text_feats],
                           dim=1)  # [1, T+1, d_model]
        memory_key_padding_mask = torch.cat([
            torch.zeros((1, 1), dtype=torch.bool, device=device),
            ~tokenized['attention_mask'].bool()
        ], dim=1)  # [1, T+1]

        colors = []
        batch_size = 1
        current_emb = self.start_embed.unsqueeze(0).expand(
            batch_size, -1, -1)  # [1, 1, d_model]

        for i in range(min(palette_size, self.max_seq_len)):

            pos = self.position_embed(torch.arange(0, current_emb.size(
                1), device=device)).unsqueeze(0)  # [1, i+1, d_model]
            decoder_in = current_emb + pos  # [1, i+1, d_model]

            output = self.decoder(
                decoder_in,
                memory,
                tgt_mask=nn.Transformer.generate_square_subsequent_mask(
                    decoder_in.size(1), device=device),
                memory_key_padding_mask=memory_key_padding_mask
            )  # [1, i+1, d_model]

            mu = torch.cat([self.out_mu_L(output[:, -1]),
                           self.out_mu_ab(output[:, -1])], dim=-1)  # [1, 3]
            logvar = self.out_logvar(output[:, -1])  # [1, 3]
            color = mu + torch.exp(0.5 * logvar) * \
                torch.randn_like(mu) * temp  # [1, 3]
            color[:, 0].clamp_(0, 1)
            color[:, 1:].clamp_(-1, 1)

            colors.append(color)

            color_emb = self.color_embed(color.unsqueeze(1))  # [1, 1, d_model]
            current_emb = torch.cat(
                [current_emb, color_emb], dim=1)  # [1, i+2, d_model]

        return torch.cat(colors, dim=0).unsqueeze(0)


model = Text2PaletteModel().to(device)
state_dict = torch.load(model_path, map_location=torch.device(device))
model.load_state_dict(state_dict['model'])
model.to(device)
model.eval()

import gradio as gr



def generate(text, palette_size=5, temp=0.5):

    html=""
    all_hex_palettes = []

    with torch.no_grad():
        generated_palette = model.generate(
            text, 
            palette_size=int(palette_size),
            temp=temp
        )
    
    lab = generated_palette[0].cpu().numpy()
    hex_palette = [rgb_to_hex(normalized_lab_to_rgb(lab_color)) for lab_color in lab]
    all_hex_palettes.append(hex_palette)
    
    html += "<div style='display: flex; flex-direction: row;align-items: center; width:100%;'>"
    
    hex_codes = []
    for i,hex_color in enumerate(hex_palette):
        hex_color = "#"+hex_color.upper()
        hex_codes.append(hex_color)
        html += f'<div style=\'margin:0;flex: 1; text-align: center;\'><div style=\'background-color: {hex_color}; width: 100%; height: 100px;border-radius:{"1em 0 0 1em" if i==0 else "0 1em 1em 0" if i==len(hex_palette)-1 else "0"}\'></div><p style=\'font-size: 14px; margin-top: 5px;\'>{hex_color}</p></div>'
    html += "</div>"

    json_output = json.dumps({"palettes": all_hex_palettes}, indent=2)
    html+=json_output

    return html



with gr.Blocks() as demo:
    gr.Markdown("<h1>Palette Generator</h1>")
    
    input = gr.Textbox(label="Input text", placeholder="Describe the palette in your mind")

    with gr.Row():
        palette_size = gr.Slider(2, 10, value=5, step=1, label="Colors")
        temp = gr.Slider(0.0, 1.0, value=0.5, step=0.1, label="Temperature")
    with gr.Row():
        with gr.Column():
            gr.Examples(
                examples=[["fries in ketchup"], ["blueberry milkshake"], ["Oreo McFlurry"]],
                inputs=[input],
                label="Food & Drinks"
            )
        with gr.Column():
            gr.Examples(
                examples=[["bonfire"], ["sheep on grass"], ["North Arctic"]],
                inputs=[input],
                label="Objects & Places"
            )
    with gr.Row():
        with gr.Column():
            gr.Examples(
                examples=[["rock climbing"], ["scuba-diving"], ["Halloween pumpkin party"]],
                inputs=[input],
                label="Activities"
            )
        with gr.Column():
            gr.Examples(
                examples=[["sweetheart"], ["sorrow"], ["murder"]],
                inputs=[input],
                label="Abstract"
            )
    generate_button = gr.Button("🎨 Generate")
    output = gr.HTML("<div style=\"height: 100px\"></div>")

    generate_button.click(
        generate,
        inputs=[input, palette_size, temp],
        outputs=output
    )


demo.launch()