Update app.py

app.py

@@ -1,146 +1,323 @@
-import gradio as gr
-import numpy as np
-import random
-from diffusers import DiffusionPipeline
+from base64 import b64encode
+
+import numpy
 import torch
+from diffusers import AutoencoderKL, LMSDiscreteScheduler, UNet2DConditionModel
+from huggingface_hub import notebook_login
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
+# For video display:
+from matplotlib import pyplot as plt
+from pathlib import Path
+from PIL import Image
+from torch import autocast
+from torchvision import transforms as tfms
+from tqdm.auto import tqdm
+from transformers import CLIPTextModel, CLIPTokenizer, logging
+import os
+import numpy as np  
 
-if torch.cuda.is_available():
-    torch.cuda.max_memory_allocated(device=device)
-    pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16", use_safetensors=True)
-    pipe.enable_xformers_memory_efficient_attention()
-    pipe = pipe.to(device)
-else: 
-    pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", use_safetensors=True)
-    pipe = pipe.to(device)
+torch.manual_seed(1)
+# if not (Path.home()/'.cache/huggingface'/'token').exists(): notebook_login()
 
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
+# Supress some unnecessary warnings when loading the CLIPTextModel
+logging.set_verbosity_error()
 
-def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
+# Set device
+torch_device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu" 
 
-    if randomize_seed:
-        seed = random.randint(0, MAX_SEED)
-        
-    generator = torch.Generator().manual_seed(seed)
-    
-    image = pipe(
-        prompt = prompt, 
-        negative_prompt = negative_prompt,
-        guidance_scale = guidance_scale, 
-        num_inference_steps = num_inference_steps, 
-        width = width, 
-        height = height,
-        generator = generator
-    ).images[0] 
-    
-    return image
-
-examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
-]
-
-css="""
-#col-container {
-    margin: 0 auto;
-    max-width: 520px;
-}
-"""
-
-if torch.cuda.is_available():
-    power_device = "GPU"
-else:
-    power_device = "CPU"
-
-with gr.Blocks(css=css) as demo:
+# Load the autoencoder model which will be used to decode the latents into image space.
+vae = AutoencoderKL.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="vae")
+
+# Load the tokenizer and text encoder to tokenize and encode the text.
+tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
+text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14")
+
+# The UNet model for generating the latents.
+unet = UNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="unet")
+
+# The noise scheduler
+scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
+
+# To the GPU we go!
+vae = vae.to(torch_device)
+text_encoder = text_encoder.to(torch_device)
+unet = unet.to(torch_device)
+token_emb_layer = text_encoder.text_model.embeddings.token_embedding
+pos_emb_layer = text_encoder.text_model.embeddings.position_embedding
+
+position_ids = text_encoder.text_model.embeddings.position_ids[:, :77]
+position_embeddings = pos_emb_layer(position_ids)
+
+
+def get_output_embeds(input_embeddings):
+    # CLIP's text model uses causal mask, so we prepare it here:
+    bsz, seq_len = input_embeddings.shape[:2]
+    causal_attention_mask = text_encoder.text_model._build_causal_attention_mask(bsz, seq_len, dtype=input_embeddings.dtype)
+
+    # Getting the output embeddings involves calling the model with passing output_hidden_states=True
+    # so that it doesn't just return the pooled final predictions:
+    encoder_outputs = text_encoder.text_model.encoder(
+        inputs_embeds=input_embeddings,
+        attention_mask=None, # We aren't using an attention mask so that can be None
+        causal_attention_mask=causal_attention_mask.to(torch_device),
+        output_attentions=None,
+        output_hidden_states=True, # We want the output embs not the final output
+        return_dict=None,
+    )
+
+    # We're interested in the output hidden state only
+    output = encoder_outputs[0]
+
+    # There is a final layer norm we need to pass these through
+    output = text_encoder.text_model.final_layer_norm(output)
+
+    # And now they're ready!
+    return output
+
+
+def set_timesteps(scheduler, num_inference_steps):
+    scheduler.set_timesteps(num_inference_steps)
+    scheduler.timesteps = scheduler.timesteps.to(torch.float32)
+
+def pil_to_latent(input_im):
+    # Single image -> single latent in a batch (so size 1, 4, 64, 64)
+    with torch.no_grad():
+        latent = vae.encode(tfms.ToTensor()(input_im).unsqueeze(0).to(torch_device)*2-1) # Note scaling
+    return 0.18215 * latent.latent_dist.sample()
+
+def latents_to_pil(latents):
+    # bath of latents -> list of images
+    latents = (1 / 0.18215) * latents
+    with torch.no_grad():
+        image = vae.decode(latents).sample
+    image = (image / 2 + 0.5).clamp(0, 1)
+    image = image.detach().cpu().permute(0, 2, 3, 1).numpy()
+    images = (image * 255).round().astype("uint8")
+    pil_images = [Image.fromarray(image) for image in images]
+    return pil_images
+
+
+def generate_with_embs(text_embeddings, text_input, seed):
+
+    height = 512                        # default height of Stable Diffusion
+    width = 512                         # default width of Stable Diffusion
+    num_inference_steps = 10            # Number of denoising steps
+    guidance_scale = 7.5                # Scale for classifier-free guidance
+    generator = torch.manual_seed(seed)   # Seed generator to create the inital latent noise
+    batch_size = 1
+
+    max_length = text_input.input_ids.shape[-1]
+    uncond_input = tokenizer(
+      [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
+    )
+    with torch.no_grad():
+        uncond_embeddings = text_encoder(uncond_input.input_ids.to(torch_device))[0]
+    text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+    # Prep Scheduler
+    set_timesteps(scheduler, num_inference_steps)
+
+    # Prep latents
+    latents = torch.randn(
+    (batch_size, unet.in_channels, height // 8, width // 8),
+    generator=generator,
+    )
+    latents = latents.to(torch_device)
+    latents = latents * scheduler.init_noise_sigma
+
+    # Loop
+    for i, t in tqdm(enumerate(scheduler.timesteps), total=len(scheduler.timesteps)):
+        # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
+        latent_model_input = torch.cat([latents] * 2)
+        sigma = scheduler.sigmas[i]
+        latent_model_input = scheduler.scale_model_input(latent_model_input, t)
+
+        # predict the noise residual
+        with torch.no_grad():
+            noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"]
+
+        # perform guidance
+        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+        # compute the previous noisy sample x_t -> x_t-1
+        latents = scheduler.step(noise_pred, t, latents).prev_sample
+
+    return latents_to_pil(latents)[0]
+
+
+def generate_with_prompt_style(prompt, style, seed = 42):
+
+    prompt = prompt + ' in style of s'
+    embed = torch.load(style)
+
+    text_input = tokenizer(prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
+    # for t in text_input['input_ids'][0][:20]: # We'll just look at the first 7 to save you from a wall of '<|endoftext|>'
+    #     print(t, tokenizer.decoder.get(int(t)))
+    input_ids = text_input.input_ids.to(torch_device)
+
+    token_embeddings = token_emb_layer(input_ids)
+    # The new embedding - our special birb word
+    replacement_token_embedding = embed[list(embed.keys())[0]].to(torch_device)
+
+    # Insert this into the token embeddings
+    token_embeddings[0, torch.where(input_ids[0]==338)] = replacement_token_embedding.to(torch_device)
+
+    # Combine with pos embs
+    input_embeddings = token_embeddings + position_embeddings
+
+    #  Feed through to get final output embs
+    modified_output_embeddings = get_output_embeds(input_embeddings)
+
+    # And generate an image with this:
+    return generate_with_embs(modified_output_embeddings, text_input, seed)
+
+
+import torch
+
+def contrast_loss(images):
+    variance = torch.var(images)
+    return -variance
+
+def generate_with_prompt_style_guidance(prompt, style, seed=42):
+
+    prompt = prompt + ' in style of s'
     
-    with gr.Column(elem_id="col-container"):
-        gr.Markdown(f"""
-        # Text-to-Image Gradio Template
-        Currently running on {power_device}.
-        """)
-        
-        with gr.Row():
-            
-            prompt = gr.Text(
-                label="Prompt",
-                show_label=False,
-                max_lines=1,
-                placeholder="Enter your prompt",
-                container=False,
-            )
-            
-            run_button = gr.Button("Run", scale=0)
-        
-        result = gr.Image(label="Result", show_label=False)
-
-        with gr.Accordion("Advanced Settings", open=False):
-            
-            negative_prompt = gr.Text(
-                label="Negative prompt",
-                max_lines=1,
-                placeholder="Enter a negative prompt",
-                visible=False,
-            )
-            
-            seed = gr.Slider(
-                label="Seed",
-                minimum=0,
-                maximum=MAX_SEED,
-                step=1,
-                value=0,
-            )
-            
-            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
-            
-            with gr.Row():
-                
-                width = gr.Slider(
-                    label="Width",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=512,
-                )
-                
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=512,
-                )
-            
-            with gr.Row():
-                
-                guidance_scale = gr.Slider(
-                    label="Guidance scale",
-                    minimum=0.0,
-                    maximum=10.0,
-                    step=0.1,
-                    value=0.0,
-                )
-                
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=12,
-                    step=1,
-                    value=2,
-                )
-        
-        gr.Examples(
-            examples = examples,
-            inputs = [prompt]
-        )
-
-    run_button.click(
-        fn = infer,
-        inputs = [prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
-        outputs = [result]
+    embed = torch.load(style)
+
+    height = 512                        # default height of Stable Diffusion
+    width = 512                         # default width of Stable Diffusion
+    num_inference_steps = 10  #           # Number of denoising steps
+    guidance_scale = 8 #               # Scale for classifier-free guidance
+    generator = torch.manual_seed(seed)   # Seed generator to create the inital latent noise
+    batch_size = 1
+    contrast_loss_scale = 200 #
+
+    # Prep text
+    text_input = tokenizer([prompt], padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
+    with torch.no_grad():
+        text_embeddings = text_encoder(text_input.input_ids.to(torch_device))[0]
+
+    input_ids = text_input.input_ids.to(torch_device)
+
+    # Get token embeddings
+    token_embeddings = token_emb_layer(input_ids)
+
+    # The new embedding - our special birb word
+    replacement_token_embedding = embed[list(embed.keys())[0]].to(torch_device)
+
+    # Insert this into the token embeddings
+    token_embeddings[0, torch.where(input_ids[0]==338)] = replacement_token_embedding.to(torch_device)
+
+    # Combine with pos embs
+    input_embeddings = token_embeddings + position_embeddings
+
+    #  Feed through to get final output embs
+    modified_output_embeddings = get_output_embeds(input_embeddings)
+
+    # And the uncond. input as before:
+    max_length = text_input.input_ids.shape[-1]
+    uncond_input = tokenizer(
+        [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
     )
+    with torch.no_grad():
+        uncond_embeddings = text_encoder(uncond_input.input_ids.to(torch_device))[0]
+
+    text_embeddings = torch.cat([uncond_embeddings, modified_output_embeddings])
+
+    # Prep Scheduler
+    scheduler.set_timesteps(num_inference_steps)
+
+    # Prep latents
+    latents = torch.randn(
+    (batch_size, unet.config.in_channels, height // 8, width // 8),
+    generator=generator,
+    )
+    latents = latents.to(torch_device)
+    latents = latents * scheduler.init_noise_sigma
+
+    # Loop
+    for i, t in tqdm(enumerate(scheduler.timesteps), total=len(scheduler.timesteps)):
+        # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
+        latent_model_input = torch.cat([latents] * 2)
+        sigma = scheduler.sigmas[i]
+        latent_model_input = scheduler.scale_model_input(latent_model_input, t)
+
+        # predict the noise residual
+        with torch.no_grad():
+            noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"]
+
+        # perform CFG
+        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+        #### ADDITIONAL GUIDANCE ###
+        if i%5 == 0:
+            # Requires grad on the latents
+            latents = latents.detach().requires_grad_()
+
+            # Get the predicted x0:
+            latents_x0 = latents - sigma * noise_pred
+            # latents_x0 = scheduler.step(noise_pred, t, latents).pred_original_sample
+
+            # Decode to image space
+            denoised_images = vae.decode((1 / 0.18215) * latents_x0).sample / 2 + 0.5 # range (0, 1)
+
+            # Calculate loss
+            loss = contrast_loss(denoised_images) * contrast_loss_scale
+
+            # # Occasionally print it out
+            # if i%10==0:
+            #     print(i, 'loss:', loss.item())
+
+            # Get gradient
+            cond_grad = torch.autograd.grad(loss, latents)[0]
+
+            # Modify the latents based on this gradient
+            latents = latents.detach() - cond_grad * sigma**2
+
+        # Now step with scheduler
+        latents = scheduler.step(noise_pred, t, latents).prev_sample
+
+
+    return latents_to_pil(latents)[0]
+
+
+import gradio as gr
+
+
+dict_styles = {'<gartic-phone>':'styles/learned_embeds_gartic-phone.bin', 
+               '<hawaiian shirt>':'styles/learned_embeds_hawaiian-shirt.bin', 
+               '<gp>': 'styles/learned_embeds_phone01.bin', 
+               '<style-spdmn>':'styles/learned_embeds_style-spdmn.bin', 
+               '<yvmqznrm>': 'styles/learned_embedssd_yvmqznrm.bin}
+# dict_styles.keys()
+
+def inference(prompt, style): 
+    
+    if prompt is not None and style is not None: 
+        style = dict_styles[style]
+        result = generate_with_prompt_style_guidance(prompt, style)
+        return np.array(result)
+    else:
+        return None
+
+title = "Stable Diffusion and Textual Inversion"
+description = "A simple Gradio interface to stylize Stable Diffusion outputs"
+examples = [['A man sipping wine wearing a spacesuit on the moon', 'Stripes']]
 
-demo.queue().launch()
+demo = gr.Interface(inference,
+                    inputs = [gr.Textbox(label='Prompt'), 
+                              gr.Dropdown(['Arcane', 'Button eyes', 'Dr Strange', 'GTA-5', 'Illustration', 
+                                           'Manga', 'Matrix', 'Oil Painting', 'Pokemon', 'Stripes'], label='Style')
+                              ],
+                    outputs = [
+                              gr.Image(label="Stable Diffusion Output"),
+                              ],
+                    title = title,
+                    description = description,
+                    # examples = examples, 
+                    # cache_examples=True
+                    )
+demo.launch()
+