test.py

import torch
from PIL import Image

from transformers import (
    CLIPTokenizer,
    CLIPTextModel
)

from diffusers import (
    UNet2DConditionModel,
    DDPMScheduler
)

from ema_pytorch import EMA


# ==========================
# CONFIG
# ==========================

CHECKPOINT = "model.pt"

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

PROMPT = "a pokemon named eevee"

STEPS = 1000


# ==========================
# TEXT
# ==========================

tokenizer = CLIPTokenizer.from_pretrained(
    "openai/clip-vit-base-patch32"
)

text_encoder = CLIPTextModel.from_pretrained(
    "openai/clip-vit-base-patch32"
)

text_encoder.to(DEVICE)
text_encoder.eval()


# ==========================
# MODEL
# ==========================

unet = UNet2DConditionModel(
    sample_size=128,

    in_channels=3,
    out_channels=3,

    layers_per_block=2,

    block_out_channels=(
        128,
        256,
        512,
        512,
    ),

    down_block_types=(
        "DownBlock2D",
        "DownBlock2D",
        "CrossAttnDownBlock2D",
        "DownBlock2D"
    ),

    up_block_types=(
        "UpBlock2D",
        "CrossAttnUpBlock2D",
        "UpBlock2D",
        "UpBlock2D"
    ),

    cross_attention_dim=512
)


# ==========================
# LOAD EMA
# ==========================

checkpoint = torch.load(
    CHECKPOINT,
    map_location=DEVICE
)


unet.load_state_dict(
    checkpoint["model"]
)


ema = EMA(
    unet,
    beta=0.9999
)

ema.load_state_dict(
    checkpoint["ema"]
)


# use EMA weights
ema.ema_model.to(DEVICE)

unet = ema.ema_model

unet.to(DEVICE)
unet.eval()



# ==========================
# SCHEDULER
# ==========================

scheduler = DDPMScheduler(
    num_train_timesteps=1000,
    beta_schedule="scaled_linear",
    prediction_type="epsilon"
)



# ==========================
# TEXT EMBEDDING
# ==========================

tokens = tokenizer(
    PROMPT,
    padding="max_length",
    max_length=77,
    truncation=True,
    return_tensors="pt"
)


tokens = {
    k:v.to(DEVICE)
    for k,v in tokens.items()
}


with torch.no_grad():

    text = text_encoder(
        **tokens
    ).last_hidden_state



# ==========================
# GENERATE
# ==========================

image = torch.randn(
    (1,3,128,128),
    device=DEVICE
)


scheduler.set_timesteps(
    STEPS
)


with torch.no_grad():

    for t in scheduler.timesteps:

        noise_pred = unet(
            image,
            t,
            encoder_hidden_states=text
        ).sample


        image = scheduler.step(
            noise_pred,
            t,
            image
        ).prev_sample



# ==========================
# SAVE
# ==========================

image = (
    image
    .clamp(-1,1)
    .add(1)
    .div(2)
)


image = (
    image[0]
    .permute(1,2,0)
    .cpu()
    .numpy()
)


image = (
    image * 255
).astype("uint8")


Image.fromarray(
    image
).save(
    "pokemon_test.png"
)


print(
    "Saved pokemon_test.png"
)