Set Things Up

Code

#### If you don't already have the libraries that will be imported in this notebook, you can uncomment the line below to install them
# %pip install torch numpy matplotlib transformers diffusers accelerate

Code

## A lot of models on hugging face are in PyTorch
import torch

### This bit of code determines whether you have an NVIDIA gpu or you want to use the mac silicon GPU.
if torch.cuda.is_available():
    device = "cuda:0"
elif torch.backends.mps.is_available():
    device = torch.device("mps")
else:
    device = "cpu"

## Other imports
import numpy as np
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')

Text Completion

These are the blenderbot models developed by facebook. The following code shows directly using the model and tokenizer objects.

Blenderbot (small)

Code

from transformers import AutoTokenizer, BlenderbotSmallForConditionalGeneration

modelname = "facebook/blenderbot_small-90M"
tokenizer = AutoTokenizer.from_pretrained(modelname)
model = BlenderbotSmallForConditionalGeneration.from_pretrained(modelname).to(device)

Code

prompt = "My dog is cute. Why don't you agree?"

inputs = tokenizer(prompt, return_tensors="pt").to(device)
reply_ids = model.generate(**inputs)
output = tokenizer.batch_decode(reply_ids, skip_special_tokens=True)[0]

print(prompt)
print("> ",output)

My dog is cute. Why don't you agree?
>  i don't know, i guess it's just that i've never had a dog before.

Blenderbot (regular)

Code

from transformers import BlenderbotTokenizer, BlenderbotForConditionalGeneration
    
modelname = "facebook/blenderbot-400M-distill"
# modelname = "facebook/blenderbot-1B-distill"
# modelname = "facebook/blenderbot-3B" ## a little too large for my laptop

tokenizer = BlenderbotTokenizer.from_pretrained(modelname)
model = BlenderbotForConditionalGeneration.from_pretrained(modelname).to(device)

Code

prompt = "My dog is cute. Why don't you agree?"

inputs = tokenizer([prompt], return_tensors="pt").to(device)
reply_ids = model.generate(**inputs)
output = tokenizer.batch_decode(reply_ids,skip_special_tokens=True)[0]

print(prompt)
print(">",output)

My dog is cute. Why don't you agree?
>  I love dogs, but I'm allergic to them. What kind of dog do you have?

Images

The following uses the stable diffusion quick rendering model. The code demonstrates the use of a pipeline aproach

Code

## Clear the previous models out of the GPU
model = None
tokenizer = None
output = None
reply_ids = None
pipe = None

Code

from diffusers import AutoPipelineForText2Image

# modelname = 'stabilityai/sd-turbo'
modelname = 'stabilityai/sdxl-turbo'

pipe = AutoPipelineForText2Image.from_pretrained(modelname, torch_dtype=torch.float16, variant="fp16")
pipe = pipe.to(device)

Loading pipeline components...: 100%|███████████████████████████████████████████████████████████████| 7/7 [00:00<00:00, 15.75it/s]

Code

prompt = "isometric science laboratory, chemistry, biology, NMR, sci-fi, 3d art, muted colors, soft lighting, ray tracing"
negprompt = 'doctor, medicine, dentist, office'

imgs = []
for _ in range(3):
    image = pipe(prompt=prompt, negative_prompt=negprompt, num_inference_steps=2, strength=0.9, guidance_scale=0.0).images[0]
    imgs.append(image)
    
fig,ax=plt.subplots(1,len(imgs),figsize=(13,4))
[ax[i].imshow(imgs[i]) for i in range(len(imgs))]
[aa.axis('off') for aa in ax]
plt.show()

100%|███████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:03<00:00,  1.54s/it]
100%|███████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:02<00:00,  1.30s/it]
100%|███████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:08<00:00,  4.25s/it]