Create app.py

app.py

@@ -0,0 +1,63 @@
+import os
+os.system('pip install git+https://github.com/huggingface/transformers --upgrade')
+
+import gradio as gr
+from transformers import ImageGPTFeatureExtractor, ImageGPTForCausalLM
+import torch
+import requests
+from PIL import Image
+import os 
+import matplotlib.pyplot as plt
+
+feature_extractor = ImageGPTFeatureExtractor.from_pretrained("openai/imagegpt-small")
+model = ImageGPTForCausalLM.from_pretrained("openai/imagegpt-small")
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model.to(device)
+
+# load image examples
+urls = ['https://assetsnffrgf-a.akamaihd.net/assets/m/502013285/univ/art/502013285_univ_sqr_xl.jpg']
+for idx, url in enumerate(urls):
+  image = Image.open(requests.get(url, stream=True).raw)
+  image.save(f"image_{idx}.png")
+
+def process_image(image):
+    # prepare 8 images, shape (8, 1024)
+    encoding = feature_extractor([image for _ in range(8)], return_tensors="pt")
+
+    # create primers
+    samples = encoding.pixel_values.numpy()
+    n_px_crop = 16
+    primers = samples.reshape(-1,n_px*n_px)[:,:n_px_crop*n_px] # crop top n_px_crop rows. These will be the conditioning tokens
+    
+    # generate (no beam search)
+    context = np.concatenate((np.full((batch_size, 1), model.config.vocab_size - 1), primers), axis=1)
+    context = torch.tensor(context).to(device)
+    output = model.generate(input_ids=context, max_length=n_px*n_px + 1, temperature=1.0, do_sample=True, top_k=40)
+
+    # decode back to images
+    samples = output[:,1:].cpu().detach().numpy()
+    samples_img = [np.reshape(np.rint(127.5 * (clusters[s] + 1.0)), [n_px, n_px, 3]).astype(np.uint8) for s in samples] # convert color cluster tokens back to pixels
+    
+    # save as list of files
+    completions = []
+    output_dir = '.'
+    for i in range(len(samples_img)):
+      fname = os.path.join(output_dir, "completion" + str(i) + ".png")
+      plt.imsave(fname=fname, arr=samples_img[i], format='png')
+      completions.append(fname)
+
+    return completions
+
+title = "Interactive demo: ImageGPT"
+description = "Demo for OpenAI's ImageGPT: Generative Pretraining from Pixels. To use it, simply upload an image or use the example image below and click 'submit'. Results will show up in a few seconds."
+article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2109.10282'>ImageGPT: Generative Pretraining from Pixels</a> | <a href='https://openai.com/blog/image-gpt/'>Official blog</a></p>"
+examples =[["image_0.png"]]
+
+iface = gr.Interface(fn=process_image, 
+                     inputs=gr.inputs.Image(type="pil"), 
+                     outputs=[gr.outputs.Image(type='file', label=f'completion_{i}') for i in range(len(samples_img))],
+                     title=title,
+                     description=description,
+                     article=article,
+                     examples=examples)
+iface.launch(debug=True)