pytorch instead of tensorflow

vit_gpt2.py

@@ -1,8 +1,7 @@
-import tensorflow as tf
-from transformers import TFVisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
+from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
 from PIL import Image
 
-model = TFVisionEncoderDecoderModel.from_pretrained("vit-gpt2-image-captioning", from_pt=True)
+model = VisionEncoderDecoderModel.from_pretrained("vit-gpt2-image-captioning")
 feature_extractor = ViTImageProcessor.from_pretrained("vit-gpt2-image-captioning")
 tokenizer = AutoTokenizer.from_pretrained("vit-gpt2-image-captioning")
 
@@ -17,7 +16,7 @@ def predict_step(img_array):
     if i_image.mode != "RGB":
         i_image = i_image.convert(mode="RGB")
 
-    pixel_values = feature_extractor(images=i_image, return_tensors="tf", do_normalize=True).pixel_values
+    pixel_values = feature_extractor(images=i_image, return_tensors="pt", do_normalize=True).pixel_values
 
     output_ids = model.generate(pixel_values, **gen_kwargs)