other

mytests/test_vision_encoder_vit8.py

@@ -0,0 +1,54 @@
+import torch
+# from .vision_encoder import get_vision_encoder
+# from dinov2.models import vision_transformer as vits
+# from megatron.model.encoders.vision.vision_encoder import get_vision_encoder
+import torch
+from PIL import Image
+import open_clip
+
+class Args:
+    def __init__(self):
+
+        pass
+
+def test_eva_clip_image_transformer_shapes():
+
+    device = "cuda"
+
+    model, preprocess = open_clip.create_model_from_pretrained('ViT-B-16','hf-hub:timm/vit_base_patch16_224.augreg_in21k')
+
+    for transform in preprocess.transforms:
+        print(transform)
+    # preprocess.transforms = preprocess.transforms[3:]
+    # print(preprocess.transforms)
+
+    visual = model.visual
+    visual.trunk.output_tokens = True
+
+    visual = visual.to(device)
+    del model
+
+    image_path = "dog.jpg"
+    image = preprocess(Image.open(image_path)).unsqueeze(0).to(device)
+    print(image.shape)
+
+    pooled = visual(image)
+
+    print(pooled.shape)
+    # print(tokens.shape)
+
+
+
+def main():
+
+
+    test_eva_clip_image_transformer_shapes()
+    # test_eva_clip_image_transformer()
+    # test_dino_image_frozen_transformer()
+    # test_dino_image_frozen_lora_transformer()
+    # test_dino_video_transformer_basic()
+
+if __name__ == "__main__":
+
+
+    main()