def test_inference(self):
model_name = "openai/clip-vit-base-patch32"
model = TFCLIPModel.from_pretrained(model_name)
processor = CLIPProcessor.from_pretrained(model_name)
image = prepare_img()
inputs = processor(text=["a photo of a cat", "a photo of a dog"],
images=image,
padding=True,
return_tensors="tf")
outputs = model(**inputs, training=False)
# verify the logits
self.assertEqual(
outputs.logits_per_image.shape,
tf.TensorShape(
(inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
)
self.assertEqual(
outputs.logits_per_text.shape,
tf.TensorShape(
(inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
)
expected_logits = tf.constant([[24.5701, 19.3049]])
tf.debugging.assert_near(outputs.logits_per_image,
expected_logits,
atol=1e-3)
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
model = TFCLIPModel(config)
result = model(input_ids, pixel_values, attention_mask, training=False)
self.parent.assertEqual(
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
)
self.parent.assertEqual(
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
)
def test_model_from_pretrained(self):
for model_name in TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
model = TFCLIPModel.from_pretrained(model_name)
self.assertIsNotNone(model)