def test_inference(self):
model = VisionTextDualEncoderModel.from_pretrained(
"clip-italian/clip-italian", logit_scale_init_value=1)
processor = VisionTextDualEncoderProcessor.from_pretrained(
"clip-italian/clip-italian")
image = Image.open(
"./tests/fixtures/tests_samples/COCO/000000039769.png")
inputs = processor(
text=["una foto di un gatto", "una foto di un cane"],
images=image,
padding=True,
return_tensors="pt")
outputs = model(**inputs)
# verify the logits
self.assertEqual(
outputs.logits_per_image.shape,
(inputs.pixel_values.shape[0], inputs.input_ids.shape[0]))
self.assertEqual(
outputs.logits_per_text.shape,
(inputs.input_ids.shape[0], inputs.pixel_values.shape[0]),
)
expected_logits = torch.tensor([[1.2284727, 0.3104122]])
self.assertTrue(
torch.allclose(outputs.logits_per_image,
expected_logits,
atol=1e-3))
def check_save_load(self,
text_config,
input_ids,
attention_mask,
vision_config,
pixel_values=None,
**kwargs):
vision_model, text_model = self.get_vision_text_model(
vision_config, text_config)
model = VisionTextDualEncoderModel(vision_model=vision_model,
text_model=text_model)
model.to(torch_device)
model.eval()
with torch.no_grad():
output = model(input_ids=input_ids,
pixel_values=pixel_values,
attention_mask=attention_mask)
out_1 = output[0].cpu().numpy()
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model = VisionTextDualEncoderModel.from_pretrained(
tmpdirname).eval()
model.to(torch_device)
after_output = model(input_ids=input_ids,
pixel_values=pixel_values,
attention_mask=attention_mask)
out_2 = after_output[0].cpu().numpy()
max_diff = np.amax(np.abs(out_2 - out_1))
self.assertLessEqual(max_diff, 1e-5)
def check_pt_flax_equivalence(self, pt_model, fx_model, input_ids,
attention_mask, pixel_values, **kwargs):
pt_model.to(torch_device)
pt_model.eval()
# prepare inputs
inputs_dict = {
"input_ids": input_ids,
"attention_mask": attention_mask,
"pixel_values": pixel_values
}
pt_inputs = inputs_dict
flax_inputs = {k: v.numpy() for k, v in pt_inputs.items()}
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
fx_outputs = fx_model(**flax_inputs).to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]):
self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
# PT -> Flax
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
fx_model_loaded = FlaxVisionTextDualEncoderModel.from_pretrained(
tmpdirname, from_pt=True)
fx_outputs_loaded = fx_model_loaded(**flax_inputs).to_tuple()
self.assertEqual(len(fx_outputs_loaded), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output_loaded, pt_output in zip(fx_outputs_loaded[:4],
pt_outputs[:4]):
self.assert_almost_equals(fx_output_loaded, pt_output.numpy(),
4e-2)
# Flax -> PT
with tempfile.TemporaryDirectory() as tmpdirname:
fx_model.save_pretrained(tmpdirname)
pt_model_loaded = VisionTextDualEncoderModel.from_pretrained(
tmpdirname, from_flax=True)
pt_model_loaded.to(torch_device)
pt_model_loaded.eval()
with torch.no_grad():
pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs_loaded),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output_loaded in zip(fx_outputs[:4],
pt_outputs_loaded[:4]):
self.assert_almost_equals(fx_output, pt_output_loaded.numpy(),
4e-2)
def test_real_model_save_load_from_pretrained(self):
model_2, inputs = self.get_pretrained_model_and_inputs()
model_2.to(torch_device)
with torch.no_grad():
outputs = model_2(**inputs)
out_2 = outputs[0].cpu().numpy()
with tempfile.TemporaryDirectory() as tmp_dirname:
model_2.save_pretrained(tmp_dirname)
model_1 = VisionTextDualEncoderModel.from_pretrained(
tmp_dirname)
model_1.to(torch_device)
after_outputs = model_1(**inputs)
out_1 = after_outputs[0].cpu().numpy()
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
