test(tailor): assure weights are preserved after calling to_embedding…

…_model (#106)

* test(tailor): assure weights are preserved

* test(tailor): add keras weight test

* test(tailor): add keras weights test

tests/unit/tailor/test_keras.py

@@ -1,5 +1,6 @@
 import pytest
 import tensorflow as tf
+import numpy as np
 
 from finetuner.tailor.keras import KerasTailor
 
@@ -121,10 +122,18 @@ def test_trim_fail_given_unexpected_layer_name(model, layer_name):
     ],
     indirect=['model'],
 )
-def test_trim(model, layer_name, expected_output_shape):
+def test_to_embedding_model(model, layer_name, expected_output_shape):
     keras_tailor = KerasTailor(model)
-    embed_model = keras_tailor.to_embedding_model(layer_name=layer_name)
-    assert embed_model.output_shape == expected_output_shape
+    model = keras_tailor.to_embedding_model(layer_name=layer_name)
+    assert model.output_shape == expected_output_shape
+
+
+def test_weights_preserved_given_pretrained_model(vgg16_cnn_model):
+    weights = vgg16_cnn_model.layers[0].get_weights()
+    keras_tailor = KerasTailor(vgg16_cnn_model)
+    vgg16_cnn_model = keras_tailor.to_embedding_model(layer_name='fc2')
+    weights_after_convert = vgg16_cnn_model.layers[0].get_weights()
+    np.testing.assert_array_equal(weights, weights_after_convert)
 
 
 @pytest.mark.parametrize(

tests/unit/tailor/test_paddle.py

@@ -1,6 +1,7 @@
 import paddle
 import paddle.nn as nn
 import pytest
+import numpy as np
 
 from finetuner.tailor.paddle import PaddleTailor
 
@@ -217,15 +218,18 @@ def test_freeze(model, layer_name, input_size, input_dtype, freeze):
     ],
     indirect=['model'],
 )
-def test_trim(
+def test_to_embedding_model(
     model, layer_name, input_size, input_, input_dtype, expected_output_shape
 ):
+    weight = model.parameters()[0].numpy()  # weight of the 0th layer
     paddle_tailor = PaddleTailor(
         model=model,
         input_size=input_size,
         input_dtype=input_dtype,
     )
     model = paddle_tailor.to_embedding_model(freeze=False, layer_name=layer_name)
+    weight_after_convert = model.parameters()[0].numpy()
+    np.testing.assert_array_equal(weight, weight_after_convert)
     out = model(paddle.cast(paddle.rand(input_), input_dtype))
     assert list(out.shape) == expected_output_shape
 

tests/unit/tailor/test_torch.py

@@ -1,6 +1,7 @@
 import pytest
 import torch
 import torch.nn as nn
+import numpy as np
 
 from finetuner.tailor.pytorch import PytorchTailor
 
@@ -187,9 +188,10 @@ def test_trim_fail_given_unexpected_layer_idx(
     ],
     indirect=['model'],
 )
-def test_trim(
+def test_to_embedding_model(
     model, layer_name, input_size, input_, input_dtype, expected_output_shape
 ):
+    weights = list(model.parameters())[0].detach().numpy()  # weights of the first layer
     pytorch_tailor = PytorchTailor(
         model=model,
         input_size=input_size,
@@ -199,6 +201,8 @@ def test_trim(
         freeze=False,
         layer_name=layer_name,
     )
+    weights_after_convert = list(model.parameters())[0].detach().numpy()
+    np.testing.assert_array_equal(weights, weights_after_convert)
     input_ = torch.rand(input_)
     if input_dtype == 'int64':
         input_ = input_.type(torch.LongTensor)