def test_gradient_pass_though(self, model, input_shape):
if model == "darknet53":
check = {
'1024': [
input_shape[0], input_shape[1] // 32, input_shape[2] // 32,
255
],
'512': [
input_shape[0], input_shape[1] // 16, input_shape[2] // 16,
255
],
'256': [
input_shape[0], input_shape[1] // 8, input_shape[2] // 8,
255
],
}
name = "spp"
elif model == "darknet_tiny":
check = {
'1024': [
input_shape[0], input_shape[1] // 32, input_shape[2] // 32,
340
],
'256': [
input_shape[0], input_shape[1] // 16, input_shape[2] // 16,
340
],
}
name = "tiny"
else:
check = None
return
loss = ks.losses.MeanSquaredError()
optimizer = ks.optimizers.SGD()
test_layer = builder.Backbone_Builder(model)
pred = Yolov3Head(name)
init = tf.random_normal_initializer()
x = tf.Variable(
initial_value=init(shape=input_shape, dtype=tf.float32))
y = {
key: tf.Variable(initial_value=init(shape=value, dtype=tf.float32))
for key, value in check.items()
}
with tf.GradientTape() as tape:
x_cent = test_layer(x)
x_hat = pred(x_cent)
losses = 0
for key in y:
grad_loss = loss(x_hat[key], y[key])
losses += grad_loss
grad = tape.gradient(losses, pred.trainable_variables)
optimizer.apply_gradients(zip(grad, pred.trainable_variables))
self.assertNotIn(None, grad)
return
def test_pass_through(self, model, input_shape):
if model == "darknet53":
check = {
'256': [
input_shape[0],
input_shape[1] // 8,
input_shape[2] // 8,
256],
'512': [
input_shape[0],
input_shape[1] // 16,
input_shape[2] // 16,
512],
'1024': [
input_shape[0],
input_shape[1] // 32,
input_shape[2] // 32,
1024],
}
elif model == "darknet_tiny":
check = {
'256': [
input_shape[0],
input_shape[1] // 16,
input_shape[2] // 16,
256],
'1024': [
input_shape[0],
input_shape[1] // 32,
input_shape[2] // 32,
1024],
}
else:
check = None
init = tf.random_normal_initializer()
x = tf.Variable(initial_value=init(
shape=input_shape, dtype=tf.float32))
y = builder.Backbone_Builder(model)(x)
y_shape = {key: value.shape.as_list() for key, value in y.items()}
self.assertAllEqual(check, y_shape)
print(y_shape, check)
return