FIX: compile triplet loss within keras model

[WindQAQ]

Closes #295.

[Squadrick]

The docstring says: y_true: 1-D integer Tensor with shape [batch_size] of multiclass integer labels. The old assertion took care of that, with the new assert, we can pass in a higher dimensional Tensor, and it'll pass that assert.

Take for example y_true passed as a 3-D Tensor.

yt_shape = tf.shape(y_true) # Tensor("Shape:0", shape=(3,), dtype=int32)
yt_shape.shape # (3,)
yt_shape.shape.rank # 1, passes the new assert

The older assert seemed to do as intended, however during compile Keras passes in y_true as the following:

Tensor("y_true:0", shape=(None, None), dtype=float32)

This is what causes the older assert to fail.

[WindQAQ]

@Squadrick, you are right. I made some mistakes on shape inference here...

[WindQAQ]

I think it would be a dummy check as there is not explicit shape or rank checking in tf.keras.losses. @facaiy and @Squadrick, what do you feel about this?

[facaiy]

labels = tf.reshape(labels, [lshape[0], 1]) will fail for invalid shape, so I think it safe to remove the check.

[WindQAQ]

Facai, I add a testcase for invalid shape with calling triplet_semihard_loss directly. (Actually, I do not know how to specify the shape of y_true for keras sequential model...)

[facaiy]

I delete the requirement for new test case, Tzu-Wei. Let's add it later if we think it necessary :-) Apologized for the misleading message.

[facaiy]

@Squadrick Hi, Dheeraj, what do you think?

[Squadrick]

Yeah, I think it's best we remove the check for now. Taking a look at tf.keras.losses, they don't seem to do a shape check either. As long as the results are numerically accurate, it should be good to go.

[WindQAQ]

I delete the requirement for new test case, Tzu-Wei. Let's add it later if we think it necessary :-) Apologized for the misleading message.

No worries :-) It makes sense to add it when necessary.

[hhhwwwuuu]

ummm. so, how should I fix ?