I am trying to postprocess my model's prediction before computing the loss function, since my true data (y_train) is the outer product of the NN output. I have followed these steps:
- I know that the operation I am trying to do usign numpy is:
nX = 201
nT = 101
nNNout = nX nT
nBatch = 32
NNout = np.random.rand(nBatch, nNNout)
f = NNout[:, :nX]
g = NNout[:,nX:]
test = np.empty([nBatch, nX*nT])
for i in range(nBatch):
test[i,:] = np.outer(f[i,:], g[i,:]).flatten('F')
where the NN output contains f and g. What I actually need is the vectorised version of the outer product of f and g for each batch instance.
- I have translated this in a compact tensorflow operation as:
test2 = tf.Variable([tf.reshape(tf.transpose(tf.tensordot(f[i,:],g[i,:], axes=0)),[nX*nT]) for i in range(nBatch)])
which I have checked that is correct and that outputs the same values than in step 1.
- Then, I am just trying to add this operation after the prediciton of my model as:
n_epochs = 20
batch_size = 32
n_steps = len(x_train) // batch_size
optimizer = keras.optimizers.Nadam(learning_rate=0.01)
loss_fn = keras.losses.mean_squared_error
mean_loss = keras.metrics.Mean()
metrics = [keras.metrics.MeanAbsoluteError()]
# ------------ Training ------------
for epoch in range(1, n_epochs 1):
print("Epoch {}/{}".format(epoch, n_epochs))
for step in range(1, n_steps 1):
X_batch, y_batch = random_batch(x_train, np.array(y_train))
with tf.GradientTape() as tape:
y_pred = model(X_batch, training=True)
u_pred = tf.Variable([tf.reshape(tf.transpose(tf.tensordot(y_pred[i, :nX], y_pred[i, nX:], axes=0)), [nX * nT]) for i in
range(batch_size)])
main_loss = tf.reduce_mean(loss_fn(y_batch, u_pred))
loss = tf.add_n([main_loss] model.losses)
gradients = tape.gradient(loss, model.trainable_variables)
My main issue is that gradients become a list of Nones when I add the operation. If I simply compute the loss function with my model's prediction (y_pred) the code is able to compute the gradients.
Could you please help me find the error I am making here?
CodePudding user response:
You are creating a new (trainable) variable in u_pred, thus breaking any dependency of u_pred on y_pred. The reason why value matches is because you initialise your new variable with the prediction, but it has no functional dependency on each other anymore, there are no gradients flowing.
I am guessing that you did that because you needed a tf.Tensor and not a list, and you ended up with type errors. You probably want to use something among the lines of tf.concatenate
and not tf.Variable
for that.