9.4. TensorFlow, PyTorch and Keras

class GRU(tf.keras.Model):
    def __init__(self, hidden_dim, output_dim):
        super().__init__()
        self.gru = tf.keras.layers.GRU(
            hidden_dim,
            activation="tanh",
            recurrent_activation="sigmoid",
            kernel_initializer="glorot_normal",
            recurrent_initializer="orthogonal",
        )
        self.dense = tf.keras.layers.Dense(output_dim, activation="linear")

    def call(self, x):
        x = self.gru(x)
        x = self.dense(x)
        return x

n_sequence = 25
n_data = 100

n_sample = n_data - n_sequence  # number of sample

sin_data = ds.create_wave(n_data, 0.05)
X, Y = ds.dataset(sin_data, n_sequence, False)

X = X.reshape(X.shape[0], X.shape[1], 1)
Y = Y.reshape(Y.shape[0], Y.shape[1])

input_units = 1
hidden_units = 32
output_units = 1

model = GRU(hidden_units, output_units)

model.build(input_shape=(None, n_sequence, input_units))
model.summary()

lr = 0.001
beta1 = 0.9
beta2 = 0.999

criterion = losses.MeanSquaredError()
optimizer = optimizers.Adam(learning_rate=lr, beta_1=beta1, beta_2=beta2)
train_loss = metrics.Mean()

# ============================
# Training
# ============================

lr = 0.001
beta1 = 0.9
beta2 = 0.999

criterion = losses.MeanSquaredError()
optimizer = optimizers.Adam(learning_rate=lr, beta_1=beta1, beta_2=beta2)
train_loss = metrics.Mean()

n_epochs = 300

for epoch in range(1, n_epochs + 1):

    with tf.GradientTape() as tape:
        preds = model(X)
        loss = criterion(Y, preds)
    grads = tape.gradient(loss, model.trainable_variables)
    optimizer.apply_gradients(zip(grads, model.trainable_variables))
    train_loss(loss)

$ python GRU_tf.py
Model: "gru"
_________________________________________________________________
 Layer (type)                Output Shape              Param #
=================================================================
 gru_1 (GRU)                 multiple                  3360

 dense (Dense)               multiple                  33

=================================================================
Total params: 3,393
Trainable params: 3,393
Non-trainable params: 0
_________________________________________________________________
epoch: 1/300, loss: 0.651

... snip ...

n_sequence = 25
n_data = 100

n_sample = n_data - n_sequence  # number of sample

sin_data = ds.create_wave(n_data, 0.05)
X, Y = ds.dataset(sin_data, n_sequence, False)

X = X.reshape(X.shape[0], X.shape[1], 1)
Y = Y.reshape(Y.shape[0], Y.shape[1])

if torch.cuda.is_available():
    device = torch.device("cuda")
elif torch.backends.mps.is_available():
    device = torch.device("mps")
else:
    device = torch.device("cpu")
print(device)


class GRU(nn.Module):
    def __init__(self, input_units, hidden_units, output_units):
        super().__init__()
        self.gru = nn.GRU(input_units, hidden_units, batch_first=True)
        self.dense = nn.Linear(hidden_units, output_units)

        nn.init.xavier_normal_(self.gru.weight_ih_l0)
        nn.init.orthogonal_(self.gru.weight_hh_l0)

    def forward(self, x):
        h, _ = self.gru(x)
        y = self.dense(h[:, -1])
        return y

input_units = 1
hidden_units = 32
output_units = 1

model = GRU(input_units, hidden_units, output_units).to(device)

summary(model=GRU(input_units, hidden_units, output_units), input_size=X.shape)

lr = 0.001
beta1 = 0.9
beta2 = 0.999

criterion = nn.MSELoss(reduction="mean")
optimizer = optimizers.Adam(
    model.parameters(), lr=lr, betas=(beta1, beta2), amsgrad=True
)

# ============================
# Training
# ============================
n_epochs = 300
history_loss = []

for epoch in range(1, n_epochs + 1):
    train_loss = 0.0

    x = torch.Tensor(X).to(device)
    y = torch.Tensor(Y).to(device)
    model.train()
    preds = model(x)
    loss = criterion(y, preds)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    train_loss += loss.item()

    history_loss.append(train_loss)

$ python GRU_pt.py
mps
==========================================================================================
Layer (type:depth-idx)                   Output Shape              Param #
==========================================================================================
GRU                                      [75, 1]                   --
├─GRU: 1-1                               [75, 25, 32]              3,360
├─Linear: 1-2                            [75, 1]                   33
==========================================================================================
Total params: 3,393
Trainable params: 3,393
Non-trainable params: 0
Total mult-adds (Units.MEGABYTES): 6.30
==========================================================================================
Input size (MB): 0.01
Forward/backward pass size (MB): 0.48
Params size (MB): 0.01
Estimated Total Size (MB): 0.50
==========================================================================================
epoch: 10/300, loss: 0.593

... snip ...

# ============================
# Model creation
# ============================

lr = 0.001

hidden_dim = 32
input_dim = 1
output_dim = 1

model = Sequential()
model.add(
    GRU(
        hidden_dim,
        activation="tanh",
        recurrent_activation="sigmoid",
        use_bias=True,
        batch_input_shape=(None, n_sequence, input_dim),
        return_sequences=False,
    )
)
model.add(Dense(output_dim, activation="linear", use_bias=True))
model.compile(loss="mean_squared_error", optimizer=Adam(lr))

# ============================
# Training
# ============================

n_epochs = 100

batch_size = 5

history_rst = model.fit(
    X,
    Y,
    batch_size=batch_size,
    epochs=n_epochs,
    validation_split=0.1,
    shuffle=True,
    verbose=1,
)

$ python GRU_keras.py

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #
=================================================================
 gru (GRU)                   (None, 32)                3360

 dense (Dense)               (None, 1)                 33

=================================================================
Total params: 3,393
Trainable params: 3,393
Non-trainable params: 0
_________________________________________________________________
Epoch 1/100

... snip ...