def do_train_noweights(model, dataset_batches):
optimizer = AdamOptimizer()
loss_history = []
for (batch, (invec, labels)) in enumerate(dataset_batches.take(1000)):
with tf.GradientTape() as tape:
logits = model(invec, training=True)
loss_value = sparse_softmax_cross_entropy(labels, logits)
loss_history.append(loss_value.numpy())
grads = tape.gradient(loss_value, model.trainable_variables)
optimizer.apply_gradients(
zip(grads, model.trainable_variables),
global_step=tf.compat.v1.train.get_or_create_global_step())
return loss_history
def do_train(model,
dataset_batches,
k,
epochs=None,
num_epochs=None,
optimizer_params=None,
saveloc=None):
optimizer = AdamOptimizer(**optimizer_params)
loss_history = []
label_losses_history = []
if num_epochs:
epochs = range(num_epochs)
#weights = tf.constant(np.ones((32, 1)))
weights_dict = {0: 1., 1: 1., 2: 1., 3: 1.}
for epoch in epochs:
for (batch, (invec, labels,
_)) in enumerate(tqdm(dataset_batches.take(k))):
if labels.shape[0] < 32:
continue
weights = np.array(
[weights_dict[labels[i].numpy()] for i in range(32)])
# NOTE: is this tape in the right place?
with tf.GradientTape() as tape:
logits = model(invec, training=True)
loss_value = sparse_softmax_cross_entropy(labels,
logits,
weights=weights)
indices_vec = [[
i for i in range(32) if labels[i].numpy() == label
] for label in [0, 1, 2, 3]]
losses = [
sparse_softmax_cross_entropy(labels.numpy()[indices],
logits.numpy()[indices],
weights=weights[indices])
for indices in indices_vec
]
weights_dict = weights_for_losses(losses)
loss_history.append(loss_value.numpy())
label_losses_history.append([x.numpy() for x in losses])
grads = tape.gradient(loss_value, model.trainable_variables)
optimizer.apply_gradients(
zip(grads, model.trainable_variables),
global_step=tf.compat.v1.train.get_or_create_global_step())
prefix = (f'{saveloc}/epoch_{str(epoch).zfill(3)}'
f'_batch_{str(batch).zfill(5)}')
if batch % 10 == 0:
save_model(model, (f'{prefix}_model.h5'))
to_json_local(loss_history,
f'{prefix}_train_loss_history.json')
to_json_local(label_losses_history,
f'{prefix}_train_label_losses_history.json')
return loss_history