s += 1
loss, xent = step(x, y)
if tf.equal(s % 10, 0):
e = ps.metric.result()
tf.print('Step:', s, ', loss:', loss, ', xent:', e)
return loss, xent
for e in range(ps.num_epochs):
loss, xent = epoch()
print(f'Epoch {e} loss:', loss.numpy(), ', xent:', xent.numpy())
params = dict(
dim_batch=2,
dim_dense=150,
dim_hidden=15,
dim_vocab=len(qd.vocab),
len_max_input=20,
loss=ks.losses.SparseCategoricalCrossentropy(from_logits=True),
metric=ks.metrics.SparseCategoricalCrossentropy(from_logits=True),
num_epochs=10,
num_shards=2,
optimizer=ks.optimizers.Adam(),
)
if __name__ == '__main__':
ps = qd.Params(**params)
# import advanced_tf.masking as qm
# qm.main_graph(ps, dset_for(ps), model_for(ps))
main_eager(ps, dset_for(ps), model_for(ps))
def main_graph(ps, ds, m):
b = pth.Path('/tmp/q')
b.mkdir(parents=True, exist_ok=True)
lp = datetime.now().strftime('%Y%m%d-%H%M%S')
lp = b / f'logs/{lp}'
c = tf.train.Checkpoint(model=m)
mp = b / 'model' / f'{m.name}'
mgr = tf.train.CheckpointManager(c, str(mp), max_to_keep=3)
# if mgr.latest_checkpoint:
# vs = tf.train.list_variables(mgr.latest_checkpoint)
# print(f'\n*** checkpoint vars: {vs}')
c.restore(mgr.latest_checkpoint).expect_partial()
class CheckpointCB(ks.callbacks.Callback):
def on_epoch_end(self, epoch, logs=None):
mgr.save()
cbs = [
CheckpointCB(),
ks.callbacks.TensorBoard(
log_dir=str(lp),
histogram_freq=1,
),
]
m.fit(ds, callbacks=cbs, epochs=ps.num_epochs)
if __name__ == '__main__':
ps = qd.Params(**qc.params)
main_graph(ps, qc.dset_for(ps), model_for(ps))