def run_epoch(self, session):
"""
- DOES: runs one epoch, i.e., for each batch it: computes the batch loss
(forwardprop), computes all gradients w.r.t to the batch loss and updates
all network variables/parameters accordingly (backprop).
"""
batch_losses = []
for step, (captions, imgs,
labels) in enumerate(train_data_iterator(self)):
# create a feed_dict with the batch data:
feed_dict = self.create_feed_dict(captions,
imgs,
labels_batch=labels,
dropout=self.config.dropout)
# compute the batch loss and compute & apply all gradients w.r.t to
# the batch loss (without self.train_op in the call, the network
# would not train, we would only compute the batch loss):
batch_loss, _ = session.run([self.loss, self.train_op],
feed_dict=feed_dict)
batch_losses.append(batch_loss)
if step % 100 == 0:
print "batch: %d | loss: %f" % (step, batch_loss)
log("batch: %d | loss: %f" % (step, batch_loss))
if step > 5 and self.debug:
break
# return a list containing the batch loss for each batch:
return batch_losses
def run_epoch(self, session):
batch_losses = []
for step, (data_batch, labels_batch,
sequence_length) in enumerate(ut.train_data_iterator(self)):
feed_dict = self.create_feed_dict(data_batch, labels_batch,
self.config.dropout,
sequence_length)
batch_loss, _ = session.run([self.loss, self.train_op],
feed_dict=feed_dict)
batch_losses.append(batch_loss)
return batch_losses