example["targets"] = tf.reshape(example["targets"],
[BATCH_SIZE, 1, 1, 1]) # Make it 4D.
loss, gv = loss_fn(example)
optimizer.apply_gradients(gv)
if count % 50 == 0:
print("Step: %d, Loss: %.3f" % (count, loss.numpy()))
if count >= NUM_STEPS:
break
model.set_mode(Modes.EVAL)
mnist_eval_dataset = mnist_problem.dataset(Modes.EVAL, data_dir)
# Create eval metric accumulators for accuracy (ACC) and accuracy in
# top 5 (ACC_TOP5)
metrics_accum, metrics_result = metrics.create_eager_metrics(
[metrics.Metrics.ACC, metrics.Metrics.ACC_TOP5])
for count, example in enumerate(tfe.Iterator(mnist_eval_dataset)):
if count >= 200:
break
# Make the inputs and targets 4D
example["inputs"] = tf.reshape(example["inputs"], [1, 28, 28, 1])
example["targets"] = tf.reshape(example["targets"], [1, 1, 1, 1])
# Call the model
predictions, _ = model(example)
# Compute and accumulate metrics
metrics_accum(predictions, example["targets"])
#print(2)
optimizer.apply_gradients(gv)
#print(3)
if count % 10 == 0:
print("Step: %d, Loss: %.3f" % (count, loss.numpy()))
if count >= NUM_STEPS:
break
model.set_mode(Modes.EVAL)
ptb_eval_dataset = ptb_problem.dataset(Modes.EVAL, data_dir)
#b_train_dataset = ptb_train_dataset.repeat(None).padded_batch(1,ptb_train_dataset.output_shapes)
# Create eval metric accumulators for accuracy (ACC) and accuracy in
# top 5 (ACC_TOP5)
metrics_accum, metrics_result = metrics.create_eager_metrics(
[metrics.Metrics.NEG_LOG_PERPLEXITY, metrics.Metrics.APPROX_BLEU])
for count, example in enumerate(tfe.Iterator(ptb_eval_dataset)):
if count >= 200:
break
# Make the inputs and targets 4D
example["targets"] = tf.reshape(example["targets"], [1, -1, 1, 1])
#example["targets"] = tf.reshape(example["targets"], [1, 1, 1, 1])
# Call the model
predictions, _ = model(example)
#predictions = decode(output)
# Compute and accumulate metrics
metrics_accum(predictions, example["targets"])