train_loss = 0
train_acc = 0
eval_acc = 0
for data, label in train_data:
data = data.as_in_context(mx_ctx)
label = label.as_in_context(mx_ctx)
with autograd.record():
output = net(data)
loss = loss_fn(output, label)
loss.backward()
trainer.step(batch_size=BATCH_SIZE)
train_loss += loss.mean().asscalar()
train_acc += acc(output, label)
for data, label in eval_data:
data = data.as_in_context(mx_ctx)
label = label.as_in_context(mx_ctx)
eval_acc += acc(net(data), label)
print("epoch {}, loss {:0.3f}, acc {:0.3f}, val_acc {:0.3f}".format(
epoch, train_loss / len(train_data), train_acc / len(train_data),
eval_acc / len(eval_data)))
print("Elapsed time {:02f} seconds".format(time.time() - t0))
# %%
# -- Save parameters
net.save_parameters('fashion_mnist_gluon.params')
train_loss /= len(train_data)
eval_loss /= len(eval_data)
_, train_acc = train_metric.get()
_, eval_acc = eval_metric.get()
img_per_sec = len(mnist_train) / chrono_epoch
print(
f"{len(mnist_train)}, epoch {epoch}, train_loss {train_loss:0.3f}, acc {train_acc:0.3f}, val_loss: {eval_loss:0.3f}, val_acc {eval_acc:0.3f}, lr: {trainer.learning_rate}, img/sec: {img_per_sec:0.2f}, chrono: {chrono_epoch:0.2f}"
)
# Add some values to mxboard
sw.add_scalar("Accuracy", ("train_acc", train_acc), global_step=epoch)
sw.add_scalar("Accuracy", ("val_acc", eval_acc), global_step=epoch)
for key, param in net.collect_params(".*weight").items():
sw.add_histogram(f"grad_{param.name}",
param.grad(ctx=mx_ctx[0]),
global_step=epoch)
# Save parameters if they are better than the previous epoch
if eval_acc > prev_val_acc:
net.save_parameters(f"{param_file_prefix}-{epoch:04d}.params")
prev_val_acc = eval_acc
print("=== Total training time: {:02f} seconds".format(time.time() - t0))
# %%
# -- Save parameters
net.save_parameters(f"{param_file_prefix}-final.params")
