fc6c75deed83ab3b85c47e53656ab85289eaea66,examples/opensets/mnist_model3.py,,,#,20

Before Change

    mnist = MNIST()

    train_template = (Pipeline(config=dict(model=VGG7))
                .init_variable("model", DenseNet121)
                .init_variable("loss_history", init_on_each_run=list)
                .init_variable("current_loss", init_on_each_run=0)
                .init_variable("pred_label", init_on_each_run=list)
                .init_model("dynamic", V("model"), "conv",
                            config={"inputs": dict(images={"shape": B("image_shape")},
                                                   labels={"classes": 10, "transform": "ohe", "name": "targets"}),
                                    "input_block/inputs": "images",
                                    //"input_block/filters": 16,
                                    //"body/block/bottleneck": 1,
                                    //"head/units": [100, 100, 10],
                                    //"nothing": F(lambda batch: batch.images.shape[1:]),
                                    //"body/block/filters": 16,
                                    //"body/block/width_factor": 2,
                                    //"body": dict(se_block=1, se_factor=4, resnext=1, resnext_factor=4, bottleneck=1),
                                    "output": dict(ops=["accuracy"])})
                .resize(shape=(64, 64))
                .train_model("conv", fetches="loss",
                                     feed_dict={"images": B("images"),
                                                "labels": B("labels")},
                             save_to=V("current_loss"), use_lock=True)
                .print(V("current_loss"), model=V("model"))
                .update_variable("loss_history", V("current_loss"), mode="a"))

    train_pp = (train_template << mnist.train)
    print("Start training...")
    t = time()
    train_pp.run(BATCH_SIZE, shuffle=True, n_epochs=1, drop_last=False, prefetch=0)
    print("End training", time() - t)


    print()
    print("Start testing...")
    t = time()
    test_pp = (mnist.test.p
                .import_model("conv", train_pp)
                .init_variable("accuracy", init_on_each_run=list)
                .predict_model("conv", fetches="output_accuracy", feed_dict={"images": B("images"),
                                                                             "labels": B("labels")},
                               save_to=V("accuracy"), mode="a")
                .run(BATCH_SIZE, shuffle=True, n_epochs=1, drop_last=True, prefetch=0))
    print("End testing", time() - t)

    accuracy = np.array(test_pp.get_variable("accuracy")).mean()
    print("Accuracy {:6.2f}".format(accuracy))


    conv = train_pp.get_model_by_name("conv")

After Change

    mnist = MNIST()

    train_template = (Pipeline(config=dict(model=VGG7))
                .init_variable("model", VGG7)
                .init_variable("loss_history", init_on_each_run=list)
                .init_variable("current_loss", init_on_each_run=0)
                .init_model("dynamic", V("model"), "conv",
                            config={"inputs": dict(images={"shape": B("image_shape")},
                                                   labels={"classes": 10, "transform": "ohe", "name": "targets"}),
                                    "input_block/inputs": "images",
                                    //"input_block/filters": 16,
                                    //"body/block/bottleneck": 1,
                                    "head/units": [100, 100, 10],
                                    //"nothing": F(lambda batch: batch.images.shape[1:]),
                                    //"body/block/filters": 16,
                                    //"body/block/width_factor": 2,
                                    //"body": dict(se_block=1, se_factor=4, resnext=1, resnext_factor=4, bottleneck=1),
                                    "output": dict(ops=["accuracy"])})
                //.resize(shape=(64, 64))
                .train_model("conv", fetches="loss",
                                     feed_dict={"images": B("images"),
                                                "labels": B("labels")},
                             save_to=V("current_loss"), use_lock=True)
                .print(V("current_loss"), model=V("model"))
                .update_variable("loss_history", V("current_loss"), mode="a"))

    train_pp = (train_template << mnist.train)
    print("Start training...")
    t = time()
    print(train_pp)
    train_pp.run(BATCH_SIZE, shuffle=True, n_epochs=1, drop_last=False, prefetch=0)
    print("End training", time() - t)


    print()
    print("Start testing...")
    t = time()
    test_pp = (mnist.test.p
                .import_model("conv", train_pp)
                .init_variable("accuracy", init_on_each_run=list)
                .predict_model("conv", fetches="output_accuracy", feed_dict={"images": B("images"),
                                                                             "labels": B("labels")},
                               save_to=V("accuracy"), mode="a")
                .run(BATCH_SIZE, shuffle=True, n_epochs=1, drop_last=True, prefetch=0))
    print("End testing", time() - t)

    accuracy = np.array(test_pp.get_variable("accuracy")).mean()
    print("Accuracy {:6.2f}".format(accuracy))


    conv = train_pp.get_model_by_name("conv")

Instances