CNN模型的预测

Question

I am trying to do a model to recognize one human movement. My activity is to recognize a person catching. I already have a record with about 260 right movements and I annoted that file with labels "catch" and "nothing" to train my model. I have another file, this file is annoted too, just to test the accuracy of the model.

I am using a CNN model based on this site . And I modified the window size to 400 (equivalent of 4s of record) So after trainning my model I have this results (my model is saved and is already trained from a previous record, that's why it gives good results in the first epochs):

Epoch:  0  Training Loss:  0.5428493594505138  Training Accuracy:  0.99394274

Epoch:  1  Training Loss:  0.5227164919283446  Training Accuracy:  0.99394274

Epoch:  2  Training Loss:  0.5037865922760709  Training Accuracy:  0.99449337

Epoch:  3  Training Loss:  0.4860136515261339  Training Accuracy:  0.99614537

Testing Accuracy: 0.5686275

My code to train is:

for epoch in range(training_epochs):
        cost_history = np.empty(shape=[1], dtype=float)
        for b in range(total_batchs):
            offset = (b * batch_size) % (train_y.shape[0] - batch_size)
            batch_x = train_x[offset:(offset + batch_size), :, :, :]
            batch_y = train_y[offset:(offset + batch_size), :]
            _, c = session.run([optimizer, loss], feed_dict={X: batch_x, Y: batch_y})
            cost_history = np.append(cost_history, c)
            print("Epoch: ", epoch, " Training Loss: ", np.mean(cost_history), " Training Accuracy: ",                session.run(accuracy, feed_dict={X: train_x, Y: train_y}))
        print("Testing Accuracy:", session.run(accuracy, feed_dict={X: test_input, Y: test_labels}))

After that, I wanna predict some results whit this:

prediction = session.run(y_, feed_dict={X: predict_input})

But the results were:

[[7.6319778e-04 9.9923682e-01]

 [3.3351363e-04 9.9966645e-01]

 [2.5510782e-04 9.9974483e-01]

...

 [2.5133172e-04 9.9974865e-01]

 [2.4705922e-04 9.9975294e-01]

 [3.0652966e-04 9.9969351e-01]

 [1.5634180e-04 9.9984360e-01]]

Questions at Hand:

• Am I doing something wrong?
• Shouldn't the values of the predictions be higher values like [0.80 0.20]?
• The window size is good with 400?
• In prediction each line corresponds to a window size of data am I right?

Answer 1

Clarifying range of prediction values

as I see it, your values are 0.007 and 0.993 or something in that Magnitude - so what exactly do you mean by higher, since the sum is always 1. So if one gets higher the other should get smaller - if I understood your Problem correctly.

Clarifying prediction correspondance

Since you labeled your data with the classes "catch" and "nothing" that's what your predictions correspond to. So an Output of (0, 1) would mean that your Network predicts the Input to be of the second class.

Clarifying window size

To solve this I first Need you to explain, what you mean by window size. The amount of recorded seconds that you use as an input? If so, I would say, that if you can recognize a "catch" within this time Frame - then your Network should also be able to do so. Other than that, this would be an appropriate case for a line search.