1
$\begingroup$

I'm trying to develop a small network for the aerial cactus identification challenge.

This is a binary classification challenge (0 no cactus, 1 is cactus), but my networks is always outputting 1.

I have identified that the network always output values >= 0.5, so the sigmoid function outputs 1. However, I can't understand why. I built a simple network, which works when implemented with keras, but I'm trying to use tf.nn for learning purposes, but can't make it work

My network architecture:

2 Conv 64x3 + Maxpooling

2 Conv 128x3 + Maxpooling

Flatten

Dense 1024

Dense 512

Dense 1

I create the layers this way:

    initializer = tf.keras.initializers.glorot_uniform()
    def new_weights(shape, name='W', glorot=False):
        content = tf.random.normal(shape, stddev=0.03)
        if glorot:
            content = initializer(shape)
        return tf.Variable(content)

    def conv_block(inputs, n_filters, n_size, weights, stage=0):
        shape = inputs.shape

        weights1 = weights[f'W_{stage}_1']
        bias = weights[f'b_{stage}_1']

        layer = tf.nn.conv2d(
            inputs,
            filters=weights1,
            strides=[1, 1, 1, 1],
            padding='VALID',
            name=f'W_{stage}_1'
        )
        layer += bias
        activated = tf.nn.relu(layer)

        weights2 = weights[f'W_{stage}_2']
        bias2 = weights[f'b_{stage}_2']

        layer = tf.nn.conv2d(
            activated,
            filters=weights2,
            strides=[1, 1, 1, 1],
            padding='VALID',
            name=f'W_{stage}_2'
        )
        layer += bias2
        activated = tf.nn.relu(layer)
        activated = tf.nn.max_pool(activated, ksize=2, strides=2,padding='VALID')
        print(stage, activated.shape)
        return activated, weights

    def flatten(conv):
        flat = tf.reshape(conv, [-1, 3200])
        print('flat', flat.shape)
        return flat

    def dense_block(flat, nb, weights, stage=0, relu=True):
        weights1 = weights[f'Wd_{stage}_1']
        bias = weights[f'bd_{stage}_1']

        res = tf.matmul(flat, weights1) + bias
        if relu:
            res = tf.nn.relu(res)
        print('d', stage, res.shape)
        return res

Here is my assembling of the layers:

    weights = {
        'W_0_1': new_weights([3, 3, 3, 64], 'W_0_1', glorot=True),
        'b_0_1': new_weights([64], 'b_0_1'),
        'W_0_2': new_weights([3, 3, 64, 64], 'W_0_2', glorot=True),
        'b_0_2': new_weights([64], 'b_0_2'),
        'W_1_1': new_weights([3, 3, 64, 128], 'W_1_1', glorot=True),
        'b_1_1': new_weights([128], 'b_1_1'),
        'W_1_2': new_weights([3, 3, 128, 128], 'W_1_2', glorot=True),
        'b_1_2': new_weights([128], 'b_1_2'),
        'Wd_0_1': new_weights([3200, 1024], 'Wd_0_1'),
        'bd_0_1': new_weights([1024], 'bd_0_1'),
        'Wd_1_1': new_weights([1024, 512], 'Wd_1_1'),
        'bd_1_1': new_weights([512], 'bd_1_1'),
        'Wd_2_1': new_weights([512, 1], 'Wd_2_1'),
        'bd_2_1': new_weights([1], 'bd_2_1'),
    }

    def process_one_batch(x, y):
        block1 = conv_block(x, 64, 3, weights, 0)

        block2, _ = conv_block(block1[0], 128, 3, weights, 1)
        flat = flatten(block2)

        dense1 = dense_block(flat, 1024, weights, 0)
        dense2 = dense_block(dense1, 512, weights, 1)
        dense2 = dense_block(dense2, 1, weights, 2, False)

        res = dense2
        print_op = tf.print("dense2:", dense2, tf.nn.sigmoid(dense2), output_stream=sys.stdout)

        #with tf.control_dependencies([print_op]):
        res = tf.nn.sigmoid(dense2)


        return res

    NB_EPOCHS = 5
    def create_dataset(X, y, batch_size=BATCH_SIZE, nb_epochs=NB_EPOCHS, batch=True):
        dataset = tf.data.Dataset.from_tensor_slices((X, y))
        dataset = dataset.map(my_process_path)
        if batch:
            dataset = dataset.batch(batch_size)
        dataset = dataset.repeat(nb_epochs)
        dataset = dataset.prefetch(buffer_size=2)
        iterator =  tf.data.make_one_shot_iterator(dataset)
        #iterator = dataset.make_one_shot_iterator()
        next_element = iterator.get_next()
        y_ = process_one_batch(next_element[0], next_element[1])
        return dataset, next_element, y_

    train_ds, (train_x, train_y), prediction = create_dataset(X_train.values, y_train.values)
    test_ds, (test_x, test_y), test_prediction = create_dataset(X_test.values, y_test.values, batch=True)

And the loss + training loop part:

    cross_entropy
    optimiser = tf.compat.v1.train.AdamOptimizer(learning_rate=0.001).minimize(cross_entropy)
    def get_acc(y_true, y_pred, threshold=0.5):
        to_check = tf.cast(tf.round(y_pred), tf.int64)
        correct_prediction = tf.equal(y_true, to_check)
        print_op = tf.print("tensors:", y_true, to_check, y_pred, correct_prediction, output_stream=sys.stdout)

        #with tf.control_dependencies([print_op]):
        accuracy_ = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        return accuracy_

    accuracy = get_acc(test_y, tf.reshape(test_prediction, [-1]))
    accuracy_train = get_acc(train_y, tf.reshape(prediction, [-1]))
    # setup the initialisation operator
    init_op = tf.global_variables_initializer()

    with tf.Session() as sess:
        train_steps = int(len(X_train.values) / BATCH_SIZE)
        val_steps = int(len(X_test.values) / BATCH_SIZE)
        # initialise the variables
        sess.run(init_op)
        print('Init')
        for epoch in range(NB_EPOCHS):
            avg_cost = 0
            train_acc = 0
            for i in range(train_steps):
                if i % 100 == 0:
                    pass
                    #print(epoch, i)
                #dbg = sess.run([next_element[0], next_element[1], y_])
                _, c, ac = sess.run([optimiser, cross_entropy,accuracy_train])
                avg_cost += c
                train_acc += ac
            avg_acc = 0
            for i in range(val_steps):
                acc = sess.run(accuracy)
                avg_acc += acc
            print('train_acc: ', train_acc/train_steps)
            #print(sess.run([accuracy]))
            print(train_steps, val_steps)
            print("Epoch:", (epoch + 1), "cost =", "{:.3f}, acc: {:.3f}".format(avg_cost / train_steps, avg_acc / val_steps ))
        print("\nTraining complete!")

I ran my previous keras model on the dataset created like this, and it runs fine, so I guess it's a mistake related to the network construction, but I can't seem to figure why are my values so high. Any help would be really appreciated.

Thanks !

$\endgroup$
8
  • $\begingroup$ Is the dataset balanced? It sounds like a typical result of an imbalanced dataset. F1 loss function works well for imbalanced datasets. $\endgroup$ Mar 17, 2020 at 18:10
  • $\begingroup$ Yes the dataset is imbalanced, but I got good results with the model built on Keras. Is Keras doing some auto balancing ? $\endgroup$
    – Papotitu
    Mar 17, 2020 at 18:32
  • $\begingroup$ Not experienced with Keras. Try the F1 loss function and see what happens. $\endgroup$ Mar 17, 2020 at 18:35
  • $\begingroup$ I'm not sure what you're referring to, the F1 I know is a model evaluation metrics, not a training loss. Could you specify your idea ? $\endgroup$
    – Papotitu
    Mar 17, 2020 at 18:45
  • $\begingroup$ You can implement the F1 as a loss function. It is common to use with imbalanced datasets. I have used it several times when accuracy fails and just predicts one class. Quick google: tensorflow.org/addons/api_docs/python/tfa/metrics/F1Score $\endgroup$ Mar 17, 2020 at 20:09

1 Answer 1

1
$\begingroup$

Most likely there is a bug in your implementation.

Your code could be revised to find bugs more easily:

  • You are reusing the same variables over and over (e.g., layer and activated). If you use unique variable names, it is easier to track the current state.

  • The functions might not be useful. If you rewrite your code inline, then it is easier to track the current state.

  • You reimplement functionality unnecessarily. For example, the TensorFlow package has an accuracy built-in tf.keras.metrics.Accuracy. If you replace your implementations with package-based code, it might eliminate the bug.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.