SimonBlanke /
Hyperactive
| @@ 24-90 (lines=67) @@ | ||
| 21 | y_test = mnist.test.labels |
|
| 22 | ||
| 23 | ||
| 24 | def cnn(para, X_train, y_train): |
|
| 25 | def conv_net(x_dict, n_classes, dropout, reuse, is_training): |
|
| 26 | with tf.variable_scope("ConvNet", reuse=reuse): |
|
| 27 | x = x_dict["images"] |
|
| 28 | x = tf.reshape(x, shape=[-1, 28, 28, 1]) |
|
| 29 | conv1 = tf.layers.conv2d(x, para["filters_0"], 5, activation=tf.nn.relu) |
|
| 30 | conv1 = tf.layers.max_pooling2d(conv1, 2, 2) |
|
| 31 | conv2 = tf.layers.conv2d(conv1, para["filters_1"], 3, activation=tf.nn.relu) |
|
| 32 | conv2 = tf.layers.max_pooling2d(conv2, 2, 2) |
|
| 33 | fc1 = tf.contrib.layers.flatten(conv2) |
|
| 34 | fc1 = tf.layers.dense(fc1, para["dense_0"]) |
|
| 35 | fc1 = tf.layers.dropout(fc1, rate=dropout, training=is_training) |
|
| 36 | out = tf.layers.dense(fc1, n_classes) |
|
| 37 | ||
| 38 | return out |
|
| 39 | ||
| 40 | def model_fn(features, labels, mode): |
|
| 41 | logits_train = conv_net( |
|
| 42 | features, num_classes, dropout, reuse=False, is_training=True |
|
| 43 | ) |
|
| 44 | logits_test = conv_net( |
|
| 45 | features, num_classes, dropout, reuse=True, is_training=False |
|
| 46 | ) |
|
| 47 | ||
| 48 | pred_classes = tf.argmax(logits_test, axis=1) |
|
| 49 | # pred_probas = tf.nn.softmax(logits_test) |
|
| 50 | ||
| 51 | if mode == tf.estimator.ModeKeys.PREDICT: |
|
| 52 | return tf.estimator.EstimatorSpec(mode, predictions=pred_classes) |
|
| 53 | ||
| 54 | loss_op = tf.reduce_mean( |
|
| 55 | tf.nn.sparse_softmax_cross_entropy_with_logits( |
|
| 56 | logits=logits_train, labels=tf.cast(labels, dtype=tf.int32) |
|
| 57 | ) |
|
| 58 | ) |
|
| 59 | optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) |
|
| 60 | train_op = optimizer.minimize(loss_op, global_step=tf.train.get_global_step()) |
|
| 61 | ||
| 62 | acc_op = tf.metrics.accuracy(labels=labels, predictions=pred_classes) |
|
| 63 | ||
| 64 | estim_specs = tf.estimator.EstimatorSpec( |
|
| 65 | mode=mode, |
|
| 66 | predictions=pred_classes, |
|
| 67 | loss=loss_op, |
|
| 68 | train_op=train_op, |
|
| 69 | eval_metric_ops={"accuracy": acc_op}, |
|
| 70 | ) |
|
| 71 | ||
| 72 | return estim_specs |
|
| 73 | ||
| 74 | model = tf.estimator.Estimator(model_fn) |
|
| 75 | ||
| 76 | input_fn = tf.estimator.inputs.numpy_input_fn( |
|
| 77 | x={"images": X_train}, |
|
| 78 | y=y_train, |
|
| 79 | batch_size=batch_size, |
|
| 80 | num_epochs=None, |
|
| 81 | shuffle=True, |
|
| 82 | ) |
|
| 83 | model.train(input_fn, steps=num_steps) |
|
| 84 | ||
| 85 | input_fn = tf.estimator.inputs.numpy_input_fn( |
|
| 86 | x={"images": X_test}, y=y_test, batch_size=batch_size, shuffle=False |
|
| 87 | ) |
|
| 88 | e = model.evaluate(input_fn) |
|
| 89 | ||
| 90 | return float(e["accuracy"]) |
|
| 91 | ||
| 92 | ||
| 93 | search_config = { |
|
| @@ 24-90 (lines=67) @@ | ||
| 21 | y_test = mnist.test.labels |
|
| 22 | ||
| 23 | ||
| 24 | def cnn(para, X_train, y_train): |
|
| 25 | def conv_net(x_dict, n_classes, dropout, reuse, is_training): |
|
| 26 | with tf.variable_scope("ConvNet", reuse=reuse): |
|
| 27 | x = x_dict["images"] |
|
| 28 | x = tf.reshape(x, shape=[-1, 28, 28, 1]) |
|
| 29 | conv1 = tf.layers.conv2d(x, para["filters_0"], 5, activation=tf.nn.relu) |
|
| 30 | conv1 = tf.layers.max_pooling2d(conv1, 2, 2) |
|
| 31 | conv2 = tf.layers.conv2d(conv1, para["filters_1"], 3, activation=tf.nn.relu) |
|
| 32 | conv2 = tf.layers.max_pooling2d(conv2, 2, 2) |
|
| 33 | fc1 = tf.contrib.layers.flatten(conv2) |
|
| 34 | fc1 = tf.layers.dense(fc1, para["dense_0"]) |
|
| 35 | fc1 = tf.layers.dropout(fc1, rate=dropout, training=is_training) |
|
| 36 | out = tf.layers.dense(fc1, n_classes) |
|
| 37 | ||
| 38 | return out |
|
| 39 | ||
| 40 | def model_fn(features, labels, mode): |
|
| 41 | logits_train = conv_net( |
|
| 42 | features, num_classes, dropout, reuse=False, is_training=True |
|
| 43 | ) |
|
| 44 | logits_test = conv_net( |
|
| 45 | features, num_classes, dropout, reuse=True, is_training=False |
|
| 46 | ) |
|
| 47 | ||
| 48 | pred_classes = tf.argmax(logits_test, axis=1) |
|
| 49 | # pred_probas = tf.nn.softmax(logits_test) |
|
| 50 | ||
| 51 | if mode == tf.estimator.ModeKeys.PREDICT: |
|
| 52 | return tf.estimator.EstimatorSpec(mode, predictions=pred_classes) |
|
| 53 | ||
| 54 | loss_op = tf.reduce_mean( |
|
| 55 | tf.nn.sparse_softmax_cross_entropy_with_logits( |
|
| 56 | logits=logits_train, labels=tf.cast(labels, dtype=tf.int32) |
|
| 57 | ) |
|
| 58 | ) |
|
| 59 | optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) |
|
| 60 | train_op = optimizer.minimize(loss_op, global_step=tf.train.get_global_step()) |
|
| 61 | ||
| 62 | acc_op = tf.metrics.accuracy(labels=labels, predictions=pred_classes) |
|
| 63 | ||
| 64 | estim_specs = tf.estimator.EstimatorSpec( |
|
| 65 | mode=mode, |
|
| 66 | predictions=pred_classes, |
|
| 67 | loss=loss_op, |
|
| 68 | train_op=train_op, |
|
| 69 | eval_metric_ops={"accuracy": acc_op}, |
|
| 70 | ) |
|
| 71 | ||
| 72 | return estim_specs |
|
| 73 | ||
| 74 | model = tf.estimator.Estimator(model_fn) |
|
| 75 | ||
| 76 | input_fn = tf.estimator.inputs.numpy_input_fn( |
|
| 77 | x={"images": X_train}, |
|
| 78 | y=y_train, |
|
| 79 | batch_size=batch_size, |
|
| 80 | num_epochs=None, |
|
| 81 | shuffle=True, |
|
| 82 | ) |
|
| 83 | model.train(input_fn, steps=num_steps) |
|
| 84 | ||
| 85 | input_fn = tf.estimator.inputs.numpy_input_fn( |
|
| 86 | x={"images": X_test}, y=y_test, batch_size=batch_size, shuffle=False |
|
| 87 | ) |
|
| 88 | e = model.evaluate(input_fn) |
|
| 89 | ||
| 90 | return float(e["accuracy"]) |
|
| 91 | ||
| 92 | ||
| 93 | search_config = { |
|
