如何使用estimator.export_savedmodel()来保存TensorFlow模型？
特别是，serving_input_receiver_fn()里应该放什么内容？
我已经基于VGGNet架构创建了一个自定义的估计器，我使用自己的图片并对图片进行了一些变换（可以在_parse_function()中看到这些变换）。
我已经阅读了这里的文档，但我并不确定具体应该为我的代码写什么（请参见下面的代码）。最终我想保存模型并使用TensorFlow Serving。
from __future__ import absolute_importfrom __future__ import divisionfrom __future__ import print_functionimport tensorflow as tfimport numpy as npfrom sklearn.model_selection import train_test_splitimport osimport matplotlib.pyplot as pltimport matplotlib.image as mpimgimport scipyfrom scipy import ndimageimport scipy.misctf.logging.set_verbosity(tf.logging.INFO)def cnn_model_fn(features, labels, mode): """Model function for CNN.""" # Input Layer input_layer = tf.reshape(features, [-1, 224, 224, 3]) # Convolutional Layer #1 conv1 = tf.layers.conv2d( inputs=input_layer, filters=64, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #2 conv2 = tf.layers.conv2d( inputs=conv1, filters=64, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Pooling Layer #1 pool1 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2) # Convolutional Layer #3 conv3 = tf.layers.conv2d( inputs=pool1, filters=128, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #4 conv4 = tf.layers.conv2d( inputs=conv3, filters=128, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) pool2 = tf.layers.max_pooling2d(inputs=conv4, pool_size=[2, 2], strides=2) # Convolutional Layer #5 conv5 = tf.layers.conv2d( inputs=pool2, filters=256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #6 conv6 = tf.layers.conv2d( inputs=conv5, filters=256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #7 conv7 = tf.layers.conv2d( inputs=conv6, filters=256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) pool3 = tf.layers.max_pooling2d(inputs=conv7, pool_size=[2, 2], strides=2) # Convolutional Layer #8 conv8 = tf.layers.conv2d( inputs=pool3, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #9 conv9 = tf.layers.conv2d( inputs=conv8, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #10 conv10 = tf.layers.conv2d( inputs=conv9, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) pool4 = tf.layers.max_pooling2d(inputs=conv10, pool_size=[2, 2], strides=2) # Convolutional Layer #11 conv11 = tf.layers.conv2d( inputs=pool4, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #12 conv12 = tf.layers.conv2d( inputs=conv11, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #12 conv13 = tf.layers.conv2d( inputs=conv12, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) pool5 = tf.layers.max_pooling2d(inputs=conv13, pool_size=[2, 2], strides=2) # Dense Layer pool5_flat = tf.reshape(pool5, [-1, 7 * 7 * 512]) dense1 = tf.layers.dense(inputs=pool5_flat, units=4096, activation=tf.nn.relu) dense2 = tf.layers.dense(inputs=dense1, units=4096, activation=tf.nn.relu) dense3 = tf.layers.dense(inputs=dense2, units=1024, activation=tf.nn.relu) dropout = tf.layers.dropout( inputs=dense3, rate=0.001, training=mode == tf.estimator.ModeKeys.TRAIN) logits1 = tf.layers.dense(inputs=dropout, units=2) logits2 = tf.layers.dense(inputs=dropout, units=4) predictions = { "classes1": tf.argmax(input=logits1, axis=1), "classes2": tf.argmax(input=logits2, axis=1), "probabilities1": tf.nn.softmax(logits1, name="softmax_tensor_1"), "probabilities2": tf.nn.softmax(logits2, name="softmax_tensor_2") } if mode == tf.estimator.ModeKeys.PREDICT: return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions) # Calculate Loss (for both TRAIN and EVAL modes) loss1 = tf.losses.sparse_softmax_cross_entropy(labels=labels[:,0], logits=logits1) loss2 = tf.losses.sparse_softmax_cross_entropy(labels=labels[:,1], logits=logits2) loss = loss1 + loss2 # Configure the Training Op (for TRAIN mode) if mode == tf.estimator.ModeKeys.TRAIN: optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01) train_op = optimizer.minimize( loss=loss, global_step=tf.train.get_global_step()) return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op) # Add evaluation metrics (for EVAL mode) eval_metric_ops = { "accuracy1": tf.metrics.accuracy( labels=labels[:,0], predictions=predictions["classes1"]), "accuracy2": tf.metrics.accuracy( labels=labels[:,1], predictions=predictions["classes2"]), "precision1": tf.metrics.precision(labels=labels[:,0], predictions=predictions["classes1"]), "precision2": tf.metrics.precision(labels=labels[:,1], predictions=predictions["classes2"]), "recall1": tf.metrics.recall(labels=labels[:,0], predictions=predictions["classes1"]), "recall2": tf.metrics.recall(labels=labels[:,1], predictions=predictions["classes2"]) } return tf.estimator.EstimatorSpec( mode=mode, loss=loss, eval_metric_ops=eval_metric_ops)def _parse_function(filename, label): image_string = tf.read_file(filename) image_decoded = tf.image.decode_image(image_string) image_typecasted = tf.cast(image_decoded, tf.float32) image_reshaped = tf.reshape(image_typecasted, [-1, 224, 224, 3]) return image_reshaped, labeldef _parse_function(filename): image_string = tf.read_file(filename) image_decoded = tf.image.decode_image(image_string) image_typecasted = tf.cast(image_decoded, tf.float32) image_reshaped = tf.reshape(image_typecasted, [-1, 224, 224, 3]) return image_reshapeddef stratified_train_test_split_(): filenamelist = [] labelslist = [] DIRECTORY = 'path_to' for filename in os.listdir(DIRECTORY): fullfilename = DIRECTORY + filename if filename.endswith('.back.0.jpg'): #back image, original orientation filenamelist.append(fullfilename) temp = [0,0] labelslist.append(temp) elif filename.endswith('.back.90.jpg'): #back image, rotated clockwise 90 filenamelist.append(fullfilename) temp = [0,1] labelslist.append(temp) elif filename.endswith('.back.180.jpg'): #back image, rotated clockwise 180 filenamelist.append(fullfilename) temp = [0,2] labelslist.append(temp) elif filename.endswith('.back.270.jpg'): #back image, rotated clockwise 270 filenamelist.append(fullfilename) temp = [0,3] labelslist.append(temp) elif filename.endswith('.front.0.jpg'): #front image, rotated clockwise 0 filenamelist.append(fullfilename) temp = [1,0] labelslist.append(temp) elif filename.endswith('.front.90.jpg'): #front image, rotated clockwise 90 filenamelist.append(fullfilename) temp = [1,1] labelslist.append(temp) elif filename.endswith('.front.180.jpg'): #front image, rotated clockwise 180 filenamelist.append(fullfilename) temp = [1,2] labelslist.append(temp) elif filename.endswith('.front.270.jpg'): #front image, rotated clockwise 270 filenamelist.append(fullfilename) temp = [1,3] labelslist.append(temp) X_train, X_test, y_train, y_test = train_test_split(filenamelist, labelslist, test_size=0.20, random_state=42, shuffle=True, stratify=labelslist) return X_train, X_test, y_train, y_testdef my_input_fn_train(X_train, y_train): filenames = tf.constant(X_train) labels = tf.constant(y_train) dataset = tf.data.Dataset.from_tensor_slices((filenames, labels)) dataset = dataset.map(_parse_function) # # Shuffle, repeat, and batch the examples. dataset = dataset.shuffle(5000).repeat().batch(64) # # Build the Iterator, and return the read end of the pipeline. return dataset.make_one_shot_iterator().get_next()def my_input_fn_test(X_test, y_test): filenames = tf.constant(X_test) labels = tf.constant(y_test) dataset = tf.data.Dataset.from_tensor_slices((filenames, labels)) dataset = dataset.map(_parse_function) # # Shuffle, repeat, and batch the examples. dataset = dataset.shuffle(5000).repeat(1).batch(64) # # Build the Iterator, and return the read end of the pipeline. return dataset.make_one_shot_iterator().get_next()def my_input_fn_predict(filename): filenames = tf.constant(filename) dataset = tf.data.Dataset.from_tensors((filenames)) dataset = dataset.map(_parse_function) return dataset.make_one_shot_iterator().get_next()def main(unused_argv): # Create the Estimator mnist_classifier = tf.estimator.Estimator( model_fn=cnn_model_fn, model_dir="path_to_model_directory", config = tf.estimator.RunConfig( save_checkpoints_steps=None, save_checkpoints_secs=600, save_summary_steps=5)) # Set up logging for predictions tensors_to_log_1 = {"probabilities1": "softmax_tensor_1"} tensors_to_log_2 = {"probabilities2": "softmax_tensor_2"} logging_hook_1 = tf.train.LoggingTensorHook( tensors=tensors_to_log_1, every_n_iter=100) logging_hook_2 = tf.train.LoggingTensorHook( tensors=tensors_to_log_2, every_n_iter=100) #Splitting the train test split seperately X_train, X_test, y_train, y_test = stratified_train_test_split_() #Removed the training, testing and prediction calls. #Code for exporting the models using def serving_input_receiver_fn(): #???? mnist_classifier.export_savedmodel(export_dir_base, serving_input_fn)if __name__ == "__main__": tf.app.run()回答：
这是上面问题的答案，我已经解决了这个问题，所以我发布这个答案以供有需要的人参考。
from __future__ import absolute_importfrom __future__ import divisionfrom __future__ import print_functionimport tensorflow as tfimport numpy as npfrom sklearn.model_selection import train_test_splitimport osimport matplotlib.pyplot as pltimport matplotlib.image as mpimgimport scipyfrom scipy import ndimageimport scipy.misctf.logging.set_verbosity(tf.logging.INFO)# Our application logic will be added heredef cnn_model_fn(features, labels, mode): """Model function for CNN.""" #tf.print("shape of features",features.shape) #print("type of feature ", type(features)) #Added this for solving the issue features = features["feature"] features = tf.Print(features, [tf.shape(features), "shape of features before second reshape"], summarize=10) #tf.print("shape of labels ", labels.shape) # Input Layer input_layer = tf.reshape(features, [-1, 224, 224, 3]) #print(type(input_layer)) input_layer = tf.Print(input_layer, [tf.shape(input_layer), "shape of input_layer after second reshape"], summarize=10) #print("Successfully performed reshape operation") # Convolutional Layer #1 conv1 = tf.layers.conv2d( inputs=input_layer, filters=64, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #2 conv2 = tf.layers.conv2d( inputs=conv1, filters=64, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) #print("output of convolutional layer 1 : ",type(conv1)) # Pooling Layer #1 pool1 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2) #print(" Successfully done with 1st layer and shape is : ", pool1.shape) # Convolutional Layer #3 conv3 = tf.layers.conv2d( inputs=pool1, filters=128, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #4 conv4 = tf.layers.conv2d( inputs=conv3, filters=128, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) pool2 = tf.layers.max_pooling2d(inputs=conv4, pool_size=[2, 2], strides=2) #print(" Successfully done with 2nd layer and shape is : ", pool2.shape) # Convolutional Layer #5 conv5 = tf.layers.conv2d( inputs=pool2, filters=256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #6 conv6 = tf.layers.conv2d( inputs=conv5, filters=256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #7 conv7 = tf.layers.conv2d( inputs=conv6, filters=256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) pool3 = tf.layers.max_pooling2d(inputs=conv7, pool_size=[2, 2], strides=2) #print(" Successfully done with 3rd layer and shape is : ", pool3.shape) # Convolutional Layer #8 conv8 = tf.layers.conv2d( inputs=pool3, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #9 conv9 = tf.layers.conv2d( inputs=conv8, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #10 conv10 = tf.layers.conv2d( inputs=conv9, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) pool4 = tf.layers.max_pooling2d(inputs=conv10, pool_size=[2, 2], strides=2) #print(" Successfully done with 4th layer and shape is : ", pool4.shape) # Convolutional Layer #11 conv11 = tf.layers.conv2d( inputs=pool4, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #12 conv12 = tf.layers.conv2d( inputs=conv11, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) # Convolutional Layer #12 conv13 = tf.layers.conv2d( inputs=conv12, filters=512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu) pool5 = tf.layers.max_pooling2d(inputs=conv13, pool_size=[2, 2], strides=2) #print(" Successfully done with 5th layer and shape is : ", pool5.shape) # Dense Layer pool5_flat = tf.reshape(pool5, [-1, 7 * 7 * 512]) dense1 = tf.layers.dense(inputs=pool5_flat, units=4096, activation=tf.nn.relu) dense2 = tf.layers.dense(inputs=dense1, units=4096, activation=tf.nn.relu) dense3 = tf.layers.dense(inputs=dense2, units=1024, activation=tf.nn.relu) #print(dense3.shape) dropout = tf.layers.dropout( inputs=dense3, rate=0.001, training=mode == tf.estimator.ModeKeys.TRAIN) #print(" Completed all the layers ") #This is the place where we will have two flows of different layers which are used for # Logits Layer for Front Back Prediction #logits = tf.layers.dense(inputs=dropout, units=2) logits1 = tf.layers.dense(inputs=dropout, units=2) logits2 = tf.layers.dense(inputs=dropout, units=4) #Modified this for solving the issue predictions1 = { # Generate predictions (for PREDICT and EVAL mode) "classes1": tf.argmax(input=logits1, axis=1), # Add `softmax_tensor` to the graph. It is used for PREDICT and by the # `logging_hook`. "probabilities1": tf.nn.softmax(logits1, name="softmax_tensor_1"), } #Modified this for solving the issue predictions2 = { # Generate predictions (for PREDICT and EVAL mode) "classes2": tf.argmax(input=logits2, axis=1), # Add `softmax_tensor` to the graph. It is used for PREDICT and by the # `logging_hook`. "probabilities2": tf.nn.softmax(logits2, name="softmax_tensor_2") } #Modified this for solving the issue if mode == tf.estimator.ModeKeys.PREDICT: return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions1, export_outputs={tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: tf.estimator.export.ClassificationOutput(classes=tf.as_string(predictions1['classes1'])), 'class2': tf.estimator.export.ClassificationOutput(classes=tf.as_string(predictions2['classes2']))}) # Calculate Loss (for both TRAIN and EVAL modes) loss1 = tf.losses.sparse_softmax_cross_entropy(labels=labels[:,0], logits=logits1) loss2 = tf.losses.sparse_softmax_cross_entropy(labels=labels[:,1], logits=logits2) loss1 = tf.Print(loss1, ["loss1 : ", loss1]) loss2 = tf.Print(loss2, ["loss2 : ", loss2]) loss = loss1 + loss2 loss = tf.Print(loss, ["loss : ", loss]) tf.summary.scalar('Loss1', loss1) tf.summary.scalar('Loss2', loss2) tf.summary.scalar('Loss', loss) #print("Loss of this step is : ",loss) # Configure the Training Op (for TRAIN mode) if mode == tf.estimator.ModeKeys.TRAIN: optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01) train_op = optimizer.minimize( loss=loss, global_step=tf.train.get_global_step()) return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op) #print("completed training") # Add evaluation metrics (for EVAL mode) eval_metric_ops = { "accuracy1": tf.metrics.accuracy( labels=labels[:,0], predictions=predictions1["classes1"]), "accuracy2": tf.metrics.accuracy( labels=labels[:,1], predictions=predictions2["classes2"]), "precision1": tf.metrics.precision(labels=labels[:,0], predictions=predictions1["classes1"]), "precision2": tf.metrics.precision(labels=labels[:,1], predictions=predictions2["classes2"]), "recall1": tf.metrics.recall(labels=labels[:,0], predictions=predictions1["classes1"]), "recall2": tf.metrics.recall(labels=labels[:,1], predictions=predictions2["classes2"]) } tf.summary.scalar('Accuracy_for_Front/Back', eval_metric_ops.get('accuracy1')) tf.summary.scalar('Accuracy_for_Orientation', eval_metric_ops.get('accuracy2')) tf.summary.scalar('Precision_for_Front/Back', eval_metric_ops.get('precision1')) tf.summary.scalar('Precision_for_Orientation', eval_metric_ops.get('precision2')) tf.summary.scalar('Recall_for_Front/Back', eval_metric_ops.get('recall1')) tf.summary.scalar('Recall_for_Orientation', eval_metric_ops.get('recall2')) #print("Accuracy of this step is : ", eval_metric_ops.get('accuracy1')) return tf.estimator.EstimatorSpec( mode=mode, loss=loss, eval_metric_ops=eval_metric_ops )def _parse_function(filename, label): filename = tf.Print(filename, [filename, " Names of files "]) image_string = tf.read_file(filename) image_decoded = tf.image.decode_image(image_string) image_typecasted = tf.cast(image_decoded, tf.float32) #print("type of the image : ",type(image_typecasted)) #image_resized = tf.image.resize_images(image_decoded, [28, 28]) image_typecasted = tf.Print(image_typecasted, [tf.shape(image_typecasted), " shape of image_typecasted before first reshape"],summarize=10) image_reshaped = tf.reshape(image_typecasted, [-1, 224, 224, 3]) image_reshaped = tf.Print(image_reshaped, [tf.shape(image_reshaped), " shape of image_reshaped after first reshape"], summarize=10) return image_reshaped, labeldef _parse_function(filename): image_string = tf.read_file(filename) image_decoded = tf.image.decode_image(image_string) image_typecasted = tf.cast(image_decoded, tf.float32) #print("type of the image : ",type(image_typecasted)) #image_resized = tf.image.resize_images(image_decoded, [28, 28]) image_reshaped = tf.reshape(image_typecasted, [-1, 224, 224, 3]) return image_reshapeddef stratified_train_test_split_(): filenamelist = [] labelslist = [] DIRECTORY = 'path_to_directory' for filename in os.listdir(DIRECTORY): fullfilename = DIRECTORY + filename if filename.endswith('.back.0.jpg'): #back image, original orientation filenamelist.append(fullfilename) temp = [0,0] labelslist.append(temp) elif filename.endswith('.back.90.jpg'): #back image, rotated clockwise 90 filenamelist.append(fullfilename) temp = [0,1] labelslist.append(temp) elif filename.endswith('.back.180.jpg'): #back image, rotated clockwise 180 filenamelist.append(fullfilename) temp = [0,2] labelslist.append(temp) elif filename.endswith('.back.270.jpg'): #back image, rotated clockwise 270 filenamelist.append(fullfilename) temp = [0,3] labelslist.append(temp) elif filename.endswith('.front.0.jpg'): #front image, rotated clockwise 0 filenamelist.append(fullfilename) temp = [1,0] labelslist.append(temp) elif filename.endswith('.front.90.jpg'): #front image, rotated clockwise 90 filenamelist.append(fullfilename) temp = [1,1] labelslist.append(temp) elif filename.endswith('.front.180.jpg'): #front image, rotated clockwise 180 filenamelist.append(fullfilename) temp = [1,2] labelslist.append(temp) elif filename.endswith('.front.270.jpg'): #front image, rotated clockwise 270 filenamelist.append(fullfilename) temp = [1,3] labelslist.append(temp) #splitting the train test data #currently we are doing stratified sampling but still we are ignoring the X_train, X_test, y_train, y_test = train_test_split(filenamelist, labelslist, test_size=0.20, random_state=42, shuffle=True, stratify=labelslist) #print("data is split into test and train set") # # Build the Iterator, and return the read end of the pipeline. return X_train, X_test, y_train, y_testdef my_input_fn_train(X_train, y_train): filenames = tf.constant(X_train) labels = tf.constant(y_train) dataset = tf.data.Dataset.from_tensor_slices((filenames, labels)) dataset = dataset.map(_parse_function) # # Shuffle, repeat, and batch the examples. dataset = dataset.shuffle(5000).repeat().batch(64) # # Build the Iterator, and return the read end of the pipeline. return dataset.make_one_shot_iterator().get_next()def my_input_fn_test(X_test, y_test): filenames = tf.constant(X_test) labels = tf.constant(y_test) dataset = tf.data.Dataset.from_tensor_slices((filenames, labels)) dataset = dataset.map(_parse_function) # # Shuffle, repeat, and batch the examples. dataset = dataset.shuffle(5000).repeat(1).batch(64) # # Build the Iterator, and return the read end of the pipeline. return dataset.make_one_shot_iterator().get_next()def my_input_fn_predict(filename): filenames = tf.constant(filename) dataset = tf.data.Dataset.from_tensors((filenames)) dataset = dataset.map(_parse_function) return dataset.make_one_shot_iterator().get_next()def main(unused_argv): # Create the Estimator mnist_classifier = tf.estimator.Estimator( model_fn=cnn_model_fn, model_dir="path_to_model_directory", config = tf.estimator.RunConfig( save_checkpoints_steps=None, save_checkpoints_secs=600, save_summary_steps=5)) export_dir_base = 'path_to_folder' # Set up logging for predictions tensors_to_log_1 = {"probabilities1": "softmax_tensor_1"} tensors_to_log_2 = {"probabilities2": "softmax_tensor_2"} logging_hook_1 = tf.train.LoggingTensorHook( tensors=tensors_to_log_1, every_n_iter=100) logging_hook_2 = tf.train.LoggingTensorHook( tensors=tensors_to_log_2, every_n_iter=100) #print("Successfully created Estimator") #Added this for solving the issue def serving_input_receiver_fn(): feature_spec = {'image/encoded': tf.FixedLenFeature(shape=[], dtype=tf.string)} serialized_tf_example = tf.placeholder(dtype=tf.string, name='input_example_tensor') receiver_tensors = {'examples': serialized_tf_example} features = tf.parse_example(serialized_tf_example, feature_spec) jpegs = features['image/encoded'] images = tf.map_fn(_parse_function, jpegs, dtype=tf.float32) return tf.estimator.export.ServingInputReceiver(images, receiver_tensors) mnist_classifier.export_savedmodel(export_dir_base, serving_input_receiver_fn)if __name__ == "__main__": tf.app.run()