烁凡个人页面-阿里云开发者社区

TensorFlow 2.0+Keras 防坑指南

TensorFlow 2.0是对1.x版本做了一次大的瘦身，Eager Execution默认开启，并且使用Keras作为默认高级API，这些改进大大降低的TensorFlow使用难度。本文主要记录了一次曲折的使用Keras+TensorFlow2.0的BatchNormalization的踩坑经历，这个坑差点要把TF2.0的新特性都毁灭殆尽，如果你在学习TF2.0的官方教程，不妨一观。问题的产生从教程[1]https://www.tensorflow.org/alpha/tutorials/images/transfer_learning?hl=zh-cn（讲述如何Transfer Learning）说起： IMG_SHAPE = (IMG_SIZE, IMG_SIZE, 3) # Create the base model from the pre-trained model MobileNet V2 base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE, include_top=False,weights='imagenet') model = tf.keras.Sequential([ base_model, tf.keras.layers.GlobalAveragePooling2D(), tf.keras.layers.Dense(NUM_CLASSES) ]) 简单的代码我们就复用了MobileNetV2的结构创建了一个分类器模型，接着我们就可以调用Keras的接口去训练模型： model.compile(optimizer=tf.keras.optimizers.RMSprop(lr=base_learning_rate), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy']) model.summary() history = model.fit(train_batches.repeat(), epochs=20, steps_per_epoch = steps_per_epoch, validation_data=validation_batches.repeat(), validation_steps=validation_steps) 输出的结果看，一起都很完美： Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= mobilenetv2_1.00_160 (Model) (None, 5, 5, 1280) 2257984 _________________________________________________________________ global_average_pooling2d (Gl (None, 1280) 0 _________________________________________________________________ dense (Dense) (None, 2) 1281 ================================================================= Total params: 2,259,265 Trainable params: 1,281 Non-trainable params: 2,257,984 _________________________________________________________________ Epoch 11/20 581/581 [==============================] - 134s 231ms/step - loss: 0.4208 - accuracy: 0.9484 - val_loss: 0.1907 - val_accuracy: 0.9812 Epoch 12/20 581/581 [==============================] - 114s 197ms/step - loss: 0.3359 - accuracy: 0.9570 - val_loss: 0.1835 - val_accuracy: 0.9844 Epoch 13/20 581/581 [==============================] - 116s 200ms/step - loss: 0.2930 - accuracy: 0.9650 - val_loss: 0.1505 - val_accuracy: 0.9844 Epoch 14/20 581/581 [==============================] - 114s 196ms/step - loss: 0.2561 - accuracy: 0.9701 - val_loss: 0.1575 - val_accuracy: 0.9859 Epoch 15/20 581/581 [==============================] - 119s 206ms/step - loss: 0.2302 - accuracy: 0.9715 - val_loss: 0.1600 - val_accuracy: 0.9812 Epoch 16/20 581/581 [==============================] - 115s 197ms/step - loss: 0.2134 - accuracy: 0.9747 - val_loss: 0.1407 - val_accuracy: 0.9828 Epoch 17/20 581/581 [==============================] - 115s 197ms/step - loss: 0.1546 - accuracy: 0.9813 - val_loss: 0.0944 - val_accuracy: 0.9828 Epoch 18/20 581/581 [==============================] - 116s 200ms/step - loss: 0.1636 - accuracy: 0.9794 - val_loss: 0.0947 - val_accuracy: 0.9844 Epoch 19/20 581/581 [==============================] - 115s 198ms/step - loss: 0.1356 - accuracy: 0.9823 - val_loss: 0.1169 - val_accuracy: 0.9828 Epoch 20/20 581/581 [==============================] - 116s 199ms/step - loss: 0.1243 - accuracy: 0.9849 - val_loss: 0.1121 - val_accuracy: 0.9875 然而这种写法还是不方便Debug，我们希望可以精细的控制迭代的过程，并能够看到中间结果，所以我们训练的过程改成了这样： optimizer = tf.keras.optimizers.RMSprop(lr=base_learning_rate) train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy') @tf.function def train_cls_step(image, label): with tf.GradientTape() as tape: predictions = model(image) loss = tf.keras.losses.SparseCategoricalCrossentropy()(label, predictions) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) train_accuracy(label, predictions) for images, labels in train_batches: train_cls_step(images,labels) 重新训练后，结果依然很完美！但是，这时候我们想对比一下Finetune和重头开始训练的差别，所以把构建模型的代码改成了这样： base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE, include_top=False,weights=None) 使得模型的权重随机生成，这时候训练结果就开始抽风了，Loss不下降，Accuracy稳定在50%附近游荡： Step #10: loss=0.6937199831008911 acc=46.5625% Step #20: loss=0.6932525634765625 acc=47.8125% Step #30: loss=0.699873685836792 acc=49.16666793823242% Step #40: loss=0.6910845041275024 acc=49.6875% Step #50: loss=0.6935917139053345 acc=50.0625% Step #60: loss=0.6965731382369995 acc=49.6875% Step #70: loss=0.6949992179870605 acc=49.19642639160156% Step #80: loss=0.6942993402481079 acc=49.84375% Step #90: loss=0.6933775544166565 acc=49.65277862548828% Step #100: loss=0.6928421258926392 acc=49.5% Step #110: loss=0.6883170008659363 acc=49.54545593261719% Step #120: loss=0.695658802986145 acc=49.453125% Step #130: loss=0.6875559091567993 acc=49.61538314819336% Step #140: loss=0.6851695775985718 acc=49.86606979370117% Step #150: loss=0.6978713274002075 acc=49.875% Step #160: loss=0.7165156602859497 acc=50.0% Step #170: loss=0.6945627331733704 acc=49.797794342041016% Step #180: loss=0.6936900615692139 acc=49.9305534362793% Step #190: loss=0.6938323974609375 acc=49.83552551269531% Step #200: loss=0.7030564546585083 acc=49.828125% Step #210: loss=0.6926192045211792 acc=49.76190185546875% Step #220: loss=0.6932414770126343 acc=49.786930084228516% Step #230: loss=0.6924526691436768 acc=49.82337188720703% Step #240: loss=0.6882281303405762 acc=49.869789123535156% Step #250: loss=0.6877702474594116 acc=49.86249923706055% Step #260: loss=0.6933954954147339 acc=49.77163314819336% Step #270: loss=0.6944763660430908 acc=49.75694274902344% Step #280: loss=0.6945018768310547 acc=49.49776840209961% 我们将predictions的结果打印出来，发现batch内每个输出都是一模一样的： 0 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 1 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 2 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 3 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 4 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 5 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 6 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 7 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 8 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 9 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32) 只是修改了初始权重，为何会产生这样的结果？问题排查实验1 是不是训练不够充分，或者learning rate设置的不合适？经过几轮调整，发现无论训练多久，learning rate变大变小，都无法改变这种结果实验2 既然是权重的问题，是不是权重随机初始化的有问题，把初始权重拿出来统计了一下，一切正常实验3 这种问题根据之前的经验，在导出Inference模型的时候BatchNormalization没有处理好会出现这种一个batch内所有结果都一样的问题。但是如何解释训练的时候为什么会出现这个问题？而且为什么Finetue不会出现问题呢？只是改了权重的初始值而已呀按照这个方向去Google的一番，发现了Keras的BatchNormalization确实有很多issue，其中一个问题是在保存模型的是BatchNormalzation的moving mean和moving variance不会被保存[6]https://github.com/tensorflow/tensorflow/issues/16455，而另外一个issue提到问题就和我们问题有关系的了：[2] https://github.com/tensorflow/tensorflow/issues/19643[3] https://github.com/tensorflow/tensorflow/issues/23873最后，这位作者找到了原因，并且总结在了这里：[4] https://pgaleone.eu/tensorflow/keras/2019/01/19/keras-not-yet-interface-to-tensorflow/ 根据这个提示，我们做了如下尝试：实验3.1 改用model.fit的写法进行训练，在最初的几个epoch里面，我们发现好的一点的是training accuracy已经开始缓慢提升了，但是validation accuracy存在原来的问题。而且通过model.predict_on_batch()拿到中间结果，发现依然还是batch内输出都一样。 Epoch 1/20 581/581 [==============================] - 162s 279ms/step - loss: 0.6768 - sparse_categorical_accuracy: 0.6224 - val_loss: 0.6981 - val_sparse_categorical_accuracy: 0.4984 Epoch 2/20 581/581 [==============================] - 133s 228ms/step - loss: 0.4847 - sparse_categorical_accuracy: 0.7684 - val_loss: 0.6931 - val_sparse_categorical_accuracy: 0.5016 Epoch 3/20 581/581 [==============================] - 130s 223ms/step - loss: 0.3905 - sparse_categorical_accuracy: 0.8250 - val_loss: 0.6996 - val_sparse_categorical_accuracy: 0.4984 Epoch 4/20 581/581 [==============================] - 131s 225ms/step - loss: 0.3113 - sparse_categorical_accuracy: 0.8660 - val_loss: 0.6935 - val_sparse_categorical_accuracy: 0.5016 但是，随着训练的深入，结果出现了逆转，开始变得正常了（tf.function的写法是无论怎么训练都不会变化，幸好没有放弃治疗）（追加：其实这里还是有问题的，继续看后面，当时就觉得怪怪的，不应该收敛这么慢） Epoch 18/20 581/581 [==============================] - 131s 226ms/step - loss: 0.0731 - sparse_categorical_accuracy: 0.9725 - val_loss: 1.4896 - val_sparse_categorical_accuracy: 0.8703 Epoch 19/20 581/581 [==============================] - 130s 225ms/step - loss: 0.0664 - sparse_categorical_accuracy: 0.9748 - val_loss: 0.6890 - val_sparse_categorical_accuracy: 0.9016 Epoch 20/20 581/581 [==============================] - 126s 217ms/step - loss: 0.0631 - sparse_categorical_accuracy: 0.9768 - val_loss: 1.0290 - val_sparse_categorical_accuracy: 0.9031 通多model.predict_on_batch()拿到的结果也和这个Accuracy也是一致的实验3.2 通过上一个实验，我们验证了确实如果只通过Keras的API去训练，是正常。更深层的原因是什么呢？是不是BatchNomalization没有update moving mean和moving variance导致的呢？答案是Yes我们分别在两中训练方法前后，打印 moving mean和moving variance的值： def get_bn_vars(collection): moving_mean, moving_variance = None, None for var in collection: name = var.name.lower() if "variance" in name: moving_variance = var if "mean" in name: moving_mean = var if moving_mean is not None and moving_variance is not None: return moving_mean, moving_variance raise ValueError("Unable to find moving mean and variance") mean, variance = get_bn_vars(model.variables) print(mean) print(variance) 我们发现，确实如果使用model.fit()进行训练，mean和variance是在update的(虽然更新的速率看着有些奇怪)，但是对于tf.function那种写法这两个值就没有被update 那这里我们也可以解释为什么Finetune不会出现问题了，因为imagenet训练的mean, variance已经是一个比较好的值了，即使不更新也可以正常使用实验3.3 是不是改成[4]里面说的方法构建动态的Input_Shape的模型就OK了呢？ class MyModel(Model): def __init__(self): super(MyModel, self).__init__() self.conv1 = Conv2D(32, 3, activation='relu') self.batch_norm1=BatchNormalization() self.flatten = Flatten() self.d1 = Dense(128, activation='relu') self.d2 = Dense(10, activation='softmax') def call(self, x): x = self.conv1(x) x = self.batch_norm1(x) x = self.flatten(x) x = self.d1(x) return self.d2(x) model = MyModel() #model.build((None,28,28,1)) model.summary() @tf.functiondef train_step(image, label): with tf.GradientTape() as tape: predictions = model(image) loss = loss_object(label, predictions) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) train_loss(loss) train_accuracy(label, predictions) 模型如下： Model: "my_model" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d (Conv2D) multiple 320 _________________________________________________________________ batch_normalization_v2 (Batc multiple 128 _________________________________________________________________ flatten (Flatten) multiple 0 _________________________________________________________________ dense (Dense) multiple 2769024 _________________________________________________________________ dense_1 (Dense) multiple 1290 ================================================================= Total params: 2,770,762 Trainable params: 2,770,698 Non-trainable params: 64 从Output Shape看，构建模型没问题跑了一遍MINST，结果也很不错！以防万一，我们同样测试了一下mean和variance是否被更新，然而结果出乎意料，并没有！也就是说[4]里面说的方案在我们这里并不可行实验3.4 既然我们定位问题是在BatchNormalization这里，所以就想到BatchNormalization的training和testing时候行为是不一致的，在testing的时候moving mean和variance是不需要update的，那么会不会是tf.function的这种写法并不会自动更改这个状态呢？查看源码，发现BatchNormalization的call()存在一个training参数，而且默认是False Call arguments: inputs: Input tensor (of any rank). training: Python boolean indicating whether the layer should behave in training mode or in inference mode. - `training=True`: The layer will normalize its inputs using the mean and variance of the current batch of inputs. - `training=False`: The layer will normalize its inputs using the mean and variance of its moving statistics, learned during training. 所以，做了如下改进： class MyModel(Model): def __init__(self): super(MyModel, self).__init__() self.conv1 = Conv2D(32, 3, activation='relu') self.batch_norm1=BatchNormalization() self.flatten = Flatten() self.d1 = Dense(128, activation='relu') self.d2 = Dense(10, activation='softmax') def call(self, x,training=True): x = self.conv1(x) x = self.batch_norm1(x,training=training) x = self.flatten(x) x = self.d1(x) return self.d2(x) model = MyModel() #model.build((None,28,28,1)) model.summary() @tf.functiondef train_step(image, label): with tf.GradientTape() as tape: predictions = model(image,training=True) loss = loss_object(label, predictions) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) train_loss(loss) train_accuracy(label, predictions) @tf.functiondef test_step(image, label): predictions = model(image,training=False) t_loss = loss_object(label, predictions) test_loss(t_loss) test_accuracy(label, predictions) 结果显示，moving mean和variance开始更新啦，测试Accuracy也是符合预期所以，我们可以确定问题的根源在于需要指定BatchNormalization是在training还是在testing！实验3.5 3.4中方法虽然解决了我们的问题，但是它是使用构建Model的subclass的方式，而我们之前的MobileNetV2是基于更加灵活Keras Functional API构建的，由于无法控制call()函数的定义，没有办法灵活切换training和testing的状态，另外用Sequential的方式构建时也是一样。[5]https://blog.keras.io/keras-as-a-simplified-interface-to-tensorflow-tutorial.html[7]https://github.com/keras-team/keras/issues/7085[8]https://github.com/keras-team/keras/issues/6752从5[8]中，我了解到两个情况， tf.keras.backend.set_learning_phase()可以改变training和testing的状态； model.updates和layer.updates 存着old_value和new_value的Assign Op 所以我首先尝试： tf.keras.backend.set_learning_phase(True) 结果，MobileNetV2构建的模型也可以正常工作了。而且收敛的速度似乎比model.fit()还快了很多，结合之前model.fit()收敛慢的困惑，这里又增加的一个实验，在model.fit()的版本里面也加上这句话，发现同样收敛速度也变快了！1个epoch就能得到不错的结果了！因此，这里又产生了一个问题model.fit()到底有没有设learning_phase状态？如果没有是怎么做moving mean和variance的update的？第二个方法，由于教程中讲述的是如何在1.x的版本构建，而在eager execution模式下，似乎没有办法去run这些Assign Operation。仅做参考吧 update_ops = [] for assign_op in model.updates: update_ops.append(assign_op)) #但是不知道拿到这些update_ops在eager execution模式下怎么处理呢？结论总结一下，我们从[4]找到了解决问题的启发点，但是最终证明[4]里面的问题和解决方法用到我们这里并不能真正解决问题，问题的关键还是在于Keras+TensorFlow2.0里面我们如何处理在training和testing状态下行为不一致的Layer；以及对于model.fit()和tf.funtion这两种训练方法的区别，最终来看model.fit()里面似乎包含很多诡异的行为。最终的使用建议如下：在使用model.fit()或者model.train_on_batch()这种Keras的API训练模型时，也推荐手动设置tf.keras.backend.set_learning_phase(True)，可以加快收敛如果使用eager execution这种方法， 1）使用构建Model的subclass，但是针对call()设置training的状态，对于BatchNoramlization，Dropout这样的Layer进行不同处理 2）使用Functional API或者Sequential的方式构建Model，设置tf.keras.backend.set_learning_phase(True)，但是注意在testing的时候改变一下状态最后，为什么TF 2.0的教程里面没有提及这些？默认你已经精通Keras了吗？[捂脸哭] 感谢感谢柏涛帆月应知老师提供的帮助 [1]https://www.tensorflow.org/alpha/tutorials/images/transfer_learning?hl=zh-cn[2] https://github.com/tensorflow/tensorflow/issues/19643[3] https://github.com/tensorflow/tensorflow/issues/23873[4] https://pgaleone.eu/tensorflow/keras/2019/01/19/keras-not-yet-interface-to-tensorflow/[5]https://blog.keras.io/keras-as-a-simplified-interface-to-tensorflow-tutorial.html[6]https://github.com/tensorflow/tensorflow/issues/16455[7]https://github.com/keras-team/keras/issues/7085[8]https://github.com/keras-team/keras/issues/6752

如何部署自己的SSD检测模型到Android TFLite上

TensorFlow Object Detection API 上提供了使用SSD部署到TFLite运行上去的方法, 可是这套API封装太死板, 如果你要自己实现了一套SSD的训练算法,应该怎么才能部署到TFLite上呢? 首先,抛开后处理的部分,你的SSD模型(无论是VGG-SSD和Mobilenet-SSD), 你最终的模型的输出是对class_predictions和bbox_predictions; 并且是encoded的 Encoding的方式: class_predictions: M个Feature Layer, Feature Layer的大小(宽高)视网络结构而定; 每个Feature Layer有Num_Anchor_Depth_of_this_layer x Num_classes个channels box_predictions: M个Feature Layer; 每个Feature Layer有Num_Anchor_Depth_of_this_layer x 4个channes 这4个channel分别代表dy,dx,h,w, 即bbox中心距离anchor中心坐标的偏移量和宽高注:通常,为了平衡loss之间的大小, 不会直接编码dy,dx,w,h的原始值,而是dy/anchor_h*scale0, dx/anchor_w*scale0, log(h/anchor_h)*scale1, log(w/anchor_w)*scale1, 也就是偏移量的绝对值除anchor宽高得到相对值,然后再乘上一个scale, 经验值 scale0取5,scale1取10; 对于h,w是对得到相对值后先取log再乘以scale, h/anchor_h的范围在1附近, 取log后可以转换到0附近;所以在解码的时候需要做对应相反的变换; 在后面TFLite_Detection_PostProcess的Op实现里就有这么一段: 然后我们需要的是做的是decode出来对每个class的confidence和location的预测值后处理在Object Detection API的 export_tflite_ssd_graph_lib.py文件中,你可以看到,它区别与直接freeze pb的操作就在于最后替换了后处理的部分; Plain Text Plain Text Bash Basic C C++ C# CSS C++ Diff Git go GraphQL HTML HTTP Java JavaScript JSON JSX Kotlin Less Makefile Markdown MATLAB Nginx Objective-C Pascal Perl PHP PowerShell Ruby Protobuf Python R Ruby Scala Shell SQL Swift TypeScript VB.net XML YAML KaTeX Mermaid PlantUML Flow Graphviz frozen_graph_def = exporter.freeze_graph_with_def_protos( input_graph_def=tf.get_default_graph().as_graph_def(), input_saver_def=input_saver_def, input_checkpoint=checkpoint_to_use, output_node_names=','.join([ 'raw_outputs/box_encodings', 'raw_outputs/class_predictions', 'anchors' ]), restore_op_name='save/restore_all', filename_tensor_name='save/Const:0', clear_devices=True, output_graph='', initializer_nodes='') # Add new operation to do post processing in a custom op (TF Lite only) if add_postprocessing_op: transformed_graph_def = append_postprocessing_op( frozen_graph_def, max_detections, max_classes_per_detection, nms_score_threshold, nms_iou_threshold, num_classes, scale_values) else: # Return frozen without adding post-processing custom op transformed_graph_def = frozen_graph_def 后处理的部分,其实看代码也很简单,就是增加了一个叫TFLite_Detection_PostProcess的node,用于解码和非极大抑制. 这个node的输入就是上面提到的box_predictions和class_predictions, 还有anchors的编码; 用这个node的目的只TFLite并不支持tf.contrib.image.non_max_surpression操作 Reshape过程: 这里需要明确,TFLite_Detection_PostProcess 这个op对raw_outputs/box_encodings, raw_outputs/class_predictions, anchors的Shape是有一个定制要求的 raw_outputs/box_encodings.shape=[1, num_anchors,4] raw_outputs/class_predictions.shape=[1, num_anchors,num_classes+1] anchors.shape=[1,num_anchors,4] 这里需要注意:1, 这三个都必须是3维的Tensor; 2.raw_outputs/class_predictions.shape的最后一个维度是包含background的classes, 也就是是num_classes+1; TFLite_Detection_PostProcess还有一个参数num_classes, 这个参数值是不包含background的, 所以也就导致TFLite_Detection_PostProcess的输出的class index是从0计数的; Plain Text Plain Text Bash Basic C C++ C# CSS C++ Diff Git go GraphQL HTML HTTP Java JavaScript JSON JSX Kotlin Less Makefile Markdown MATLAB Nginx Objective-C Pascal Perl PHP PowerShell Ruby Protobuf Python R Ruby Scala Shell SQL Swift TypeScript VB.net XML YAML KaTeX Mermaid PlantUML Flow Graphviz with tf.variable_scope('raw_outputs'): cls_pred = [tf.reshape(pred, [-1, num_classes]) for pred in cls_pred] location_pred = [tf.reshape(pred, [-1, 4]) for pred in location_pred] cls_pred = tf.concat(cls_pred, axis=0) location_pred = tf.expand_dims(tf.concat(location_pred, axis=0),0, name='box_encodings') cls_pred=tf.nn.softmax(cls_pred) tf.identity(tf.expand_dims(cls_pred,0), name='class_predictions') 这段代码就是用来reshape成要求的输入的, 需要注意的是对class_prediction需要做依次softmax或者sigmoid, 具体选择哪种取决于你是否允许一个anchor对应多个类; 对于anchors, 这其实是一constant的值: Plain Text Plain Text Bash Basic C C++ C# CSS C++ Diff Git go GraphQL HTML HTTP Java JavaScript JSON JSX Kotlin Less Makefile Markdown MATLAB Nginx Objective-C Pascal Perl PHP PowerShell Ruby Protobuf Python R Ruby Scala Shell SQL Swift TypeScript VB.net XML YAML KaTeX Mermaid PlantUML Flow Graphviz num_anchors = anchor_cy.get_shape().as_list() with tf.Session() as sess: y_out, x_out, h_out, w_out = sess.run([anchor_cy, anchor_cx, anchor_h, anchor_w]) encoded_anchors = tf.constant( np.transpose(np.stack((y_out, x_out, h_out, w_out))), dtype=tf.float32, shape=[num_anchors[0], 4]) 注意: 之前我使用tf.stack合成这个值的时候发现,TFLite只支持axis=0的时候的tf.stack, 否则就会转换是吧导出pb 添加完后处理,既可以导出一个带有后处理功能的pb文件了; 如果你不添加后处理,把它放在CPU上后续去做,其实也可以省去不少麻烦; Plain Text Plain Text Bash Basic C C++ C# CSS C++ Diff Git go GraphQL HTML HTTP Java JavaScript JSON JSX Kotlin Less Makefile Markdown MATLAB Nginx Objective-C Pascal Perl PHP PowerShell Ruby Protobuf Python R Ruby Scala Shell SQL Swift TypeScript VB.net XML YAML KaTeX Mermaid PlantUML Flow Graphviz binary_graph = os.path.join(output_dir, 'tflite_graph.pb') with tf.gfile.GFile(binary_graph, 'wb') as f: f.write(transformed_graph_def.SerializeToString()) txt_graph = os.path.join(output_dir, 'tflite_graph.pbtxt') with tf.gfile.GFile(txt_graph, 'w') as f: f.write(str(transformed_graph_def)) 注意: 导出的pb如果包含后处理, 是没办法用正常的TF执行的,必须转成tflite执行导出tflite Plain Text Plain Text Bash Basic C C++ C# CSS C++ Diff Git go GraphQL HTML HTTP Java JavaScript JSON JSX Kotlin Less Makefile Markdown MATLAB Nginx Objective-C Pascal Perl PHP PowerShell Ruby Protobuf Python R Ruby Scala Shell SQL Swift TypeScript VB.net XML YAML KaTeX Mermaid PlantUML Flow Graphviz bazel run --config=opt tensorflow/contrib/lite/toco:toco -- \ --input_file=$OUTPUT_DIR/tflite_graph.pb \ --output_file=$OUTPUT_DIR/detect.tflite \ --input_shapes=1,300,300,3 \ --input_arrays=normalized_input_image_tensor \ --output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3' \ --inference_type=QUANTIZED_UINT8 \ --mean_values=128 \ --std_values=128 \ --change_concat_input_ranges=false \ --allow_custom_ops or bazel run -c opt tensorflow/lite/toco:toco -- \ --input_file=$OUTPUT_DIR/tflite_graph.pb \ --output_file=$OUTPUT_DIR/detect.tflite \ --input_shapes=1,300,300,3 \ --input_arrays=normalized_input_image_tensor \ --output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3' \ --inference_type=FLOAT \ --allow_custom_ops 导出的过程中,可能遇到Converting unsupported operation: TFLite_Detection_PostProcess 这个提示, 正常如果是TF在1.10以上就忽略这个提示好了然后你可以先用python的程序加载这个tflite去测试一下注意: 这时候会发现一个问题, TFLite_Detection_PostProcess的NMS操作是忽略类标签的,如果你设置max_classes_per_detection=1; 但是如果你设置成>1的值, 会发现它吧background的标签也算进来了, 导致出来很多误检测的bbox; 部署Android 然后,你可以尝试部署到Android上, 在不使用NNAPI的时候正常,但是如果是NNAPI就需要自己实现相关操作了,否则会crash掉

烁凡个人页面-阿里云开发者社区

个人介绍

擅长的技术

深度学习经典算法PPO的通俗理解

简单玩玩TensorFlow的Post Training Quantization

树莓派3B+上编译TFLite 1.14踩坑记

TensorFlow 2.0+Keras 防坑指南

如何部署自己的SSD检测模型到Android TFLite上

简单玩玩TensorFlow的Post Training Quantization

树莓派3B+上编译TFLite 1.14踩坑记

深度学习经典算法PPO的通俗理解

简单玩玩TensorFlow的Post Training Quantization

树莓派3B+上编译TFLite 1.14踩坑记

TensorFlow 2.0+Keras 防坑指南

如何部署自己的SSD检测模型到Android TFLite上

TensorFlow量化训练

SLAM+AR 技术与应用的一些思考