如何在 Tensorflow 中更改保存的模型输入形状? [英] How to change a saved model input shape in Tensorflow?
问题描述
我想制作这个 repo https://github.com/ildoonet/tf-pose-estimation 与 Intel Movidius 一起运行,所以我尝试使用 mvNCCompile 转换 pb 模型.
I want to make this repo https://github.com/ildoonet/tf-pose-estimation run with Intel Movidius, so I tried convert the pb model using mvNCCompile.
问题是 mvNCCompile 需要固定的输入形状,但我拥有的模型是动态的.
The problem is mvNCCompile require a fixed input shape but the model I have is a dynamic one.
我试过了
graph_path = 'models/graph/mobilenet_thin/graph_opt.pb'
with tf.gfile.GFile(graph_path, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
graph = tf.get_default_graph()
tf.import_graph_def(graph_def, name='TfPoseEstimator')
x = graph.get_tensor_by_name('TfPoseEstimator/image:0')
x.set_shape([1, 368, 368, 3])
x = graph.get_tensor_by_name('TfPoseEstimator/MobilenetV1/Conv2d_0/Conv2D:0')
x.set_shape([1, 368, 368, 24])
得到这个
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_0/weights:0' shape=(3, 3, 3, 24) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/image:0' shape=(1, 368, 368, 3) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_0/Conv2D:0' shape=(1, 368, 368, 24) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_0/Conv2D_bn_offset:0' shape=(24,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_0/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 24) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_0/Relu:0' shape=(?, ?, ?, 24) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_1_depthwise/depthwise_weights:0' shape=(3, 3, 24, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_1_pointwise/weights:0' shape=(1, 1, 24, 48) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_1_depthwise/depthwise:0' shape=(?, ?, ?, 24) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_1_pointwise/Conv2D:0' shape=(?, ?, ?, 48) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_1_pointwise/Conv2D_bn_offset:0' shape=(48,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_1_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 48) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_1_pointwise/Relu:0' shape=(?, ?, ?, 48) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_2_depthwise/depthwise_weights:0' shape=(3, 3, 48, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_2_pointwise/weights:0' shape=(1, 1, 48, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_2_depthwise/depthwise:0' shape=(?, ?, ?, 48) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_2_pointwise/Conv2D:0' shape=(?, ?, ?, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_2_pointwise/Conv2D_bn_offset:0' shape=(96,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_2_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_2_pointwise/Relu:0' shape=(?, ?, ?, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_3_depthwise/depthwise_weights:0' shape=(3, 3, 96, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_3_pointwise/weights:0' shape=(1, 1, 96, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_3_depthwise/depthwise:0' shape=(?, ?, ?, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_3_pointwise/Conv2D:0' shape=(?, ?, ?, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_3_pointwise/Conv2D_bn_offset:0' shape=(96,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_3_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_3_pointwise/Relu:0' shape=(?, ?, ?, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_4_depthwise/depthwise_weights:0' shape=(3, 3, 96, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_4_pointwise/weights:0' shape=(1, 1, 96, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_4_depthwise/depthwise:0' shape=(?, ?, ?, 96) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_4_pointwise/Conv2D:0' shape=(?, ?, ?, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_4_pointwise/Conv2D_bn_offset:0' shape=(192,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_4_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_4_pointwise/Relu:0' shape=(?, ?, ?, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_5_depthwise/depthwise_weights:0' shape=(3, 3, 192, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_5_pointwise/weights:0' shape=(1, 1, 192, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_5_depthwise/depthwise:0' shape=(?, ?, ?, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_5_pointwise/Conv2D:0' shape=(?, ?, ?, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_5_pointwise/Conv2D_bn_offset:0' shape=(192,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_5_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_5_pointwise/Relu:0' shape=(?, ?, ?, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_6_depthwise/depthwise_weights:0' shape=(3, 3, 192, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_6_pointwise/weights:0' shape=(1, 1, 192, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_6_depthwise/depthwise:0' shape=(?, ?, ?, 192) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_6_pointwise/Conv2D:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_6_pointwise/Conv2D_bn_offset:0' shape=(384,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_6_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_6_pointwise/Relu:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_7_depthwise/depthwise_weights:0' shape=(3, 3, 384, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_7_pointwise/weights:0' shape=(1, 1, 384, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_7_depthwise/depthwise:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_7_pointwise/Conv2D:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_7_pointwise/Conv2D_bn_offset:0' shape=(384,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_7_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_7_pointwise/Relu:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_8_depthwise/depthwise_weights:0' shape=(3, 3, 384, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_8_pointwise/weights:0' shape=(1, 1, 384, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_8_depthwise/depthwise:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_8_pointwise/Conv2D:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_8_pointwise/Conv2D_bn_offset:0' shape=(384,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_8_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_8_pointwise/Relu:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_9_depthwise/depthwise_weights:0' shape=(3, 3, 384, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_9_pointwise/weights:0' shape=(1, 1, 384, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_9_depthwise/depthwise:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_9_pointwise/Conv2D:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_9_pointwise/Conv2D_bn_offset:0' shape=(384,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_9_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_9_pointwise/Relu:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_10_depthwise/depthwise_weights:0' shape=(3, 3, 384, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_10_pointwise/weights:0' shape=(1, 1, 384, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_10_depthwise/depthwise:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_10_pointwise/Conv2D:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_10_pointwise/Conv2D_bn_offset:0' shape=(384,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_10_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_10_pointwise/Relu:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_11_depthwise/depthwise_weights:0' shape=(3, 3, 384, 1) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_11_pointwise/weights:0' shape=(1, 1, 384, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_11_depthwise/depthwise:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_11_pointwise/Conv2D:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_11_pointwise/Conv2D_bn_offset:0' shape=(384,) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_11_pointwise/BatchNorm/FusedBatchNorm:0' shape=(?, ?, ?, 384) dtype=float32>,)
(<tf.Tensor 'TfPoseEstimator/MobilenetV1/Conv2d_11_pointwise/Relu:0' shape=(?, ?, ?, 384) dtype=float32>,)
除了 TfPoseEstimator/image:0
和 TfPoseEstimator/MobilenetV1/Conv2d_0/Conv2D:0
之外的其他层仍然具有 ?
形状.
Another layers beside TfPoseEstimator/image:0
and TfPoseEstimator/MobilenetV1/Conv2d_0/Conv2D:0
still have ?
shape.
我是 Tensorflow 的新手,所以这可能是一个愚蠢的问题,但如何更改已保存模型的输入形状?
I'm very new in Tensorflow so this might be a stupid question, but how to change the input shape of a saved model?
推荐答案
我设法用这个解决了这个问题.
I manage to solve this problem using this.
import tensorflow as tf
if __name__ == '__main__':
graph_path = 't/tf_model.pb'
with tf.gfile.GFile(graph_path, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
graph = tf.get_default_graph()
tf_new_image = tf.placeholder(shape=(1, 368, 368, 3), dtype='float32', name='new_image')
tf.import_graph_def(graph_def, name='TfPoseEstimator', input_map={"image:0": tf_new_image})
tf.train.write_graph(graph, "t", "mobilenet_thin_model.pb", as_text=False)
这篇关于如何在 Tensorflow 中更改保存的模型输入形状?的文章就介绍到这了,希望我们推荐的答案对大家有所帮助,也希望大家多多支持IT屋!