我使用 keras（張量流后端）訓(xùn)練了一個網(wǎng)絡(luò)，并將模型保存為 json，權(quán)重保存為 h5。我現(xiàn)在嘗試將其轉(zhuǎn)換為單個張量流 pb 文件，它抱怨輸出節(jié)點(diǎn)的名稱。系統(tǒng)信息：Tensorflow 2.3.0 Keras 2.4.3 Cuda 10.1 Cudnn 7轉(zhuǎn)換腳本非常簡單：import jsonfrom tensorflow import kerasfrom keras import backend as Kimport tensorflow as tfjson_file = "my-trained-model.json"h5_file = "my-trained-model.h5"Output_Path = "./trained_models/"Frozen_pb_File = "my-trained-model.pb"def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True): from tensorflow.python.framework.graph_util import convert_variables_to_constants graph = session.graph with graph.as_default(): freeze_var_names = list(set(v.op.name for v in tf.compat.v1.global_variables()).difference(keep_var_names or [])) output_names = output_names or [] output_names += [v.op.name for v in tf.compat.v1.global_variables()] # Graph -> GraphDef ProtoBuf input_graph_def = graph.as_graph_def() if clear_devices: for node in input_graph_def.node: node.device = "" frozen_graph = convert_variables_to_constants(session, input_graph_def, output_names, freeze_var_names) return frozen_graphwith open(json_file, 'r') as json_file: model = keras.models.model_from_json(json_file.read())model.load_weights(h5_file)model.summary()# get output node namesOutputNames = [out.op.name for out in model.outputs]print("\nOutput Names:\n", OutputNames) # this prints "concatenate/concat" as the only output node name# freeze the modelfrozen_graph = freeze_session(tf.compat.v1.keras.backend.get_session(), output_names=OutputNames)# save the output files# this is the .pb file (a binary file)tf.io.write_graph(frozen_graph, Output_Path, Frozen_pb_File, as_text=False)當(dāng)我運(yùn)行這個時，AssertionError: concatenate/concat is not in graph因此，由于某種原因，它正在讀取“concatenate/concat”的輸出節(jié)點(diǎn)名稱。下面給出模型總結(jié)，可以看到輸出節(jié)點(diǎn)是“concatenate”；但是，即使我將輸出節(jié)點(diǎn)名稱硬編碼為“連接”，我也會收到類似的斷言錯誤：AssertionError: concatenate is not in graph
1 回答
繁華開滿天機(jī)
TA貢獻(xiàn)1816條經(jīng)驗(yàn) 獲得超4個贊
看起來這都是由于嘗試凍結(jié) TensorFlow 2.3 模型而引起的。顯然，Tensorflow 2.0+ 已棄用“凍結(jié)”概念，轉(zhuǎn)而采用“保存模型”概念。一旦發(fā)現(xiàn)這一點(diǎn)，我就能夠立即將 h5/json 保存到已保存的模型 pb 中。
我仍然不確定這種格式是否針對推理進(jìn)行了優(yōu)化，所以我將對此進(jìn)行一些跟進(jìn)，但由于我的問題是關(guān)于我看到的錯誤，我想我會發(fā)布導(dǎo)致問題的原因。
作為參考，這是我的 python 腳本，用于將 keras h5/json 文件轉(zhuǎn)換為 Tensorflow 保存的模型格式。
import os
from keras.models import model_from_json
import tensorflow as tf
import genericpath
from genericpath import *
def splitext(p):
p = os.fspath(p)
if isinstance(p, bytes):
sep = b'/'
extsep = b'.'
else:
sep = '/'
extsep = '.'
return genericpath._splitext(p, sep, None, extsep)
def load_model(path,custom_objects={},verbose=0):
from keras.models import model_from_json
path = splitext(path)[0]
with open('%s.json' % path,'r') as json_file:
model_json = json_file.read()
model = model_from_json(model_json, custom_objects=custom_objects)
model.load_weights('%s.h5' % path)
# if verbose: print 'Loaded from %s' % path
return model
json_file = "model.json" # the h5 file should be "model.h5"
model = load_model(json_file) # load the json/h5 pair
model.save('my_saved_model') # this is a directory name to store the saved model
添加回答
舉報(bào)
0/150
提交
取消