基于keras的模型迁移无法使用npu进行加速

一、问题现象（附报错日志上下文）：
使用自动迁移将模型进行迁移后，发现只增加了from npu_bridge.npu_init import *，原代码是基于keras实现，并且是使用原生Keras API实现的，并非使用Tensorflow的Keras API而实现的，https://support.huaweicloud.com/tfmigr-cann503alpha2training/atlasmprtg_13_0012.html 中提到目前的迁移不支持原生Keras API，在进行自动迁移的过程中，模型代码的迁移报告：
迁移报告1
模型迁移修改内容
并且由于模型代码和调用模型的代码只加了一行from npu_bridge.npu_init import *，导致在服务器上只能用cpu跑模型，不能用npu跑。

二、软件版本:
-- CANN 版本：ascend-share/5.0.3.alpha002_tensorflow-ascend910-cp37-euleros2.8-aarch64-training:1.15.0-21.0.2_0916
--Tensorflow：1.15 :
--Python：3.7.5:

模型代码：

import sys,os
import numpy as np
import scipy.ndimage
import tensorflow as tf

import keras
from keras import backend as K
from keras.initializers import glorot_normal,glorot_normal
from keras.models import Model
from keras.layers import Input,Conv2D,MaxPooling2D,UpSampling2D,Conv2DTranspose,ZeroPadding2D
from keras.layers import Flatten,Dense,Dropout,Activation,RepeatVector,Lambda,Reshape,Subtract
from keras.layers import Concatenate
from keras.layers import Layer
from keras.layers import merge
from keras.layers.normalization import BatchNormalization
from keras.layers.advanced_activations import LeakyReLU
from keras.regularizers import l2
from keras import losses
from keras import optimizers
from keras.callbacks import TensorBoard, ModelCheckpoint
import pix2pose_model.resnet50_mod as resnet
import math

class transformer_loss(Layer):
    def __init__(self,sym=0,**kwargs):
        self.sym=sym

        super(transformer_loss,self).__init__(**kwargs)
    def build(self,input_shape):
        super(transformer_loss,self).build(input_shape)
    def call(self,x):
        y_pred=x[0]
        y_recont_gt=x[1] 
        y_prob_pred=tf.squeeze(x[2],axis=3) 
        y_prob_gt=x[3]
        visible = tf.cast(y_prob_gt > 0.5,y_pred.dtype)
        visible = tf.squeeze(visible,axis=3) 
        #generate transformed values using sym
        if(len(self.sym)>1):
            #if(True):
            for sym_id,transform in enumerate(self.sym): #3x3 matrix
                tf_mat=tf.convert_to_tensor(transform,y_recont_gt.dtype)
                y_gt_transformed = tf.transpose(tf.matmul(tf_mat,tf.transpose(tf.reshape(y_recont_gt,[-1,3]))))
                y_gt_transformed = tf.reshape(y_gt_transformed,[-1,128,128,3])
                loss_xyz_temp = K.sum(K.abs(y_gt_transformed-y_pred),axis=3)/3
                loss_sum=K.sum(loss_xyz_temp,axis=[1,2])
                if(sym_id>0):
                    loss_sums = tf.concat([loss_sums,tf.expand_dims(loss_sum,axis=0)],axis=0)
                    loss_xyzs=  tf.concat([loss_xyzs,tf.expand_dims(loss_xyz_temp,axis=0)],axis=0)
                else:
                    loss_sums = tf.expand_dims(loss_sum,axis=0) 
                    loss_xyzs = tf.expand_dims(loss_xyz_temp,axis=0)
            
            min_values = tf.reduce_min(loss_sums,axis=0,keepdims=True) 
            loss_switch = tf.cast(tf.equal(loss_sums,min_values),y_pred.dtype)
            loss_xyz = tf.expand_dims(tf.expand_dims(loss_switch,axis=2),axis=3)*loss_xyzs
            loss_xyz = K.sum(loss_xyz,axis=0) 
        else:
            loss_xyz = K.sum(K.abs(y_recont_gt-y_pred),axis=3)/3
        prob_loss = K.square(y_prob_pred-K.minimum(loss_xyz,1)) 
        loss_invisible = (1-visible)*loss_xyz
        loss_visible = visible*loss_xyz
        loss = loss_visible*3 + loss_invisible+ 0.5*prob_loss 
        loss = K.mean(loss,axis=[1,2])
        return loss
    def compute_output_shape(self,input_shape):
        return (tuple([input_shape[0][0],1]))


def aemodel_unet_prob(p=0.5):
    input_img = Input(shape=(128, 128, 3))
    bn_axis= 3

    f1_1 = Conv2D(64, (5, 5), strides=(2, 2), name='conv1_1',padding='same')(input_img)
    f1_1 = BatchNormalization(axis=bn_axis)(f1_1)

    #f1_1 = Activation('relu')(f1_1) #64x64x64
    f1_1 = LeakyReLU()(f1_1)
    f1_2 = Conv2D(64, (5, 5), strides=(2, 2), name='conv1_2',padding='same')(input_img)
    f1_2 = BatchNormalization(axis=bn_axis)(f1_2)
    f1_2 = LeakyReLU()(f1_2)
    f1 = Concatenate()([f1_1,f1_2]) #64x64x128

    f2_1 = Conv2D(128, (5, 5), strides=(2, 2), name='conv2_1',padding='same')(f1)
    f2_1 = BatchNormalization(axis=bn_axis)(f2_1)
    f2_1 = LeakyReLU()(f2_1)
    f2_2 = Conv2D(128, (5, 5), strides=(2, 2), name='conv2_2',padding='same')(f1)
    f2_2 = BatchNormalization(axis=bn_axis)(f2_2)
    f2_2 = LeakyReLU()(f2_2)
    f2 = Concatenate()([f2_1,f2_2]) #32x32x256

    f3_1 = Conv2D(128, (5, 5), strides=(2, 2), name='conv3_1',padding='same')(f2)
    f3_1 = BatchNormalization(axis=bn_axis)(f3_1)
    f3_1 = LeakyReLU()(f3_1)
    f3_2 = Conv2D(128, (5, 5), strides=(2, 2), name='conv3_2',padding='same')(f2)
    f3_2 = BatchNormalization(axis=bn_axis)(f3_2)
    f3_2 = LeakyReLU()(f3_2)
    f3 = Concatenate()([f3_1,f3_2]) #16x16x256

    f4_1 = Conv2D(256, (5, 5), strides=(2, 2), name='conv4_1',padding='same')(f3)
    f4_1 = BatchNormalization(axis=bn_axis)(f4_1)
    f4_1 = LeakyReLU()(f4_1)
    f4_2 = Conv2D(256, (5, 5), strides=(2, 2), name='conv4_2',padding='same')(f3)
    f4_2 = BatchNormalization(axis=bn_axis)(f4_2)
    f4_2 = LeakyReLU()(f4_2)
    f4 = Concatenate()([f4_1,f4_2]) #8x8x512

    x= Flatten()(f4)
    encoded= Dense(256)(x) #128:default, 256:large #bottle
    d1= Dense(8*8*256)(encoded)
    d1= Reshape( (8,8,-1) )(d1) #8x8x256
    d1 = Conv2DTranspose(256,kernel_size=(5,5),strides=(2,2),padding='same')(d1) #16x16x256
    d1 = BatchNormalization(axis=bn_axis)(d1)
    #d1 = Dropout(p)(d1)
    d1 = LeakyReLU()(d1)

    d1_uni = Concatenate()([d1,f3_2]) #16x16x256
    d1_uni = Conv2D(256, (5, 5), strides=(1, 1), name='deconv1',padding='same')(d1_uni) #16x16x256
    d1_uni = BatchNormalization(axis=bn_axis)(d1_uni)#
    #d1_uni = Dropout(p)(d1_uni)
    d1_uni = LeakyReLU()(d1_uni)#

    d2 = Conv2DTranspose(128,kernel_size=(5,5),strides=(2,2),padding='same')(d1_uni) #32x32x128
    d2 = BatchNormalization(axis=bn_axis)(d2)
    #d2 = Dropout(p)(d2)
    d2 = LeakyReLU()(d2)
    d2_uni=  Concatenate()([d2,f2_2]) #32x32x256
    d2_uni = Conv2D(256, (5, 5), strides=(1, 1), name='deconv2',padding='same')(d2_uni) #32x32x256
    d2_uni = BatchNormalization(axis=bn_axis)(d2_uni) #
    d2_uni = LeakyReLU()(d2_uni)#
    #to 32x32x256

    d3 = Conv2DTranspose(64,kernel_size=(5,5),strides=(2,2),padding='same')(d2_uni) #64x64x64
    d3 = BatchNormalization(axis=bn_axis)(d3) #
    #d3 = Dropout(p)(d3)
    d3 = LeakyReLU()(d3)
    d3_uni=  Concatenate()([d3,f1_2]) #64x64x128
    d3_uni = Conv2D(128, (5, 5), strides=(1, 1), name='deconv3',padding='same')(d3_uni) #64x64x128
    d3_uni = BatchNormalization(axis=bn_axis)(d3_uni) #
    d3_uni = LeakyReLU()(d3_uni)
    #to 64x64x128

    decoded = Conv2DTranspose(3,kernel_size=(5,5),strides=(2,2),padding='same')(d3_uni) #128x128x3
    decoded = Activation('tanh')(decoded) #8x8x256
    pixel_prob = Conv2DTranspose(1,kernel_size=(5,5),strides=(2,2),padding='same')(d3_uni) #128x128x3
    pixel_prob = Activation('sigmoid')(pixel_prob) #8x8x256
    #has to be sigmoid..
    generator_train = Model(inputs=[input_img],outputs=[decoded,pixel_prob])

    return generator_train


def DCGAN_discriminator():
    nb_filters = 64
    nb_conv = int(np.floor(np.log(128) / np.log(2)))
    list_filters = [nb_filters * min(8, (2 ** i)) for i in range(nb_conv)]

    input_img = Input(shape=(128, 128, 3))
    x = Conv2D(list_filters[0], (3, 3), strides=(2, 2), name="disc_conv2d_1", padding="same")(input_img)
    x = BatchNormalization(axis=-1)(x)
    x = LeakyReLU(0.2)(x)
    # Next convs
    for i, f in enumerate(list_filters[1:]):
        name = "disc_conv2d_%s" % (i + 2)
        x = Conv2D(f, (3, 3), strides=(2, 2), name=name, padding="same")(x)
        x = BatchNormalization(axis=-1)(x)
        x = LeakyReLU(0.2)(x)

    x_flat = Flatten()(x)
    x_out = Dense(1, activation="sigmoid", name="disc_dense")(x_flat)
    discriminator_model = Model(inputs=input_img, outputs=[x_out])
    return discriminator_model


def aemodel_unet_resnet50(p=0.5):
    bn_axis= 3
    input_img = Input(shape=(128,128,3))    
    resnet_model = resnet.ResNet50(include_top=False, weights='imagenet',input_shape=(None,None,3))
    resnet_part = Model(inputs=resnet_model.input, 
                    outputs=[resnet_model.get_layer('act_conv1').output,
                            resnet_model.get_layer('act2c_branch').output,
                            resnet_model.get_layer('act3d_branch').output])

    f1,f2,f3 =resnet_part(input_img)

    f1_2 = Lambda(lambda x : x[:,:,:,:32])(f1)
    f2_2 = Lambda(lambda x : x[:,:,:,:128])(f2)
    f3_2 = Lambda(lambda x : x[:,:,:,:128])(f3)

    f4_1 = Conv2D(256, (5, 5), strides=(2, 2), name='conv4_1',padding='same')(f3)
    f4_1 = BatchNormalization(axis=bn_axis)(f4_1)
    f4_1 = LeakyReLU()(f4_1)
    f4_2 = Conv2D(256, (5, 5), strides=(2, 2), name='conv4_2',padding='same')(f3)
    f4_2 = BatchNormalization(axis=bn_axis)(f4_2)
    f4_2 = LeakyReLU()(f4_2)
    f4 = Concatenate()([f4_1,f4_2]) #8x8x512

    x= Flatten()(f4)
    encoded= Dense(256)(x) #128:default, 256:large #bottle
    d1= Dense(8*8*256)(encoded)
    d1= Reshape( (8,8,-1) )(d1) #8x8x256
    d1 = Conv2DTranspose(256,kernel_size=(5,5),strides=(2,2),padding='same')(d1) #16x16x256
    d1 = BatchNormalization(axis=bn_axis)(d1)
    #d1 = Dropout(p)(d1)
    d1 = LeakyReLU()(d1)

    d1_uni = Concatenate()([d1,f3_2]) #16x16x256
    d1_uni = Conv2D(256, (5, 5), strides=(1, 1), name='deconv1',padding='same')(d1_uni) #16x16x256
    d1_uni = BatchNormalization(axis=bn_axis)(d1_uni)#
    #d1_uni = Dropout(p)(d1_uni)
    d1_uni = LeakyReLU()(d1_uni)#

    d2 = Conv2DTranspose(128,kernel_size=(5,5),strides=(2,2),padding='same')(d1_uni) #32x32x128
    d2 = BatchNormalization(axis=bn_axis)(d2)
    #d2 = Dropout(p)(d2)
    d2 = LeakyReLU()(d2)
    d2_uni=  Concatenate()([d2,f2_2]) #32x32x256
    d2_uni = Conv2D(256, (5, 5), strides=(1, 1), name='deconv2',padding='same')(d2_uni) #32x32x256
    d2_uni = BatchNormalization(axis=bn_axis)(d2_uni) #
    d2_uni = LeakyReLU()(d2_uni)#
    #to 32x32x256

    d3 = Conv2DTranspose(64,kernel_size=(5,5),strides=(2,2),padding='same')(d2_uni) #64x64x64
    d3 = BatchNormalization(axis=bn_axis)(d3) #
    #d3 = Dropout(p)(d3)
    d3 = LeakyReLU()(d3)
    d3_uni=  Concatenate()([d3,f1_2]) #64x64x128
    d3_uni = Conv2D(128, (5, 5), strides=(1, 1), name='deconv3',padding='same')(d3_uni) #64x64x128
    d3_uni = BatchNormalization(axis=bn_axis)(d3_uni) #
    d3_uni = LeakyReLU()(d3_uni)
    #to 64x64x128

    decoded = Conv2DTranspose(3,kernel_size=(5,5),strides=(2,2),padding='same')(d3_uni) #128x128x3
    decoded = Activation('tanh')(decoded) #8x8x256
    pixel_prob = Conv2DTranspose(1,kernel_size=(5,5),strides=(2,2),padding='same')(d3_uni) #128x128x3
    pixel_prob = Activation('sigmoid')(pixel_prob) #8x8x256
    #has to be sigmoid..
    generator_train = Model(inputs=[input_img],outputs=[decoded,pixel_prob])

    return generator_train

你好，keras的模型迁移当前仅支持通过Tensorflow的Keras API编写的训练脚本输入图片说明

那我的这种情况该怎么迁移呢？

您好，请您将脚本修改为使用Tensorflow的Keras API且在cpu上跑通再作迁移。

1、同学参考手工迁移指导书中tf keras的迁移步骤，已经迁移成功。
2、NPU执行遇到动态shape问题，引入原因是keras model.predict()函数，如果batch_size参数不传，默认值为32，和同学实际batch_size=50不一致。设置predict的batch_size=50后，NPU训练功能通过。

和同学确认，当前ISSUE关闭，动态shape支持度问题在后续版本优化。

Ascend / modelzoo

内容风险标识

评论 (4)

Ascend / modelzoo .gitee-modal { width: 500px !important; }

内容风险标识