convert tensor to numpy array

Question

I write the following code for extract features from two images with deep CNN usinf tensorflow:

# -*- coding: utf-8 -*-
# Implementation of Wang et al 2017: Automatic Brain Tumor Segmentation using Cascaded Anisotropic Convolutional Neural Networks. https://arxiv.org/abs/1709.00382

# Author: Guotai Wang
# Copyright (c) 2017-2018 University College London, United Kingdom. All rights reserved.
# http://cmictig.cs.ucl.ac.uk
#
# Distributed under the BSD-3 licence. Please see the file licence.txt
# This software is not certified for clinical use.
#
from __future__ import absolute_import, print_function
import numpy as np
from scipy import ndimage
import time
import os
import sys
import pickle
import tensorflow as tf
from tensorflow.contrib.data import Iterator
from util.data_loader import *
from util.data_process import *
from util.train_test_func import *
from util.parse_config import parse_config
from train import NetFactory
print("import finished")
def test(config_file):
    # 1, load configure file
    config = parse_config(config_file)
    config_data = config['data']
    config_net1 = config.get('network1', None)
    config_net2 = config.get('network2', None)
    config_net3 = config.get('network3', None)
    config_test = config['testing']  
    batch_size  = config_test.get('batch_size', 5)
    print("configure file loaded")

    # 2.1, network for whole tumor
    if(config_net1):
        net_type1    = config_net1['net_type']
        net_name1    = config_net1['net_name']
        data_shape1  = config_net1['data_shape']
        label_shape1 = config_net1['label_shape']
        class_num1   = config_net1['class_num']
        print("configure file of whole tumor is loaded")

        # construct graph for 1st network
        full_data_shape1 = [batch_size] + data_shape1
        x1 = tf.placeholder(tf.float32, shape = full_data_shape1)          
        net_class1 = NetFactory.create(net_type1)
        net1 = net_class1(num_classes = class_num1,w_regularizer = None,
                    b_regularizer = None, name = net_name1)
        net1.set_params(config_net1)
        predicty1, caty1 = net1(x1, is_training = True)
        proby1 = tf.nn.softmax(predicty1)
    else:
        config_net1ax = config['network1ax']
        config_net1sg = config['network1sg']
        config_net1cr = config['network1cr']
        print("configure files of whole tumor in three planes are loaded")

        # construct graph for 1st network axial
        net_type1ax    = config_net1ax['net_type']
        net_name1ax    = config_net1ax['net_name']
        data_shape1ax  = config_net1ax['data_shape']
        label_shape1ax = config_net1ax['label_shape']
        class_num1ax   = config_net1ax['class_num']

        full_data_shape1ax = [batch_size] + data_shape1ax
        x1ax = tf.placeholder(tf.float32, shape = full_data_shape1ax)          
        net_class1ax = NetFactory.create(net_type1ax)
        net1ax = net_class1ax(num_classes = class_num1ax,w_regularizer = None,
                    b_regularizer = None, name = net_name1ax)
        net1ax.set_params(config_net1ax)
        predicty1ax, caty1ax = net1ax(x1ax, is_training = True)
        proby1ax = tf.nn.softmax(predicty1ax)
        print("graph for 1st network1ax is constructed")

        # construct graph for 1st network sagittal
        net_type1sg    = config_net1sg['net_type']
        net_name1sg    = config_net1sg['net_name']
        data_shape1sg  = config_net1sg['data_shape']
        label_shape1sg = config_net1sg['label_shape']
        class_num1sg   = config_net1sg['class_num']

        full_data_shape1sg = [batch_size] + data_shape1sg
        x1sg = tf.placeholder(tf.float32, shape = full_data_shape1sg)          
        net_class1sg = NetFactory.create(net_type1sg)
        net1sg = net_class1sg(num_classes = class_num1sg,w_regularizer = None,
                    b_regularizer = None, name = net_name1sg)
        net1sg.set_params(config_net1sg)
        predicty1sg, caty1sg = net1sg(x1sg, is_training = True)
        proby1sg = tf.nn.softmax(predicty1sg)
        print("graph for 1st network1sg is constructed")

        # construct graph for 1st network coronal
        net_type1cr    = config_net1cr['net_type']
        net_name1cr    = config_net1cr['net_name']
        data_shape1cr  = config_net1cr['data_shape']
        label_shape1cr = config_net1cr['label_shape']
        class_num1cr   = config_net1cr['class_num']

        full_data_shape1cr = [batch_size] + data_shape1cr
        x1cr = tf.placeholder(tf.float32, shape = full_data_shape1cr)          
        net_class1cr = NetFactory.create(net_type1cr)
        net1cr = net_class1cr(num_classes = class_num1cr,w_regularizer = None,
                    b_regularizer = None, name = net_name1cr)
        net1cr.set_params(config_net1cr)
        predicty1cr, caty1cr = net1cr(x1cr, is_training = True)
        proby1cr = tf.nn.softmax(predicty1cr)
        print("graph for 1st network1cr is constructed")

    # 3, create session and load trained models
    all_vars = tf.global_variables()
    sess = tf.InteractiveSession()   
    sess.run(tf.global_variables_initializer())  
    if(config_net1):
        net1_vars = [x for x in all_vars if x.name[0:len(net_name1) + 1]==net_name1 + '/']
        saver1 = tf.train.Saver(net1_vars)
        saver1.restore(sess, config_net1['model_file'])
    else:
        net1ax_vars = [x for x in all_vars if x.name[0:len(net_name1ax) + 1]==net_name1ax + '/']
        saver1ax = tf.train.Saver(net1ax_vars)
        saver1ax.restore(sess, config_net1ax['model_file'])
        net1sg_vars = [x for x in all_vars if x.name[0:len(net_name1sg) + 1]==net_name1sg + '/']
        saver1sg = tf.train.Saver(net1sg_vars)
        saver1sg.restore(sess, config_net1sg['model_file'])     
        net1cr_vars = [x for x in all_vars if x.name[0:len(net_name1cr) + 1]==net_name1cr + '/']
        saver1cr = tf.train.Saver(net1cr_vars)
        saver1cr.restore(sess, config_net1cr['model_file'])
        print("all variables of net1 is saved")

    # 4, load test images
    dataloader = DataLoader(config_data)
    dataloader.load_data()
    image_num = dataloader.get_total_image_number()

    # 5, start to test
    test_slice_direction = config_test.get('test_slice_direction', 'all')
    save_folder = config_data['save_folder']
    test_time = []
    struct = ndimage.generate_binary_structure(3, 2)
    margin = config_test.get('roi_patch_margin', 5)

    x=['x1','x2']
    paddings=tf.constant([[0,0],[0,0],[10,10],[0,0],[0,0]])
    for i in range(image_num):
        [temp_imgs, temp_weight, temp_name, img_names, temp_bbox, temp_size] = dataloader.get_image_data_with_name(i)
        t0 = time.time()
        # 5.1, test of 1st network
        if(config_net1):
            data_shapes  = [ data_shape1[:-1],  data_shape1[:-1],  data_shape1[:-1]]
            label_shapes = [label_shape1[:-1], label_shape1[:-1], label_shape1[:-1]]
            nets = [net1, net1, net1]
            outputs = [proby1, proby1, proby1]
            inputs =  [x1, x1, x1]
            class_num = class_num1
        else:
            data_shapes  = [ data_shape1ax[:-1],  data_shape1sg[:-1],  data_shape1cr[:-1]]
            label_shapes = [label_shape1ax[:-1], label_shape1sg[:-1], label_shape1cr[:-1]]
            nets = [net1ax, net1sg, net1cr]
            outputs = [proby1ax, proby1sg, proby1cr]
            inputs =  [x1ax, x1sg, x1cr]
            class_num = class_num1ax
        predi=tf.concat([predicty1ax,tf.reshape(predicty1sg,[5,11,180,160,2]),tf.pad(predicty1cr,paddings,"CONSTANT")],0)
        cati=tf.concat([caty1ax,tf.reshape(caty1sg,[5,11,180,160,14]),tf.pad(caty1cr,paddings,"CONSTANT")],0)
        prob1 = test_one_image_three_nets_adaptive_shape(temp_imgs, data_shapes, label_shapes, data_shape1ax[-1], class_num,
                   batch_size, sess, nets, outputs, inputs, shape_mode = 0)
        pred1 =  np.asarray(np.argmax(prob1, axis = 3), np.uint16)
        pred1 = pred1 * temp_weight
        print("net1 is tested")
        globals()[x[i]]=predi
        test_time.append(time.time() - t0)
        print(temp_name)
    test_time = np.asarray(test_time)
    print('test time', test_time.mean())
    np.savetxt(save_folder + '/test_time.txt', test_time)

if __name__ == '__main__':
    if(len(sys.argv) != 2):
        print('Number of arguments should be 2. e.g.')
        print('    python test.py config17/test_all_class.txt')
        exit()
    config_file = str(sys.argv[1])
    assert(os.path.isfile(config_file))
    test(config_file)
    y=tf.stack([x1,x2],0)
    z=tf.Session().run(y)

the output is a tensor(y) that I want to convert it to numpy array using tf.Session().run() but I get this error:

InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [5,19,180,160,4] [[Node: Placeholder = Placeholderdtype=DT_FLOAT, shape=[5,19,180,160,4], _device="/job:localhost/replica:0/task:0/device:GPU:0"]]

Answer 1

Note, this answer is based on a deep look in the crystal ball, predicting the code, which seems to be classified -- at least not written in the question itself.

Have a look at the error message:

InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor

This is exactly, what is wrong with your code. Trimming down, your code is essentially just (there are a lot of issues):

import tensorflow as tf

x1 = tf.placeholder(tf.float32, [None, 3])
y = tf.layers.dense(x1, 2)

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())

print(tf.Session().run(y))

The output tensor y cannot be evaluated without knowing the value of x1, since it depends on this value.

1. Fix use proper naming

import tensorflow as tf

x1 = tf.placeholder(tf.float32, [None, 3], name='my_input')
y = tf.layers.dense(x1, 2, name='fc1')

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())

print(tf.Session().run(y))

Now the error-message becomes much clearer

tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'my_input' with dtype float and shape [?,3]

2. Fix: provide a feed_dict

To let TensorFlow know, which value the computation of y should be based on, you need to feed it into the graph:

import tensorflow as tf

x1 = tf.placeholder(tf.float32, [None, 3], name='my_input')
y = tf.layers.dense(x1, 2, name='fc1')

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
np_result = tf.Session().run(y, feed_dict={x1: [[42, 43, 44]]})

Now, this reveals the second issue with your code. You have 2 sessions:

• sess = tf.InteractiveSession() (session_a)
• tf.Session() in tf.Session().run() (session_b)

Now, session_a get all initialized variables, since your code contains

sess.run(tf.global_variables_initializer())

But, during tf.Session().run(...) another session is created, leaving a new error message:

FailedPreconditionError (see above for traceback): Attempting to use uninitialized value ...

3. Fix: use just one session

import tensorflow as tf

x1 = tf.placeholder(tf.float32, [None, 3], name='my_input')
y = tf.layers.dense(x1, 2, name='fc1')

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
np_result = sess.run(y, feed_dict={x1: [[42, 43, 44]]})

And to provide, the best possible solution:

import tensorflow as tf

# construct graph somewhere
x1 = tf.placeholder(tf.float32, [None, 3], name='my_input')
y = tf.layers.dense(x1, 2, name='fc1')

with tf.Session() as sess:
    # init variables / or load them
    sess.run(tf.global_variables_initializer())
    # make sure, that no operations willl be added to the graph
    sess.graph.finalize()

    # fetch result as numpy array
    np_result = sess.run(y, feed_dict={x1: [[42, 43, 44]]})

The code you either wrote yourself or copied from somewhere is the best demonstration of "How to not write in tensorflow."

last remark:

TensorFlow forces you to create a clean structure. This is important. It should become a habit to follow this structure. After a while, you see these parts immediately, which smells like bad code.

If you use an entire network, then just replace tf.layers.dense with my_network_definition and

def my_network_definition(x1):
    output = ...
    return output

In pytorch, you can write in the arbitrary style like you provided in the question. Not saying, you should do that. But it is possible. So then, try to follow the structure TensorFlow expects from you.

Dear pytorch users, I am looking forward to your feedback.