Run ROLO on tensorflow 1.1.8

[shuchitagupta]

Could someone help me run ROLO on newer versions of tensorflow? Thanks.

[wanjinchang]

I modifiy the LSTM_Single function to with tf.device('/gpu:0'): # X, input shape: (batch_size, time_step_size, input_vec_size) # XT shape: (time_step_size, batch_size, input_vec_size) _X = tf.transpose(_X, [1, 0, 2]) # permute time_step_size and batch_size # Reshape to prepare input to hidden activation _X = tf.reshape(_X, [self.num_steps * self.batch_size, self.num_input]) # (num_steps*batch_size, num_input) # Split data because rnn cell needs a list of inputs for the RNN inner loop # Each array shape: (batch_size, num_input) _X = tf.split(_X, self.num_steps, 0) # n_steps * (batch_size, num_input) print(_X)

    cell = tf.nn.rnn_cell.LSTMCell(self.num_input, self.num_input)
    state = cell.zero_state(self.batch_size, dtype=tf.float32)
    outputs, state = tf.nn.static_rnn(cell, _X, initial_state=state, dtype=tf.float32)
    tf.get_variable_scope().reuse_variables()

it run correctly,but the result run on pretrainde model is wrong, cannot reproduce the result mentioned by author...

[wei-1234567]

请问你在跑的过程中遇到 ValueError：使用无效保存路径调用恢复/u03/Guanghan/dev/ROLO-dev/output/ROLO_model/model_step3_exp1_old.ckpt 问题吗，是怎样解决的

[wanjinchang]

你把这个路径改成你本地模型的路径就可以啦

[dvidal8]

@wanjinchang what model .ckpt did you use? I got the next error:

NotFoundError (see above for traceback): Tensor name "rnn/lstm_cell/bias" not found in checkpoint files /home/dvn/ROLO/models/model_demo.ckpt

And the error shows up with any of the uploaded models...

[wanjinchang]

The TensorFlow version of this project maybe 0.x,if you use newer version upper to 1.x, the api of the lstm have been changed.You should use the follow code for reference to convert the model to fix this problem. OLD_CHECKPOINT_FILE = "your path/model_step3_exp3.ckpt" NEW_CHECKPOINT_FILE = "your path/model_step3_exp3_new.ckpt"

import tensorflow as tf vars_to_rename = { "RNN/LSTMCell/W_0": "rnn/lstm_cell/kernel", "RNN/LSTMCell/B": "rnn/lstm_cell/bias", } new_checkpoint_vars = {} reader = tf.train.NewCheckpointReader(OLD_CHECKPOINT_FILE) for old_name in reader.get_variable_to_shape_map(): print(old_name) if old_name in vars_to_rename: new_name = vars_to_rename[old_name] else: new_name = old_name new_checkpoint_vars[new_name] = tf.Variable(reader.get_tensor(old_name))

init = tf.global_variables_initializer() saver = tf.train.Saver(new_checkpoint_vars)

with tf.Session() as sess: sess.run(init) saver.save(sess, NEW_CHECKPOINT_FILE)

[bluetooth12]

I tried the code above (see below) and I get three files: model_step3_exp3_new.ckpt.data-00000-of-00001 model_step3_exp3_new.ckpt.index model_step3_exp3_new.ckpt.meta

What do I put in ROLO_network_test_all.py (approx line 264)? self.rolo_weights_file= 'C:/Users/el006794/ROLO-master/model_step3_exp3.ckpt'

OLD_CHECKPOINT_FILE = "C:/Users/el006794/ROLO-master/model_step3_exp3.ckpt" NEW_CHECKPOINT_FILE = "C:/Users/el006794/ROLO-master/model_step3_exp3_new.ckpt"

import tensorflow as tf vars_to_rename = { "RNN/LSTMCell/W_0": "rnn/lstm_cell/kernel", "RNN/LSTMCell/B": "rnn/lstm_cell/bias", } new_checkpoint_vars = {} reader = tf.train.NewCheckpointReader(OLD_CHECKPOINT_FILE) for old_name in reader.get_variable_to_shape_map(): print(old_name) if old_name in vars_to_rename: new_name = vars_to_rename[old_name] else: new_name = old_name new_checkpoint_vars[new_name] = tf.Variable(reader.get_tensor(old_name))

init = tf.global_variables_initializer() saver = tf.train.Saver(new_checkpoint_vars)

with tf.Session() as sess: sess.run(init) saver.save(sess, NEW_CHECKPOINT_FILE) #saver.restore(sess, NEW_CHECKPOINT_FILE)

with tf.Session() as sess: saver = tf.train.import_meta_graph('model_step3_exp3_new.ckpt.meta') #saver.restore(sess, "model_step3_exp3_new.ckpt.data-00000-of-00001") saver.restore(sess, NEW_CHECKPOINT_FILE)

#with tf.Session() as sess:

saver = tf.train.Saver()

saver.restore(sess, NEW_CHECKPOINT_FILE)

[pari93411]

Hi! How can I get these three files? I run the code but it gives me the error about the restore file.

[GarryLau]

@bluetooth12 Looking at my rewrite file 'ROLO_network_test_single.py':

import sys,os
CURRENT_DIR=os.path.abspath('.')
print(CURRENT_DIR + '/utils' )
sys.path.append(CURRENT_DIR + '/utils' )

MODEl_DIR=os.path.abspath('../../../tensorflow')
print(MODEl_DIR)
sys.path.append(MODEl_DIR)

# Imports
import ROLO_utils as utils

import tensorflow as tf
from tensorflow.python.ops import rnn, rnn_cell
import cv2

import numpy as np
import os.path
import time
import random


class ROLO_TF:
    disp_console = True
    restore_weights = True#False

    # YOLO parameters
    fromfile = None
    tofile_img = 'test/output.jpg'
    tofile_txt = 'test/output.txt'
    imshow = True
    filewrite_img = False
    filewrite_txt = False
    disp_console = True
    yolo_weights_file = 'weights/YOLO_small.ckpt'
    alpha = 0.1
    threshold = 0.2
    iou_threshold = 0.5
    num_class = 20
    num_box = 2
    grid_size = 7
    classes =  ["aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train","tvmonitor"]
    w_img, h_img = [352, 240]

    # ROLO Network Parameters
    model_path = '/home/lg/projects/ROLO/new'
    # rolo_weights_file = '/u03/Guanghan/dev/ROLO-dev/model_dropout_30.ckpt'
    meta_path = '/home/lg/projects/ROLO/new/model_demo_new.meta'

    lstm_depth = 3
    num_steps = 3  # number of frames as an input sequence
    num_feat = 4096
    num_predict = 6 # final output of LSTM 6 loc parameters
    num_gt = 4
    num_input = num_feat + num_predict # data input: 4096+6= 5002

    # ROLO Parameters
    batch_size = 1
    display_step = 1

    # tf Graph input
    x = tf.placeholder("float32", [None, num_steps, num_input])
    istate = tf.placeholder("float32", [None, 2*num_input]) #state & cell => 2x num_input
    y = tf.placeholder("float32", [None, num_gt])

    # Define weights
    weights = {
        'out': tf.Variable(tf.random_normal([num_input, num_predict]))
    }
    biases = {
        'out': tf.Variable(tf.random_normal([num_predict]))
    }


    def __init__(self,argvs = []):
        print("ROLO init")
        self.ROLO(argvs)


    def LSTM_single(self, name,  _X, _istate, _weights, _biases):
        _X = tf.transpose(_X, [1, 0, 2]) # permute time_step_size and batch_size
        _X = tf.reshape(_X, [self.num_steps * self.batch_size, self.num_input]) # (num_steps*batch_size, num_input)
        _X = tf.split(_X, self.num_steps, 0) # n_steps * (batch_size, num_input)
        cell = tf.nn.rnn_cell.LSTMCell(self.num_input, self.num_input)
        state = cell.zero_state(self.batch_size, dtype=tf.float32)
        outputs, state = tf.nn.static_rnn(cell, _X, initial_state=state, dtype=tf.float32)
        tf.get_variable_scope().reuse_variables()   
        return outputs

        # Experiment with dropout
    def dropout_features(self, feature, prob):
        num_drop = int(prob * 4096)
        drop_index = random.sample(xrange(4096), num_drop)
        for i in range(len(drop_index)):
            index = drop_index[i]
            feature[index] = 0
        return feature
    '''---------------------------------------------------------------------------------------'''
    def build_networks(self):
        if self.disp_console : print ("Building ROLO graph...")

        # Build rolo layers
        self.lstm_module = self.LSTM_single('lstm_test', self.x, self.istate, self.weights, self.biases)
        self.ious= tf.Variable(tf.zeros([self.batch_size]), name="ious")
        self.sess = tf.Session()
        self.sess.run(tf.global_variables_initializer())
        self.saver = tf.train.import_meta_graph(self.meta_path)       #loading map
        self.saver.restore(self.sess, tf.train.latest_checkpoint(self.model_path))
        if self.disp_console : print ("Loading complete!" + '\n')


    def testing(self, x_path, y_path):
        total_loss = 0

        print("TESTING ROLO...")
        # Use rolo_input for LSTM training
        pred = self.LSTM_single('lstm_train', self.x, self.istate, self.weights, self.biases)
        print("pred: ", pred)
        self.pred_location = pred[0][:, 4097:4101]
        print("pred_location: ", self.pred_location)
        print("self.y: ", self.y)

        self.correct_prediction = tf.square(self.pred_location - self.y)
        print("self.correct_prediction: ", self.correct_prediction)
        self.accuracy = tf.reduce_mean(self.correct_prediction) * 100
        print("self.accuracy: ", self.accuracy)

        # Initializing the variables
        init = tf.initialize_all_variables()

        # Launch the graph
        with tf.Session() as sess:
            if (self.restore_weights == True):
                sess.run(init)
                self.saver.restore(self.sess, tf.train.latest_checkpoint(self.model_path))
                print ("Loading complete!" + '\n')
            else:
                sess.run(init)

            id = 0 #don't change this

            # Keep training until reach max iterations
            while id < self.testing_iters - self.num_steps:
                # Load training data & ground truth
                batch_xs = self.rolo_utils.load_yolo_output_test(x_path, self.batch_size, self.num_steps, id) # [num_of_examples, num_input] (depth == 1)

                # Apply dropout to batch_xs
                #for item in range(len(batch_xs)):
                #    batch_xs[item] = self.dropout_features(batch_xs[item], 0.4)

                batch_ys = self.rolo_utils.load_rolo_gt_test(y_path, self.batch_size, self.num_steps, id)
                print("Batch_ys_initial: ", batch_ys)
                batch_ys = utils.locations_from_0_to_1(self.w_img, self.h_img, batch_ys)


                # Reshape data to get 3 seq of 5002 elements
                batch_xs = np.reshape(batch_xs, [self.batch_size, self.num_steps, self.num_input])
                batch_ys = np.reshape(batch_ys, [self.batch_size, 4])
                print("Batch_ys: ", batch_ys)

                pred_location= sess.run(self.pred_location,feed_dict={self.x: batch_xs, self.y: batch_ys, self.istate: np.zeros((self.batch_size, 2*self.num_input))})
                print("ROLO Pred: ", pred_location)
                print("ROLO Pred in pixel: ", pred_location[0][0]*self.w_img, pred_location[0][1]*self.h_img, pred_location[0][2]*self.w_img, pred_location[0][3]*self.h_img)


                # Save pred_location to file
                utils.save_rolo_output_test(self.output_path, pred_location, id, self.num_steps, self.batch_size)

                #sess.run(optimizer, feed_dict={self.x: batch_xs, self.y: batch_ys, self.istate: np.zeros((self.batch_size, 2*self.num_input))})

                if id % self.display_step == 0:
                    # Calculate batch loss
                    loss = sess.run(self.accuracy, feed_dict={self.x: batch_xs, self.y: batch_ys, self.istate: np.zeros((self.batch_size, 2*self.num_input))})
                    print ("Iter " + str(id*self.batch_size) + ", Minibatch Loss= " + "{:.6f}".format(loss)) #+ "{:.5f}".format(self.accuracy)
                    total_loss += loss
                id += 1
                print(id)

            print ("Testing Finished!")
            avg_loss = total_loss/id
            print ("Avg loss: " + str(avg_loss))

        return None

    def ROLO(self, argvs):

            self.rolo_utils= utils.ROLO_utils()
            self.rolo_utils.loadCfg()
            self.params = self.rolo_utils.params

            arguments = self.rolo_utils.argv_parser(argvs)

            if self.rolo_utils.flag_train is True:
                self.training(utils.x_path, utils.y_path)
            elif self.rolo_utils.flag_track is True:
                self.build_networks()
                self.track_from_file(utils.file_in_path)
            elif self.rolo_utils.flag_detect is True:
                self.build_networks()
                self.detect_from_file(utils.file_in_path)
            else:
                print ("Default: running ROLO test.")
                self.build_networks()

                test= 8
                [self.w_img, self.h_img, sequence_name, dummy_1, self.testing_iters] = utils.choose_video_sequence(test)

                x_path = os.path.join('../DATA', sequence_name, 'yolo_out/')
                y_path = os.path.join('../DATA', sequence_name, 'groundtruth_rect.txt')
                self.output_path = os.path.join('../DATA', sequence_name, 'rolo_out_test/')
                utils.createFolder(self.output_path)

                self.testing(x_path, y_path)

    '''----------------------------------------main-----------------------------------------------------'''
def main(argvs):
        ROLO_TF(argvs)

if __name__=='__main__':
        main(' ')