How to pass update parameters to my SSD object?

[JcmeLs]

Hi ,@balancap

Thanks for you hard work to implement SSD on TensorFlow!

And now I want to train my own model on my own Dataset:

1. I create a Dataset like Pascal VOC 2007.
2. Because I want to Identify the text area,so I only need 2 classes:Background ( 0 ) and text ( 1 ).I change the num_classes and no_annotation_label both set to 2.
3. I train Fine-tuning a network trained on ImageNet.This is my train script

DATASET_DIR=./tfrecords
TRAIN_DIR=./log/
CHECKPOINT_PATH=./checkpoints/vgg_16.ckpt
python train_ssd_network.py \
    --train_dir=${TRAIN_DIR} \
    --dataset_dir=${DATASET_DIR} \
    --dataset_name=pascalvoc_2007 \
    --dataset_split_name=train \
    --model_name=ssd_300_vgg \
    --checkpoint_path=${CHECKPOINT_PATH} \
    --checkpoint_model_scope=vgg_16 \
    --checkpoint_exclude_scopes=ssd_300_vgg/conv6,ssd_300_vgg/conv7,ssd_300_vgg/block8,ssd_300_vgg/block9,ssd_300_vgg/block10,ssd_300_vgg/block11,ssd_300_vgg/block4_box,ssd_300_vgg/block7_box,ssd_300_vgg/block8_box,ssd_300_vgg/block9_box,ssd_300_vgg/block10_box,ssd_300_vgg/block11_box \
    --trainable_scopes=ssd_300_vgg/conv6,ssd_300_vgg/conv7,ssd_300_vgg/block8,ssd_300_vgg/block9,ssd_300_vgg/block10,ssd_300_vgg/block11,ssd_300_vgg/block4_box,ssd_300_vgg/block7_box,ssd_300_vgg/block8_box,ssd_300_vgg/block9_box,ssd_300_vgg/block10_box,ssd_300_vgg/block11_box \
    --save_summaries_secs=60 \
    --save_interval_secs=600 \
    --weight_decay=0.0005 \
    --optimizer=adam \
    --learning_rate=0.0001 \
    --learning_rate_decay_factor=0.94 \
    --batch_size=32

And I use CPU to train 1 day.Now total loss is 2.1. So, I want to eval my model.This is my script:

EVAL_DIR=${TRAIN_DIR}/eval
python eval_ssd_network.py \
    --eval_dir=${EVAL_DIR} \
    --dataset_dir=${DATASET_DIR} \
    --dataset_name=pascalvoc_2007 \
    --dataset_split_name=test \
    --model_name=ssd_300_vgg \
    --checkpoint_path=${TRAIN_DIR} \
    --wait_for_checkpoints=True \
    --batch_size=1 \
    --max_num_batches=500

But it had a InvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match. lhs shape= [84] rhs shape= [8]. And I found issues,you say that

when you change the number of classes or/and the aspects ratios, you need to pass update parameters to the SSD object you're creating. That's what I had forgotten in the convertion script.

But I don't know how to updata.Can you help me? thanks.

Best Regards, Jcmels

[Cuky88]

Did you find a solution?

[JcmeLs]

@Cuky88 I use Google's model --object_detection to train it

[cxzhou95]

@Cuky88 @JcmeLs In nets file ssd_vgg_300.py modify net param num_class=21 -->num_class=2

[Janezzliu]

@JcmeLs Did you find the method to updata it?Or did you solve the problem:Assign requires shapes of both tensors to match. lhs shape= [84] rhs shape= [8].?

[Janezzliu]

I have solve this problem today. You should add --num_classes=2 into the following codes: EVAL_DIR=${TRAIN_DIR}/eval python eval_ssd_network.py
--eval_dir=${EVAL_DIR}
--dataset_dir=${DATASET_DIR}
--dataset_name=pascalvoc_2007
--dataset_split_name=test
--model_name=ssd_300_vgg
--checkpoint_path=${TRAIN_DIR}
--wait_for_checkpoints=True
--batch_size=1
--max_num_batches=500 That is to say,you should run like this: EVAL_DIR=${TRAIN_DIR}/eval python eval_ssd_network.py
--eval_dir=${EVAL_DIR}
--dataset_dir=${DATASET_DIR}
--dataset_name=pascalvoc_2007
--dataset_split_name=test
--model_name=ssd_300_vgg
--checkpoint_path=${TRAIN_DIR}
--wait_for_checkpoints=True
--batch_size=1
--max_num_batches=500
----num_classes=2 I do like this and it works.Hope it can help others.

[magicxiaobai]

@Janezzliu can you tell me where you edit?i meet the same question ,but i do not solve it ,thank you very much.