the pspnet_v1_50.ckpt file?

[CharlesShang]

Just found this awesome project. Would you mind releasing the pspnet_v1_50.ckpt file?

[gperezs]

Dear @pudae, thanks for the implementation, really easy to use. Will you upload the pspnet_v1_50.ckpt model??

[TgithubJ]

Very helpful codes. @pudae , Could you also provide it to me?

[sc4321]

Very helpful codes. @pudae , Could you also provide it to me?

[sc4321]

Guys I'm new in this things. Can anyone please explain what I need to do in order to load an image and see the results of semantic classification ? I have Tensorflow working. At least the "hello world" is displaying (the version with GPU support). Thanks in advance

[HXKwindwizard]

Dear @pudae, thanks for the implementation. Could you explain how to use your codes ? Will you upload the trained model??

[JackHenry1992]

Guys , have you got the pspnet_v1_50.ckpt model? I need it also, Thks a lot.

[pudae]

Sorry for too late answer. I have the pspnet_v1_50.ckpt model. But... because this code is not compete reimplementation of original PSPNet,

• no auxiliary loss
• batch normalization without gathering from multiple GPU
• different data augmentation
• and something else..

it's very poor result, only about 33 mIOU on ADE20k validation set. If you guys need such a poor result, I'll upload it.

[howard-mahe]

Hi @pudae, thanks a lot for this well-written code.

Regarding the _pspnet_v150.ckpt model, I don't know if this refers to (1) the pre-trained model on ImageNet or (2) the one fine-tuned on ADE20K. Anyway, I think model (1) would be useful. Even 33% mIoU is better than FCN, SegNet or DilatedNet. It would be also very useful if you could upload model (2) pre-trained on ImageNet because the root_block is different from the original ResNet50 and I believe that your layers naming convention is also different. Model (2) would make fine-tuning on other dataset possible with Tensorflow.

[pudae]

Sorry to late upload the checkpoint file. I have had one, but I lost it, so I retrained. This is the new checkpoint file.

• gdrive

This result is a little better, about 0.349 mIOU.

[gperezs]

Thanks @pudae !

[Heidisnaps]

@pudae hi. how to make checkpoint file to ckpt file?

[howard-mahe]

@Heidisnaps You don't really need .ckpt file. One can just restore variables using the checkpoint.

• Create a file checkpoint in /path/to/checkpoint/ directory and add the following lines inside:

  model_checkpoint_path: "/path/to/checkpoint/model.ckpt-150000"
  all_model_checkpoint_paths: "/path/to/checkpoint/model.ckpt-150000"

• Restore variables in your evaluation code:

  # Define your graph
  # ...
  # Load variables from the provided checkpoint
  restore_var = ... # list of variables you want to restore
  loader = tf.train.Saver(var_list=restore_var)
  loader = tf.train.import_meta_graph('/path/to/checkpoint/model.ckpt-150000.meta')
  loader.restore(sess, tf.train.latest_checkpoint(/path/to/checkpoint/))

[Heidisnaps]

@howard-mahe thanks for answer. what's mean:

Restore variables in your evaluation code: Define your graph ... Load variables from the provided checkpoint restore_var = ... # list of variables you want to restore loader = tf.train.Saver(var_list=restore_var) loader = tf.train.import_meta_graph('/path/to/checkpoint/model.ckpt-150000.meta') loader.restore(sess, tf.train.latest_checkpoint(/path/to/checkpoint/))

Do you mean to create a new xxx.py? Or do you mean to modify the existing py file (test_model.py or something)?

[howard-mahe]

I gave a quick look at test_model.py, I think you should better use test_segmentation.

There are already in this script some useful lines to restore a checkpoint, like:

variables_to_restore = slim.get_variables_to_restore()
...
saver = tf.train.Saver(variables_to_restore)
...
saver.restore(sess, checkpoint_path)

If this script doesn't run in your first attempt, just adapt a little bit how to restore the checkpoint (using code I provided in my last answer)