add ADE20K dataset #3853
add ADE20K dataset #3853
Conversation
| FLAGS = tf.app.flags.FLAGS | ||
| flags = tf.app.flags | ||
|
|
||
| tf.app.flags.DEFINE_string( |
There was a problem hiding this comment.
Ah.. just some ancient code. I will remove it ;-)
| import random | ||
| import string | ||
| import sys | ||
| from PIL import Image |
There was a problem hiding this comment.
There was a problem hiding this comment.
This is legacy code. I will remove it.
| RuntimeError: If loaded image and label have different shape. | ||
| """ | ||
|
|
||
| img_names = glob.glob(os.path.join(dataset_dir, '*.jpg')) |
| seg_names.append(seg) | ||
|
|
||
| num_images = len(img_names) | ||
| num_per_shard = int(math.ceil(num_images) / float(_NUM_SHARDS)) |
There was a problem hiding this comment.
math.ceil(num_images / float(_NUM_SHARDS)) ?
|
|
||
| CURRENT_DIR=$(pwd) | ||
| WORK_DIR="./ADE20K" | ||
| mkdir -p ${WORK_DIR} |
There was a problem hiding this comment.
"${WORK_DIR}", please also fix variables below.
There was a problem hiding this comment.
I am just trying to "stay consistent with existing code ;-)". Anyway, I will fix those.
| _NUM_SHARDS = 4 | ||
|
|
||
| def _convert_dataset(dataset_split, dataset_dir, dataset_label_dir): | ||
| """ Convert the ADE20k dataset into into tfrecord format (SSTable). |
There was a problem hiding this comment.
"" Convert -> """Converts
| """ Convert the ADE20k dataset into into tfrecord format (SSTable). | ||
|
|
||
| Args: | ||
| dataset_split: dataset split (e.g., train, val) |
There was a problem hiding this comment.
Make first word capital.
There was a problem hiding this comment.
And put a period (i.e., '.') at the end of each sentence.
| splits_to_sizes = { | ||
| 'train': 20210, # num of samples in images/training | ||
| 'val': 2000, # num of samples in images/validation | ||
| 'eval': 2, |
There was a problem hiding this comment.
What is this split? Add comment for it? It seems that this split is not generated?
There was a problem hiding this comment.
Oh yes, this split is used for my own testing. I just forgot to remove it. I will remove it.
| @@ -99,7 +99,7 @@ def get_model_init_fn(train_logdir, | |||
| tf.logging.info('Initializing model from path: %s', tf_initial_checkpoint) | |||
|
|
|||
| # Variables that will not be restored. | |||
| exclude_list = ['global_step'] | |||
| exclude_list = ['global_step', 'logits'] | |||
There was a problem hiding this comment.
There are cases where you do want to restore `logits' weights (e.g., further fine-tuning them on validation set).
My proposal will be: Could you please
- Modify get_extra_layer_scopes in model.py as follows.
def get_extra_layer_scopes(last_layers_contain_logits_only=False):
"""Gets the scopes for extra layers.
Args:
last_layers_contain_logits_only: Boolean, True if only consider logits
as the last layers (i.e., exclude ASPP module, decoder module and so on).
Returns:
A list of scopes for extra layers.
"""
if last_layers_contain_logits_only:
return [_LOGITS_SCOPE_NAME]
else:
return [
_LOGITS_SCOPE_NAME,
_IMAGE_POOLING_SCOPE,
_ASPP_SCOPE,
_CONCAT_PROJECTION_SCOPE,
_DECODER_SCOPE,
]
- Add a flag in train.py.
flags.DEFINE_boolean('last_layers_contain_logits_only', False,
'Only consider logits as last layers or not.')
- Modify line 295 in train.py as follows.
last_layers = model.get_extra_layer_scopes(
FLAGS.last_layers_contain_logits_only)
research/deeplab/g3doc/ade20k.md
Outdated
| fine_tune_batch_norm = False. | ||
|
|
||
| 2. User should fine tune the `min_resize_value` and `max_resize_value` to get | ||
| better result. Note that `resize_factor` has to equals to `output_stride`. |
There was a problem hiding this comment.
has to equals -> has to be equal
|
@YknZhu @aquariusjay All fixed. Please review. |
|
Hello @walkerlala , could you share a script similar to local_test.sh with all steps to download, convert and retrain a deeplab model with ADE20K dataset, instead of Pascal dataset like in local_test.sh ? |
This PR provide scripts/documents for downloading/converting the ADE20K dataset and training deeplabv3 on it. One particular thing to note about is the
exclude_listvariable inutils/train_utils.py. I think it is OK to do this, but it might hurt performance of other models.