Mask_rcnn: Tensorflow 2.0 support

Created on 2 Oct 2019  ·  32Comments  ·  Source: matterport/Mask_RCNN

Tensorflow 2.0 has released yesterday . I try to change some API in Tensorflow 2.0 , but not well . When Mask_RCNN support Tensorflow 2.0 ?

The problem is

Screen Shot 2019-10-02 at 9 48 51 PM

picture lokinfey
👍14 🚀1 ❤1

Most helpful comment

For #1775,
I think the "layer.output" is not really under the tf2.0, so it can't be used as bool.
I use the code as below, 

def compile(self, learning_rate, momentum):
    """Gets the model ready for training. Adds losses, regularization, and
    metrics. Then calls the Keras compile() function.
    """
    # Optimizer object
    optimizer = keras.optimizers.SGD(
        lr=learning_rate, momentum=momentum,
        clipnorm=self.config.GRADIENT_CLIP_NORM)
    # Add Losses
    # First, clear previously set losses to avoid duplication
    # self.keras_model._losses = []
    # self.keras_model._per_input_losses = {}
    loss_names = [
        "rpn_class_loss", "rpn_bbox_loss",
        "mrcnn_class_loss", "mrcnn_bbox_loss", "mrcnn_mask_loss"]
    added_loss_name = []
    for name in loss_names:
        layer = self.keras_model.get_layer(name)

        if layer.output.name in added_loss_name:
            # if layer.output in self.keras_model.losses:
            continue
        loss = (
                tf.math.reduce_mean(layer.output, keepdims=True)
                * self.config.LOSS_WEIGHTS.get(name, 1.))
        self.keras_model.add_loss(loss)

        added_loss_name.append(layer.output.name)

It is not really solve the problem. My code is here

picture koso019003 on 16 Dec 2019
👍3 🚀1 ❤1 🎉1

All 32 comments

I've met the same problems too, And I'm trying to use the lower version of keras and tensorflow. Hope it will work.

Rorywh picture Rorywh on 18 Oct 2019

@lokinfey did you resolve this?

tattonchantry picture tattonchantry on 19 Oct 2019

Hi,
I am going to work on a university project about semantic segmentation and on of the requirements is that we use the new Tensorflow 2.0. Is there currently a way to use Mask RCNN with Tensorflow 2.0 or will it be possible in the near future?

HansWurst90 picture HansWurst90 on 20 Oct 2019

@lokinfey did you resolve this?

hi i finished tf 2.0 without gpu ，but i need to more test gpu version

lokinfey picture lokinfey on 23 Oct 2019
👍3

@lokinfey I noticed that Custom Layers output shapes are different in tf.keras and keras.
Mask RCNN use Custom Layers like: PyramidROIAlign, DetectionTargetLayer, etc.

More info: https://github.com/tensorflow/tensorflow/issues/33785

kiflowb777 picture kiflowb777 on 4 Nov 2019

Any update on that?

JacobDomagala picture JacobDomagala on 7 Dec 2019
👍4

Another issue is:

AttributeError                            Traceback (most recent call last)
<ipython-input-55-0da0f5bdf6f6> in <module>
----> 1 model = modellib.MaskRCNN(mode='training', config=config, model_dir=ROOT_DIR)
      2 
      3 model.load_weights(COCO_WEIGHTS_PATH, by_name=True, exclude=[
      4     'mrcnn_class_logits', 'mrcnn_bbox_fc', 'mrcnn_bbox', 'mrcnn_mask'])

~/VC/backend/Mask_RCNN/mrcnn/model.py in __init__(self, mode, config, model_dir)
   1835         self.model_dir = model_dir
   1836         self.set_log_dir()
-> 1837         self.keras_model = self.build(mode=mode, config=config)
   1838 
   1839     def build(self, mode, config):

~/VC/backend/Mask_RCNN/mrcnn/model.py in build(self, mode, config)
   1988             rois, target_class_ids, target_bbox, target_mask =\
   1989                 DetectionTargetLayer(config, name="proposal_targets")([
-> 1990                     target_rois, input_gt_class_ids, gt_boxes, input_gt_masks])
   1991 
   1992             # Network Heads

~/VC/backend/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py in symbolic_fn_wrapper(*args, **kwargs)
     73         if _SYMBOLIC_SCOPE.value:
     74             with get_graph().as_default():
---> 75                 return func(*args, **kwargs)
     76         else:
     77             return func(*args, **kwargs)

~/VC/backend/lib/python3.7/site-packages/keras/engine/base_layer.py in __call__(self, inputs, **kwargs)
    487             # Actually call the layer,
    488             # collecting output(s), mask(s), and shape(s).
--> 489             output = self.call(inputs, **kwargs)
    490             output_mask = self.compute_mask(inputs, previous_mask)
    491 

~/VC/backend/Mask_RCNN/mrcnn/model.py in call(self, inputs)
    662             lambda w, x, y, z: detection_targets_graph(
    663                 w, x, y, z, self.config),
--> 664             self.config.IMAGES_PER_GPU, names=names)
    665         return outputs
    666 

~/VC/backend/Mask_RCNN/mrcnn/utils.py in batch_slice(inputs, graph_fn, batch_size, names)
    818     for i in range(batch_size):
    819         inputs_slice = [x[i] for x in inputs]
--> 820         output_slice = graph_fn(*inputs_slice)
    821         if not isinstance(output_slice, (tuple, list)):
    822             output_slice = [output_slice]

~/VC/backend/Mask_RCNN/mrcnn/model.py in <lambda>(w, x, y, z)
    661             [proposals, gt_class_ids, gt_boxes, gt_masks],
    662             lambda w, x, y, z: detection_targets_graph(
--> 663                 w, x, y, z, self.config),
    664             self.config.IMAGES_PER_GPU, names=names)
    665         return outputs

~/VC/backend/Mask_RCNN/mrcnn/model.py in detection_targets_graph(proposals, gt_class_ids, gt_boxes, gt_masks, config)
    551     positive_count = int(config.TRAIN_ROIS_PER_IMAGE *
    552                          config.ROI_POSITIVE_RATIO)
--> 553     positive_indices = tf.random_shuffle(positive_indices)[:positive_count]
    554     positive_count = tf.shape(positive_indices)[0]
    555     # Negative ROIs. Add enough to maintain positive:negative ratio.

AttributeError: module 'tensorflow' has no attribute 'random_shuffle'
tomgross picture tomgross on 8 Dec 2019
👀3

@tomgross I had the same issue. Did you find the solution?

neoyang0620 picture neoyang0620 on 10 Dec 2019

@neoyang0620 For this issue yes. See PR #1896

tomgross picture tomgross on 10 Dec 2019

I am also facing the same problem as #1775 ! Do anyone solved it?

vincent0924 picture vincent0924 on 14 Dec 2019

I am also facing the same problem as #1775 ! Do anyone solved it?

Tensorflow 2.0 has released yesterday . I try to change some API in Tensorflow 2.0 , but not well . When Mask_RCNN support Tensorflow 2.0 ?

The problem is

Screen Shot 2019-10-02 at 9 48 51 PM

Hi, lokinfey
how u solve this issue?

vincent0924 picture vincent0924 on 14 Dec 2019

Another issue is:

AttributeError                            Traceback (most recent call last)
<ipython-input-55-0da0f5bdf6f6> in <module>
----> 1 model = modellib.MaskRCNN(mode='training', config=config, model_dir=ROOT_DIR)
      2 
      3 model.load_weights(COCO_WEIGHTS_PATH, by_name=True, exclude=[
      4     'mrcnn_class_logits', 'mrcnn_bbox_fc', 'mrcnn_bbox', 'mrcnn_mask'])

~/VC/backend/Mask_RCNN/mrcnn/model.py in __init__(self, mode, config, model_dir)
   1835         self.model_dir = model_dir
   1836         self.set_log_dir()
-> 1837         self.keras_model = self.build(mode=mode, config=config)
   1838 
   1839     def build(self, mode, config):

~/VC/backend/Mask_RCNN/mrcnn/model.py in build(self, mode, config)
   1988             rois, target_class_ids, target_bbox, target_mask =\
   1989                 DetectionTargetLayer(config, name="proposal_targets")([
-> 1990                     target_rois, input_gt_class_ids, gt_boxes, input_gt_masks])
   1991 
   1992             # Network Heads

~/VC/backend/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py in symbolic_fn_wrapper(*args, **kwargs)
     73         if _SYMBOLIC_SCOPE.value:
     74             with get_graph().as_default():
---> 75                 return func(*args, **kwargs)
     76         else:
     77             return func(*args, **kwargs)

~/VC/backend/lib/python3.7/site-packages/keras/engine/base_layer.py in __call__(self, inputs, **kwargs)
    487             # Actually call the layer,
    488             # collecting output(s), mask(s), and shape(s).
--> 489             output = self.call(inputs, **kwargs)
    490             output_mask = self.compute_mask(inputs, previous_mask)
    491 

~/VC/backend/Mask_RCNN/mrcnn/model.py in call(self, inputs)
    662             lambda w, x, y, z: detection_targets_graph(
    663                 w, x, y, z, self.config),
--> 664             self.config.IMAGES_PER_GPU, names=names)
    665         return outputs
    666 

~/VC/backend/Mask_RCNN/mrcnn/utils.py in batch_slice(inputs, graph_fn, batch_size, names)
    818     for i in range(batch_size):
    819         inputs_slice = [x[i] for x in inputs]
--> 820         output_slice = graph_fn(*inputs_slice)
    821         if not isinstance(output_slice, (tuple, list)):
    822             output_slice = [output_slice]

~/VC/backend/Mask_RCNN/mrcnn/model.py in <lambda>(w, x, y, z)
    661             [proposals, gt_class_ids, gt_boxes, gt_masks],
    662             lambda w, x, y, z: detection_targets_graph(
--> 663                 w, x, y, z, self.config),
    664             self.config.IMAGES_PER_GPU, names=names)
    665         return outputs

~/VC/backend/Mask_RCNN/mrcnn/model.py in detection_targets_graph(proposals, gt_class_ids, gt_boxes, gt_masks, config)
    551     positive_count = int(config.TRAIN_ROIS_PER_IMAGE *
    552                          config.ROI_POSITIVE_RATIO)
--> 553     positive_indices = tf.random_shuffle(positive_indices)[:positive_count]
    554     positive_count = tf.shape(positive_indices)[0]
    555     # Negative ROIs. Add enough to maintain positive:negative ratio.

AttributeError: module 'tensorflow' has no attribute 'random_shuffle'

try random.shuffle instead of random_shuffle. Fixed my issue

austingreisman picture austingreisman on 15 Dec 2019

Hi :

I tried to downgrade the tensorflow2.0  to 1.4 and keras to 2.2.4 , And it worked. You can have a try. Good Luck.

在 2019-12-14 09:17:08，"vincent0924" notifications@github.com 写道：

I am also facing the same problem as #1775 ! Do anyone solved it?

Tensorflow 2.0 has released yesterday . I try to change some API in Tensorflow 2.0 , but not well . When Mask_RCNN support Tensorflow 2.0 ?

The problem is

Hi, lokinfey
how u solve this issue?

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub, or unsubscribe.

Rorywh picture Rorywh on 16 Dec 2019
👍1

For #1775,
I think the "layer.output" is not really under the tf2.0, so it can't be used as bool.
I use the code as below, 

def compile(self, learning_rate, momentum):
    """Gets the model ready for training. Adds losses, regularization, and
    metrics. Then calls the Keras compile() function.
    """
    # Optimizer object
    optimizer = keras.optimizers.SGD(
        lr=learning_rate, momentum=momentum,
        clipnorm=self.config.GRADIENT_CLIP_NORM)
    # Add Losses
    # First, clear previously set losses to avoid duplication
    # self.keras_model._losses = []
    # self.keras_model._per_input_losses = {}
    loss_names = [
        "rpn_class_loss", "rpn_bbox_loss",
        "mrcnn_class_loss", "mrcnn_bbox_loss", "mrcnn_mask_loss"]
    added_loss_name = []
    for name in loss_names:
        layer = self.keras_model.get_layer(name)

        if layer.output.name in added_loss_name:
            # if layer.output in self.keras_model.losses:
            continue
        loss = (
                tf.math.reduce_mean(layer.output, keepdims=True)
                * self.config.LOSS_WEIGHTS.get(name, 1.))
        self.keras_model.add_loss(loss)

        added_loss_name.append(layer.output.name)

It is not really solve the problem. My code is here

koso019003 picture koso019003 on 16 Dec 2019
👍3 🚀1 ❤1 🎉1

@Rorywh You meant Tensorflow 1.14.0

matthewmav picture matthewmav on 18 Dec 2019

Yes, Tensorflow 1.14,

在 2019-12-19 00:43:49，"matthewmav" notifications@github.com 写道：

@Rorywh You meant Tensorflow 1.14

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or unsubscribe.

Rorywh picture Rorywh on 19 Dec 2019

Does anyone know how to fix this issue without changing the tensorflow and/or keras version?? Only through code??

advaitkumar3107 picture advaitkumar3107 on 24 Dec 2019

Hi all

I publish mask-rcnn for tensorflow 2.0 https://github.com/lokinfey/MaskRCNNTF2.0 , you can download it to use in your tensorflow 2.0

lokinfey picture lokinfey on 24 Dec 2019
👎16 ❤2

Hi @lokinfey Thanks for joining the attempt to bring tf 2.0 compatiblity to Mask_RCNN.

  • What's the reason behind starting a new repository? This way we loose all the history and tracking changes is harder. If this repository is unmaintained a fork would be the way to go IMO.
  • I don't think your version will work OOTB right now. Several issues I adressed in #1896 are still present 🤔 and keras is still standalone and not the included one in tensorflow. Did you run all included examples successfully?

I suggesst you help fixing #1896 and bring it to merge or put it at a new maintained location. What do you think?

tomgross picture tomgross on 24 Dec 2019

Hi @lokinfey Thanks for joining the attempt to bring tf 2.0 compatiblity to Mask_RCNN.

  • What's the reason behind starting a new repository? This way we loose all the history and tracking changes is harder. If this repository is unmaintained a fork would be the way to go IMO.
  • I don't think your version will work OOTB right now. Several issues I adressed in #1896 are still present 🤔 and keras is still standalone and not the included one in tensorflow. Did you run all included examples successfully?

I suggesst you help fixing #1896 and bring it to merge or put it at a new maintained location. What do you think?

Okay , i will hide this repository and help fixing #1896

lokinfey picture lokinfey on 24 Dec 2019
❤2 👍1

This is great news @tomgross @lokinfey . @Rorywh @advaitkumar3107 for now, I changed the TensorFlow and Keras lines in my Mask R-CNN requirements.txt file so I wouldn't use TF 2.0 by accident in new projects:
tensorflow>=1.3.0,<2.0
keras>=2.0.8,<2.3

garciart picture garciart on 15 Jan 2020
🎉4

Another issue is:

AttributeError                            Traceback (most recent call last)
<ipython-input-55-0da0f5bdf6f6> in <module>
----> 1 model = modellib.MaskRCNN(mode='training', config=config, model_dir=ROOT_DIR)
      2 
      3 model.load_weights(COCO_WEIGHTS_PATH, by_name=True, exclude=[
      4     'mrcnn_class_logits', 'mrcnn_bbox_fc', 'mrcnn_bbox', 'mrcnn_mask'])

~/VC/backend/Mask_RCNN/mrcnn/model.py in __init__(self, mode, config, model_dir)
   1835         self.model_dir = model_dir
   1836         self.set_log_dir()
-> 1837         self.keras_model = self.build(mode=mode, config=config)
   1838 
   1839     def build(self, mode, config):

~/VC/backend/Mask_RCNN/mrcnn/model.py in build(self, mode, config)
   1988             rois, target_class_ids, target_bbox, target_mask =\
   1989                 DetectionTargetLayer(config, name="proposal_targets")([
-> 1990                     target_rois, input_gt_class_ids, gt_boxes, input_gt_masks])
   1991 
   1992             # Network Heads

~/VC/backend/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py in symbolic_fn_wrapper(*args, **kwargs)
     73         if _SYMBOLIC_SCOPE.value:
     74             with get_graph().as_default():
---> 75                 return func(*args, **kwargs)
     76         else:
     77             return func(*args, **kwargs)

~/VC/backend/lib/python3.7/site-packages/keras/engine/base_layer.py in __call__(self, inputs, **kwargs)
    487             # Actually call the layer,
    488             # collecting output(s), mask(s), and shape(s).
--> 489             output = self.call(inputs, **kwargs)
    490             output_mask = self.compute_mask(inputs, previous_mask)
    491 

~/VC/backend/Mask_RCNN/mrcnn/model.py in call(self, inputs)
    662             lambda w, x, y, z: detection_targets_graph(
    663                 w, x, y, z, self.config),
--> 664             self.config.IMAGES_PER_GPU, names=names)
    665         return outputs
    666 

~/VC/backend/Mask_RCNN/mrcnn/utils.py in batch_slice(inputs, graph_fn, batch_size, names)
    818     for i in range(batch_size):
    819         inputs_slice = [x[i] for x in inputs]
--> 820         output_slice = graph_fn(*inputs_slice)
    821         if not isinstance(output_slice, (tuple, list)):
    822             output_slice = [output_slice]

~/VC/backend/Mask_RCNN/mrcnn/model.py in <lambda>(w, x, y, z)
    661             [proposals, gt_class_ids, gt_boxes, gt_masks],
    662             lambda w, x, y, z: detection_targets_graph(
--> 663                 w, x, y, z, self.config),
    664             self.config.IMAGES_PER_GPU, names=names)
    665         return outputs

~/VC/backend/Mask_RCNN/mrcnn/model.py in detection_targets_graph(proposals, gt_class_ids, gt_boxes, gt_masks, config)
    551     positive_count = int(config.TRAIN_ROIS_PER_IMAGE *
    552                          config.ROI_POSITIVE_RATIO)
--> 553     positive_indices = tf.random_shuffle(positive_indices)[:positive_count]
    554     positive_count = tf.shape(positive_indices)[0]
    555     # Negative ROIs. Add enough to maintain positive:negative ratio.

AttributeError: module 'tensorflow' has no attribute 'random_shuffle'

try to use random.shuffle

rrm003 picture rrm003 on 22 Jan 2020

I'm running in to the same issue described in the first comment in this issue:

`OperatorNotAllowedInGraphError            Traceback (most recent call last)
<ipython-input-18-16fc152be760> in <module>()
      3             learning_rate=config.LEARNING_RATE,
      4             epochs=3,
----> 5             layers='all')
      6 history = model.keras_model.history.history

4 frames
/content/drive/My Drive/Colab Notebooks/Mask_RCNN_2.0/mrcnn/model.py in train(self, train_dataset, val_dataset, learning_rate, epochs, layers, augmentation, custom_callbacks, no_augmentation_sources)
   2342         log("Checkpoint Path: {}".format(self.checkpoint_path))
   2343         self.set_trainable(layers)
-> 2344         self.compile(learning_rate, self.config.LEARNING_MOMENTUM)
   2345 
   2346         # Work-around for Windows: Keras fails on Windows when using

/content/drive/My Drive/Colab Notebooks/Mask_RCNN_2.0/mrcnn/model.py in compile(self, learning_rate, momentum)
   2160         for name in loss_names:
   2161             layer = self.keras_model.get_layer(name)
-> 2162             if layer.output in self.keras_model.losses:
   2163                 continue
   2164             loss = (

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/framework/ops.py in __bool__(self)
    763       `TypeError`.
    764     """
--> 765     self._disallow_bool_casting()
    766 
    767   def __nonzero__(self):

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/framework/ops.py in _disallow_bool_casting(self)
    532     else:
    533       # Default: V1-style Graph execution.
--> 534       self._disallow_in_graph_mode("using a `tf.Tensor` as a Python `bool`")
    535 
    536   def _disallow_iteration(self):

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/framework/ops.py in _disallow_in_graph_mode(self, task)
    521     raise errors.OperatorNotAllowedInGraphError(
    522         "{} is not allowed in Graph execution. Use Eager execution or decorate"
--> 523         " this function with @tf.function.".format(task))
    524 
    525   def _disallow_bool_casting(self):

OperatorNotAllowedInGraphError: using a `tf.Tensor` as a Python `bool` is not allowed in Graph execution. Use Eager execution or decorate this function with @tf.function.`

I've been working with tf 2.1 and tf.keras 2.2.4. However, after downgrading to tf2.0 the issue persisted. I've looked around the Mask_RCNN issues section and I've seen this issue pop up, but haven't seen a clear solution. Has anybody figured out what needs to be changed in model.py to resolve this issue?
Thanks

ncate picture ncate on 13 Mar 2020

Hi all

I publish mask-rcnn for tensorflow 2.0 https://github.com/lokinfey/MaskRCNNTF2.0 , you can download it to use in your tensorflow 2.0

Page not found

Omar-Belghaouti picture Omar-Belghaouti on 20 Mar 2020

Any news guys ?

Altimis picture Altimis on 22 Apr 2020

@lokinfey did you resolve this?

hi i finished tf 2.0 without gpu ，but i need to more test gpu version

Hi
I have convert to tf.2.2 ,can you help me with few queries>
Thanks

kartik4949 picture kartik4949 on 26 May 2020

random_sh

Hi all
I publish mask-rcnn for tensorflow 2.0 https://github.com/lokinfey/MaskRCNNTF2.0 , you can download it to use in your tensorflow 2.0

Page not found

Another issue is:

AttributeError                            Traceback (most recent call last)
<ipython-input-55-0da0f5bdf6f6> in <module>
----> 1 model = modellib.MaskRCNN(mode='training', config=config, model_dir=ROOT_DIR)
      2 
      3 model.load_weights(COCO_WEIGHTS_PATH, by_name=True, exclude=[
      4     'mrcnn_class_logits', 'mrcnn_bbox_fc', 'mrcnn_bbox', 'mrcnn_mask'])

~/VC/backend/Mask_RCNN/mrcnn/model.py in __init__(self, mode, config, model_dir)
   1835         self.model_dir = model_dir
   1836         self.set_log_dir()
-> 1837         self.keras_model = self.build(mode=mode, config=config)
   1838 
   1839     def build(self, mode, config):

~/VC/backend/Mask_RCNN/mrcnn/model.py in build(self, mode, config)
   1988             rois, target_class_ids, target_bbox, target_mask =\
   1989                 DetectionTargetLayer(config, name="proposal_targets")([
-> 1990                     target_rois, input_gt_class_ids, gt_boxes, input_gt_masks])
   1991 
   1992             # Network Heads

~/VC/backend/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py in symbolic_fn_wrapper(*args, **kwargs)
     73         if _SYMBOLIC_SCOPE.value:
     74             with get_graph().as_default():
---> 75                 return func(*args, **kwargs)
     76         else:
     77             return func(*args, **kwargs)

~/VC/backend/lib/python3.7/site-packages/keras/engine/base_layer.py in __call__(self, inputs, **kwargs)
    487             # Actually call the layer,
    488             # collecting output(s), mask(s), and shape(s).
--> 489             output = self.call(inputs, **kwargs)
    490             output_mask = self.compute_mask(inputs, previous_mask)
    491 

~/VC/backend/Mask_RCNN/mrcnn/model.py in call(self, inputs)
    662             lambda w, x, y, z: detection_targets_graph(
    663                 w, x, y, z, self.config),
--> 664             self.config.IMAGES_PER_GPU, names=names)
    665         return outputs
    666 

~/VC/backend/Mask_RCNN/mrcnn/utils.py in batch_slice(inputs, graph_fn, batch_size, names)
    818     for i in range(batch_size):
    819         inputs_slice = [x[i] for x in inputs]
--> 820         output_slice = graph_fn(*inputs_slice)
    821         if not isinstance(output_slice, (tuple, list)):
    822             output_slice = [output_slice]

~/VC/backend/Mask_RCNN/mrcnn/model.py in <lambda>(w, x, y, z)
    661             [proposals, gt_class_ids, gt_boxes, gt_masks],
    662             lambda w, x, y, z: detection_targets_graph(
--> 663                 w, x, y, z, self.config),
    664             self.config.IMAGES_PER_GPU, names=names)
    665         return outputs

~/VC/backend/Mask_RCNN/mrcnn/model.py in detection_targets_graph(proposals, gt_class_ids, gt_boxes, gt_masks, config)
    551     positive_count = int(config.TRAIN_ROIS_PER_IMAGE *
    552                          config.ROI_POSITIVE_RATIO)
--> 553     positive_indices = tf.random_shuffle(positive_indices)[:positive_count]
    554     positive_count = tf.shape(positive_indices)[0]
    555     # Negative ROIs. Add enough to maintain positive:negative ratio.

AttributeError: module 'tensorflow' has no attribute 'random_shuffle'

This is still an issue with mask-rcnn 2.1 and tensorflow 2.0.

nchlis picture nchlis on 2 Jun 2020

Hi all,

I have been facing the same issue #1775 for a while and I just found a workaround that works for me. I am working with Tensorflow 2.2.0 and Keras 2.3.1 with GPU. This solution was proposed by @tmedeirosClostra on the issue #1896 which consists in modifying the following piece of code:

for name in loss_names:
            layer = self.keras_model.get_layer(name)
            if layer.output in self.keras_model.losses:
                continue
            loss = (
                tf.reduce_mean(layer.output, keepdims=True)
                * self.config.LOSS_WEIGHTS.get(name, 1.))
            self.keras_model.add_loss(loss)

to

output_names = []
for name in loss_names:
    layer = self.keras_model.get_layer(name)
    if layer.output.name in output_names:
        continue
    loss = (
                tf.reduce_mean(input_tensor=layer.output, keepdims=True)
                * self.config.LOSS_WEIGHTS.get(name, 1.))
    self.keras_model.add_loss(loss)
    output_names.append(layer.output.name)

Then you will be likely to face another problem like AttributeError: 'Model' object has no attribute 'metrics_tensors', which is referenced in the issue #1754 and the solution proposed by @mffigueroa. It consists in changing the line
self.keras_model.metrics_tensors.append(loss)
to
self.keras_model.add_metric(loss, name)

Hope this can solve your problem

IgnacioAmat picture IgnacioAmat on 24 Jun 2020
👍3 ❤1

Hi all,

I have been facing the same issue #1775 for a while and I just found a workaround that works for me. I am working with Tensorflow 2.2.0 and Keras 2.3.1 with GPU. This solution was proposed by @tmedeirosClostra on the issue #1896 which consists in modifying the following piece of code:

for name in loss_names:
            layer = self.keras_model.get_layer(name)
            if layer.output in self.keras_model.losses:
                continue
            loss = (
                tf.reduce_mean(layer.output, keepdims=True)
                * self.config.LOSS_WEIGHTS.get(name, 1.))
            self.keras_model.add_loss(loss)

to

output_names = []
for name in loss_names:
    layer = self.keras_model.get_layer(name)
    if layer.output.name in output_names:
        continue
    loss = (
                tf.reduce_mean(input_tensor=layer.output, keepdims=True)
                * self.config.LOSS_WEIGHTS.get(name, 1.))
    self.keras_model.add_loss(loss)
    output_names.append(layer.output.name)

Then you will be likely to face another problem like AttributeError: 'Model' object has no attribute 'metrics_tensors', which is referenced in the issue #1754 and the solution proposed by @mffigueroa. It consists in changing the line
self.keras_model.metrics_tensors.append(loss)
to
self.keras_model.add_metric(loss, name)

Hope this can solve your problem

problem have solve!!!!

zengjianyou picture zengjianyou on 1 Jul 2020

Hi all,
I have been facing the same issue #1775 for a while and I just found a workaround that works for me. I am working with Tensorflow 2.2.0 and Keras 2.3.1 with GPU. This solution was proposed by @tmedeirosClostra on the issue #1896 which consists in modifying the following piece of code:

for name in loss_names:
            layer = self.keras_model.get_layer(name)
            if layer.output in self.keras_model.losses:
                continue
            loss = (
                tf.reduce_mean(layer.output, keepdims=True)
                * self.config.LOSS_WEIGHTS.get(name, 1.))
            self.keras_model.add_loss(loss)

to

output_names = []
for name in loss_names:
    layer = self.keras_model.get_layer(name)
    if layer.output.name in output_names:
        continue
    loss = (
                tf.reduce_mean(input_tensor=layer.output, keepdims=True)
                * self.config.LOSS_WEIGHTS.get(name, 1.))
    self.keras_model.add_loss(loss)
    output_names.append(layer.output.name)

Then you will be likely to face another problem like AttributeError: 'Model' object has no attribute 'metrics_tensors', which is referenced in the issue #1754 and the solution proposed by @mffigueroa. It consists in changing the line
self.keras_model.metrics_tensors.append(loss)
to
self.keras_model.add_metric(loss, name)
Hope this can solve your problem

problem have solve!!!!

Can you push your code ? Thanks

ily666666 picture ily666666 on 5 Jul 2020

@ily666666 The changes I made in the model.py file are referenced in #2278 , hope this helps you solve your problem !

IgnacioAmat picture IgnacioAmat on 9 Jul 2020
👍1

Hi! kindly see this issue https://github.com/matterport/Mask_RCNN/issues/2312 there is a pointer to the same implementation that supports TensorFlow 2=> (with GPU). If this solves your problem kindly close the issue so as others can navigate to other issues easier

Ab-Abdurrahman picture Ab-Abdurrahman on 23 Aug 2020
Was this page helpful?
0 / 5 - 0 ratings

Related issues

LifeBeyondExpectations picture LifeBeyondExpectations  ·  4Comments
PaulChongPeng picture PaulChongPeng  ·  4Comments
AndreaPi picture AndreaPi  ·  3Comments
apptech-evan-huang picture apptech-evan-huang  ·  3Comments
chrispolo picture chrispolo  ·  4Comments