Saving broken between tf.keras and Keras Core

[tirthasheshpatel]

Saving doesn't work between tf.keras and keras_core. More specifically, weights saved in keras_core don't load in tf.keras and vice versa.

Here's a MRE:

Saving Script:

import keras_core as keras

class MyModel(keras.models.Model):
    def __init__(self, **kwargs):
        x = keras.Input((None, None, 3))
        x_out = keras.layers.Conv2D(1, 1)(x)
        super().__init__(inputs=x, outputs=x_out, **kwargs)

model = MyModel()
model.summary()
model.save_weights("model.weights.h5")

Model summary before saving:

Model: "my_model"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃ Layer (type)                       ┃ Output Shape                  ┃     Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│ input_layer (InputLayer)           │ (None, None, None, 3)         │           0 │
├────────────────────────────────────┼───────────────────────────────┼─────────────┤
│ conv2d (Conv2D)                    │ (None, None, None, 1)         │           4 │
└────────────────────────────────────┴───────────────────────────────┴─────────────┘
 Total params: 4 (16.00 B)
 Trainable params: 4 (16.00 B)
 Non-trainable params: 0 (0.00 B)

Loading script:

import tensorflow as tf
from tensorflow import keras

tf.debugging.disable_traceback_filtering()

class MyModel(keras.models.Model):
    def __init__(self, **kwargs):
        x = keras.Input((None, None, 3))
        x_out = keras.layers.Conv2D(1, 1)(x)
        super().__init__(inputs=x, outputs=x_out, **kwargs)

model = MyModel()
model.summary()
model.load_weights("model.weights.h5")

Model summary before loading:

Model: "my_model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_1 (InputLayer)        [(None, None, None, 3)]   0         

 conv2d (Conv2D)             (None, None, None, 1)     4         

=================================================================
Total params: 4 (16.00 Byte)
Trainable params: 4 (16.00 Byte)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________

Error:

Traceback (most recent call last):
  File "/home/tirthasheshpatel/oss/keras-cv/mre.py", line 15, in <module>
    model.load_weights("model.weights.h5")
  File "/home/tirthasheshpatel/oss/virtualenvs/keras-dev/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py", line 61, in error_handler
    return fn(*args, **kwargs)
  File "/home/tirthasheshpatel/oss/virtualenvs/keras-dev/lib/python3.10/site-packages/keras/src/engine/training.py", line 3132, in load_weights
    return saving_api.load_weights(
  File "/home/tirthasheshpatel/oss/virtualenvs/keras-dev/lib/python3.10/site-packages/keras/src/saving/saving_api.py", line 267, in load_weights
    saving_lib.load_weights_only(
  File "/home/tirthasheshpatel/oss/virtualenvs/keras-dev/lib/python3.10/site-packages/keras/src/saving/saving_lib.py", line 323, in load_weights_only
    _load_state(
  File "/home/tirthasheshpatel/oss/virtualenvs/keras-dev/lib/python3.10/site-packages/keras/src/saving/saving_lib.py", line 456, in _load_state
    _load_container_state(
  File "/home/tirthasheshpatel/oss/virtualenvs/keras-dev/lib/python3.10/site-packages/keras/src/saving/saving_lib.py", line 513, in _load_container_state
    _load_state(
  File "/home/tirthasheshpatel/oss/virtualenvs/keras-dev/lib/python3.10/site-packages/keras/src/saving/saving_lib.py", line 425, in _load_state
    trackable.load_own_variables(weights_store.get(inner_path))
  File "/home/tirthasheshpatel/oss/virtualenvs/keras-dev/lib/python3.10/site-packages/keras/src/engine/base_layer.py", line 3539, in load_own_variables
    raise ValueError(
ValueError: Layer 'conv2d' expected 2 variables, but received 0 variables during loading. Expected: ['conv2d/kernel:0', 'conv2d/bias:0']

Versions:

Keras Core version: 0.1.5
TensorFlow Keras version: 2.13.1

[fchollet]

Can you print model.summary() on each side (before saving/loading)?

[tirthasheshpatel]

Here's the output of model summary.

Model summary before saving:

Model: "my_model"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃ Layer (type)                       ┃ Output Shape                  ┃     Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│ input_layer (InputLayer)           │ (None, None, None, 3)         │           0 │
├────────────────────────────────────┼───────────────────────────────┼─────────────┤
│ conv2d (Conv2D)                    │ (None, None, None, 1)         │           4 │
└────────────────────────────────────┴───────────────────────────────┴─────────────┘
 Total params: 4 (16.00 B)
 Trainable params: 4 (16.00 B)
 Non-trainable params: 0 (0.00 B)

Model summary before loading:

Model: "my_model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_1 (InputLayer)        [(None, None, None, 3)]   0         

 conv2d (Conv2D)             (None, None, None, 1)     4         

=================================================================
Total params: 4 (16.00 Byte)
Trainable params: 4 (16.00 Byte)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________

[fchollet]

This is probably caused by a naming discrepancy. We should fix it. @nkovela1 do you have cycles to take a look?

[fchollet]

Also -- as a workaround, you can do weights = model.get_weights(); new_model.set_weights(weights). This will work.

[fchollet]

I found the cause -- basically some change in TF seems to have broken the save file namespace in tf.keras (introduction of a _layer_checkpoint_dependencies path which overrides layers. We have to fix this on the tf.keras side. Things are nominal in Keras Core.

[nkovela1]

Sure thing, I just created a Colab to inspect the h5 file and found that same discrepancy with _layer_checkpoint_dependencies: https://colab.sandbox.google.com/drive/1Ir1AQp6DUtYXk-nomRVgjM11ukTPXnnt

[fchollet]

This is now fixed at HEAD but the issue will persist in TF 2.13 and TF 2.14. Use get_weights()/set_weights() as a workaround.

[tirthasheshpatel]

Sounds good, thanks for the quick fix!

[AYREB]

I'm not following. Would you mind elaborating on where I should insert the code for the get_weights()/set_weights() workaround. I am using a script to load and use the model and that's wear the issue is happening.

`import keras import numpy as np import tensorflow as tf from keras.models import load_model import matplotlib.pyplot as plt

class_names = ['apple', 'banana', 'beetroot', 'bell pepper', 'cabbage', 'capsicum', 'carrot', 'cauliflower', 'chilli pepper', 'corn', 'cucumber', 'eggplant', 'garlic', 'ginger', 'grapes', 'jalepeno', 'kiwi', 'lemon', 'lettuce', 'mango', 'onion', 'orange', 'paprika', 'pear', 'peas', 'pineapple', 'pomegranate', 'potato', 'raddish', 'soy beans', 'spinach', 'sweetcorn', 'sweetpotato', 'tomato', 'turnip', 'watermelon']

image_path = "download.jpg"

model = keras.models.load_model("TrainedModel.keras")

img = tf.keras.utils.load_img( image_path, target_size=(180, 180) ) img_array = tf.keras.utils.img_to_array(img) img_array = tf.expand_dims(img_array, 0) # Create a batch

predictions = model.predict(img_array) score = tf.nn.softmax(predictions[0])

print( "This image most likely belongs to {} with a {:.2f} percent confidence." .format(class_names[np.argmax(score)], 100 * np.max(score)) ) `