thegodone
commented
3 years ago did you get / solve this issue ?
Open mahenning opened 3 years ago
thegodone
commented
3 years ago did you get / solve this issue ?
mahenning
commented
3 years ago Unfortunatelly not yet. I tried to run a sequential model with it, but for that it has to be an instance of class Layer. For now I test different possibilities, as the error doesn't give me a starting point for fixing. But if I have a solution for it, I'll post it here.
mahenning
commented
3 years ago Regarding the first error
'Cannot convert a symbolic Keras input/output to a numpy array. ' TypeError: Cannot convert a symbolic Keras input/output to a numpy array. This error may indicate that you're trying to pass a symbolic value to a NumPy call, which is not supported. Or, you may be trying to pass Keras symbolic inputs/outputs to a TF API that does not register dispatching, preventing Keras from automatically converting the API call to a lambda layer in the Functional Model.
In the Keras API docs I found this:
Note that even if eager execution is enabled, Input produces a symbolic tensor-like object (i.e. a placeholder). This symbolic tensor-like object can be used with lower-level TensorFlow ops that take tensors as inputs, as such: this is a logistic regression in Keras x = Input(shape=(32,)) y = Dense(16, activation='softmax')(x) model = Model(x, y)
(This behavior does not work for higher-order TensorFlow APIs such as control flow and being directly watched by a tf.GradientTape).
The last line in which the error message is thrown is the ..keras/layers/core.py:1359, which is wrapped backprop.GradientTape.
with backprop.GradientTape(watch_accessed_variables=True) as tape, \ variable_scope.variable_creator_scope(_variable_creator):
We explicitly drop
namearguments here,# to guard against the case where an op explicitly has a # `name` passed (which is susceptible to producing # multiple ops w/ the same name when the layer is reused) kwargs.pop('name', None) result = self.function(*args, **kwargs)
Does that mean that the haste layers can't be used with an keras Input layer? If so, how can I create a model with them if neither the sequential nor the functional api (requires the input layer AFAIK) can be used? In your example on the main page the layer gets a tensor directly. Do I have to handle batches manually and input the data directly instead of letting the fit function work?
thegodone
commented
3 years ago import numpy as np import tensorflow as tf import haste_tf as haste
train_x = np.random.rand(500,40,20) train_y = np.random.rand(500,40,1)
inputs = tf.keras.Input(shape=(train_x.shape[1], train_x.shape[2])) haste1 = haste.LayerNormGRU(20, direction='unidirectional', zoneout=0.1, dropout=0.1) fc1 = tf.keras.layers.Dense(60, activation='relu', kernel_initializer='he_uniform') dr1 = tf.keras.layers.Dropout(0.2) fc2 = tf.keras.layers.Dense(1)
x, state = haste1(inputs, training=True) x = fc1(inputs) x = dr1(x) outputs = fc2(x)
model = tf.keras.Model(inputs=inputs, outputs=outputs)
print(model.summary())
opt = tf.keras.optimizers.Adam(learning_rate=0.01) model.compile(loss='categorical_crossentropy', optimizer=opt)
model_hist = model.fit(train_x, train_y, epochs=10, batch_size=32, verbose=1)
this code is working for me
mahenning
commented
3 years ago For me the error still remains. What python/Tensorflow version did you use for testing?
thegodone
commented
3 years ago it's a jupyter notebook on a Docker image:
GPU: "NVIDIA-SMI 418.116.00 Driver Version: 418.116.00 CUDA Version: 10.2"
print(tf.version) 2.3.0 from platform import python_version print(python_version()) 3.7.10 haste.version print(haste.version) 0.5.0-rc0
NAME="Ubuntu" VERSION="18.04.4 LTS (Bionic Beaver)" ID=ubuntu ID_LIKE=debian PRETTY_NAME="Ubuntu 18.04.4 LTS" VERSION_ID="18.04" HOME_URL="https://www.ubuntu.com/" SUPPORT_URL="https://help.ubuntu.com/" BUG_REPORT_URL="https://bugs.launchpad.net/ubuntu/" PRIVACY_POLICY_URL="https://www.ubuntu.com/legal/terms-and-policies/privacy-policy" VERSION_CODENAME=bionic UBUNTU_CODENAME=bionic
I have also a version tf 2.4 / py 3.8 but not tested on this image
thegodone
commented
3 years ago Does it works for you ?
mahenning
commented
3 years ago No. Unfortunately I can't test this in another environment, as haste does not install the libhaste_tf.so when I run pip(3) install haste_tf on neither a second PC (py 3.6) nor my PC with a fresh venv with python 3.8. I tested a haste reinstall in my conda venv with python 3.7 and there the libhaste_tf.so is created (I also checked if it's getting deleted in between). Any idea on that? I understand that this is an unrelated problem to my initial post, so another issue? I really don't get what I'm doing wrong to get all this errors.
mahenning
commented
3 years ago I tested it in a google colab (https://colab.research.google.com/drive/1mHuT35cnW5uOgulDjHGPf88gm74O9Caz?usp=sharing) and there the same error occurs.
max1mn
commented
3 years ago hello, everyone. actually, the 'real' bug here is this - 'TypeError: haste_lstm() missing 1 required positional argument: 'training'
it happens because keras 'wraps' this call in https://github.com/lmnt-com/haste/blob/459608cfdb4de4d28d2213df8e71f005be8d0f35/frameworks/tf/lstm.py#L211 into TFOpLambda
LIB.haste_lstm expects 'training' argument, but when this call is wrapped, this argument is lost because in constructor of TFOpLambda the flag '_expects_training_arg' explicitly set to False and because of that 'training' is removed
this is what happens in tf 2.5.0
p.s. obvious immediate fix - just comment this line in keras sources
mahenning
commented
3 years ago Hey, thank you for figuring that out, I'll try that! But is there hope to get that fixed officially, without changing the keras core file?
mw66
commented
1 year ago import numpy as np import tensorflow as tf import haste_tf as haste
train_x = np.random.rand(500,40,20) train_y = np.random.rand(500,40,1)
inputs = tf.keras.Input(shape=(train_x.shape[1], train_x.shape[2])) haste1 = haste.LayerNormGRU(20, direction='unidirectional', zoneout=0.1, dropout=0.1) fc1 = tf.keras.layers.Dense(60, activation='relu', kernel_initializer='he_uniform') dr1 = tf.keras.layers.Dropout(0.2) fc2 = tf.keras.layers.Dense(1)
x, state = haste1(inputs, training=True) x = fc1(inputs) x = dr1(x) outputs = fc2(x)
model = tf.keras.Model(inputs=inputs, outputs=outputs)
print(model.summary())
opt = tf.keras.optimizers.Adam(learning_rate=0.01) model.compile(loss='categorical_crossentropy', optimizer=opt)
model_hist = model.fit(train_x, train_y, epochs=10, batch_size=32, verbose=1)
this code is working for me
@thegodone actually, can you add this code as an example of how to mix haste layers into tf.keras model, I think new user will appreciate it.
Hello,
I know this seems more of a debugging problem/problem on my side, but get the following error message when running my code, and it only appears when running it with a haste layer:
I construct the model with the following code:
train_x numpy array shape is (21788, 1000, 4) OS: Ubuntu 20.04 Python version: 3.7 Keras: 2.4.3 Tensorflow: 2.4.1 numpy: 1.19.5 GPU: GTX 1060 CUDA: 11.2
Normally I wouldn't post those error messages on github, but as the code would run without the haste layer, I suspect that the cause of the error lies somewhere close to it, and this repo seems to be the best place to ask and I didn't find any solutions elsewhere. I hope you can help me, I'd really like to try out your implementation for my dataset.