OliverJMG
commented
9 months ago Changing this line
patch[0].append(tf.argmax(prediction, 1))
to
patch[0].append(tf.argmax(prediction[0]))
resolves the issue.
The prediction variable is a 2D array with only one row, so tf.argmax(prediction, 1) returns a 1D array with only one element as output. This means patch[0] ends up as a list of arrays rather than a list of integers which causes the problem. Passing the only row of prediction to argmax instead means you get an integer output and the code works.
jdbcode
Hi there,
I am running this code in a Jupyter Notebook on a VertexAI instance, n1-highmem-16 with a GPU attached, using the pre-built TF instance.
I think there may be an issue with: https://github.com/google/earthengine-community/blob/master/guides/linked/TF_demo1_keras.ipynb
Running the "Write the predictions to a TFRecord file" code snippet generates the error:
TypeError: only integer scalar arrays can be converted to a scalar index for line : int64_list=tf.train.Int64List( value=patch[0])),I have attempted to resolve this by implementing:
int_predictions_p0 = np.array(patch[0], dtype=np.int64)...feature={ 'prediction': tf.train.Feature( int64_list=tf.train.Int64List( value=int_predictions_p0)),In order to set the dtype to the desired one, noting that the other three classes/patches are requesting float lists and therefore should be fine. Apolgoies, I cannot yet tell you if this fix worked as it is taking time to run! And I am not sure if that is as intended or due to what I have done here...
EDIT: three hours later can confirm that my attempted fix did not work. No data has been written to the output file either during the opertaion or upon terminating the process.