Related
I made a model using BERT, for a NLI problem, the algorithm ran without problems, however, when I wanted to adapt it to RoBERTa, and I use strategy.scope (), it generates an error that I don't know how to solve, I appreciate any indication.
´´´
max_len1 = 515 # 128*4 de premisa mas 128*4 de hipotesis
def build_model1():
input_word_ids = tf.keras.Input(shape=(max_len1,), dtype=tf.int32,name="input_word_ids")
input_mask = tf.keras.Input(shape = (max_len1,),dtype=tf.int32,name = "input_mask")
input_type_ids = tf.keras.Input(shape = (max_len1,),dtype=tf.int32,name="input_type_ids")
embedding = model([input_word_ids,input_mask,input_type_ids])[0]
output = tf.keras.layers.Dense(3,activation='softmax')(embedding[:,0,:])
model3 = tf.keras.Model(inputs=[input_word_ids, input_mask, input_type_ids], outputs=output)
model3.compile(tf.keras.optimizers.Adam(lr=1e-5),
loss = 'sparse_categorical_crossentropy', metrics= ['accuracy'])
return model3
with strategy.scope():
model3 = build_model1()
model3.summary()
WARNING:tensorflow:The parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).
WARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f2425631d00>> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method
To silence this warning, decorate the function with #tf.autograph.experimental.do_not_convert
WARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f2425631d00>> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method
To silence this warning, decorate the function with #tf.autograph.experimental.do_not_convert
WARNING: AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f2425631d00>> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method
To silence this warning, decorate the function with #tf.autograph.experimental.do_not_convert
WARNING:tensorflow:The parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).
WARNING:tensorflow:AutoGraph could not transform <function wrap at 0x7f243c214d40> and will run it as-is.
Cause: while/else statement not yet supported
To silence this warning, decorate the function with #tf.autograph.experimental.do_not_convert
WARNING:tensorflow:AutoGraph could not transform <function wrap at 0x7f243c214d40> and will run it as-is.
Cause: while/else statement not yet supported
To silence this warning, decorate the function with #tf.autograph.experimental.do_not_convert
WARNING:tensorflow:The parameter `return_dict` cannot be set in graph mode and will always be set to `True`.
WARNING: AutoGraph could not transform <function wrap at 0x7f243c214d40> and will run it as-is.
Cause: while/else statement not yet supported
To silence this warning, decorate the function with #tf.autograph.experimental.do_not_convert
WARNING:tensorflow:The parameter `return_dict` cannot be set in graph mode and will always be set to `True`.
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-24-e91a2e7e4b41> in <module>()
1 with strategy.scope():
----> 2 model3 = build_model1()
3 model3.summary()
2 frames
/usr/local/lib/python3.7/dist-packages/tensorflow/python/keras/engine/training.py in _validate_compile(self, optimizer, metrics, **kwargs)
2533 'with strategy.scope():\n'
2534 ' model=_create_model()\n'
-> 2535 ' model.compile(...)' % (v, strategy))
2536
2537 # Model metrics must be created in the same distribution strategy scope
ValueError: Variable (<tf.Variable 'tfxlm_roberta_model/roberta/encoder/layer_._0/attention/self/query/kernel:0' shape=(1024, 1024) dtype=float32, numpy=
array([[-0.00294119, -0.00129846, 0.00517603, ..., 0.03835522,
0.0218797 , 0.02100084],
[-0.00933813, -0.05062149, 0.01634834, ..., -0.02387142,
0.0113477 , -0.02262339],
[-0.02023344, -0.04181184, -0.00581416, ..., -0.00609464,
0.00801133, 0.00512151],
...,
[-0.02129102, -0.03157991, -0.04071935, ..., 0.04682101,
0.01948426, 0.00312433],
[-0.04902648, -0.01055507, 0.01377375, ..., 0.00845209,
0.01616496, -0.01041171],
[ 0.00759454, -0.00162496, -0.00215843, ..., -0.03199947,
-0.03871808, 0.04949447]], dtype=float32)>) was not created in the distribution strategy scope
of (<tensorflow.python.distribute.tpu_strategy.TPUStrategy object at 0x7f21fcbbb210>). It is most
likely due to not all layers or the model or optimizer being created outside the distribution
strategy scope. Try to make sure your code looks similar to the following.
with strategy.scope():
model=_create_model()
model.compile(...)
´´´
The same code, as I said above, works perfectly for BERT, obviously, for RoBERTa I made the changes in the tokenizer and the loading of the model
I managed to solve it, investigating, I reached that the implementation of roberta went beyond just calling the model
Update #1 (original question and details below):
As per the suggestion of #MatthijsHollemans below I've tried to run this by removing dynamic_axes from the initial create_onnx step below. This removed both:
Description of image feature 'input_image' has missing or non-positive width 0.
and
Input 'input_image' of layer '63' not found in any of the outputs of the preceeding layers.
Unfortunately this opens up two sub-questions:
I still want to have a functional ONNX model. Is there a more appropriate way to make H and W dynamic? Or should I be saving two versions of the ONNX model, one without dynamic_axes for the CoreML conversion, and one with for use as a valid ONNX model?
Although this solves the compilation error in xcode (specified below) it introduces the following runtime issues:
Finalizing CVPixelBuffer 0x282f4c5a0 while lock count is 1.
[espresso] [Espresso::handle_ex_plan] exception=Invalid X-dimension 1/480 status=-7
[coreml] Error binding image input buffer input_image: -7
[coreml] Failure in bindInputsAndOutputs.
I am calling this the same way I was calling the fixed size model, which does still work fine. The image dimensions are 640 x 480.
As specified below the model should accept any image between 64x64 and higher.
For flexible shape models, do I need to provide an input differently in xcode?
Original Question (parts still relevant)
I have been slowly working on converting a style transfer model from pytorch > onnx > coreml. One of the issues that has been a struggle is flexible/dynamic input + output shape.
This method (besides i/o renaming) has worked well on iOS 12 & 13 when using a static input shape.
I am using the following code to do the onnx > coreml conversion:
def create_coreml(name):
mlmodel = convert(
model="onnx/" + name + ".onnx",
preprocessing_args={'is_bgr': True},
deprocessing_args={'is_bgr': True},
image_input_names=['input_image'],
image_output_names=['stylized_image'],
minimum_ios_deployment_target='13'
)
spec = mlmodel.get_spec()
img_size_ranges = flexible_shape_utils.NeuralNetworkImageSizeRange()
img_size_ranges.add_height_range((64, -1))
img_size_ranges.add_width_range((64, -1))
flexible_shape_utils.update_image_size_range(
spec,
feature_name='input_image',
size_range=img_size_ranges)
flexible_shape_utils.update_image_size_range(
spec,
feature_name='stylized_image',
size_range=img_size_ranges)
mlmodel = coremltools.models.MLModel(spec)
mlmodel.save("mlmodel/" + name + ".mlmodel")
Although the conversion 'succeeds' there are a couple of warnings (spaces added for readability):
Translation to CoreML spec completed. Now compiling the CoreML model.
/usr/local/lib/python3.7/site-packages/coremltools/models/model.py:111:
RuntimeWarning: You will not be able to run predict() on this Core ML model. Underlying exception message was:
Error compiling model:
"Error reading protobuf spec. validator error: Description of image feature 'input_image' has missing or non-positive width 0.".
RuntimeWarning)
Model Compilation done.
/usr/local/lib/python3.7/site-packages/coremltools/models/model.py:111:
RuntimeWarning: You will not be able to run predict() on this Core ML model. Underlying exception message was:
Error compiling model:
"compiler error: Input 'input_image' of layer '63' not found in any of the outputs of the preceeding layers.
".
RuntimeWarning)
If I ignore these warnings and try to compile the model for latest targets (13.0) I get the following error in xcode:
coremlc: Error: compiler error: Input 'input_image' of layer '63' not found in any of the outputs of the preceeding layers.
Here is what the problematic area appears to look like in netron:
My main question is how can I get these two warnings out of the way?
Happy to provide any other details.
Thanks for any advice!
Below is my pytorch > onnx conversion:
def create_onnx(name):
prior = torch.load("pth/" + name + ".pth")
model = transformer.TransformerNetwork()
model.load_state_dict(prior)
dummy_input = torch.zeros(1, 3, 64, 64) # I wasn't sure what I would set the H W to here?
torch.onnx.export(model, dummy_input, "onnx/" + name + ".onnx",
verbose=True,
opset_version=10,
input_names=["input_image"], # These are being renamed from garbled originals.
output_names=["stylized_image"], # ^
dynamic_axes={'input_image':
{2: 'height', 3: 'width'},
'stylized_image':
{2: 'height', 3: 'width'}}
)
onnx.save_model(original_model, "onnx/" + name + ".onnx")
I thought that setting a fixed number of decimal points to all numbers of an array of Decimals, and the new arrays resulting from operations thereof, could be achieved by simply doing:
from decimal import *
getcontext().prec = 5 # 4 decimal points
v = Decimal(0.005)
print(v)
0.005000000000000000104083408558608425664715468883514404296875
However, I get spurious results that I know are the consequence of the contribution of these extra decimals to the calculations. Therefore, as a workaround, I used the round() function like this:
C_subgrid= [Decimal('33.340'), Decimal('33.345'), Decimal('33.350'), Decimal('33.355'), Decimal('33.360'), Decimal('33.365'), Decimal('33.370'), Decimal('33.375'), Decimal('33.380'), Decimal('33.385'), Decimal('33.390'), Decimal('33.395'), Decimal('33.400'), Decimal('33.405'), Decimal('33.410'), Decimal('33.415'), Decimal('33.420'), Decimal('33.425'), Decimal('33.430'), Decimal('33.435'), Decimal('33.440'), Decimal('33.445'), Decimal('33.450'), Decimal('33.455'), Decimal('33.460'), Decimal('33.465'), Decimal('33.470'), Decimal('33.475'), Decimal('33.480'), Decimal('33.485'), Decimal('33.490'), Decimal('33.495'), Decimal('33.500'), Decimal('33.505'), Decimal('33.510'), Decimal('33.515'), Decimal('33.520'), Decimal('33.525'), Decimal('33.530'), Decimal('33.535'), Decimal('33.540'), Decimal('33.545'), Decimal('33.550'), Decimal('33.555'), Decimal('33.560'), Decimal('33.565'), Decimal('33.570'), Decimal('33.575'), Decimal('33.580'), Decimal('33.585'), Decimal('33.590'), Decimal('33.595'), Decimal('33.600'), Decimal('33.605'), Decimal('33.610'), Decimal('33.615'), Decimal('33.620'), Decimal('33.625'), Decimal('33.630'), Decimal('33.635'), Decimal('33.640'), Decimal('33.645'), Decimal('33.650'), Decimal('33.655'), Decimal('33.660'), Decimal('33.665'), Decimal('33.670'), Decimal('33.675'), Decimal('33.680'), Decimal('33.685'), Decimal('33.690'), Decimal('33.695'), Decimal('33.700'), Decimal('33.705'), Decimal('33.710'), Decimal('33.715'), Decimal('33.720'), Decimal('33.725'), Decimal('33.730'), Decimal('33.735'), Decimal('33.740'), Decimal('33.745'), Decimal('33.750'), Decimal('33.755'), Decimal('33.760'), Decimal('33.765'), Decimal('33.770'), Decimal('33.775'), Decimal('33.780'), Decimal('33.785'), Decimal('33.790'), Decimal('33.795'), Decimal('33.800'), Decimal('33.805'), Decimal('33.810'), Decimal('33.815'), Decimal('33.820'), Decimal('33.825'), Decimal('33.830'), Decimal('33.835'), Decimal('33.840'), Decimal('33.845'), Decimal('33.850'), Decimal('33.855'), Decimal('33.860'), Decimal('33.865'), Decimal('33.870'), Decimal('33.875'), Decimal('33.880'), Decimal('33.885'), Decimal('33.890'), Decimal('33.895'), Decimal('33.900'), Decimal('33.905'), Decimal('33.910'), Decimal('33.915'), Decimal('33.920'), Decimal('33.925'), Decimal('33.930'), Decimal('33.935'), Decimal('33.940'), Decimal('33.945'), Decimal('33.950'), Decimal('33.955'), Decimal('33.960'), Decimal('33.965'), Decimal('33.970'), Decimal('33.975'), Decimal('33.980'), Decimal('33.985'), Decimal('33.990'), Decimal('33.995'), Decimal('34.000'), Decimal('34.005'), Decimal('34.010'), Decimal('34.015'), Decimal('34.020'), Decimal('34.025'), Decimal('34.030'), Decimal('34.035'), Decimal('34.040'), Decimal('34.045'), Decimal('34.050'), Decimal('34.055'), Decimal('34.060'), Decimal('34.065'), Decimal('34.070'), Decimal('34.075'), Decimal('34.080'), Decimal('34.085'), Decimal('34.090'), Decimal('34.095'), Decimal('34.100'), Decimal('34.105'), Decimal('34.110'), Decimal('34.115'), Decimal('34.120'), Decimal('34.125'), Decimal('34.130'), Decimal('34.135'), Decimal('34.140')]
C_subgrid = [round(v, 4) for v in C_subgrid]
I got the values of C_subgrid list by printing it out during execution of my code, and I pasted it here. Not sure where the single quotes come from. This code snipped worked fine in Python2.7, but when I upgraded to Python 3.7 it started raising this error:
File "/home2/thomas/Documents/4D-CHAINS_dev/lib/peak.py", line 301, in <listcomp>
C_subgrid = [round(v, 4) for v in C_subgrid] # convert all values to fixed decimal length floats!
decimal.InvalidOperation: [<class 'decimal.InvalidOperation'>]
Strangely, if I run it within ipython it works fine, only within my code it creates problems. Can anybody think of any possible reason?
I have a dataframe which is of length 177 and I want to calculate and plot the partial auto-correlation function (PACF).
I have the data imported etc and I do:
from statsmodels.tsa.stattools import pacf
ys = pacf(data[key][array].diff(1).dropna(), alpha=0.05, nlags=176, method="ywunbiased")
xs = range(lags+1)
plt.figure()
plt.scatter(xs,ys[0])
plt.grid()
plt.vlines(xs, 0, ys[0])
plt.plot(ys[1])
The method used results in numbers greater than 1 for very long lags (90ish) which is incorrect and I get a RuntimeWarning: invalid value encountered in sqrtreturn rho, np.sqrt(sigmasq) but since I can't see their source code I don't know what this means.
To be honest, when I search for PACF, all the examples only carry out PACF up to 40 lags or 60 or so and they never have any significant PACF after lag=2 and so I couldn't compare to other examples either.
But when I use:
method="ols"
# or
method="ywmle"
the numbers are corrected. So it must be the algo they use to solve it.
I tried importing inspect and getsource method but its useless it just shows that it uses another package and I can't find that.
If you also know where the problem arises from, I would really appreciate the help.
For your reference, the values for data[key][array] are:
[1131.130005, 1144.939941, 1126.209961, 1107.300049, 1120.680054, 1140.839966, 1101.719971, 1104.23999, 1114.579956, 1130.199951, 1173.819946, 1211.920044, 1181.27002, 1203.599976, 1180.589966, 1156.849976, 1191.5, 1191.329956, 1234.180054, 1220.329956, 1228.810059, 1207.01001, 1249.47998, 1248.290039, 1280.079956, 1280.660034, 1294.869995, 1310.609985, 1270.089966, 1270.199951, 1276.660034, 1303.819946, 1335.849976, 1377.939941, 1400.630005, 1418.300049, 1438.23999, 1406.819946, 1420.859985, 1482.369995, 1530.619995, 1503.349976, 1455.27002, 1473.98999, 1526.75, 1549.380005, 1481.140015, 1468.359985, 1378.550049, 1330.630005, 1322.699951, 1385.589966, 1400.380005, 1280.0, 1267.380005, 1282.829956, 1166.359985, 968.75, 896.23999, 903.25, 825.880005, 735.090027, 797.869995, 872.8099980000001, 919.1400150000001, 919.320007, 987.4799800000001, 1020.6199949999999, 1057.079956, 1036.189941, 1095.630005, 1115.099976, 1073.869995, 1104.48999, 1169.430054, 1186.689941, 1089.410034, 1030.709961, 1101.599976, 1049.329956, 1141.199951, 1183.26001, 1180.550049, 1257.640015, 1286.119995, 1327.219971, 1325.829956, 1363.609985, 1345.199951, 1320.640015, 1292.280029, 1218.890015, 1131.420044, 1253.300049, 1246.959961, 1257.599976, 1312.410034, 1365.680054, 1408.469971, 1397.910034, 1310.329956, 1362.160034, 1379.319946, 1406.579956, 1440.670044, 1412.160034, 1416.180054, 1426.189941, 1498.109985, 1514.680054, 1569.189941, 1597.569946, 1630.73999, 1606.280029, 1685.72998, 1632.969971, 1681.550049, 1756.540039, 1805.810059, 1848.359985, 1782.589966, 1859.449951, 1872.339966, 1883.949951, 1923.569946, 1960.22998, 1930.6700440000002, 2003.369995, 1972.290039, 2018.050049, 2067.560059, 2058.899902, 1994.9899899999998, 2104.5, 2067.889893, 2085.51001, 2107.389893, 2063.110107, 2103.840088, 1972.180054, 1920.030029, 2079.360107, 2080.409912, 2043.939941, 1940.2399899999998, 1932.22998, 2059.73999, 2065.300049, 2096.949951, 2098.860107, 2173.600098, 2170.949951, 2168.27002, 2126.149902, 2198.810059, 2238.830078, 2278.8701170000004, 2363.639893, 2362.719971, 2384.199951, 2411.800049, 2423.409912, 2470.300049, 2471.649902, 2519.360107, 2575.26001, 2584.840088, 2673.610107, 2823.810059, 2713.830078, 2640.8701170000004, 2648.050049, 2705.27002, 2718.3701170000004, 2816.290039, 2901.52002, 2913.97998]
Your time series is pretty clearly not stationary, so that Yule-Walker assumptions are violated.
More generally, PACF is usually appropriate with stationary time series. You might difference your data first, before considering the partial autocorrelations.
I'm implementing some RL in PyTorch and had to write my own mse_loss function (which I found on Stackoverflow ;) ).
The loss function is:
def mse_loss(input_, target_):
return torch.sum(
(input_ - target_) * (input_ - target_)) / input_.data.nelement()
Now, in my training loop, the first input is something like:
tensor([-1.7610e+10]), tensor([-6.5097e+10])
With this input I'll get the error:
Unable to get repr for <class 'torch.Tensor'>
Computing a = (input_ - target_) works fine, while b = a * a respectively b = torch.pow(a, 2) will fail with the error metioned above.
Does anyone know a fix for this?
Thanks a lot!
Update:
I just tried using torch.nn.functional.mse_loss which will result in the same error..
I had the same error,when I use the below code
criterion = torch.nn.CrossEntropyLoss().cuda()
output=output.cuda()
target=target.cuda()
loss=criterion(output, target)
but I finally found my wrong：output is like tensor([[0.5746,0.4254]]) and target is like tensor([2]),the number 2 is out of indice of output
when I not use GPU,this error message is:
RuntimeError: Assertion `cur_target >= 0 && cur_target < n_classes' failed. at /opt/conda/conda-bld/pytorch-nightly_1547458468907/work/aten/src/THNN/generic/ClassNLLCriterion.c:93
Are you using a GPU ?
I had simillar problem (but I was using gather operations), and when I moved my tensors to CPU I could get a proper error message. I fixed the error, switched back to GPU and it was alright.
Maybe pytorch has trouble outputing the correct error when it comes from inside the GPU.