I'm trying to export a PyTorch model to TorchScript via scripting and I am stuck. I've created a toy class to showcase the issue:
import torch
from torch import nn
class SadModule(nn.Module):
"""Takes a (*, 2) input and runs it thorugh a linear layer. Can optionally
use a skip connection. The usage of the skip connection or not is an
architectural choice.
"""
def __init__(self, use_skip: bool):
nn.Module.__init__(self)
self.use_skip = use_skip
self.layer = nn.Linear(2, 2)
def forward(self, x):
if self.use_skip:
x_input = x
x = self.layer(x)
if self.use_skip:
x = x + x_input
return x
It basically consists of only a linear layer and an optional skip connection. If I try to script the model using
mod1 = SadModule(False)
scripted_mod1 = torch.jit.script(mod)
I get the following error:
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-10-a7ebc7af32c7> in <module>
----> 1 scripted_mod1 = torch.jit.script(mod)
~/Software/miniconda3/envs/pytorch3d/lib/python3.8/site-packages/torch/jit/_script.py in script(obj, optimize, _frames_up, _rcb)
895
896 if isinstance(obj, torch.nn.Module):
--> 897 return torch.jit._recursive.create_script_module(
898 obj, torch.jit._recursive.infer_methods_to_compile
899 )
~/Software/miniconda3/envs/pytorch3d/lib/python3.8/site-packages/torch/jit/_recursive.py in create_script_module(nn_module, stubs_fn, share_types)
350 check_module_initialized(nn_module)
351 concrete_type = get_module_concrete_type(nn_module, share_types)
--> 352 return create_script_module_impl(nn_module, concrete_type, stubs_fn)
353
354 def create_script_module_impl(nn_module, concrete_type, stubs_fn):
~/Software/miniconda3/envs/pytorch3d/lib/python3.8/site-packages/torch/jit/_recursive.py in create_script_module_impl(nn_module, concrete_type, stubs_fn)
408 # Compile methods if necessary
409 if concrete_type not in concrete_type_store.methods_compiled:
--> 410 create_methods_and_properties_from_stubs(concrete_type, method_stubs, property_stubs)
411 torch._C._run_emit_module_hook(cpp_module)
412 concrete_type_store.methods_compiled.add(concrete_type)
~/Software/miniconda3/envs/pytorch3d/lib/python3.8/site-packages/torch/jit/_recursive.py in create_methods_and_properties_from_stubs(concrete_type, method_stubs, property_stubs)
302 property_rcbs = [p.resolution_callback for p in property_stubs]
303
--> 304 concrete_type._create_methods_and_properties(property_defs, property_rcbs, method_defs, method_rcbs, method_defaults)
305
306
RuntimeError:
x_input is not defined in the false branch:
File "<ipython-input-7-d08ed7ff42ec>", line 12
def forward(self, x):
if self.use_skip:
~~~~~~~~~~~~~~~~~
x_input = x
~~~~~~~~~~~ <--- HERE
x = self.layer(x)
if self.use_skip:
and was used here:
File "<ipython-input-7-d08ed7ff42ec>", line 16
x = self.layer(x)
if self.use_skip:
x = x + x_input
~~~~~~~ <--- HERE
return x
So, basically TorchScript isn't able to recognise that for mod1 the True branch of either if statement won't ever be used. Moreover, if we create an instance that actually uses the skip connection,
mod2 = SadModule(True)
scripted_mod2 = torch.jit.script(mod2)
we will get another error:
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-21-b5ca61d8aa73> in <module>
----> 1 scripted_mod2 = torch.jit.script(mod2)
~/Software/miniconda3/envs/pytorch3d/lib/python3.8/site-packages/torch/jit/_script.py in script(obj, optimize, _frames_up, _rcb)
895
896 if isinstance(obj, torch.nn.Module):
--> 897 return torch.jit._recursive.create_script_module(
898 obj, torch.jit._recursive.infer_methods_to_compile
899 )
~/Software/miniconda3/envs/pytorch3d/lib/python3.8/site-packages/torch/jit/_recursive.py in create_script_module(nn_module, stubs_fn, share_types)
350 check_module_initialized(nn_module)
351 concrete_type = get_module_concrete_type(nn_module, share_types)
--> 352 return create_script_module_impl(nn_module, concrete_type, stubs_fn)
353
354 def create_script_module_impl(nn_module, concrete_type, stubs_fn):
~/Software/miniconda3/envs/pytorch3d/lib/python3.8/site-packages/torch/jit/_recursive.py in create_script_module_impl(nn_module, concrete_type, stubs_fn)
408 # Compile methods if necessary
409 if concrete_type not in concrete_type_store.methods_compiled:
--> 410 create_methods_and_properties_from_stubs(concrete_type, method_stubs, property_stubs)
411 torch._C._run_emit_module_hook(cpp_module)
412 concrete_type_store.methods_compiled.add(concrete_type)
~/Software/miniconda3/envs/pytorch3d/lib/python3.8/site-packages/torch/jit/_recursive.py in create_methods_and_properties_from_stubs(concrete_type, method_stubs, property_stubs)
302 property_rcbs = [p.resolution_callback for p in property_stubs]
303
--> 304 concrete_type._create_methods_and_properties(property_defs, property_rcbs, method_defs, method_rcbs, method_defaults)
305
306
RuntimeError:
x_input is not defined in the false branch:
File "<ipython-input-18-ac8b9713c789>", line 17
def forward(self, x):
if self.use_skip:
~~~~~~~~~~~~~~~~~
x_input = x
~~~~~~~~~~~ <--- HERE
x = self.layer(x)
if self.use_skip:
and was used here:
File "<ipython-input-18-ac8b9713c789>", line 21
x = self.layer(x)
if self.use_skip:
x = x + x_input
~~~~~~~ <--- HERE
return x
So in this case TorchScript doesn't understand that both ifs will always be true and that in fact x_input is well defined.
To avoid the issue, I could split the class into two subclasses, as in:
class SadModuleNoSkip(nn.Module):
"""Takes a (*, 2) input and runs it thorugh a linear layer. Can optionally
use a skip connection. The usage of the skip connection or not is an
architectural choice.
"""
def __init__(self):
nn.Module.__init__(self)
self.layer = nn.Linear(2, 2)
def forward(self, x):
x = self.layer(x)
return x
class SadModuleSkip(nn.Module):
"""Takes a (*, 2) input and runs it thorugh a linear layer. Can optionally
use a skip connection. The usage of the skip connection or not is an
architectural choice.
"""
def __init__(self):
nn.Module.__init__(self)
self.layer = nn.Linear(2, 2)
def forward(self, x):
x_input = x
x = self.layer(x)
x = x + x_input
return x
However, I am working on a huge code base and I would have to repeat the process for many classes, which is time consuming and could introduce bugs. Moreover, often the modules I'm working on are huge convolutional nets and the ifs just control the presence of an additional batch normalization. It seems to me undesirable to have to classes that are identical in 99% of the blocks, save for a single batch norm layer.
Is there a way in which I can help TorchScript with its handling of branches?
Edit: added a minimum viable example.
Update: doesn't work even if I type hint use_skip as constant
from typing import Final
class SadModule(nn.Module):
use_skip: Final[bool]
...
I've opened an issue on GitHub. The project maintainers explained that using Final is the way to go. Be careful though, because as of today (May 7, 2021) this feature is still in development (abeit in its final stages, see here for the feature tracker).
Even though it's not yet available in the official releases, it is present in the nightly versions of PyTorch, so you can either install the pytorch-nighly builds as explained in the website (scroll down to Install PyTorch, then choose Preview (Nightly), or wait for the next release.
For anybody reading this answer a few months from now, this feature should be already integrated in the main releases of PyTorch.
Related
Currently, I'm exploring machine learning interpretability tools for one of my project. I found Shapash quite a new tool and many people suggesting to use it to create a few easily interpretable charts for ML model. When I tried it with RandomForestClassifier it worked fine and generate a webpage full of different charts but the same I cannot achieve while using SVM(just exploring this library, not focusing on the perfect ML model for a problem).
Note - using Shapash link here
#Fit blackbox model
svc = svm.SVC()
svc.fit(X_train_smote, y_train_smote)
y_pred = svc.predict(X_test)
print(f"F1 Score {f1_score(y_test, y_pred, average='macro')}")
print(f"Accuracy {accuracy_score(y_test, y_pred)}")
from shapash import SmartExplainer
xpl = SmartExplainer(model=svc)
error which I'm getting -
---------------------------------------------------------------------------
Exception Traceback (most recent call last)
/tmp/ipykernel_13648/1233939729.py in <module>
----> 1 xpl = SmartExplainer(model=svc)
~/Python_AI/ai_env/lib/python3.8/site-packages/shapash/explainer/smart_explainer.py in __init__(self, model, backend, preprocessing, postprocessing, features_groups, features_dict, label_dict, title_story, palette_name, colors_dict, **kwargs)
194 if isinstance(backend, str):
195 backend_cls = get_backend_cls_from_name(backend)
--> 196 self.backend = backend_cls(
197 model=self.model, preprocessing=preprocessing, **kwargs)
198 elif isinstance(backend, BaseBackend):
~/Python_AI/ai_env/lib/python3.8/site-packages/shapash/backend/shap_backend.py in __init__(self, model, preprocessing, explainer_args, explainer_compute_args)
16 self.explainer_args = explainer_args if explainer_args else {}
17 self.explainer_compute_args = explainer_compute_args if explainer_compute_args else {}
---> 18 self.explainer = shap.Explainer(model=model, **self.explainer_args)
19
20 def run_explainer(self, x: pd.DataFrame) -> dict:
~/Python_AI/ai_env/lib/python3.8/site-packages/shap/explainers/_explainer.py in __init__(self, model, masker, link, algorithm, output_names, feature_names, **kwargs)
166 # if we get here then we don't know how to handle what was given to us
167 else:
--> 168 raise Exception("The passed model is not callable and cannot be analyzed directly with the given masker! Model: " + str(model))
169
170 # build the right subclass
Exception: The passed model is not callable and cannot be analyzed directly with the given masker! Model: SVC()
hello I've trained a PPO model from stabel_baselines3 on collab I saved it
model.save("model")
but when I tried loading it I got the following error:
m = PPO.load("model", env=env)
AttributeError Traceback (most recent call last)
/tmp/ipykernel_25649/121834194.py in <module>
2 env = e.MinitaurBulletEnv(render=False)
3 env.reset()
----> 4 m2 = PPO.load("model", env=env)
5 for episode in range(1, 6):
6 obs = env.reset()
~/anaconda3/lib/python3.8/site-packages/stable_baselines3/common/base_class.py in load(cls, path, env, device, custom_objects, **kwargs)
668 env = cls._wrap_env(env, data["verbose"])
669 # Check if given env is valid
--> 670 check_for_correct_spaces(env, data["observation_space"], data["action_space"])
671 else:
672 # Use stored env, if one exists. If not, continue as is (can be used for predict)
~/anaconda3/lib/python3.8/site-packages/stable_baselines3/common/utils.py in check_for_correct_spaces(env, observation_space, action_space)
217 :param action_space: Action space to check against
218 """
--> 219 if observation_space != env.observation_space:
220 raise ValueError(f"Observation spaces do not match: {observation_space} != {env.observation_space}")
221 if action_space != env.action_space:
~/anaconda3/lib/python3.8/site-packages/gym/spaces/box.py in __eq__(self, other)
138
139 def __eq__(self, other):
--> 140 return isinstance(other, Box) and (self.shape == other.shape) and np.allclose(self.low, other.low) and np.allclose(self.high, other.high)
AttributeError: 'Box' object has no attribute 'shape'
knowing that the env is a box env from pybullet
import pybullet_envs.bullet.minitaur_gym_env as e
import gym
env = e.MinitaurBulletEnv(render=False)
env.reset()
additional info is that the model loaded perfectly in collab
From your question, I can't tell if you are or aren't working on Google Colab, but if you are, I think you should definitely include the whole path to the saved model when you load it. Maybe you need to do this even if not in Colab.
What I mean is that your line of code should probably look something like this when you're loading the model:
m = PPO.load("./model.zip/", env=env)
I hope this helps!
I wrote a program using keras that detects real texts from fake (I used 5000 training data and 10,000 test data), I used Transformer and 'distilbert-base-uncased' model for detection. Now I decide to hyperparameters tuning using the grid search , which I encountered the following error:
TuneError Traceback (most recent call last)
<ipython-input-15-c4a44a2180d8> in <module>()
156 tune_iris,
157 verbose=1,
--> 158 config=hyperparameter_space,
159 )
160
/usr/local/lib/python3.6/dist-packages/ray/tune/tune.py in run(run_or_experiment, name, stop, config, resources_per_trial, num_samples, local_dir, upload_dir, trial_name_creator, loggers, sync_to_cloud, sync_to_driver, checkpoint_freq, checkpoint_at_end, sync_on_checkpoint, keep_checkpoints_num, checkpoint_score_attr, global_checkpoint_period, export_formats, max_failures, fail_fast, restore, search_alg, scheduler, with_server, server_port, verbose, progress_reporter, resume, queue_trials, reuse_actors, trial_executor, raise_on_failed_trial, return_trials, ray_auto_init)
354 if incomplete_trials:
355 if raise_on_failed_trial:
--> 356 raise TuneError("Trials did not complete", incomplete_trials)
357 else:
358 logger.error("Trials did not complete: %s", incomplete_trials)
TuneError: ('Trials did not complete', [tune_iris_83131_00000, tune_iris_83131_00001, tune_iris_83131_00002, tune_iris_83131_00003, tune_iris_83131_00004, tune_iris_83131_00005, tune_iris_83131_00006, tune_iris_83131_00007, tune_iris_83131_00008, tune_iris_83131_00009, tune_iris_83131_00010, tune_iris_83131_00011, tune_iris_83131_00012, tune_iris_83131_00013, tune_iris_83131_00014, tune_iris_83131_00015, tune_iris_83131_00016, tune_iris_83131_00017])
The program I wrote is as follows:
data = pd.concat([train_webtext,train_gen,valid_webtext,valid_gen])
sentences=data['text']
labels=labels1+labels2
len(sentences),len(labels)
DistilBertTokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-cased",do_lower_case=False)
input_ids=[]
attention_masks=[]
for sent in sentences:
bert_inp=DistilBertTokenizer.encode_plus(sent,add_special_tokens = True,max_length =64,pad_to_max_length = True,return_attention_mask = True)
input_ids.append(bert_inp['input_ids'])
attention_masks.append(bert_inp['attention_mask'])
input_ids=np.asarray(input_ids)
attention_masks=np.array(attention_masks)
labels=np.array(labels)
class TuneReporterCallback(keras.callbacks.Callback):
"""Tune Callback for Keras.
The callback is invoked every epoch.
"""
def __init__(self, logs={}):
self.iteration = 0
super(TuneReporterCallback, self).__init__()
def on_epoch_end(self, batch, logs={}):
self.iteration += 1
tune.report(keras_info=logs, mean_accuracy=logs.get("accuracy"), mean_loss=logs.get("loss"))
def tune_gpt(config):
train_inp,val_inp,train_label,val_label,train_mask,val_mask=train_test_split(input_ids,labels,attention_masks,test_size=0.6666666666666666)
DistilBert_model = TFDistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased',num_labels=2)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
optimizer = tf.keras.optimizers.Adam(learning_rate=config["learning_rate"],epsilon=1e-08)
DistilBert_model.compile(loss=loss,optimizer=optimizer,metrics=[metric])
checkpoint_callback = [tf.keras.callbacks.ModelCheckpoint( "DistilBert_model.h5",monitor='val_loss',mode='min',save_best_only=True)]
callbacks = [checkpoint_callback, TuneReporterCallback()]
history=DistilBert_model.fit([train_inp,train_mask],train_label,batch_size=config["batch_size"],epochs=config["epochs"],validation_data=([val_inp,val_mask],val_label),callbacks=callbacks)
assert len(inspect.getargspec(tune_gpt).args) == 1, "The `tune_gpt` function needs to take in the arg `config`."
hyperparameter_space ={
"batch_size": tune.grid_search([16, 32]),
"learning_rate": tune.grid_search([2e-5, 3e-5, 5e-5]),
"epochs": tune.grid_search([2, 3, 4])
}
analysis = tune.run(
tune_gpt,
verbose=1,
config=hyperparameter_space,
)
It seems your code has some errors, but the detailed error messages do not appear due to the verbose option.
Please change the verbose option
verbose=1
to
verbose=3
to see the detailed error.
( Verbosity mode. 0 = silent, 1 = only status updates, 2 = status and brief trial results, 3 = status and detailed trial results. Defaults to 3.)
I want to compile my python3 code on ubuntu, and also want to know about the problem, such that i can handle that in future.
It seems there is some problem with the imported library function.
## sample code
1 import numpy as np
2 x = np.array([[-1,-1],[-2,-1],[-3,-2],[1,1],[2,1],[3,2]])
3 y = np.array([1,1,1,2,2,2])
4 from sklearn.naive_bayes import GaussianNB
5 clf = GaussianNB(x, y)
6 clf = clf.fit(x,y) ###showing error on compiling
7 print(clf.predict([[-2,1]]))
## output shown
Traceback (most recent call last):
File "naive.py", line 7, in <module>
clf = clf.fit(x,y)
File "/home/abhihsek/.local/lib/python3.6/site-
packages/sklearn/naive_bayes.py", line 192, in fit
sample_weight=sample_weight)
File "/home/abhihsek/.local/lib/python3.6/site-
packages/sklearn/naive_bayes.py", line 371, in _partial_fit
raise ValueError('Number of priors must match number of'
ValueError: Number of priors must match number of classes.
## code of library function line 192
190 X, y = check_X_y(X, y)
191 return self._partial_fit(X, y, np.unique(y),
_refit=True,
192
sample_weight=sample_weight)
## code of library function line 371
369 # Check that the provide prior match the number of classes
370 if len(priors) != n_classes:
371 raise ValueError('Number of priors must
match
number of'
372 ' classes.')
373 # Check that the sum is 1
As #Suvan Pandey mentioned, then the code won't give any error when writing clf = GaussianNB() instead of clf = GaussianNB(x, y).
If we look at the GaussianNB class then the __init__() can take these parameters:
def __init__(self, priors=None, var_smoothing=1e-9): # <-- these have a default value
self.priors = priors
self.var_smoothing = var_smoothing
The documentation about the two parameters:
priors – Prior probabilities of the classes. If specified the priors are not adjusted according to the data.
var_smoothing – Portion of the largest variance of all features that is added to variances for calculation stability.
As your x and y variables both return an array object then they don't fit the parameters of the __init__(...).
I am trying to recreate this example of bayesian PK/PD modelling using pymc3.....
The video shows the WinBUGS code and I am trying to convert to pymc3
https://www.youtube.com/watch?v=AQDXRoBan6Y
model here....
http://imgur.com/ckoKPRF
WinBUGS code is here ....
http://imgur.com/TsViyBC
My code is ....
from pymc3 import Model, Normal, Lognormal, Uniform
import numpy as np
import pandas as pd
data = pd.read_csv('/Users/Home/Documents/pymc3/fxa.data.csv' )
cobs = np.array(data['cobs'])
fxa = np.array(data['fxa.inh.obs'])
pkpd_model = Model()
with pkpd_model:
# Priors for unknown model parameters
emax = Uniform ('emax', lower =0, upper =100)
ec50 = Lognormal('ec50', mu=0, tau = 100000)
gamma = Uniform('gamma', lower=0, upper =10)
sigma = Uniform('sigma', lower = 0, upper = 1000 )
# Expected value of outcome
fxaMean = emax*(np.power(cobs, gamma)) / (np.power(ec50, gamma) + np.power(cobs, gamma))
# Likelihood (sampling distribution) of observations
fxa = Normal('fxa', mu=fxaMean, sd=sigma, observed=fxa )
But when I run the code I get the following error, which seems to relate to the way theano is interpreting the np.power function.
I am not sure how to proceed as I am a noob to pymc3 and theano and PK/PD modelling too!
Thanks in advance
Applied interval-transform to emax and added transformed emax_interval to model.
Applied log-transform to ec50 and added transformed ec50_log to model.
Applied interval-transform to gamma and added transformed gamma_interval to model.
Applied interval-transform to sigma and added transformed sigma_interval to model.
---------------------------------------------------------------------------
AsTensorError Traceback (most recent call last)
<ipython-input-28-1fa311a15ed0> in <module>()
14
15 # Likelihood (sampling distribution) of observations
---> 16 fxa = Normal('fxa', mu=fxaMean, sd=sigma, observed=fxa )
//anaconda/lib/python2.7/site-packages/pymc3/distributions/distribution.pyc in __new__(cls, name, *args, **kwargs)
23 data = kwargs.pop('observed', None)
24 dist = cls.dist(*args, **kwargs)
---> 25 return model.Var(name, dist, data)
26 elif name is None:
27 return object.__new__(cls) # for pickle
//anaconda/lib/python2.7/site-packages/pymc3/model.pyc in Var(self, name, dist, data)
282 self.named_vars[v.name] = v
283 else:
--> 284 var = ObservedRV(name=name, data=data, distribution=dist, model=self)
285 self.observed_RVs.append(var)
286 if var.missing_values:
//anaconda/lib/python2.7/site-packages/pymc3/model.pyc in __init__(self, type, owner, index, name, data, distribution, model)
556 self.missing_values = data.missing_values
557
--> 558 self.logp_elemwiset = distribution.logp(data)
559 self.model = model
560 self.distribution = distribution
//anaconda/lib/python2.7/site-packages/pymc3/distributions/continuous.pyc in logp(self, value)
191 sd = self.sd
192 mu = self.mu
--> 193 return bound((-tau * (value - mu)**2 + T.log(tau / np.pi / 2.)) / 2.,
194 tau > 0, sd > 0)
195
//anaconda/lib/python2.7/site-packages/theano/tensor/var.pyc in __radd__(self, other)
232 # ARITHMETIC - RIGHT-OPERAND
233 def __radd__(self, other):
--> 234 return theano.tensor.basic.add(other, self)
235
236 def __rsub__(self, other):
//anaconda/lib/python2.7/site-packages/theano/gof/op.pyc in __call__(self, *inputs, **kwargs)
609 """
610 return_list = kwargs.pop('return_list', False)
--> 611 node = self.make_node(*inputs, **kwargs)
612
613 if config.compute_test_value != 'off':
//anaconda/lib/python2.7/site-packages/theano/tensor/elemwise.pyc in make_node(self, *inputs)
541 using DimShuffle.
542 """
--> 543 inputs = list(map(as_tensor_variable, inputs))
544 shadow = self.scalar_op.make_node(
545 *[get_scalar_type(dtype=i.type.dtype).make_variable()
//anaconda/lib/python2.7/site-packages/theano/tensor/basic.pyc in as_tensor_variable(x, name, ndim)
206 except Exception:
207 str_x = repr(x)
--> 208 raise AsTensorError("Cannot convert %s to TensorType" % str_x, type(x))
209
210 # this has a different name, because _as_tensor_variable is the
AsTensorError: ('Cannot convert [Elemwise{mul,no_inplace}.0 Elemwise{mul,no_inplace}.0\n Elemwise{mul,no_inplace}.0 ..., Elemwise{mul,no_inplace}.0\n Elemwise{mul,no_inplace}.0 Elemwise{mul,no_inplace}.0] to TensorType', <type 'numpy.ndarray'>)
Doh - replaced np.power with ** ! working fine!