I am training a VAE (using federated learning, but that is not so important) and wanted to keep the loss and train functions simple to exchange. The initial approach was to have a tf.function as loss function and a tf.function as train function as follows:
#tf.function
def kl_reconstruction_loss(model, model_input, beta):
x, y = model_input
mean, logvar = model.encode(x, y)
z = model.reparameterize(mean, logvar)
x_logit = model.decode(z, y)
cross_ent = tf.nn.sigmoid_cross_entropy_with_logits(logits=x_logit, labels=x)
reconstruction_loss = tf.reduce_mean(tf.reduce_sum(cross_ent, axis=[1, 2, 3]), axis=0)
kl_loss = tf.reduce_mean(0.5 * tf.reduce_sum(tf.exp(logvar) + tf.square(mean) - 1. - logvar, axis=-1), axis=0)
loss = reconstruction_loss + beta * kl_loss
return loss, kl_loss, reconstruction_loss
#tf.function
def train_fn(model: tf.keras.Model, batch, optimizer, kl_beta):
"""Trains the model on a single batch.
Args:
model: The VAE model.
batch: A batch of inputs [images, labels] for the vae.
optimizer: The optimizer to train the model.
beta: Weighting of KL loss
Returns:
The loss.
"""
def vae_loss():
"""Does the forward pass and computes losses for the generator."""
# N.B. The complete pass must be inside loss() for gradient tracing.
return kl_reconstruction_loss(model, batch, kl_beta)
with tf.GradientTape() as tape:
loss, kl_loss, rc_loss = vae_loss()
grads = tape.gradient(loss, model.trainable_variables)
grads_and_vars = zip(grads, model.trainable_variables)
optimizer.apply_gradients(grads_and_vars)
return loss
For my dataset this results in an epoch duration of approx. 25 seconds. However, since I have to call those functions directly in my code, I would have to enter different ones if I would want to try out different loss/train functions.
So, alternatively, I followed https://github.com/google-research/federated/tree/master/gans and wrapped the loss function in a class and the train function in another function. Now I have:
class VaeKlReconstructionLossFns(AbstractVaeLossFns):
#tf.function
def vae_loss(self, model, model_input, labels, global_round):
# KL Reconstruction loss
mean, logvar = model.encode(model_input, labels)
z = model.reparameterize(mean, logvar)
x_logit = model.decode(z, labels)
cross_ent = tf.nn.sigmoid_cross_entropy_with_logits(logits=x_logit, labels=model_input)
reconstruction_loss = tf.reduce_mean(tf.reduce_sum(cross_ent, axis=[1, 2, 3]), axis=0)
kl_loss = tf.reduce_mean(0.5 * tf.reduce_sum(tf.exp(logvar) + tf.square(mean) - 1. - logvar, axis=-1), axis=0)
loss = reconstruction_loss + self._get_beta(global_round) * kl_loss
if model.losses:
loss += tf.add_n(model.losses)
return loss, kl_loss, reconstruction_loss
def create_train_vae_fn(
vae_loss_fns: vae_losses.AbstractVaeLossFns,
vae_optimizer: tf.keras.optimizers.Optimizer):
"""Create a function that trains VAE, binding loss and optimizer.
Args:
vae_loss_fns: Instance of gan_losses.AbstractVAELossFns interface,
specifying the VAE training loss.
vae_optimizer: Optimizer for training the VAE.
Returns:
Function that executes one step of VAE training.
"""
# We check that the optimizer has not been used previously, which ensures
# that when it is bound the train fn isn't holding onto a different copy of
# the optimizer variables then the copy that is being exchanged b/w server and
# clients.
if vae_optimizer.variables():
raise ValueError(
'Expected vae_optimizer to not have been used previously, but '
'variables were already initialized.')
#tf.function
def train_vae_fn(model: tf.keras.Model,
model_inputs,
labels,
global_round,
new_optimizer_state=None):
"""Trains the model on a single batch.
Args:
model: The VAE model.
model_inputs: A batch of inputs (usually images) for the VAE.
labels: A batch of labels corresponding to the inputs.
global_round: The current glob al FL round for beta calculation
new_optimizer_state: A possible optimizer state to overwrite the current one with.
Returns:
The number of examples trained on.
The loss.
The updated optimizer state.
"""
def vae_loss():
"""Does the forward pass and computes losses for the generator."""
# N.B. The complete pass must be inside loss() for gradient tracing.
return vae_loss_fns.vae_loss(model, model_inputs, labels, global_round)
# Set optimizer vars
optimizer_state = get_optimizer_state(vae_optimizer)
if new_optimizer_state is not None:
# if optimizer is uninitialised, initialise vars
try:
tf.nest.assert_same_structure(optimizer_state, new_optimizer_state)
except ValueError:
initialize_optimizer_vars(vae_optimizer, model)
optimizer_state = get_optimizer_state(vae_optimizer)
tf.nest.assert_same_structure(optimizer_state, new_optimizer_state)
tf.nest.map_structure(lambda a, b: a.assign(b), optimizer_state, new_optimizer_state)
with tf.GradientTape() as tape:
loss, kl_loss, rc_loss = vae_loss()
grads = tape.gradient(loss, model.trainable_variables)
grads_and_vars = zip(grads, model.trainable_variables)
vae_optimizer.apply_gradients(grads_and_vars)
return tf.shape(model_inputs)[0], loss, optimizer_state
return train_vae_fn
This new formulation takes about 86 seconds per epoch.
I am struggling to understand why the second version performs so much worse than the first one. Does anyone have a good explanation for this?
Thanks in advance!
EDIT: My Tensorflow version is 2.5.0
I'm trying to learn more about pyhf and my understanding of what the goals are might be limited. I would love to fit my HEP data outside of ROOT, but I could be imposing expectations on pyhf which are not what the authors intended for it's use.
I'd like to write myself a hello-world example, but I might just not know what I'm doing. My misunderstanding could also be gaps in my statistical knowledge.
With that preface, let me explain what I'm trying to explore.
I have some observed set of events for which I calculate some observable and make a binned histogram of that data. I hypothesize that there are two contributing physics processes, which I call signal and background. I generate some Monte Carlo samples for these processes and the theorized total number of events is close to, but not exactly what I observe.
I would like to:
Fit the data to this two process hypothesis
Get from the fit the optimal values for the number of events for each process
Get the uncertainties on these fitted values
If appropriate, calculate an upper limit on the number of signal events.
My starter code is below, where all I'm doing is an ML fit but I'm not sure where to go. I know it's not set up to do what I want, but I'm getting lost in the examples I find on RTD. I'm sure it's me, this is not a criticism of the documentation.
import pyhf
import numpy as np
import matplotlib.pyplot as plt
nbins = 15
# Generate a background and signal MC sample`
MC_signal_events = np.random.normal(5,1.0,200)
MC_background_events = 10*np.random.random(1000)
signal_data = np.histogram(MC_signal_events,bins=nbins)[0]
bkg_data = np.histogram(MC_background_events,bins=nbins)[0]
# Generate an observed dataset with a slightly different
# number of events
signal_events = np.random.normal(5,1.0,180)
background_events = 10*np.random.random(1050)
observed_events = np.array(signal_events.tolist() + background_events.tolist())
observed_sample = np.histogram(observed_events,bins=nbins)[0]
# Plot these samples, if you like
plt.figure(figsize=(12,4))
plt.subplot(1,3,1)
plt.hist(observed_events,bins=nbins,label='Observations')
plt.legend()
plt.subplot(1,3,2)
plt.hist(MC_signal_events,bins=nbins,label='MC signal')
plt.legend()
plt.subplot(1,3,3)
plt.hist(MC_background_events,bins=nbins,label='MC background')
plt.legend()
# Use a very naive estimate of the background
# uncertainties
bkg_uncerts = np.sqrt(bkg_data)
print("Defining the PDF.......")
pdf = pyhf.simplemodels.hepdata_like(signal_data=signal_data.tolist(), \
bkg_data=bkg_data.tolist(), \
bkg_uncerts=bkg_uncerts.tolist())
print("Fit.......")
data = pyhf.tensorlib.astensor(observed_sample.tolist() + pdf.config.auxdata)
bestfit_pars, twice_nll = pyhf.infer.mle.fit(data, pdf, return_fitted_val=True)
print(bestfit_pars)
print(twice_nll)
plt.show()
Note: this answer is based on pyhf v0.5.2.
Alright, so it looks like you've managed to figure most of the big pieces for sure. However, there's two different ways to do this depending on how you prefer to set things up. In both cases, I assume you want an unconstrained fit and you want to...
fit your signal+background model to observed data
fit your background model to observed data
First, let's discuss uncertainties briefly. At the moment, we default to numpy for the tensor background and scipy for the optimizer. See documentation:
numpy backend
scipy optimizer
However, one unfortunate drawback right now with the scipy optimizer is that it cannot return the uncertainties. What you need to do anywhere in your code before the fit (although we generally recommend as early as possible) is to use the minuit optimizer instead:
pyhf.set_backend('numpy', 'minuit')
This will get you the nice features of being able to get the correlation matrix, the uncertainties on the fitted parameters, and the hessian -- amongst other things. We're working to make this consistent for scipy as well, but this is not ready right now.
All optimizations go through our optimizer API which you can currently view through the mixin here in our documentation. Specifically, the signature is
minimize(
objective,
data,
pdf,
init_pars,
par_bounds,
fixed_vals=None,
return_fitted_val=False,
return_result_obj=False,
do_grad=None,
do_stitch=False,
**kwargs)
There are a lot of options here. Let's just focus on the fact that one of the keyword arguments we can pass through is return_uncertainties which will change the bestfit parameters by adding a column for the fitted parameter uncertainty which you want.
1. Signal+Background
In this case, we want to just use the default model
result, twice_nll = pyhf.infer.mle.fit(
data,
pdf,
return_uncertainties=True,
return_fitted_val=True
)
bestfit_pars, errors = result.T
2. Background-Only
In this case, we need to turn off the signal. The way we do this is by setting the parameter of interest (POI) fixed to 0.0. Then we can get the fitted parameters for the background-only model in a similar way, but using fixed_poi_fit instead of an unconstrained fit:
result, twice_nll = pyhf.infer.mle.fixed_poi_fit(
0.0,
data,
pdf,
return_uncertainties=True,
return_fitted_val=True
)
bestfit_pars, errors = result.T
Note that this is quite simply a quick way of doing the following unconstrained fit
bkg_params = pdf.config.suggested_init()
fixed_params = pdf.config.suggested_fixed()
bkg_params[pdf.config.poi_index] = 0.0
fixed_params[pdf.config.poi_index] = True
result, twice_nll = pyhf.infer.mle.fit(
data,
pdf,
init_pars=bkg_params,
fixed_params=fixed_params,
return_uncertainties=True,
return_fitted_val=True
)
bestfit_pars, errors = result.T
Hopefully that clarifies things up more!
Giordon's solution should answer all of your question, but I thought I'd also write out the code to basically address everything we can.
I also take the liberty of changing some of your values a bit so that the signal isn't so strong that the observed CLs value isn't far off to the right of the Brazil band (the results aren't wrong obviously, but it probably makes more sense to be talking about using the discovery test statistic at that point then setting limits. :))
Environment
For this example I'm going to setup a clean Python 3 virtual environment and then install the dependencies (here we're going to be using pyhf v0.5.2)
$ python3 -m venv "${HOME}/.venvs/question"
$ . "${HOME}/.venvs/question/bin/activate"
(question) $ cat requirements.txt
pyhf[minuit,contrib]~=0.5.2
black
(question) $ python -m pip install -r requirements.txt
Code
While we can't easily get the best fit value for both the number of signal events as well as the background events we definitely can do inference to get the best fit value for the signal strength.
The following chunk of code (which is long only because of the visualization) should address all of the points of your question.
# answer.py
import numpy as np
import pyhf
import matplotlib.pyplot as plt
import pyhf.contrib.viz.brazil
# Goals:
# - Fit the model to the observed data
# - Infer the best fit signal strength given the model
# - Get the uncertainties on the best fit signal strength
# - Calculate an 95% CL upper limit on the signal strength
def plot_hist(ax, bins, data, bottom=0, color=None, label=None):
bin_width = bins[1] - bins[0]
bin_leftedges = bins[:-1]
bin_centers = [edge + bin_width / 2.0 for edge in bin_leftedges]
ax.bar(
bin_centers, data, bin_width, bottom=bottom, alpha=0.5, color=color, label=label
)
def plot_data(ax, bins, data, label="Data"):
bin_width = bins[1] - bins[0]
bin_leftedges = bins[:-1]
bin_centers = [edge + bin_width / 2.0 for edge in bin_leftedges]
ax.scatter(bin_centers, data, color="black", label=label)
def invert_interval(test_mus, hypo_tests, test_size=0.05):
# This will be taken care of in v0.5.3
cls_obs = np.array([test[0] for test in hypo_tests]).flatten()
cls_exp = [
np.array([test[1][idx] for test in hypo_tests]).flatten() for idx in range(5)
]
crossing_test_stats = {"exp": [], "obs": None}
for cls_exp_sigma in cls_exp:
crossing_test_stats["exp"].append(
np.interp(
test_size, list(reversed(cls_exp_sigma)), list(reversed(test_mus))
)
)
crossing_test_stats["obs"] = np.interp(
test_size, list(reversed(cls_obs)), list(reversed(test_mus))
)
return crossing_test_stats
def main():
np.random.seed(0)
pyhf.set_backend("numpy", "minuit")
observable_range = [0.0, 10.0]
bin_width = 0.5
_bins = np.arange(observable_range[0], observable_range[1] + bin_width, bin_width)
n_bkg = 2000
n_signal = int(np.sqrt(n_bkg))
# Generate simulation
bkg_simulation = 10 * np.random.random(n_bkg)
signal_simulation = np.random.normal(5, 1.0, n_signal)
bkg_sample, _ = np.histogram(bkg_simulation, bins=_bins)
signal_sample, _ = np.histogram(signal_simulation, bins=_bins)
# Generate observations
signal_events = np.random.normal(5, 1.0, int(n_signal * 0.8))
bkg_events = 10 * np.random.random(int(n_bkg + np.sqrt(n_bkg)))
observed_events = np.array(signal_events.tolist() + bkg_events.tolist())
observed_sample, _ = np.histogram(observed_events, bins=_bins)
# Visualize the simulation and observations
fig, ax = plt.subplots()
fig.set_size_inches(7, 5)
plot_hist(ax, _bins, bkg_sample, label="Background")
plot_hist(ax, _bins, signal_sample, bottom=bkg_sample, label="Signal")
plot_data(ax, _bins, observed_sample)
ax.legend(loc="best")
ax.set_ylim(top=np.max(observed_sample) * 1.4)
ax.set_xlabel("Observable")
ax.set_ylabel("Count")
fig.savefig("components.png")
# Build the model
bkg_uncerts = np.sqrt(bkg_sample)
model = pyhf.simplemodels.hepdata_like(
signal_data=signal_sample.tolist(),
bkg_data=bkg_sample.tolist(),
bkg_uncerts=bkg_uncerts.tolist(),
)
data = pyhf.tensorlib.astensor(observed_sample.tolist() + model.config.auxdata)
# Perform inference
fit_result = pyhf.infer.mle.fit(data, model, return_uncertainties=True)
bestfit_pars, par_uncerts = fit_result.T
print(
f"best fit parameters:\
\n * signal strength: {bestfit_pars[0]} +/- {par_uncerts[0]}\
\n * nuisance parameters: {bestfit_pars[1:]}\
\n * nuisance parameter uncertainties: {par_uncerts[1:]}"
)
# Perform hypothesis test scan
_start = 0.0
_stop = 5
_step = 0.1
poi_tests = np.arange(_start, _stop + _step, _step)
print("\nPerforming hypothesis tests\n")
hypo_tests = [
pyhf.infer.hypotest(
mu_test,
data,
model,
return_expected_set=True,
return_test_statistics=True,
qtilde=True,
)
for mu_test in poi_tests
]
# Upper limits on signal strength
results = invert_interval(poi_tests, hypo_tests)
print(f"Observed Limit on µ: {results['obs']:.2f}")
print("-----")
for idx, n_sigma in enumerate(np.arange(-2, 3)):
print(
"Expected {}Limit on µ: {:.3f}".format(
" " if n_sigma == 0 else "({} σ) ".format(n_sigma),
results["exp"][idx],
)
)
# Visualize the "Brazil band"
fig, ax = plt.subplots()
fig.set_size_inches(7, 5)
ax.set_title("Hypothesis Tests")
ax.set_ylabel(r"$\mathrm{CL}_{s}$")
ax.set_xlabel(r"$\mu$")
pyhf.contrib.viz.brazil.plot_results(ax, poi_tests, hypo_tests)
fig.savefig("brazil_band.png")
if __name__ == "__main__":
main()
which when run gives
(question) $ python answer.py
best fit parameters:
* signal strength: 1.5884737977889158 +/- 0.7803435235862329
* nuisance parameters: [0.99020988 1.06040191 0.90488207 1.03531383 1.09093327 1.00942088
1.07789316 1.01125627 1.06202964 0.95780043 0.94990993 1.04893286
1.0560711 0.9758487 0.93692481 1.04683181 1.05785515 0.92381263
0.93812855 0.96751869]
* nuisance parameter uncertainties: [0.06966439 0.07632218 0.0611428 0.07230328 0.07872258 0.06899675
0.07472849 0.07403246 0.07613661 0.08606657 0.08002775 0.08655314
0.07564512 0.07308117 0.06743479 0.07383134 0.07460864 0.06632003
0.06683251 0.06270965]
Performing hypothesis tests
/home/stackoverflow/.venvs/question/lib/python3.7/site-packages/pyhf/infer/calculators.py:229: RuntimeWarning: invalid value encountered in double_scalars
teststat = (qmu - qmu_A) / (2 * self.sqrtqmuA_v)
Observed Limit on µ: 2.89
-----
Expected (-2 σ) Limit on µ: 0.829
Expected (-1 σ) Limit on µ: 1.110
Expected Limit on µ: 1.542
Expected (1 σ) Limit on µ: 2.147
Expected (2 σ) Limit on µ: 2.882
Let us know if you have any further questions!
When I run the following code
from transformers import AutoTokenizer, AutoModelForQuestionAnswering
import torch
tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
model = AutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
text = r"""
As checked Dis is not yet on boarded to ARB portal, hence we cannot upload the invoices in portal
"""
questions = [
"Dis asked if it is possible to post the two invoice in ARB.I have not access so I wanted to check if you would be able to do it.",
]
for question in questions:
inputs = tokenizer.encode_plus(question, text, add_special_tokens=True, return_tensors="pt")
input_ids = inputs["input_ids"].tolist()[0]
text_tokens = tokenizer.convert_ids_to_tokens(input_ids)
answer_start_scores, answer_end_scores = model(**inputs)
answer_start = torch.argmax(
answer_start_scores
) # Get the most likely beginning of answer with the argmax of the score
answer_end = torch.argmax(answer_end_scores) + 1 # Get the most likely end of answer with the argmax of the score
answer = tokenizer.convert_tokens_to_string(tokenizer.convert_ids_to_tokens(input_ids[answer_start:answer_end]))
print(f"Question: {question}")
print(f"Answer: {answer}\n")
The answer that I get here is:
Question: Dis asked if it is possible to post the two invoice in ARB.I have not access so I wanted to check if you would be able to do it.
Answer: dis is not yet on boarded to ARB portal
How do I get a score for this answer? Score here is very similar to what is I get when I run Question-Answer pipeline .
I have to take this approach since Question-Answer pipeline when used is giving me Key Error for the below code
from transformers import pipeline
nlp = pipeline("question-answering")
context = r"""
As checked Dis is not yet on boarded to ARB portal, hence we cannot upload the invoices in portal.
"""
print(nlp(question="Dis asked if it is possible to post the two invoice in ARB?", context=context))
This is my attempt to get the score. It appears that I cannot figure out what feature.p_mask. So I could not remove the non-context indexes that contribute to the softmax at the moment.
# ... assuming imports and question and context
model_name="deepset/roberta-base-squad2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForQuestionAnswering.from_pretrained(model_name)
inputs = tokenizer(question, context,
add_special_tokens=True,
return_tensors='pt')
input_ids = inputs['input_ids'].tolist()[0]
outputs = model(**inputs)
# used to compute score
start = outputs.start_logits.detach().numpy()
end = outputs.end_logits.detach().numpy()
# from source code
# Ensure padded tokens & question tokens cannot belong to the set of candidate answers.
#?? undesired_tokens = np.abs(np.array(feature.p_mask) - 1) & feature.attention_mask
# Generate mask
undesired_tokens = inputs['attention_mask']
undesired_tokens_mask = undesired_tokens == 0.0
# Make sure non-context indexes in the tensor cannot contribute to the softmax
start_ = np.where(undesired_tokens_mask, -10000.0, start)
end_ = np.where(undesired_tokens_mask, -10000.0, end)
# Normalize logits and spans to retrieve the answer
start_ = np.exp(start_ - np.log(np.sum(np.exp(start_), axis=-1, keepdims=True)))
end_ = np.exp(end_ - np.log(np.sum(np.exp(end_), axis=-1, keepdims=True)))
# Compute the score of each tuple(start, end) to be the real answer
outer = np.matmul(np.expand_dims(start_, -1), np.expand_dims(end_, 1))
# Remove candidate with end < start and end - start > max_answer_len
max_answer_len = 15
candidates = np.tril(np.triu(outer), max_answer_len - 1)
scores_flat = candidates.flatten()
idx_sort = [np.argmax(scores_flat)]
start, end = np.unravel_index(idx_sort, candidates.shape)[1:]
end += 1
score = candidates[0, start, end-1]
start, end, score = start.item(), end.item(), score.item()
print(tokenizer.decode(input_ids[start:end]))
print(score)
See more source code
I am reviewing huggingface's version of Albert.
However, I cannot find any code or comment about SOP.
I can find NSP(Next Sentence Prediction) implementation from modeling_from src/transformers/modeling_bert.py.
if masked_lm_labels is not None and next_sentence_label is not None:
loss_fct = CrossEntropyLoss()
masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
total_loss = masked_lm_loss + next_sentence_loss
outputs = (total_loss,) + outputs
Is SOP inherited from here with SOP-style labeling? or Is there anything I am missing?
The sentence order loss is here:
sentence_order_loss = loss_fn(y_true=sentence_order_label, y_pred=sentence_order_reduced_logits)
It's just a cross entropy loss.
I am currently runnuing training in matlab on a matrix of logspecrum samples I am constantly dealing with underflow problems.I understood that I need to work with log's in order to deal with underflowing.
I am still strugling with uderflow though , when i calculate the mean (mue) bucause it is negetive i cant work with logs so i need the real values that underflow.
These are equasions i am working with:
In MATLAB code i calulate log_tau in oreder avoid underflow but when calulating mue i need exp(log(tau)) which goes to zero.
I am attaching relevent MATLAB code
**in the code i called the variable alpha is tau ...
for i = 1 : 50
log_c = Logsum(log_alpha,1) - log(N);
c = exp(log_c);
mue = DataMat*alpha./(repmat(exp(Logsum(log_alpha,1)),FrameSize,1));
log_abs_mue = log(abs(mue));
log_SigmaSqr = log((DataMat.^2)*alpha) - repmat(Logsum(log_alpha,1),FrameSize,1) - 2*log_abs_mue;
SigmaSqr = exp(log_SigmaSqr);
for j=1:N
rep_DataMat(:,:,j) = repmat(DataMat(:,j),1,M);
log_gamma(j,:) = log_c - 0.5*(FrameSize*log(2*pi)+sum(log_SigmaSqr)) + sum((rep_DataMat(:,:,j) - mue).^2./(2*SigmaSqr));
end
log_alpha = log_gamma - repmat(Logsum(log_gamma,2),1,M);
alpha = exp(log_alpha);
end
c = exp(log_c);
SigmaSqr = exp(log_SigmaSqr);
does any one see how i can avoid this? or what needs to be fixed in code?
What i did was add this line to the MATLAB code:
mue(isnan(mue))=0; %fix 0/0 problem
and this one:
SigmaSqr(SigmaSqr==0)=1;%fix if mue_k = x_k
not sure if this is the best solution but is seems to work...
any have a better idea?