AP metrics on new data in object detection - deep-learning

I'm using a dataset in COCO format in order to train a net of type "R_101_FPN_3x". The AP dataset is shown from evaluator in the validation set.
How can I have the AP evaluation on new input images? At the moment I get the predicted boxes on new images doing:
for images in os.listdir(folder_dir):
print("inference on: ", images)
im = cv2.imread(folder_dir+"/"+images)
outputs = predictor(im)
v = Visualizer(im[:, :, ::-1], MetadataCatalog.get(cfg.DATASETS.TRAIN[0]), scale=1.2)
out = v.draw_instance_predictions(outputs["instances"].to("cpu"))
cv2.imwrite("out_inference/"+"out"+images, out.get_image()[:, :, ::-1])
Or, in general, how can I evaluate quality of prediction on new data in a mathematical manner?

Related

Why does my ML model always show the same result?

I've already trained several models for a binary classification problem, basing my election on F-Score and AUC. The code used has been the following:
svm = StandardScaler()
svm.fit(feat_train)
feat_train_std = svm.transform(feat_train)
feat_test_std = svm.transform(feat_test)
model_10= BalancedBaggingClassifier(base_estimator=SVC(C=1.0, random_state=1, kernel='linear'),
sampling_strategy='auto',
replacement=False,
random_state=0)
model_10.fit(feat_train_std, target_train)
pred_target_10 = model_10.predict(feat_test)
mostrar_resultados(target_test, pred_target_10)
pred_target_10 = model_10.predict_proba(feat_test)[:, 1]
average_precision_10 = average_precision_score(target_test, pred_target_10)
precision_10, recall_10, thresholds = precision_recall_curve(target_test, pred_target_10)
auc_precision_recall_10 = auc(recall_10, precision_10)
disp_10 = plot_precision_recall_curve(model_10, feat_test, target_test)
disp_10.ax_.set_title('Binary class Precision-Recall curve: '
'AUC={0:0.2f}'.format(auc_precision_recall_10))
Afterwards, I load the model as follows:
modelo_pickle = 'modelo_pickle.pkl'
joblib.dump(model_10,modelo_pickle)
loaded_model = joblib.load(modelo_pickle)
Then, the aim is to load a new dataset, which columns are the same as the model's variables, and make a prediction for each line:
lista_x=x.to_numpy().tolist()
resultados=[]
for i in lista_x:
pred = loaded_model.predict([i])
resultados.append(pred)
print(resultados)
However, every single result is equal to 1, which does not make any sense. Would anyone tell me what am I missing, please?
Thank you in advance.
Regards,
Previously described.

Measurement-repeated ANCOVA in 2x2 Mixed Design

I am calculating in R an ANOVA with repeated measures in 2x2 mixed design. For this I use one of the following inputs in R:
(1)
res.aov <- anova_test(data = datac, dv = Stress, wid = REF,between = Gruppe, within = time )
get_anova_table(res.aov)
(2)
aov <- datac %>%
anova_test(dv = Stress, wid = REF, between = Gruppe, within = time, type = 3)
aov
Both lead to the same results. Now I want to add a covariate from the 1st measurement time point. So far I could not find a suitable R input for the ANCOVA for this repeated measures design.
Does anyone of you perhaps have an idea?
Many greetings
ANOVA <- aov(Stress~time+covariate, data = data)
summary(ANOVA)
For a simple ANCOVA, the R input of an ANOVA with the addition of the covariate applies. Unfortunately, I have no idea how this works in the repeated measures design.

Sequence to Sequence Loss

I'm trying to figure out how sequence to sequence loss is calculated. I am using the huggingface transformers library in this case, but this might actually be relevant to other DL libraries.
So to get the required data we can do:
from transformers import EncoderDecoderModel, BertTokenizer
import torch
import torch.nn.functional as F
torch.manual_seed(42)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
MAX_LEN = 128
tokenize = lambda x: tokenizer(x, max_length=MAX_LEN, truncation=True, padding=True, return_tensors="pt")
model = EncoderDecoderModel.from_encoder_decoder_pretrained('bert-base-uncased', 'bert-base-uncased') # initialize Bert2Bert from pre-trained checkpoints
input_seq = ["Hello, my dog is cute", "my cat cute"]
output_seq = ["Yes it is", "ok"]
input_tokens = tokenize(input_seq)
output_tokens = tokenize(output_seq)
outputs = model(
input_ids=input_tokens["input_ids"],
attention_mask=input_tokens["attention_mask"],
decoder_input_ids=output_tokens["input_ids"],
decoder_attention_mask=output_tokens["attention_mask"],
labels=output_tokens["input_ids"],
return_dict=True)
idx = output_tokens["input_ids"]
logits = F.log_softmax(outputs["logits"], dim=-1)
mask = output_tokens["attention_mask"]
Edit 1
Thanks to #cronoik I was able to replicate the loss calculated by huggingface as being:
output_logits = logits[:,:-1,:]
output_mask = mask[:,:-1]
label_tokens = output_tokens["input_ids"][:, 1:].unsqueeze(-1)
select_logits = torch.gather(output_logits, -1, label_tokens).squeeze()
huggingface_loss = -select_logits.mean()
However, since the last two tokens of the second input is just padding, shouldn't we calculate the loss to be:
seq_loss = (select_logits * output_mask).sum(dim=-1, keepdims=True) / output_mask.sum(dim=-1, keepdims=True)
seq_loss = -seq_loss.mean()
^This takes into account the length of the sequence of each row of outputs, and the padding by masking it out. Think this is especially useful when we have batches of varying length outputs.
ok I found out where I was making the mistakes. This is all thanks to this thread in the HuggingFace forum.
The output labels need to have -100 for the masked version. The transoformers library does not do it for you.
One silly mistake I made was with the mask. It should have been output_mask = mask[:, 1:] instead of :-1.
1. Using Model
We need to set the masks of output to -100. It is important to use clone as shown below:
labels = output_tokens["input_ids"].clone()
labels[output_tokens["attention_mask"]==0] = -100
outputs = model(
input_ids=input_tokens["input_ids"],
attention_mask=input_tokens["attention_mask"],
decoder_input_ids=output_tokens["input_ids"],
decoder_attention_mask=output_tokens["attention_mask"],
labels=labels,
return_dict=True)
2. Calculating Loss
So the final way to replicate it is as follows:
idx = output_tokens["input_ids"]
logits = F.log_softmax(outputs["logits"], dim=-1)
mask = output_tokens["attention_mask"]
# shift things
output_logits = logits[:,:-1,:]
label_tokens = idx[:, 1:].unsqueeze(-1)
output_mask = mask[:,1:]
# gather the logits and mask
select_logits = torch.gather(output_logits, -1, label_tokens).squeeze()
-select_logits[output_mask==1].mean(), outputs["loss"]
The above however ignores the fact that this comes from two different lines. So an alternate way of calculating loss could be:
seq_loss = (select_logits * output_mask).sum(dim=-1, keepdims=True) / output_mask.sum(dim=-1, keepdims=True)
seq_loss.mean()
thanks for sharing. However, the new version of transformers as of today actually does not "shift" anymore. The following is not needed.
#shift things
output_logits = logits[:,:-1,:]
label_tokens = idx[:, 1:].unsqueeze(-1)
output_mask = mask[:,1:

Meta-model of the field function OpenTurns 1.16rc1

After updating Openturns from 1.15 to 1.16rc1 I have the following issue with building the meta-model of the field function:
to reduce the computational burden:
ot.ResourceMap.SetAsUnsignedInteger("FittingTest-KolmogorovSamplingSize", 1)
algo = ot.FunctionalChaosAlgorithm(sample_X, outputSampleChaos)
algo.run()
metaModel = ot.PointToFieldConnection(postProcessing, algo.getResult().getMetaModel())
The "FittingTest-KolmogorovSamplingSize" was removed from OpenTurns 1.16rc1 and when I try to replace the fitting test with:
ot.ResourceMap.SetAsUnsignedInteger("FittingTest-LillieforsMaximumSamplingSize", 10)
Or with
ot.ResourceMap.SetAsUnsignedInteger("FittingTest-LillieforsMinimumSamplingSize", 1)
The code is freezing. Is there any solution for this?
The proposed solution is simply to use another distribution to model your data. You could have used any other multivariate continuous distribution of proper dimension. IMO it is not a valid answer as the distribution has no link to your data.
After inspection, it appears that the problem has nothing to do with Lilliefors's test. In OT 1.15 we were using this test (under the wrong name of Kolmogorov) to select automatically a distribution suited to the input sample, but we switched to a more sophisticated selection algorithm (see MetaModelAlgorithm::BuildDistribution). It is based on a first pass using the raw Kolomgorov test (thus ignoring the fact that parameters have been estimated) then an information-based criterion is used to select the most relevant model (AIC, AICC, BIC depending on the value of the "MetaModelAlgorithm-ModelSelectionCriterion" key in ResourceMap. The problem is caused by the TrapezoidalFactory class during the Kolmogorov phase. I will provide a fix ASAP in OpenTURNS master. In the mean time, I have adapted the proposed solution to something more adapted to your data:
degree = 6
dimension_xi_X = 3
dimension_xi_Y = 450
enumerateFunction = ot.HyperbolicAnisotropicEnumerateFunction(dimension_xi_X, 0.8)
basis = ot.OrthogonalProductPolynomialFactory(
[ot.StandardDistributionPolynomialFactory(ot.HistogramFactory().build(sample_X[:,i])) for i in range(dimension_xi_X)], enumerateFunction)
basisSize = enumerateFunction.getStrataCumulatedCardinal(degree)
#basis = ot.OrthogonalProductPolynomialFactory(
# [ot.HermiteFactory()] * dimension_xi_X, enumerateFunction)
#basisSize = 450#enumerateFunction.getStrataCumulatedCardinal(degree)
adaptive = ot.FixedStrategy(basis, basisSize)
projection = ot.LeastSquaresStrategy(
ot.LeastSquaresMetaModelSelectionFactory(ot.LARS(), ot.CorrectedLeaveOneOut()))
ot.ResourceMap.SetAsScalar("LeastSquaresMetaModelSelection-ErrorThreshold", 1.0e-7)
algo_chaos = ot.FunctionalChaosAlgorithm(sample_X,
outputSampleChaos,basis.getMeasure(), adaptive, projection)
algo_chaos.run()
result_chaos = algo_chaos.getResult()
meta_model = result_chaos.getMetaModel()
metaModel = ot.PointToFieldConnection(postProcessing,
algo_chaos.getResult().getMetaModel())
I also implemented a quick and dirty estimator of the L2-error:
# Meta_model validation
iMax = 5
# Input values
sample_X_validation = ot.Sample(np.array(month_1_parameters_MSE.iloc[:iMax,0:3]))
print("sample size=", sample_X_validation.getSize())
# sample_X = ot.Sample(month_1_parameters_MSE[['Rseries','Rsh','Isc']])
# output values
#month_1_simulated.iloc[0:1].transpose()
Field = ot.Field(mesh,np.array(month_1_simulated.iloc[0:1]).transpose())
sample_Y_validation = ot.ProcessSample(1,Field)
for k in range(1,iMax):
sample_Y_validation.add( np.array(month_1_simulated.iloc[k:k+1]).transpose() )
# In[18]:
graph = sample_Y_validation.drawMarginal(0)
graph.setColors(['red'])
drawables = graph.getDrawables()
graph2 = metaModel(sample_X_validation).drawMarginal(0)
graph2.setColors(['blue'])
drawables = graph2.getDrawables()
graph.add(graph2)
graph.setTitle('Model/Metamodel Validation')
graph.setXTitle(r'$t$')
graph.setYTitle(r'$z$')
drawables = graph.getDrawables()
L2_error = 0.0
for i in range(iMax):
L2_error = (drawables[i].getData()[:,1]-drawables[iMax+i].getData()[:,1]).computeRawMoment(2)[0]
print("L2_error=", L2_error)
You get an error of 79.488 with the previous answer and 1.3994 with the new proposal. Here is a graphical comparison.
Comparison between test data & previous answer
Comparison between test data & new proposal
The solution is to use:
degree = 1
dimension_xi_X = 3
dimension_xi_Y = 450
enumerateFunction = ot.LinearEnumerateFunction(dimension_xi_X)
basis = ot.OrthogonalProductPolynomialFactory(
[ot.HermiteFactory()] * dimension_xi_X, enumerateFunction)
basisSize =450 #enumerateFunction.getStrataCumulatedCardinal(degree)
adaptive = ot.FixedStrategy(basis, basisSize)
projection = ot.LeastSquaresStrategy(
ot.LeastSquaresMetaModelSelectionFactory(ot.LARS(), ot.CorrectedLeaveOneOut()))
ot.ResourceMap.SetAsScalar("LeastSquaresMetaModelSelection-ErrorThreshold", 1.0e-7)
algo_chaos = ot.FunctionalChaosAlgorithm(sample_X,
outputSampleChaos,basis.getMeasure(), adaptive, projection)
algo_chaos.run()
result_chaos = algo_chaos.getResult()
meta_model = result_chaos.getMetaModel()
metaModel1 = ot.PointToFieldConnection(postProcessing,
algo_chaos.getResult().getMetaModel())

Keras sequence-to-sequence encoder-decoder part-of-speech tagging example with attention mechanism

I have a sequence of indexed words w_1, ..., w_n. Since I'm new to deep learning, I am looking for a simple implementation of a seq2seq pos tagging model in Keras which uses attention mechanism and produces a sequence of POS tags t_1, ..., t_n out of my word sequence.
To be specific, I don't know how to gather the outputs of LSTM hidden layers of the encoder (since they are TimeDistributed) and how to feed the decoder LSTM layer for each timestamp with the outputs of time "t-1" for generating the output "t".
The model I'm thinking about is looking like the one in this paper http://arxiv.org/abs/1409.0473.
I think this issue will help you, though there is no attention mechanism.
https://github.com/fchollet/keras/issues/2654
The code included in the issue is as follows.
B = self.igor.batch_size
R = self.igor.rnn_size
S = self.igor.max_sequence_len
V = self.igor.vocab_size
E = self.igor.embedding_size
emb_W = self.igor.embeddings.astype(theano.config.floatX)
## dropout parameters
p_emb = self.igor.p_emb_dropout
p_W = self.igor.p_W_dropout
p_U = self.igor.p_U_dropout
p_dense = self.igor.p_dense_dropout
w_decay = self.igor.weight_decay
M = Sequential()
M.add(Embedding(V, E, batch_input_shape=(B,S),
W_regularizer=l2(w_decay),
weights=[emb_W], mask_zero=True, dropout=p_emb))
#for i in range(self.igor.num_lstms):
M.add(LSTM(R, return_sequences=True, dropout_W=p_W, dropout_U=p_U,
U_regularizer=l2(w_decay), W_regularizer=l2(w_decay)))
M.add(Dropout(p_dense))
M.add(LSTM(R*int(1/p_dense), return_sequences=True, dropout_W=p_W, dropout_U=p_U))
M.add(Dropout(p_dense))
M.add(TimeDistributed(Dense(V, activation='softmax',
W_regularizer=l2(w_decay), b_regularizer=l2(w_decay))))
print("compiling")
optimizer = Adam(self.igor.LR, clipnorm=self.igor.max_grad_norm,
clipvalue=5.0)
#optimizer = SGD(lr=0.01, momentum=0.5, decay=0.0, nesterov=True)
M.compile(loss='categorical_crossentropy', optimizer=optimizer,
metrics=['accuracy', 'perplexity'])
print("compiled")
self.model = M