kumar-abhishek

import torch as th                                                                 
                                                                                   
                                                                                   
class NLL_OHEM(th.nn.NLLLoss):                                                     
    """ Online hard example mining. 
    Needs input from nn.LogSotmax() """                                             
                                                                                   
    def __init__(self, ratio):      
        super(NLL_OHEM, self).__init__(None, True)                                 
        self.ratio = ratio                                                         

kumar-abhishek

kumar-abhishek

| Model           | Size   | Latency tp99 (top 99 %-ile) | Accuracy           |
|-----------------|--------|-----------------------------|--------------------|
| BERT            | 712 MB | >500 ms                     | 91.14              |
| DistilBERT      | 528 MB | 164 ms                      | 90.48              |
| DistilBERT+ONNX | 516 MB | 116 ms                      | same as DistilBERT |

kumar-abhishek

|     Label    | Num requests |  Avg  | tp90 (ms) |  tp99  |  Min  |   Max  | throughput |
|--------------|--------------|-------|-----------|--------|-------|--------|------------|
| HTTP Request |      60      | 77.25 |   96.75   | 110.67 | 68.25 | 126.75 |   12.71    |

kumar-abhishek

• Accuracy: #sr
  • Formula:
    • total number of correctly classified points/ total number of points
    • = TP+TN/(TP+FP+TN+FN)
  • When not to use accuracy:
    • in case of 90% imbalanced dataset: with a dumb model(which always returns True or returns False), you can still get 90% accuracy. This is bad
• AUC #sr
  • What does AUC actually mean? #sr
    • AUC measures how well the classifier separates the classes. Higher AUC means classifier performs better in separating the classes.
• What does high value of AUC mean? #sr

kumar-abhishek

kumar-abhishek

# Evaluate the model on the test data using `evaluate`
print('\n# Evaluate on test data')
results = model.evaluate(x_test, y_test, batch_size=128)
print('test loss, test acc:', results)

kumar-abhishek

from keras.models import Sequential
from keras.layers import GRU, Dense

n_features = 4
model_gru = Sequential()
model_gru.add(TimeDistributed(Dense(128), input_shape=(None, n_features)))
model_gru.add(GRU(64, activation='tanh', input_shape=(None, n_features), return_sequences=True))
model_gru.add(BatchNormalization())

model_gru.add(GRU(32 , activation = 'tanh'))

kumar-abhishek

from keras.layers import Dense, SimpleRNN

model_rnn = Sequential()
model_rnn.add(TimeDistributed(Dense(128), input_shape=(None, n_features)))
model_rnn.add(SimpleRNN(100, input_shape=[None, n_features], return_sequences=True ))
model_rnn.add(SimpleRNN(100))
model_rnn.add(Dense(n_features, activation='softmax'))

model_rnn.compile(optimizer='adam', loss='categorical_crossentropy')
model_rnn.fit_generator(generator, epochs=500, validation_data=validation_generator)

kumar-abhishek

n_steps = None
# convert into input/output
i=1
n=1
x_input = np.array([dataXScaler[0]])

x_input = x_input.reshape((1, len(x_input), n_features))
while i<len(dataXScaler):
  # demonstrate prediction
  print('Input: ')