Keras DNN outputs the same value over and over
Keras DNN outputs the same value over and over
Loading saved threads...
quasimodo · External communityPost link
External question — Data Science Stack Exchange
Author: quasimodo
Original post: https://datascience.stackexchange.com/questions/126497
License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/
Adaptation: HTML converted to plain text; contact email addresses removed.
I'm creating an ensemble of NNs with the same architecture, but each NN only outputs one value when given X_test data. The data (continuous values transformed to be [-1,1]) yields results as expected when put through other ML algos (RF and XGB), so I suspect this has something to do with my model construction. Here is the entire code, I hope it provides some clues.
import pandas as pd
import numpy as np
from scipy.stats import randint
from sklearn.tree import export_graphviz
from IPython.display import Image
import graphviz
import matplotlib.pyplot as plt
from keras.models import Sequential
from keras.layers import Dense, BatchNormalization
from keras.optimizers import Adam
from keras.callbacks import EarlyStopping
from keras.regularizers import l1
import tensorflow as tf
from keras.models import load_model
## sp500 benchmark data
sp = pd.read_csv("sp500.csv")
# sp data we need to calculate monthly sp500 returns to benchmark the OOS data
sp = sp[(sp['Date'] >= '1986-12-01') & (sp['Date'] <= '2022-01-01')]
# calculate the monthly pct change
test_sp = sp['Close'].reset_index(drop=True).pct_change().dropna()
## stock market data
data = pd.read_csv("stock_char_monthly.csv")
data['month'] = pd.to_datetime(data['month'])
data['year'] = data['month'].dt.year
# drop the bottom 20% observations by market capitalization (their snr is lowest, the stocks are illiquid, Fama: 20%:3%)
percentile_cutoff = 0.2
bottom_20 = data.groupby('month')['mktcap_lag'].transform(lambda x: x.quantile(percentile_cutoff))
data = (data[data['mktcap_lag'] > bottom_20]).reset_index(drop=True)
# target and features
target = data.ret_excess
returns = data.ret_adj
feat_data = data.filter(regex="char_")
years = data.year
train_start = 1965
test_start = 1987
train_idxs = years.loc[(years >= train_start) & (years < test_start)].index
test_idxs = years.loc[years == test_start].index
# train indices
train_beg = train_idxs[0]
train_end = train_idxs[-1]+1
# test indices
test_beg = test_idxs[0]
test_end = test_idxs[-1]+1
# split
X_train = feat_data[train_beg:train_end]
y_train = target[train_beg:train_end]
X_test = feat_data[test_beg:test_end]
y_test = target[test_beg:test_end]
# test returns
test_ret=returns[test_beg:test_end]
## NN 3 hidden layers:
# layers: 94,32,16,8,1
# activation function: ReLU
# l1 penalty: 10e-3
# learning rate: 0.001
# batch size: 10,000
# epochs: 100
# patience: 5
# adam: default
# ensemble: 5
# batch normalization between all layers
n_members = 5
models = list()
for _ in range(n_members):
# specify model architecture
model = Sequential()
model.add(Dense(94, activation='relu',kernel_regularizer=l1(0.01)))
model.add(BatchNormalization())
model.add(Dense(32, activation='relu',kernel_regularizer=l1(0.01)))
model.add(BatchNormalization())
model.add(Dense(16, activation='relu',kernel_regularizer=l1(0.01)))
model.add(BatchNormalization())
model.add(Dense(8, activation='relu',kernel_regularizer=l1(0.01)))
model.add(BatchNormalization())
model.add(Dense(1), activation='linear')
model.compile(loss='mean_squared_error', optimizer=Adam(learning_rate=0.001), metrics=['mean_squared_error'])
early_stopping = EarlyStopping(monitor='val_loss', patience=5)
#model.summary()
# fit the model
model.fit(x=X_train,
y=y_train,
batch_size = 10000,
epochs = 100,
validation_split = 0.4,
verbose = 1,
callbacks = early_stopping)
models.append(model)
yhats = [model.predict(X_test) for model in models]
yhats = np.array(yhats)
print(yhats[0])
For example, the first model only outputs 0.0098383.
[[0.0098383 ]
[0.0098383 ]
[0.0098383 ]
...
[0.0098383 ]
[0.00983831]
[0.0098383 ]]
Quote
Report
Post Reply
Quoted from Forex.com.bd-Editorial External question — Data Science Stack Exchange Author: quasimodo Source score (net votes, not local likes): 0 Original post: https://datascience.stackexchange.com/questions/126497 License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/ Adaptation: HTML converted to plain text; contact email addresses removed. I'm creating an ensemble of NNs with the same architecture, but each NN only outputs one value when given X_test data. The data (continuous values transformed to be [-1,1]) yields results as expected when put through other ML algos (RF and XGB), so I suspect this has something to do with my model construction. Here is the entire code, I hope it provides some clues. import pandas as pd import numpy as np from scipy.stats import randint from sklearn.tree import export_graphviz from IPython.display import Image import graphviz import matplotlib.pyplot as plt from keras.models import Sequential from keras.layers import Dense, BatchNormalization from keras.optimizers import Adam from keras.callbacks import EarlyStopping from keras.regularizers import l1 import tensorflow as tf from keras.models import load_model ## sp500 benchmark data sp = pd.read_csv("sp500.csv") # sp data we need to calculate monthly sp500 returns to benchmark the OOS data sp = sp[(sp['Date'] >= '1986-12-01') & (sp['Date'] <= '2022-01-01')] # calculate the monthly pct change test_sp = sp['Close'].reset_index(drop=True).pct_change().dropna() ## stock market data data = pd.read_csv("stock_char_monthly.csv") data['month'] = pd.to_datetime(data['month']) data['year'] = data['month'].dt.year # drop the bottom 20% observations by market capitalization (their snr is lowest, the stocks are illiquid, Fama: 20%:3%) percentile_cutoff = 0.2 bottom_20 = data.groupby('month')['mktcap_lag'].transform(lambda x: x.quantile(percentile_cutoff)) data = (data[data['mktcap_lag'] > bottom_20]).reset_index(drop=True) # target and features target = data.ret_excess returns = data.ret_adj feat_data = data.filter(regex="char_") years = data.year train_start = 1965 test_start = 1987 train_idxs = years.loc[(years >= train_start) & (years < test_start)].index test_idxs = years.loc[years == test_start].index # train indices train_beg = train_idxs[0] train_end = train_idxs[-1]+1 # test indices test_beg = test_idxs[0] test_end = test_idxs[-1]+1 # split X_train = feat_data[train_beg:train_end] y_train = target[train_beg:train_end] X_test = feat_data[test_beg:test_end] y_test = target[test_beg:test_end] # test returns test_ret=returns[test_beg:test_end] ## NN 3 hidden layers: # layers: 94,32,16,8,1 # activation function: ReLU # l1 penalty: 10e-3 # learning rate: 0.001 # batch size: 10,000 # epochs: 100 # patience: 5 # adam: default # ensemble: 5 # batch normalization between all layers n_members = 5 models = list() for _ in range(n_members): # specify model architecture model = Sequential() model.add(Dense(94, activation='relu',kernel_regularizer=l1(0.01))) model.add(BatchNormalization()) model.add(Dense(32, activation='relu',kernel_regularizer=l1(0.01))) model.add(BatchNormalization()) model.add(Dense(16, activation='relu',kernel_regularizer=l1(0.01))) model.add(BatchNormalization()) model.add(Dense(8, activation='relu',kernel_regularizer=l1(0.01))) model.add(BatchNormalization()) model.add(Dense(1), activation='linear') model.compile(loss='mean_squared_error', optimizer=Adam(learning_rate=0.001), metrics=['mean_squared_error']) early_stopping = EarlyStopping(monitor='val_loss', patience=5) #model.summary() # fit the model model.fit(x=X_train, y=y_train, batch_size = 10000, epochs = 100, validation_split = 0.4, verbose = 1, callbacks = early_stopping) models.append(model) yhats = [model.predict(X_test) for model in models] yhats = np.array(yhats) print(yhats[0]) For example, the first model only outputs 0.0098383. [[0.0098383 ] [0.0098383 ] [0.0098383 ] ... [0.0098383 ] [0.00983831] [0.0098383 ]]
Checking account access…