In [1]:
import numpy as np
from keras.models import Model
from keras.layers import Input
from keras.layers.recurrent import SimpleRNN
from keras import backend as K
import json
from collections import OrderedDict
In [2]:
def format_decimal(arr, places=6):
return [round(x * 10**places) / 10**places for x in arr]
In [3]:
DATA = OrderedDict()
[recurrent.SimpleRNN.0] units=4, activation='tanh'
Note dropout_W and dropout_U are only applied during training phase
In [4]:
data_in_shape = (3, 6)
rnn = SimpleRNN(4, activation='tanh')
layer_0 = Input(shape=data_in_shape)
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3400 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U', 'b']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.0'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
[recurrent.SimpleRNN.1] units=5, activation='sigmoid'
Note dropout_W and dropout_U are only applied during training phase
In [5]:
data_in_shape = (8, 5)
rnn = SimpleRNN(5, activation='sigmoid')
layer_0 = Input(shape=data_in_shape)
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3500 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U', 'b']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.1'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
[recurrent.SimpleRNN.2] units=4, activation='tanh', return_sequences=True
Note dropout_W and dropout_U are only applied during training phase
In [6]:
data_in_shape = (7, 6)
rnn = SimpleRNN(4, activation='tanh', return_sequences=True)
layer_0 = Input(shape=data_in_shape)
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3600 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U', 'b']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.2'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
[recurrent.SimpleRNN.3] units=4, activation='tanh', return_sequences=False, go_backwards=True
Note dropout_W and dropout_U are only applied during training phase
In [7]:
data_in_shape = (7, 6)
rnn = SimpleRNN(4, activation='tanh', return_sequences=False, go_backwards=True)
layer_0 = Input(shape=data_in_shape)
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3700 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U', 'b']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.3'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
[recurrent.SimpleRNN.4] units=4, activation='tanh', return_sequences=True, go_backwards=True
Note dropout_W and dropout_U are only applied during training phase
In [8]:
data_in_shape = (7, 6)
rnn = SimpleRNN(4, activation='tanh', return_sequences=True, go_backwards=True)
layer_0 = Input(shape=data_in_shape)
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3800 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U', 'b']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.4'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
[recurrent.SimpleRNN.5] units=4, activation='tanh', return_sequences=False, go_backwards=False, stateful=True
Note dropout_W and dropout_U are only applied during training phase
To test statefulness, model.predict is run twice
In [9]:
data_in_shape = (7, 6)
rnn = SimpleRNN(4, activation='tanh', return_sequences=False, go_backwards=False, stateful=True)
layer_0 = Input(batch_shape=(1, *data_in_shape))
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3800 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U', 'b']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.5'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
[recurrent.SimpleRNN.6] units=4, activation='tanh', return_sequences=True, go_backwards=False, stateful=True
Note dropout_W and dropout_U are only applied during training phase
To test statefulness, model.predict is run twice
In [10]:
data_in_shape = (7, 6)
rnn = SimpleRNN(4, activation='tanh', return_sequences=True, go_backwards=False, stateful=True)
layer_0 = Input(batch_shape=(1, *data_in_shape))
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3810 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U', 'b']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.6'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
[recurrent.SimpleRNN.7] units=4, activation='tanh', return_sequences=False, go_backwards=True, stateful=True
Note dropout_W and dropout_U are only applied during training phase
To test statefulness, model.predict is run twice
In [11]:
data_in_shape = (7, 6)
rnn = SimpleRNN(4, activation='tanh', return_sequences=False, go_backwards=True, stateful=True)
layer_0 = Input(batch_shape=(1, *data_in_shape))
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3820 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U', 'b']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.7'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
[recurrent.SimpleRNN.8] units=4, activation='tanh', use_bias=False, return_sequences=True, go_backwards=True, stateful=True
Note dropout_W and dropout_U are only applied during training phase
To test statefulness, model.predict is run twice
In [12]:
data_in_shape = (7, 6)
rnn = SimpleRNN(4, activation='tanh', use_bias=False, return_sequences=True, go_backwards=True, stateful=True)
layer_0 = Input(batch_shape=(1, *data_in_shape))
layer_1 = rnn(layer_0)
model = Model(inputs=layer_0, outputs=layer_1)
# set weights to random (use seed for reproducibility)
weights = []
for i, w in enumerate(model.get_weights()):
np.random.seed(3830 + i)
weights.append(2 * np.random.random(w.shape) - 1)
model.set_weights(weights)
weight_names = ['W', 'U']
for w_i, w_name in enumerate(weight_names):
print('{} shape:'.format(w_name), weights[w_i].shape)
print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))
data_in = 2 * np.random.random(data_in_shape) - 1
result = model.predict(np.array([data_in]))
result = model.predict(np.array([data_in]))
data_out_shape = result[0].shape
data_in_formatted = format_decimal(data_in.ravel().tolist())
data_out_formatted = format_decimal(result[0].ravel().tolist())
print('')
print('in shape:', data_in_shape)
print('in:', data_in_formatted)
print('out shape:', data_out_shape)
print('out:', data_out_formatted)
DATA['recurrent.SimpleRNN.8'] = {
'input': {'data': data_in_formatted, 'shape': data_in_shape},
'weights': [{'data': format_decimal(w.ravel().tolist()), 'shape': w.shape} for w in weights],
'expected': {'data': data_out_formatted, 'shape': data_out_shape}
}
In [13]:
print(json.dumps(DATA))
In [ ]: