I have trained an RNN model with pytorch. I need to use the model for prediction in an environment where I'm unable to install pytorch because of some strange dependency issue with glibc. However, I can install numpy and scipy and other libraries. So, I want to use the trained model, with the network definition, without pytorch.
I have the weights of the model as I save the model with its state dict and weights in the standard way, but I can also save it using just json/pickle files or similar.
I also have the network definition, which depends on pytorch in a number of ways. This is my RNN network definition.
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import random
torch.manual_seed(1)
random.seed(1)
device = torch.device('cpu')
class RNN(nn.Module):
def __init__(self, input_size, hidden_size, output_size,num_layers, matching_in_out=False, batch_size=1):
super(RNN, self).__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
self.num_layers = num_layers
self.batch_size = batch_size
self.matching_in_out = matching_in_out #length of input vector matches the length of output vector
self.lstm = nn.LSTM(input_size, hidden_size,num_layers)
self.hidden2out = nn.Linear(hidden_size, output_size)
self.hidden = self.init_hidden()
def forward(self, feature_list):
feature_list=torch.tensor(feature_list)
if self.matching_in_out:
lstm_out, _ = self.lstm( feature_list.view(len( feature_list), 1, -1))
output_space = self.hidden2out(lstm_out.view(len( feature_list), -1))
output_scores = torch.sigmoid(output_space) #we'll need to check if we need this sigmoid
return output_scores #output_scores
else:
for i in range(len(feature_list)):
cur_ft_tensor=feature_list[i]#.view([1,1,self.input_size])
cur_ft_tensor=cur_ft_tensor.view([1,1,self.input_size])
lstm_out, self.hidden = self.lstm(cur_ft_tensor, self.hidden)
outs=self.hidden2out(lstm_out)
return outs
def init_hidden(self):
#return torch.rand(self.num_layers, self.batch_size, self.hidden_size)
return (torch.rand(self.num_layers, self.batch_size, self.hidden_size).to(device),
torch.rand(self.num_layers, self.batch_size, self.hidden_size).to(device))
I am aware of this question, but I'm willing to go as low level as possible. I can work with numpy array instead of tensors, and reshape instead of view, and I don't need a device setting.
Based on the class definition above, what I can see here is that I only need the following components from torch to get an output from the forward function:
nn.LSTM
nn.Linear
torch.sigmoid
I think I can easily implement the sigmoid function using numpy. However, can I have some implementation for the nn.LSTM and nn.Linear using something not involving pytorch? Also, how will I use the weights from the state dict into the new class?
So, the question is, how can I "translate" this RNN definition into a class that doesn't need pytorch, and how to use the state dict weights for it?
Alternatively, is there a "light" version of pytorch, that I can use just to run the model and yield a result?
EDIT
I think it might be useful to include the numpy/scipy equivalent for both nn.LSTM and nn.linear. It would help us compare the numpy output to torch output for the same code, and give us some modular code/functions to use. Specifically, a numpy equivalent for the following would be great:
rnn = nn.LSTM(10, 20, 2)
input = torch.randn(5, 3, 10)
h0 = torch.randn(2, 3, 20)
c0 = torch.randn(2, 3, 20)
output, (hn, cn) = rnn(input, (h0, c0))
and also for linear:
m = nn.Linear(20, 30)
input = torch.randn(128, 20)
output = m(input)
You should try to export the model using torch.onnx. The page gives you an example that you can start with.
An alternative is to use TorchScript, but that requires torch libraries.
Both of these can be run without python. You can load torchscript in a C++ application https://pytorch.org/tutorials/advanced/cpp_export.html
ONNX is much more portable and you can use in languages such as C#, Java, or Javascript
https://onnxruntime.ai/ (even on the browser)
A running example
Just modifying a little your example to go over the errors I found
Notice that via tracing any if/elif/else, for, while will be unrolled
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import random
torch.manual_seed(1)
random.seed(1)
device = torch.device('cpu')
class RNN(nn.Module):
def __init__(self, input_size, hidden_size, output_size,num_layers, matching_in_out=False, batch_size=1):
super(RNN, self).__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
self.num_layers = num_layers
self.batch_size = batch_size
self.matching_in_out = matching_in_out #length of input vector matches the length of output vector
self.lstm = nn.LSTM(input_size, hidden_size,num_layers)
self.hidden2out = nn.Linear(hidden_size, output_size)
def forward(self, x, h0, c0):
lstm_out, (hidden_a, hidden_b) = self.lstm(x, (h0, c0))
outs=self.hidden2out(lstm_out)
return outs, (hidden_a, hidden_b)
def init_hidden(self):
#return torch.rand(self.num_layers, self.batch_size, self.hidden_size)
return (torch.rand(self.num_layers, self.batch_size, self.hidden_size).to(device).detach(),
torch.rand(self.num_layers, self.batch_size, self.hidden_size).to(device).detach())
# convert the arguments passed during onnx.export call
class MWrapper(nn.Module):
def __init__(self, model):
super(MWrapper, self).__init__()
self.model = model;
def forward(self, kwargs):
return self.model(**kwargs)
Run an example
rnn = RNN(10, 10, 10, 3)
X = torch.randn(3,1,10)
h0,c0 = rnn.init_hidden()
print(rnn(X, h0, c0)[0])
Use the same input to trace the model and export an onnx file
torch.onnx.export(MWrapper(rnn), {'x':X,'h0':h0,'c0':c0}, 'rnn.onnx',
dynamic_axes={'x':{1:'N'},
'c0':{1: 'N'},
'h0':{1: 'N'}
},
input_names=['x', 'h0', 'c0'],
output_names=['y', 'hn', 'cn']
)
Notice that you can use symbolic values for the dimensions of some axes of some inputs. Unspecified dimensions will be fixed with the values from the traced inputs. By default LSTM uses dimension 1 as batch.
Next we load the ONNX model and pass the same inputs
import onnxruntime
ort_model = onnxruntime.InferenceSession('rnn.onnx')
print(ort_model.run(['y'], {'x':X.numpy(), 'c0':c0.numpy(), 'h0':h0.numpy()}))
Basically implementing it in numpy and copying weights from your pytorch model can do the trick. For your usecase you will only need to do a forward pass so we just need to implement that only
#Set Parameters for a small LSTM network
input_size = 2 # size of one 'event', or sample, in our batch of data
hidden_dim = 3 # 3 cells in the LSTM layer
output_size = 1 # desired model output
num_layers=3
torch_lstm = RNN( input_size,
hidden_dim ,
output_size,
num_layers,
matching_in_out=True
)
state = torch_lstm.state_dict() # state will capture the weights of your model
Now for LSTM in numpy these functions will be used:
got the below code from this link: https://towardsdatascience.com/the-lstm-reference-card-6163ca98ae87
### NOT MY CODE
import numpy as np
from scipy.special import expit as sigmoid
def forget_gate(x, h, Weights_hf, Bias_hf, Weights_xf, Bias_xf, prev_cell_state):
forget_hidden = np.dot(Weights_hf, h) + Bias_hf
forget_eventx = np.dot(Weights_xf, x) + Bias_xf
return np.multiply( sigmoid(forget_hidden + forget_eventx), prev_cell_state )
def input_gate(x, h, Weights_hi, Bias_hi, Weights_xi, Bias_xi, Weights_hl, Bias_hl, Weights_xl, Bias_xl):
ignore_hidden = np.dot(Weights_hi, h) + Bias_hi
ignore_eventx = np.dot(Weights_xi, x) + Bias_xi
learn_hidden = np.dot(Weights_hl, h) + Bias_hl
learn_eventx = np.dot(Weights_xl, x) + Bias_xl
return np.multiply( sigmoid(ignore_eventx + ignore_hidden), np.tanh(learn_eventx + learn_hidden) )
def cell_state(forget_gate_output, input_gate_output):
return forget_gate_output + input_gate_output
def output_gate(x, h, Weights_ho, Bias_ho, Weights_xo, Bias_xo, cell_state):
out_hidden = np.dot(Weights_ho, h) + Bias_ho
out_eventx = np.dot(Weights_xo, x) + Bias_xo
return np.multiply( sigmoid(out_eventx + out_hidden), np.tanh(cell_state) )
We would need the sigmoid function as well so
def sigmoid(x):
return 1/(1 + np.exp(-x))
Because pytorch stores weights in stacked manner so we need to break it up for that we would need the below function
def get_slices(hidden_dim):
slices=[]
breaker=(hidden_dim*4)
slices=[[i,i+3] for i in range(0, breaker, breaker//4)]
return slices
Now we have the functions ready for lstm, now we create an lstm class to copy the weights from pytorch class and get the output from it.
class numpy_lstm:
def __init__( self, layer_num=0, hidden_dim=1, matching_in_out=False):
self.matching_in_out=matching_in_out
self.layer_num=layer_num
self.hidden_dim=hidden_dim
def init_weights_from_pytorch(self, state):
slices=get_slices(self.hidden_dim)
print (slices)
#Event (x) Weights and Biases for all gates
lstm_weight_ih='lstm.weight_ih_l'+str(self.layer_num)
self.Weights_xi = state[lstm_weight_ih][slices[0][0]:slices[0][1]].numpy() # shape [h, x]
self.Weights_xf = state[lstm_weight_ih][slices[1][0]:slices[1][1]].numpy() # shape [h, x]
self.Weights_xl = state[lstm_weight_ih][slices[2][0]:slices[2][1]].numpy() # shape [h, x]
self.Weights_xo = state[lstm_weight_ih][slices[3][0]:slices[3][1]].numpy() # shape [h, x]
lstm_bias_ih='lstm.bias_ih_l'+str(self.layer_num)
self.Bias_xi = state[lstm_bias_ih][slices[0][0]:slices[0][1]].numpy() #shape is [h, 1]
self.Bias_xf = state[lstm_bias_ih][slices[1][0]:slices[1][1]].numpy() #shape is [h, 1]
self.Bias_xl = state[lstm_bias_ih][slices[2][0]:slices[2][1]].numpy() #shape is [h, 1]
self.Bias_xo = state[lstm_bias_ih][slices[3][0]:slices[3][1]].numpy() #shape is [h, 1]
lstm_weight_hh='lstm.weight_hh_l'+str(self.layer_num)
#Hidden state (h) Weights and Biases for all gates
self.Weights_hi = state[lstm_weight_hh][slices[0][0]:slices[0][1]].numpy() #shape is [h, h]
self.Weights_hf = state[lstm_weight_hh][slices[1][0]:slices[1][1]].numpy() #shape is [h, h]
self.Weights_hl = state[lstm_weight_hh][slices[2][0]:slices[2][1]].numpy() #shape is [h, h]
self.Weights_ho = state[lstm_weight_hh][slices[3][0]:slices[3][1]].numpy() #shape is [h, h]
lstm_bias_hh='lstm.bias_hh_l'+str(self.layer_num)
self.Bias_hi = state[lstm_bias_hh][slices[0][0]:slices[0][1]].numpy() #shape is [h, 1]
self.Bias_hf = state[lstm_bias_hh][slices[1][0]:slices[1][1]].numpy() #shape is [h, 1]
self.Bias_hl = state[lstm_bias_hh][slices[2][0]:slices[2][1]].numpy() #shape is [h, 1]
self.Bias_ho = state[lstm_bias_hh][slices[3][0]:slices[3][1]].numpy() #shape is [h, 1]
def forward_lstm_pass(self,input_data):
h = np.zeros(self.hidden_dim)
c = np.zeros(self.hidden_dim)
output_list=[]
for eventx in input_data:
f = forget_gate(eventx, h, self.Weights_hf, self.Bias_hf, self.Weights_xf, self.Bias_xf, c)
i = input_gate(eventx, h, self.Weights_hi, self.Bias_hi, self.Weights_xi, self.Bias_xi,
self.Weights_hl, self.Bias_hl, self.Weights_xl, self.Bias_xl)
c = cell_state(f,i)
h = output_gate(eventx, h, self.Weights_ho, self.Bias_ho, self.Weights_xo, self.Bias_xo, c)
if self.matching_in_out: # doesnt make sense but it was as it was in main code :(
output_list.append(h)
if self.matching_in_out:
return output_list
else:
return h
Similarly for fully connected layer,
class fully_connected_layer:
def __init__(self,state, dict_name='fc', ):
self.fc_Weight = state[dict_name+'.weight'][0].numpy()
self.fc_Bias = state[dict_name+'.bias'][0].numpy() #shape is [,output_size]
def forward(self,lstm_output, is_sigmoid=True):
res=np.dot(self.fc_Weight, lstm_output)+self.fc_Bias
print (res)
if is_sigmoid:
return sigmoid(res)
else:
return res
Now we would need one class to call all of them together and generalise them with respect to multiple layers
You can modify the below class if you need more Fully connected layers or want to set false condition for sigmoid etc.
class RNN_model_Numpy:
def __init__(self, state, input_size, hidden_dim, output_size, num_layers, matching_in_out=True):
self.lstm_layers=[]
for i in range(0, num_layers):
lstm_layer_obj=numpy_lstm(layer_num=i, hidden_dim=hidden_dim, matching_in_out=True)
lstm_layer_obj.init_weights_from_pytorch(state)
self.lstm_layers.append(lstm_layer_obj)
self.hidden2out=fully_connected_layer(state, dict_name='hidden2out')
def forward(self, feature_list):
for x in self.lstm_layers:
lstm_output=x.forward_lstm_pass(feature_list)
feature_list=lstm_output
return self.hidden2out.forward(feature_list, is_sigmoid=False)
Sanity check on a numpy variable:
data = np.array(
[[1,1],
[2,2],
[3,3]])
check=RNN_model_Numpy(state, input_size, hidden_dim, output_size, num_layers)
check.forward(data)
EXPLANATION:
Since we just need forward pass, we would need certain functions that are required in LSTM, for that we have the forget gate, input gate, cell gate and output gate. They are just some operations that are done on the input that you give.
For get_slices function, this is used to break down the weight matrix that we get from pytorch state dictionary (state dictionary) is the dictionary which contains the weights of all the layers that we have in our network.
For LSTM particularly have it in this order ignore, forget, learn, output. So for that we would need to break it up for different LSTM cells.
For numpy_lstm class, we have init_weights_from_pytorch function which must be called, what it will do is that it will extract the weights from state dictionary which we got earlier from pytorch model object and then populate the numpy array weights with the pytorch weights. You can first train your model and then save the state dictionary through pickle and then use it.
The fully connected layer class just implements the hidden2out neural network.
Finally our rnn_model_numpy class is there to ensure that if you have multiple layers then it is able to send the output of one layer of lstm to other layer of lstm.
Lastly there is a small sanity check on data variable.
IMPORTANT NOTE: PLEASE NOTE THAT YOU MIGHT GET DIMENSION ERROR AS PYTORCH WAY OF HANDLING INPUT IS COMPLETELY DIFFERENT SO PLEASE ENSURE THAT YOU INPUT NUMPY IS OF SIMILAR SHAPE AS DATA VARIABLE.
Important references:
https://pytorch.org/docs/stable/generated/torch.nn.LSTM.html
https://christinakouridi.blog/2019/06/19/backpropagation-lstm/
Related
I am more or less new to the field of neural networks and python, just a couple of months of work.
I am interested in this case developed in matlab https://it.mathworks.com/help/images/image-processing-operator-approximation-using-deep-learning.html
However, I would like to try to implement this using Keras.
I have three questions regarding the two custom layers this net uses, whose codes are found here:
https://github.com/catsymptote/Salsa_cryptanalysis/blob/master/matlab/workspace/adaptiveNormalizationMu.m
https://github.com/catsymptote/Salsa_cryptanalysis/blob/master/matlab/workspace/adaptiveNormalizationLambda.m
I have not really/deeply understood what these layers actually do
Is my temptative implementation of adaptiveNormalizationMu correct on Keras? Based on what I
understood, this layer just multiplies the output of the BN layer for an adaptive scale
parameter, mu. I wrote the code following the example reported here
https://www.tutorialspoint.com/keras/keras_customized_layer.htm
I am struggling with the variables input_shape and output_shape of the code I wrote following the tutorial.
Considering batch size BS, images with dimensions dim1 and dim2, 1 channel, I would love the input to have dimension (BS, dim1, dim2, 1), and output to have the same, since it is a mere scaling. How to be coherent with the code written in matlab in the mathworks example, where the only input argument is numberOfFilters? I don't know where to introduce this parameter in the code I am trying to write. I would love not to fix the input dimension, so that I can re-use this layer at different depths of the network, but correctly choose the "depht" (like the number of filters for a standard conv2D layer)
Thank you so much for the help
F.
###
from keras import backend as K
from keras.layers import Layer
class MyAdaptiveNormalizationMu(Layer):
def __init__(self, output_dim, **kwargs):
self.output_dim = output_dim
super(MyAdaptiveNormalizationMu, self).__init__(**kwargs)
def build(self, input_shape):
self.mu = self.add_weight(name = 'mu',
shape = (input_shape[1], self.output_dim),
initializer = 'random_normal', trainable = True)
super(MyAdaptiveNormalizationMu, self).build(input_shape)
def call(self, input_data):
return input_data * self.mu
def compute_output_shape(self, input_shape): return (input_shape[0], self.output_dim)
from keras.models import Sequential
batch_size = 16
dim1 = 8
dim2 = 8
channels = 1
input_shape = (batch_size, dim1, dim2, channels)
output_shape = input_shape
model = Sequential()
model.add(MyAdaptiveNormalizationMu(output_dim=?, input_shape=?))
EDIT: I provide a second realization attempt, which seems to compile. It should do what I think adaptiveNormalizationLambda and adaptiveNormalizationMu do: multiply the input for a learnable weight matrix. However, i am still unsure if the layer is doing what it is supposed to, and if I got correctly the sense of those layers.
from keras.layers import Layer, Input
from keras.models import Model
import numpy as np
class Multiply_Weights(Layer):
def __init__(self, **kwargs):
super(Multiply_Weights, self).__init__(**kwargs)
def build(self, input_shape):
# Create a trainable weight variable for this layer.
self.kernel = self.add_weight(name='kernel',
shape=(input_shape[1], input_shape[2]),
initializer='RandomNormal',
trainable=True)
super(Multiply_Weights, self).build(input_shape)
def call(self, x, **kwargs):
# Implicit broadcasting occurs here.
# Shape x: (BATCH_SIZE, N, M)
# Shape kernel: (N, M)
# Shape output: (BATCH_SIZE, N, M)
return x * self.kernel
def compute_output_shape(self, input_shape):
return input_shape
N = 3
M = 4
BATCH_SIZE = 1
a = Input(shape=(N, M))
layer = Multiply_Weights()(a)
model = Model(inputs=a,
outputs=layer)
a = np.ones(shape=(BATCH_SIZE, N, M))
pred = model.predict(a)
print(pred)
I am new to TensorFlow and I am trying to make a custom RNN cell that behaves like a stacked LSTM cell, with residual connection between LSTM layers. Each LSTM layer also has drop-out implemented. So far I've tried to subclass keras.layers.AbstractRNNCell, but during training, the drop-out mask of the cell is not reset after every batches when I iterate the cell with keras.layers.RNN. I've read the source code of keras.layers.RNN and I found out that it only resets the drop-out mask for cells that are instance of DropoutRNNCellMixin via method _maybe_reset_cell_dropout_mask.
My question is: How do I reset the drop-out mask of my custom cell after every calls of keras.layers.RNN?
The following is the code that I wrote and its behavior. I'm using Tensorflow version 2.5.0:
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
class StackedLSTMCell(layers.AbstractRNNCell):
def __init__(self, params, **kwargs):
super(StackedLSTMCell, self).__init__(**kwargs)
self.params = params
self.depth = params['depth']
self.units = [layers.LSTMCell(params['hidden_dim'],
dropout=params['dropout_rate_input'],
recurrent_dropout=params['dropout_rate_output'],
unit_forget_bias=True,
name=f'LSTM_{i}')
for i in range(params['depth'])]
#property
def state_size(self):
return tuple(self.units[i].state_size for i in range(self.depth))
#property
def output_size(self):
return self.units[-1].output_size
def get_config(self):
return {'params': self.params}
def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
return tuple(self.units[i].get_initial_state(inputs, batch_size, dtype) \
for i in range(self.depth))
def call(self, token, inp_state):
input_sum = token
out_state = []
for i in range(self.depth):
h, h_and_c = self.units[i](token, inp_state[i])
if i+1<self.depth:
token = layers.Add()([input_sum, h])
input_sum = layers.Add()([input_sum, token])
out_state.append(h_and_c)
return h, tuple(out_state)
params = {
'hidden_dim':128,
'dropout_rate_input':0.2,
'dropout_rate_output':0.2,
'depth':3}
cell = StackedLSTMCell(params)
RNN = layers.RNN(cell, return_sequences=True)
inputs = tf.ones((1, 1, 128)) #input shape format is (batch_size, time_steps, features)
print(set([RNN(inputs, training=True).numpy()[0, 0, 0] for _ in range(5)]))
#return: {-0.19621503}
#RNN returns same output every calls.
Here is the desired behavior, taking the example of layers.LSTM:
RNN = layers.LSTM(
params['hidden_dim'],
dropout=params['dropout_rate_input'],
recurrent_dropout=params['dropout_rate_output'],
return_sequences=True)
print(set([RNN(inputs, training=True).numpy()[0, 0, 0] for _ in range(5)]))
#return: {-0.053168092, -0.016183555, -0.024903715, -0.040428974, 0.025961103}
#RNN returns different outputs after each call.
This is the source code of keras.layers.RNN:
https://github.com/tensorflow/tensorflow/blob/v2.5.0/tensorflow/python/keras/layers/recurrent.py#L198-L1002
Thank you for spending time for my problem. Any idea is appreciated.
I have three series of observations, namely Y, T, and X. I would like to study the differences between the predicted values of the two models. The first model is to learn g such that Y=g(T, X). The second model is to learn L and f such that Y=L(T)f(X). I have no problem in learning the first model using the PyTorch package or the Tensorflow package. However, I am not sure how to learn L and f. In using the PyTorch package, I can set up two feedforward MLPs with different hidden layers and inputs. For simplicity, I define a Feedforward MLP class as follows:
class Feedforward(t.nn.Module): # the definition of a feedforward neural network
# Basic definition
def __init__(self, input_size, hidden_size):
super(Feedforward, self).__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.fc1 = t.nn.Linear(self.input_size, self.hidden_size)
self.relu = t.nn.ReLU()
self.fc2 = t.nn.Linear(self.hidden_size, 1)
self.sigmoid = t.nn.Sigmoid()
# Advance definition
def forward(self, x):
hidden = self.fc1(x)
relu = self.relu(hidden)
output = self.fc2(relu)
output = self.sigmoid(output)
return output
Suppose L=Feedforward(2,10) and L=Feedforward(3,9). From my understanding, I can only learn either L or f, but not both simultaneously. Is it possible to learn L and f simultaneously using Y, T, and X?
I may be missing something, but I think you can :
L = Feedforward(2,10)
f = Feedforward(3,9)
L_opt = Adam(L.parameters(), lr=...)
f_opt = Adam(f.parameters(), lr=...)
for (x,t,y) in dataset:
L.zero_grad()
f.zero_grad()
y_pred = L(t)*f(x)
loss = (y-y_pred)**2
loss.backward()
L_opt.step()
f_opt.step()
You can also fuse them together in one single model :
class ProductModel(t.nn.Module):
def __init__(self, L, f):
self.L = L
self.f = f
def forward(self, x,t):
return self.L(t)*self.f(x)
and then train this model like you trained g
I am trying to implement a hierarchical transformer for document classification in Keras/tensorflow, in which:
(1) a word-level transformer produces a representation of each sentence, and attention weights for each word, and,
(2) a sentence-level transformer uses the outputs from (1) to produce a representation of each document, and attention weights for each sentence, and finally,
(3) the document representations produced by (2) are used to classify documents (in the following example, as belonging or not belonging to a given class).
I am attempting to model the classifier on Yang et al.'s approach here (https://www.cs.cmu.edu/~./hovy/papers/16HLT-hierarchical-attention-networks.pdf), but replacing the GRU and attention layers with transformers.
I am using Apoorv Nandan's transformer implementation from https://keras.io/examples/nlp/text_classification_with_transformer/.
I have two issues for which I would be grateful for the community's help:
(1) I get an error in the upper (sentence) level model that I can't resolve (details and code below)
(2) I don't know how to extract the word- and sentence-level attention weights, and value advice on how best to do this.
I am new to both Keras and this forum, so apologies for obvious mistakes and thank you in advance for any help.
Here is a reproducible example, indicating where I encounter errors:
First, establish the multi-head attention, transformer, and token/position embedding layers, after Nandan.
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import pandas as pd
import numpy as np
class MultiHeadSelfAttention(layers.Layer):
def __init__(self, embed_dim, num_heads=8):
super(MultiHeadSelfAttention, self).__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
if embed_dim % num_heads != 0:
raise ValueError(
f"embedding dimension = {embed_dim} should be divisible by number of heads = {num_heads}"
)
self.projection_dim = embed_dim // num_heads
self.query_dense = layers.Dense(embed_dim)
self.key_dense = layers.Dense(embed_dim)
self.value_dense = layers.Dense(embed_dim)
self.combine_heads = layers.Dense(embed_dim)
def attention(self, query, key, value):
score = tf.matmul(query, key, transpose_b=True)
dim_key = tf.cast(tf.shape(key)[-1], tf.float32)
scaled_score = score / tf.math.sqrt(dim_key)
weights = tf.nn.softmax(scaled_score, axis=-1)
output = tf.matmul(weights, value)
return output, weights
def separate_heads(self, x, batch_size):
x = tf.reshape(x, (batch_size, -1, self.num_heads, self.projection_dim))
return tf.transpose(x, perm=[0, 2, 1, 3])
def call(self, inputs):
# x.shape = [batch_size, seq_len, embedding_dim]
batch_size = tf.shape(inputs)[0]
query = self.query_dense(inputs) # (batch_size, seq_len, embed_dim)
key = self.key_dense(inputs) # (batch_size, seq_len, embed_dim)
value = self.value_dense(inputs) # (batch_size, seq_len, embed_dim)
query = self.separate_heads(
query, batch_size
) # (batch_size, num_heads, seq_len, projection_dim)
key = self.separate_heads(
key, batch_size
) # (batch_size, num_heads, seq_len, projection_dim)
value = self.separate_heads(
value, batch_size
) # (batch_size, num_heads, seq_len, projection_dim)
attention, weights = self.attention(query, key, value)
attention = tf.transpose(
attention, perm=[0, 2, 1, 3]
) # (batch_size, seq_len, num_heads, projection_dim)
concat_attention = tf.reshape(
attention, (batch_size, -1, self.embed_dim)
) # (batch_size, seq_len, embed_dim)
output = self.combine_heads(
concat_attention
) # (batch_size, seq_len, embed_dim)
return output
class TransformerBlock(layers.Layer):
def __init__(self, embed_dim, num_heads, ff_dim, dropout_rate, name=None):
super(TransformerBlock, self).__init__(name=name)
self.att = MultiHeadSelfAttention(embed_dim, num_heads)
self.ffn = keras.Sequential(
[layers.Dense(ff_dim, activation="relu"), layers.Dense(embed_dim),]
)
self.layernorm1 = layers.LayerNormalization(epsilon=1e-6)
self.layernorm2 = layers.LayerNormalization(epsilon=1e-6)
self.dropout1 = layers.Dropout(dropout_rate)
self.dropout2 = layers.Dropout(dropout_rate)
def call(self, inputs, training):
attn_output = self.att(inputs)
attn_output = self.dropout1(attn_output, training=training)
out1 = self.layernorm1(inputs + attn_output)
ffn_output = self.ffn(out1)
ffn_output = self.dropout2(ffn_output, training=training)
return self.layernorm2(out1 + ffn_output)
class TokenAndPositionEmbedding(layers.Layer):
def __init__(self, maxlen, vocab_size, embed_dim, name=None):
super(TokenAndPositionEmbedding, self).__init__(name=name)
self.token_emb = layers.Embedding(input_dim=vocab_size, output_dim=embed_dim)
self.pos_emb = layers.Embedding(input_dim=maxlen, output_dim=embed_dim)
def call(self, x):
maxlen = tf.shape(x)[-1]
positions = tf.range(start=0, limit=maxlen, delta=1)
positions = self.pos_emb(positions)
x = self.token_emb(x)
return x + positions
For the purpose of this example, the data are 10,000 documents, each truncated to 15 sentences, each sentence with a maximum of 60 words, which are already converted to integer tokens 1-1000.
X is a 3-D tensor (10000, 15, 60) containing these tokens. y is a 1-D tensor containing the classes of the documents (1 or 0). For the purpose of this example there is no relation between X and y.
The following produces the example data:
max_docs = 10000
max_sentences = 15
max_words = 60
X = tf.random.uniform(shape=(max_docs, max_sentences, max_words), minval=1, maxval=1000, dtype=tf.dtypes.int32, seed=1)
y = tf.random.uniform(shape=(max_docs,), minval=0, maxval=2, dtype=tf.dtypes.int32, seed=1)
Here I attempt to construct the word level encoder, after https://keras.io/examples/nlp/text_classification_with_transformer/:
# Lower level (produce a representation of each sentence):
embed_dim = 100 # Embedding size for each token
num_heads = 2 # Number of attention heads
ff_dim = 64 # Hidden layer size in feed forward network inside transformer
L1_dense_units = 100 # Size of the sentence-level representations output by the word-level model
dropout_rate = 0.1
vocab_size=1000
word_input = layers.Input(shape=(max_words,), name='word_input')
word_embedding = TokenAndPositionEmbedding(maxlen=max_words, vocab_size=vocab_size,
embed_dim=embed_dim, name='word_embedding')(word_input)
word_transformer = TransformerBlock(embed_dim=embed_dim, num_heads=num_heads, ff_dim=ff_dim,
dropout_rate=dropout_rate, name='word_transformer')(word_embedding)
word_pool = layers.GlobalAveragePooling1D(name='word_pooling')(word_transformer)
word_drop = layers.Dropout(dropout_rate,name='word_drop')(word_pool)
word_dense = layers.Dense(L1_dense_units, activation="relu",name='word_dense')(word_drop)
word_encoder = keras.Model(word_input, word_dense)
word_encoder.summary()
It looks as though this word encoder works as intended to produce a representation of each sentence. Here, run on the 1st document, it produces a tensor of shape (15, 100), containing the vectors representing each of 15 sentences:
word_encoder(X[0]).shape
My problem is in connecting this to the higher (sentence) level model, to produce document representations.
I get error "NotImplementedError" when trying to apply the word encoder to each sentence in a document. I would be grateful for any help in fixing this issue, since the error message is not informative as to the specific problem.
After applying the word encoder to each sentence, the goal is to apply another transformer to produce attention weights for each sentence, and a document-level representation with which to perform classification. I can't determine whether this part of the model will work because of the error above.
Finally, I would like to extract word- and sentence-level attention weights for each document, and would be grateful for advice on how to do so.
Thank you in advance for any insight.
# Upper level (produce a representation of each document):
L2_dense_units = 100
sentence_input = layers.Input(shape=(max_sentences, max_words), name='sentence_input')
# This is the line producing "NotImplementedError":
sentence_encoder = tf.keras.layers.TimeDistributed(word_encoder, name='sentence_encoder')(sentence_input)
sentence_transformer = TransformerBlock(embed_dim=L1_dense_units, num_heads=num_heads, ff_dim=ff_dim,
dropout_rate=dropout_rate, name='sentence_transformer')(sentence_encoder)
sentence_dense = layers.TimeDistributed(Dense(int(L2_dense_units)),name='sentence_dense')(sentence_transformer)
sentence_out = layers.Dropout(dropout_rate)(sentence_dense)
preds = layers.Dense(1, activation='sigmoid', name='sentence_output')(sentence_out)
model = keras.Model(sentence_input, preds)
model.summary()
I got NotImplementedError as well while trying to do the same thing as you. The thing is Keras's TimeDistributed layer needs to know its inner custom layer's output shapes. So you should add compute_output_shape method to your custom layers.
In your case MultiHeadSelfAttention, TransformerBlock and TokenAndPositionEmbedding layers should include:
class MultiHeadSelfAttention(layers.Layer):
...
def compute_output_shape(self, input_shape):
# it does not change the shape of its input
return input_shape
class TransformerBlock(layers.Layer):
...
def compute_output_shape(self, input_shape):
# it does not change the shape of its input
return input_shape
class TokenAndPositionEmbedding(layers.Layer):
...
def compute_output_shape(self, input_shape):
# it changes the shape from (batch_size, maxlen) to (batch_size, maxlen, embed_dim)
return input_shape + (self.pos_emb.output_dim,)
After you add these methods you should be able to run your code.
As for your second question, I am not sure but maybe you can return the "weights" variable that is returned from MultiHeadSelfAttention's attention method in call methods of both MultiHeadSelfAttention and TransformerBlock. So that you can access it where you build your model.
I'd like to initialize the parameters of RNN with np arrays.
In the following example, I want to pass w to the parameters of rnn. I know pytorch provides many initialization methods like Xavier, uniform, etc., but is there way to initialize the parameters by passing numpy arrays?
import numpy as np
import torch as nn
rng = np.random.RandomState(313)
w = rng.randn(input_size, hidden_size).astype(np.float32)
rnn = nn.RNN(input_size, hidden_size, num_layers)
First, let's note that nn.RNN has more than one weight variable, c.f. the documentation:
Variables:
weight_ih_l[k] – the learnable input-hidden weights of the k-th layer, of shape (hidden_size * input_size) for k = 0. Otherwise,
the shape is (hidden_size * hidden_size)
weight_hh_l[k] – the learnable hidden-hidden weights of the k-th layer, of shape (hidden_size * hidden_size)
bias_ih_l[k] – the learnable input-hidden bias of the k-th layer, of shape (hidden_size)
bias_hh_l[k] – the learnable hidden-hidden bias of the k-th layer, of shape (hidden_size)
Now, each of these variables (Parameter instances) are attributes of your nn.RNN instance. You can access them, and edit them, two ways, as show below:
Solution 1: Accessing all the RNN Parameter attributes by name (rnn.weight_hh_lK, rnn.weight_ih_lK, etc.):
import torch
from torch import nn
import numpy as np
input_size, hidden_size, num_layers = 3, 4, 2
use_bias = True
rng = np.random.RandomState(313)
rnn = nn.RNN(input_size, hidden_size, num_layers, bias=use_bias)
def set_nn_parameter_data(layer, parameter_name, new_data):
param = getattr(layer, parameter_name)
param.data = new_data
for i in range(num_layers):
weights_hh_layer_i = rng.randn(hidden_size, hidden_size).astype(np.float32)
weights_ih_layer_i = rng.randn(hidden_size, hidden_size).astype(np.float32)
set_nn_parameter_data(rnn, "weight_hh_l{}".format(i),
torch.from_numpy(weights_hh_layer_i))
set_nn_parameter_data(rnn, "weight_ih_l{}".format(i),
torch.from_numpy(weights_ih_layer_i))
if use_bias:
bias_hh_layer_i = rng.randn(hidden_size).astype(np.float32)
bias_ih_layer_i = rng.randn(hidden_size).astype(np.float32)
set_nn_parameter_data(rnn, "bias_hh_l{}".format(i),
torch.from_numpy(bias_hh_layer_i))
set_nn_parameter_data(rnn, "bias_ih_l{}".format(i),
torch.from_numpy(bias_ih_layer_i))
Solution 2: Accessing all the RNN Parameter attributes through rnn.all_weights list attribute:
import torch
from torch import nn
import numpy as np
input_size, hidden_size, num_layers = 3, 4, 2
use_bias = True
rng = np.random.RandomState(313)
rnn = nn.RNN(input_size, hidden_size, num_layers, bias=use_bias)
for i in range(num_layers):
weights_hh_layer_i = rng.randn(hidden_size, hidden_size).astype(np.float32)
weights_ih_layer_i = rng.randn(hidden_size, hidden_size).astype(np.float32)
rnn.all_weights[i][0].data = torch.from_numpy(weights_ih_layer_i)
rnn.all_weights[i][1].data = torch.from_numpy(weights_hh_layer_i)
if use_bias:
bias_hh_layer_i = rng.randn(hidden_size).astype(np.float32)
bias_ih_layer_i = rng.randn(hidden_size).astype(np.float32)
rnn.all_weights[i][2].data = torch.from_numpy(bias_ih_layer_i)
rnn.all_weights[i][3].data = torch.from_numpy(bias_hh_layer_i)
As a detailed answer is provided, I just to add one more sentence. The parameters of an nn.Module are Tensors (previously, it used to be autograd variables, which is deperecated in Pytorch 0.4). So, essentially you need to use the torch.from_numpy() method to convert the Numpy array to Tensor and then use them to initialize the nn.Module parameters.