tensorflow: Obtain RNN hidden states gradients with respect to input

Question

My model consists of an Embedding layer and a SimpleRNN layer. I have obtained the hidden states at all steps with model.predict, and plotted them against the steps. I find that the hidden states converge to zero but I am not sure if I can infer anything from that. Therefore plotting their gradients with respect to the model inputs might provide me some further insights. I would like some help with obtaining these gradients.

My model:

batch_size = 9600   # batch size can take a smaller value, e.g. 100
inp= Input(batch_shape= (batch_size, input_length), name= 'input') 
emb_out= Embedding(input_dim, output_dim, input_length= input_length, 
                         weights= [Emat], trainable= False, name= 'embedding')(inp)
rnn= SimpleRNN(200, return_sequences= True, return_state= False, stateful= True,
               batch_size= (batch_size, input_length, 100), name= 'simpleRNN')

h0 = tf.random.uniform((batch_size, 200))
rnn_allstates = rnn(emb_out, initial_state=h0)
print(rnn_allstates.shape)   # (9600, 1403, 200)
model_rnn = Model(inputs=inp, outputs= rnn_allstates, name= 'model_rnn')
model_rnn.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['acc'])
model_rnn.summary()

>>>
Layer (type)                 Output Shape              Param #   
=================================================================
input (InputLayer)           [(9600, 1403)]            0         
_________________________________________________________________
embedding (Embedding)        (9600, 1403, 100)         4348900   
_________________________________________________________________
simpleRNN (SimpleRNN)        (9600, 1403, 200)         60200     
=================================================================

Obtaining the hidden states:

rnn_ht = model_rnn.predict(xtr_pad)   # xtr_pad.shape = (9600,1403)
rnn_ht_red= np.mean(rnn_ht, 2)
rnn_ht_red= np.mean(rnn_ht_red,0)
steps= [t for t in range(1403)]
plt.plot(steps, rnn_ht_red, linestyle= 'dotted')

Attempt to obtain gradients:

sess= k.get_session()
# The hidden states tf.Variable shaped (n_samples = 9600, n_steps = 1403, n_units = 200):
states_var= model_rnn.output  
# A list of hidden states variable for all time steps, aggregated over samples and RNN units:
ht_vars= [states_var[:, t, :] for t in range(1403)]        # each item in list has shape (9600, 200)
ht_vars_agg= [tf.reduce_mean(ht,[0,1]) for ht in ht_vars]  # each item in list has shape (), because I wish to obtain a SINGLE gradient value at each time step.

# Create gradient function and feed data:
dhtdx_vars= [k.gradients(ht, model_rnn.input) for ht in ht_vars_agg]
dhtdx= [sess.run(pd, feed_dict={model_rnn.input: xtr_pad} ) for pd in dhtdx_vars  ]

The following error points to the last line above

TypeError: Fetch argument None has invalid type <class 'NoneType'>

Every backend gradient item in dhtdx_vars is [None]. When I remove the aggregation line the same error still persists.

An attempt with gradient tape also returns None Error in the computed gradient.

with tf.GradientTape() as tape:
    x= model_rnn.input
    ht = model_rnn(x)
grad = tape.gradient(ht, model_rnn.input)

Thanks in advance for any help.

looks like someone has had this None error in gradient before, but no answers yet: https://stackoverflow.com/questions/50103231/how-to-calculate-gradients-for-rnn-hidden-states-with-tensorarray — siegfried, Feb 14 '22 at 05:08
@AloneTogether Thx again for your help. yeah it doesn't make sense... I was having those incompatibility errors using a smaller batch size when feeding the whole input into the model API: ```model(full_X)```. But in fact ```model.predict(full_X), model.fit(full_X, full_y)``` both work (after adding a Dense(1) layer), and batch size can actually take e.g. 100 — siegfried, Feb 14 '22 at 09:48

score 1 · Accepted Answer · answered Feb 14 '22 at 08:13

The problem is tf.GradientTape() doesn't propagate the gradients through integer inputs. That is probably the reason you are getting None gradients. What you can do is calculate the gradients with respect to the output of the Embedding layer like this:

import tensorflow as tf

input_length = 1403
inp= tf.keras.layers.Input(shape= (input_length,)) 
emb_out= tf.keras.layers.Embedding(500, 100, input_length= input_length, trainable= False)(inp)
rnn_out = tf.keras.layers.SimpleRNN(200,  return_sequences = True)(emb_out)

model = tf.keras.Model(inputs=inp, outputs=rnn_out)
model.summary()


xte_pad = tf.random.uniform((10, 1403), maxval=500, dtype=tf.int32)
y = tf.random.normal((10, 1403, 200))
ds = tf.data.Dataset.from_tensor_slices((xte_pad, y)).batch(5)

embedding_layer = model.layers[1]
rnn_layer = model.layers[2]
epochs = 1
for epoch in range(epochs):
  for step, (x_batch_train, y_batch_train) in enumerate(ds):
    with tf.GradientTape() as tape:
        embedded_x = embedding_layer(x_batch_train)
        tape.watch(embedded_x)
        y = rnn_layer(embedded_x)

    grads = tape.gradient(y, embedded_x)
    tf.print(grads.shape)

Model: "model_10"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_13 (InputLayer)       [(None, 1403)]            0         
                                                                 
 embedding_12 (Embedding)    (None, 1403, 100)         50000     
                                                                 
 simple_rnn_12 (SimpleRNN)   (None, 1403, 200)         60200     
                                                                 
=================================================================
Total params: 110,200
Trainable params: 60,200
Non-trainable params: 50,000
_________________________________________________________________
TensorShape([5, 1403, 100])
TensorShape([5, 1403, 100])

That's totally fine. The embedding layer outputs are actually the more relevant quantities in terms of inspecting hidden states gradients behaviour. — siegfried, Feb 14 '22 at 11:24

tensorflow: Obtain RNN hidden states gradients with respect to input

1 Answers1

Linked