shady-cs15 / slash Goto Github PK

License + Documentation

• Adam (Better accuracy but memory state is always negative i.e. global context becomes 0)
• SGD (Less accuracy but memory state is ~ 50 % +ve, i.e. 50 % active after ReLU)

• Save raw wave and masks in uint8 in stead of float32.
• No need to save labels
• read directly from data/abc.wav and create tmp/abc_x.npy and tmp/abc_m.npy

one batch corresponds to one clip.
select batch_size number of anchor pointers in the clip.
use masking as required.

check if the mlp stack pays attention to global context as well.
make down_sampl = tf.zeros(..)
and check generate audio

• train sample rnn original code (theano) and test generative capabilities of the same.
• train custom arch on same dataset for same iterations and test generative capabilities.
• Let model overfit if need be

update batch_generate.py to generate batches spanning the entire dataset.

check generation phase of gen-script.py.

format print statements

• floats up to 4 decimal places
• print global step number in training
• remove song #
• print epoch # once for every song not for every iter

• Need to warm up the rnn for at least first 2 seconds, provide correct input.
• For the last 3 seconds, provide the sampled output

check generative capabilities at every validation step.

clip_iter=20
n_epochs = 10
If accuracy doesn't improve, use constant divisor for normalizing in stead of tf.reduce_sum.

Sample for validation

dev-script.py

line 29

is_training=False

line 52

bptt_batch_loss, acc, np_state, op, out = 
sess.run([t_model.loss, t_model.mean_acc, t_model.final_state, optimizer, t_model.outputs],
feed_dict={input:bptt_batch_x, label:bptt_batch_y, t_model.initial_state[0]:np_state[0], t_model.initial_state[1]:np_state[1]})
				

state includes

• best validation loss
• iter before saving last params
• epoch (ep) before saving last params
• clip # (i) before saving last params

add summaries

• Use one hot encoding in the computation graph
• Feed only integers quantized to range(0-15)
• For masks use vector of shape [batch_size, n_steps] in stead of [batch_size, n_steps, 16]. Reason follows.

>>> x = tf.placeholder('float32',[2, 3])
>>> y = tf.placeholder('float32',[2, 1])
>>> z = x*y
>>> with tf.Session as s:
...     X = [[1., 1., 1.], [1., 1., 1.]]
...     Y = [[1.], [0.]]
...     Z = s.run([z], feed_dict={x:X, y:Y})
...     print Z

• In stead of looping, use linear downsampling (from lstm_dim to global_context_size); or repeat the tensor along one axis
• for each local context, segment from downsample and add in stead of concatenating, if need be. Logically though, concatenation preserves the local context and global context separately and hence is more rich in information.
• build computation graph faster

Roughly 100 audio clips needed

Needed for generation