shithub: opus

Info  •  Files  •  Log  •  Branches

Larger range of quantizers

--- a/dnn/training_tf2/decode_rdovae.py

+++ b/dnn/training_tf2/decode_rdovae.py

@@ -83,7 +83,7 @@

 print(bits.shape)

 lambda_val = 0.0007 * np.ones((nb_sequences, sequence_size//2, 1))

-quant_id = np.round(10*np.log(lambda_val/.0007)).astype('int16')

+quant_id = np.round(10*np.log(lambda_val/.0002)).astype('int16')

 quant_id = quant_id[:,:,0]

 quant_embed = qembedding(quant_id)

 quant_scale = tf.math.softplus(quant_embed[:,:,:nbits])

@@ -98,8 +98,8 @@

 state = np.reshape(state, (nb_sequences, sequence_size//2, 24))

 state = state[:,-1,:]

-#state = pvq_quantize(state, 30)

-state = state/(1e-15+tf.norm(state, axis=-1,keepdims=True))

+state = pvq_quantize(state, 30)

+#state = state/(1e-15+tf.norm(state, axis=-1,keepdims=True))

 print("shapes are:")

 print(bits.shape)

--- a/dnn/training_tf2/encode_rdovae.py

+++ b/dnn/training_tf2/encode_rdovae.py

@@ -104,8 +104,8 @@

 nbits=80

 bits.astype('float32').tofile(args.output + "-syms.f32")

-lambda_val = 0.0007 * np.ones((nb_sequences, sequence_size//2, 1))

-quant_id = np.round(10*np.log(lambda_val/.0007)).astype('int16')

+lambda_val = 0.001 * np.ones((nb_sequences, sequence_size//2, 1))

+quant_id = np.round(10*np.log(lambda_val/.0002)).astype('int16')

 quant_id = quant_id[:,:,0]

 quant_embed = qembedding(quant_id)

 quant_scale = tf.math.softplus(quant_embed[:,:,:nbits])

@@ -122,4 +122,4 @@

 print(dec_out.shape)

-dec_out.numpy().astype('float32').tofile(args.output + "-unquant_out.f32")

+dec_out.numpy().astype('float32').tofile(args.output + "-quant_out.f32")

--- a/dnn/training_tf2/train_rdovae.py

+++ b/dnn/training_tf2/train_rdovae.py

@@ -121,10 +121,11 @@

 print(features.shape)

 features = features[:, :, :nb_used_features]

-#lambda_val = np.random.uniform(.0007, .002, (features.shape[0], features.shape[1], 1))

-lambda_val = np.repeat(np.random.uniform(.0007, .002, (features.shape[0], 1, 1)), features.shape[1]//2, axis=1)

-#lambda_val = 0*lambda_val + .001

-quant_id = np.round(10*np.log(lambda_val/.0007)).astype('int16')

+#lambda_val = np.repeat(np.random.uniform(.0007, .002, (features.shape[0], 1, 1)), features.shape[1]//2, axis=1)

+#quant_id = np.round(10*np.log(lambda_val/.0007)).astype('int16')

+#quant_id = quant_id[:,:,0]

+quant_id = np.repeat(np.random.randint(39, size=(features.shape[0], 1, 1), dtype='int16'), features.shape[1]//2, axis=1)

+lambda_val = .0002*np.exp(quant_id/10.)

 quant_id = quant_id[:,:,0]

 # dump models to disk as we go

-- 

⑨