nested_chordnames = d["chordnames"] flat_idx = d["flat_idx"] labels = d["labels"] train_labels = labels[:-1000] valid_labels = labels[-1000:] n_labels = len( set(list(np.unique(train_labels)) + list(np.unique(valid_labels)))) print("n_labels {}".format(n_labels)) train_itr_random_state = np.random.RandomState(1122) val_itr_random_state = np.random.RandomState(1) train_itr = list_iterator([train_image_data, train_labels], 50, random_state=train_itr_random_state) val_itr = list_iterator([val_image_data, valid_labels], 50, random_state=val_itr_random_state) random_state = np.random.RandomState(1999) # only 6x6 in the first place... kernel_size0 = (5, 5) kernel_size1 = (3, 3) n_channels = 64 n_layers = 15 def create_pixel_cnn(inp, lbl):
batch_size = 64 n_syms = len(vocab) shuffled_inds = list(range(len(word_inds))) train_itr_random_state.shuffle(shuffled_inds) split = 250000 train_inds = shuffled_inds[:split] valid_inds = shuffled_inds[split:] train_word_inds = [word_inds[i] for i in train_inds] train_rev_word_inds = [rev_word_inds[i] for i in train_inds] valid_word_inds = [word_inds[i] for i in valid_inds] valid_rev_word_inds = [rev_word_inds[i] for i in valid_inds] train_itr = list_iterator([train_word_inds, train_rev_word_inds], batch_size, random_state=train_itr_random_state) valid_itr = list_iterator([valid_word_inds, valid_rev_word_inds], batch_size, random_state=valid_itr_random_state) with tf.Session(config=config) as sess: saver = tf.train.import_meta_graph(model_path + '.meta') saver.restore(sess, model_path) fields = [ "inputs", "outputs", "outputs_masks", "pred_logits", "enc_atts_0", "enc_atts_1", "enc_atts_2", "dec_atts_0", "dec_atts_1", "dec_atts_2" ] vs = namedtuple('Params', fields)(*[tf.get_collection(name)[0] for name in fields]) x, y = valid_itr.next_batch()
valid_sines = sines[:, 1::2] valid_sines = [valid_sines[:, i] for i in range(valid_sines.shape[1])] """ f, axarr = plt.subplots(4, 1) axarr[0].plot(train_sines[0].ravel()) axarr[1].plot(valid_sines[0].ravel()) axarr[2].plot(train_sines[1].ravel()) axarr[3].plot(valid_sines[0].ravel()) plt.savefig("tmp") """ train_itr_random_state = np.random.RandomState(1122) valid_itr_random_state = np.random.RandomState(12) batch_size = 10 train_itr = list_iterator([train_sines], batch_size, random_state=train_itr_random_state) valid_itr = list_iterator([valid_sines], batch_size, random_state=valid_itr_random_state) random_state = np.random.RandomState(1999) n_hid = 100 n_emb = 512 rnn_init = None #"truncated_normal" forward_init = None #"truncated_normal" def create_model(inp_tm1, inp_t, h1_init, c1_init, h1_q_init, c1_q_init): p_tm1 = Linear([inp_tm1], [1],
from tfbldr.datasets import list_iterator from tfbldr import get_params_dict from tfbldr import run_loop import tensorflow as tf import numpy as np from collections import namedtuple mnist = fetch_mnist() image_data = mnist["images"] / 255. # save last 10k to validate on train_image_data = image_data[:-10000] val_image_data = image_data[-10000:] train_itr_random_state = np.random.RandomState(1122) val_itr_random_state = np.random.RandomState(1) train_itr = list_iterator([train_image_data], 64, random_state=train_itr_random_state) val_itr = list_iterator([val_image_data], 64, random_state=val_itr_random_state) random_state = np.random.RandomState(1999) l1_dim = (16, 4, 4, 2) l2_dim = (32, 4, 4, 2) l3_dim = (64, 1, 1, 1) embedding_dim = 512 l_dims = [l1_dim, l2_dim, l3_dim] stride_div = np.prod([ld[-1] for ld in l_dims]) bpad = 1
axarr.imshow(arr, interpolation=None, cmap=viridis_cm) #axarr.set_yaxis("off") plt.axis("off") x1 = arr.shape[0] y1 = arr.shape[1] #asp = autoaspect(x1, y1) #axarr.set_aspect(asp) plt.savefig("tmp") """ # each "image" is 100 by 257 # save last 20 to validate on train_data = joined_spec val_data = joined_spec train_itr_random_state = np.random.RandomState(1122) val_itr_random_state = np.random.RandomState(1) train_itr = list_iterator([train_data], 25, random_state=train_itr_random_state) val_itr = list_iterator([val_data], 25, random_state=val_itr_random_state) random_state = np.random.RandomState(1999) l1_dim = (32, 4, 257, [1, 2, 1, 1]) l2_dim = (64, 4, 1, [1, 2, 1, 1]) l3_dim = (128, 4, 1, [1, 2, 1, 1]) l4_dim = (128, 1, 1, [1, 1, 1, 1]) embedding_dim = 1024 l_dims = [l1_dim, l2_dim, l3_dim, l4_dim] stride_div = np.prod([ld[-1] for ld in l_dims]) ebpad = [0, 4 // 2 - 1, 0, 0] dbpad = [0, 4 // 2 - 1, 0, 0] def create_encoder(inp, bn_flag): l1 = Conv2d([inp], [1], l_dims[0][0], kernel_size=l_dims[0][1:3], name="enc1",
#nested_chordnames = d["chordnames"] #labels = d["labels"] #train_labels = labels[:-1000] #valid_labels = labels[-1000:] #n_labels = len(set(list(np.unique(train_labels)) + list(np.unique(valid_labels)))) #print("n_labels {}".format(n_labels)) train_itr_random_state = np.random.RandomState(1122) val_itr_random_state = np.random.RandomState(1) #train_itr = list_iterator([train_image_data, train_labels], 50, random_state=train_itr_random_state) #val_itr = list_iterator([val_image_data, valid_labels], 50, random_state=val_itr_random_state) train_itr = list_iterator([train_image_data, train_conditions], 50, random_state=train_itr_random_state) val_itr = list_iterator([val_image_data, val_conditions], 50, random_state=val_itr_random_state) random_state = np.random.RandomState(1999) # only 6x6 in the first place... kernel_size0 = (5, 5) kernel_size1 = (3, 3) n_channels = 64 n_layers = 15 def create_pixel_cnn(inp, cond):
s = s[:len(s) - len(s) % step] starts = np.arange(0, len(s) - cut + step, step) for st in starts: real_final.append(s[st:st + cut][None, :, None]) return real_final cut = 256 step = 256 train_audio = _cuts(train_data, cut, step) valid_audio = _cuts(valid_data, cut, step) train_itr_random_state = np.random.RandomState(1122) valid_itr_random_state = np.random.RandomState(12) train_itr = list_iterator([train_audio], 50, random_state=train_itr_random_state) valid_itr = list_iterator([valid_audio], 50, random_state=valid_itr_random_state) random_state = np.random.RandomState(1999) l1_dim = (64, 1, 4, [1, 1, 2, 1]) l2_dim = (128, 1, 4, [1, 1, 2, 1]) l3_dim = (256, 1, 4, [1, 1, 2, 1]) l3_dim = (257, 1, 4, [1, 1, 2, 1]) l4_dim = (256, 1, 4, [1, 1, 2, 1]) l5_dim = (257, 1, 1, [1, 1, 1, 1]) embedding_dim = 512 n_components = 10 dmol_proj = 64