def _loop(function, itr): prev_h1, prev_h2, prev_h3 = [ np_zeros((minibatch_size, n_hid)) for i in range(3) ] prev_kappa = np_zeros((minibatch_size, att_size)) prev_w = np_zeros((minibatch_size, n_chars)) X_mb, X_mb_mask, c_mb, c_mb_mask = next(itr) n_cuts = len(X_mb) // cut_len + 1 partial_costs = [] for n in range(n_cuts): start = n * cut_len stop = (n + 1) * cut_len if len(X_mb[start:stop]) < cut_len: new_len = cut_len - len(X_mb) % cut_len zeros = np.zeros((new_len, X_mb.shape[1], X_mb.shape[2])) zeros = zeros.astype(X_mb.dtype) mask_zeros = np.zeros((new_len, X_mb_mask.shape[1])) mask_zeros = mask_zeros.astype(X_mb_mask.dtype) X_mb = np.concatenate((X_mb, zeros), axis=0) X_mb_mask = np.concatenate((X_mb_mask, mask_zeros), axis=0) assert len(X_mb[start:stop]) == cut_len assert len(X_mb_mask[start:stop]) == cut_len rval = function(X_mb[start:stop], X_mb_mask[start:stop], c_mb, c_mb_mask, prev_h1, prev_h2, prev_h3, prev_kappa, prev_w) current_cost = rval[0] prev_h1, prev_h2, prev_h3 = rval[1:4] prev_h1 = prev_h1[-1] prev_h2 = prev_h2[-1] prev_h3 = prev_h3[-1] prev_kappa = rval[4][-1] prev_w = rval[5][-1] partial_costs.append(current_cost) return partial_costs
def _loop(function, itr): prev_h1, prev_h2 = [np_zeros((minibatch_size, n_hid)) for i in range(2)] prev_kappa = np_zeros((minibatch_size, att_size)) prev_w = np_zeros((minibatch_size, n_chars)) X_mb, X_mb_mask, c_mb, c_mb_mask = next(itr) n_cuts = len(X_mb) // cut_len + 1 partial_costs = [] for n in range(n_cuts): start = n * cut_len stop = (n + 1) * cut_len if len(X_mb[start:stop]) < cut_len: new_len = cut_len - len(X_mb) % cut_len zeros = np.zeros((new_len, X_mb.shape[1], X_mb.shape[2])) zeros = zeros.astype(X_mb.dtype) mask_zeros = np.zeros((new_len, X_mb_mask.shape[1])) mask_zeros = mask_zeros.astype(X_mb_mask.dtype) X_mb = np.concatenate((X_mb, zeros), axis=0) X_mb_mask = np.concatenate((X_mb_mask, mask_zeros), axis=0) assert len(X_mb[start:stop]) == cut_len assert len(X_mb_mask[start:stop]) == cut_len rval = function(X_mb[start:stop], X_mb_mask[start:stop], c_mb, c_mb_mask, prev_h1, prev_h2, prev_kappa, prev_w) current_cost = rval[0] prev_h1, prev_h2 = rval[1:3] prev_h1 = prev_h1[-1] prev_h2 = prev_h2[-1] prev_kappa = rval[3][-1] prev_w = rval[4][-1] partial_costs.append(current_cost) return partial_costs
def _loop(function, itr): prev_h1, prev_h2, prev_h3 = [np_zeros((minibatch_size, n_hid)) for i in range(3)] X_mb, X_mb_mask, y_mb, y_mb_mask = next(itr) n_cuts = len(X_mb) // cut_len + 1 partial_costs = [] for n in range(n_cuts): start = n * cut_len stop = (n + 1) * cut_len if len(X_mb[start:stop]) < cut_len: # skip end edge case break new_len = cut_len - len(X_mb) % cut_len zeros = np.zeros((new_len, X_mb.shape[1], X_mb.shape[2])) zeros = zeros.astype(X_mb.dtype) mask_zeros = np.zeros((new_len, X_mb_mask.shape[1])) mask_zeros = mask_zeros.astype(X_mb_mask.dtype) X_mb = np.concatenate((X_mb, zeros), axis=0) X_mb_mask = np.concatenate((X_mb_mask, mask_zeros), axis=0) assert len(X_mb[start:stop]) == cut_len assert len(X_mb_mask[start:stop]) == cut_len rval = function(X_mb[start:stop], X_mb_mask[start:stop], y_mb[start:stop], y_mb_mask[start:stop], prev_h1, prev_h2, prev_h3) current_cost = rval[0] prev_h1, prev_h2, prev_h3 = rval[1:4] prev_h1 = prev_h1[-1] prev_h2 = prev_h2[-1] prev_h3 = prev_h3[-1] partial_costs.append(current_cost) return partial_costs
def _loop(function, itr): prev_h1, prev_h2, prev_h3 = [np_zeros((minibatch_size, n_hid)) for i in range(3)] X_mb, X_mb_mask = next(itr) # sanity check that masking code is OK assert X_mb_mask.min() > 1E-6 n_cuts = len(X_mb) // cut_len + 1 partial_costs = [] for n in range(n_cuts): if n % 100 == 0: print("step %i" % n, end="") else: print(".", end="") start = n * cut_len stop = (n + 1) * cut_len if len(X_mb[start:stop]) < cut_len: # skip end edge case break rval = function(X_mb[start:stop], X_mb_mask[start:stop], prev_h1, prev_h2, prev_h3) current_cost = rval[0] prev_h1, prev_h2, prev_h3 = rval[1:4] prev_h1 = prev_h1[-1] prev_h2 = prev_h2[-1] prev_h3 = prev_h3[-1] partial_costs.append(current_cost) print("") return partial_costs
def _loop(function, itr): prev_h1, prev_h2, prev_h3 = [ np_zeros((minibatch_size, n_hid)) for i in range(3) ] X_mb, X_mb_mask = next(itr) # sanity check that masking code is OK assert X_mb_mask.min() > 1E-6 n_cuts = len(X_mb) // cut_len + 1 partial_costs = [] for n in range(n_cuts): if n % 100 == 0: print("step %i" % n, end="") else: print(".", end="") start = n * cut_len stop = (n + 1) * cut_len if len(X_mb[start:stop]) < cut_len: # skip end edge case break rval = function(X_mb[start:stop], X_mb_mask[start:stop], prev_h1, prev_h2, prev_h3) current_cost = rval[0] prev_h1, prev_h2, prev_h3 = rval[1:4] prev_h1 = prev_h1[-1] prev_h2 = prev_h2[-1] prev_h3 = prev_h3[-1] partial_costs.append(current_cost) print("") return partial_costs
checkpoint_file = args.plot if not os.path.exists(checkpoint_file): raise ValueError("Checkpoint file path %s" % checkpoint_file, " does not exist!") print(checkpoint_file) checkpoint_dict = load_checkpoint(checkpoint_file) train_costs = checkpoint_dict["train_costs"] valid_costs = checkpoint_dict["valid_costs"] plt.plot(train_costs) plt.plot(valid_costs) plt.savefig("costs.png") X_mb, X_mb_mask, c_mb, c_mb_mask = next(valid_itr) valid_itr.reset() prev_h1, prev_h2, prev_h3 = [ np_zeros((minibatch_size, n_hid)) for i in range(3) ] prev_kappa = np_zeros((minibatch_size, att_size)) prev_w = np_zeros((minibatch_size, n_chars)) if args.sample is not None: predict_function = checkpoint_dict["predict_function"] attention_function = checkpoint_dict["attention_function"] sample_function = checkpoint_dict["sample_function"] if args.write is not None: sample_string = args.write print("Sampling using sample string %s" % sample_string) oh = dense_to_one_hot( np.array([vocabulary[c] for c in sample_string]), vocabulary_size) c_mb = np.zeros( (len(oh), minibatch_size, oh.shape[-1])).astype(c_mb.dtype)
checkpoint_file = args.plot if not os.path.exists(checkpoint_file): raise ValueError("Checkpoint file path %s" % checkpoint_file, " does not exist!") print(checkpoint_file) checkpoint_dict = load_checkpoint(checkpoint_file) train_costs = checkpoint_dict["overall_train_costs"] valid_costs = checkpoint_dict["overall_valid_costs"] plt.plot(train_costs) plt.plot(valid_costs) plt.savefig("costs.png") X_mb, X_mb_mask, c_mb, c_mb_mask = next(valid_itr) valid_itr.reset() prev_h1, prev_h2, prev_h3 = [np_zeros((minibatch_size, n_hid)) for i in range(3)] prev_kappa = np_zeros((minibatch_size, att_size)) prev_w = np_zeros((minibatch_size, n_chars)) bias = args.bias if args.sample is not None: predict_function = checkpoint_dict["predict_function"] attention_function = checkpoint_dict["attention_function"] sample_function = checkpoint_dict["sample_function"] if args.write is not None: sample_string = args.write print("Sampling using sample string %s" % sample_string) oh = dense_to_one_hot( np.array([vocabulary[c] for c in sample_string]), vocabulary_size) c_mb = np.zeros(
else: checkpoint_file = args.plot if not os.path.exists(checkpoint_file): raise ValueError("Checkpoint file path %s" % checkpoint_file, " does not exist!") print(checkpoint_file) checkpoint_dict = load_checkpoint(checkpoint_file) train_costs = checkpoint_dict["train_costs"] valid_costs = checkpoint_dict["valid_costs"] plt.plot(train_costs) plt.plot(valid_costs) plt.savefig("costs.png") X_mb, X_mb_mask, c_mb, c_mb_mask = next(valid_itr) valid_itr.reset() prev_h1, prev_h2, prev_h3 = [np_zeros((minibatch_size, n_hid)) for i in range(3)] prev_kappa = np_zeros((minibatch_size, att_size)) prev_w = np_zeros((minibatch_size, n_chars)) if args.sample is not None: predict_function = checkpoint_dict["predict_function"] attention_function = checkpoint_dict["attention_function"] sample_function = checkpoint_dict["sample_function"] if args.write is not None: sample_string = args.write print("Sampling using sample string %s" % sample_string) oh = dense_to_one_hot( np.array([vocabulary[c] for c in sample_string]), vocabulary_size) c_mb = np.zeros( (len(oh), minibatch_size, oh.shape[-1])).astype(c_mb.dtype)
checkpoint_file = args.plot if not os.path.exists(checkpoint_file): raise ValueError("Checkpoint file path %s" % checkpoint_file, " does not exist!") print(checkpoint_file) checkpoint_dict = load_checkpoint(checkpoint_file) train_costs = checkpoint_dict["overall_train_costs"] valid_costs = checkpoint_dict["overall_valid_costs"] plt.plot(train_costs) plt.plot(valid_costs) plt.savefig("costs.png") X_mb, X_mb_mask, c_mb, c_mb_mask = next(train_itr) train_itr.reset() prev_h1, prev_h2, prev_h3 = [ np_zeros((minibatch_size, n_hid)) for i in range(3) ] prev_kappa = np_zeros((minibatch_size, att_size)) prev_w = np_zeros((minibatch_size, n_chars)) bias = args.bias if args.sample is not None: predict_function = checkpoint_dict["predict_function"] attention_function = checkpoint_dict["attention_function"] sample_function = checkpoint_dict["sample_function"] if args.write is not None: sample_string = args.write print("Sampling using sample string %s" % sample_string) oh = dense_to_one_hot( np.array([vocabulary[c] for c in sample_string]), vocabulary_size) c_mb = np.zeros(
import matplotlib.pyplot as plt checkpoint_file = args.sample if not os.path.exists(checkpoint_file): raise ValueError("Checkpoint file path %s" % checkpoint_file, " does not exist!") print(checkpoint_file) checkpoint_dict = load_checkpoint(checkpoint_file) train_costs = checkpoint_dict["train_costs"] valid_costs = checkpoint_dict["valid_costs"] plt.plot(train_costs) plt.plot(valid_costs) plt.savefig("costs.png") X_mb, X_mb_mask, y_mb, y_mb_mask = next(train_itr) train_itr.reset() prev_h1, prev_h2, prev_h3 = [np_zeros((minibatch_size, n_hid)) for i in range(3)] predict_function = checkpoint_dict["predict_function"] sample_function = checkpoint_dict["sample_function"] if args.sample_length is None: raise ValueError("NYI - use -sl or --sample_length ") else: fixed_steps = args.sample_length completed = [] init_x = np.zeros_like(X_mb[0]) + int(n_bins // 2) for i in range(fixed_steps): if i % 100 == 0: print("Sampling step %i" % i) # remove second init_x later
required=False) args = parser.parse_args() if args.sample is not None: import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt checkpoint_file = args.sample if not os.path.exists(checkpoint_file): raise ValueError("Checkpoint file path %s" % checkpoint_file, " does not exist!") print(checkpoint_file) checkpoint_dict = load_checkpoint(checkpoint_file) X_mb, X_mb_mask = next(train_itr) train_itr.reset() prev_h1, prev_h2, prev_h3 = [np_zeros((minibatch_size, n_hid)) for i in range(3)] predict_function = checkpoint_dict["predict_function"] sample_function = checkpoint_dict["sample_function"] if args.temperature is None: args.temperature = 1. if args.sample_length is None: raise ValueError("NYI - use -sl or --sample_length ") else: fixed_steps = args.sample_length temperature = args.temperature completed = [] # 0 is in the middle # CANNOT BE 1 timestep - will get floating point exception!
required=False) args = parser.parse_args() if args.sample is not None: import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt checkpoint_file = args.sample if not os.path.exists(checkpoint_file): raise ValueError("Checkpoint file path %s" % checkpoint_file, " does not exist!") print(checkpoint_file) checkpoint_dict = load_checkpoint(checkpoint_file) X_mb, X_mb_mask, c_mb, c_mb_mask = next(train_itr) train_itr.reset() prev_h1, prev_h2, prev_h3 = [np_zeros((minibatch_size, n_hid)) for i in range(3)] prev_kappa = np_zeros((minibatch_size, n_att_size)) prev_w = np_zeros((minibatch_size, n_chars)) predict_function = checkpoint_dict["predict_function"] sample_function = checkpoint_dict["sample_function"] if args.temperature is None: args.temperature = 1. if args.sample_length is None: raise ValueError("NYI - use -sl or --sample_length ") else: sample_string = 'apple' print("Sampling using sample string %s" % sample_string) oh = speech_ds.dense_to_one_hot(
args = parser.parse_args() if args.sample is not None: import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt checkpoint_file = args.sample if not os.path.exists(checkpoint_file): raise ValueError("Checkpoint file path %s" % checkpoint_file, " does not exist!") print(checkpoint_file) checkpoint_dict = load_checkpoint(checkpoint_file) X_mb, X_mb_mask = next(train_itr) train_itr.reset() prev_h1, prev_h2, prev_h3 = [ np_zeros((minibatch_size, n_hid)) for i in range(3) ] predict_function = checkpoint_dict["predict_function"] sample_function = checkpoint_dict["sample_function"] if args.temperature is None: args.temperature = 1. if args.sample_length is None: raise ValueError("NYI - use -sl or --sample_length ") else: fixed_steps = args.sample_length temperature = args.temperature completed = [] # 0 is in the middle # CANNOT BE 1 timestep - will get floating point exception!