code
stringlengths
17
6.64M
def extract_stats(output): regex = ' ([^:]*):(.*)' out = {} matches = re.finditer(regex, output.decode('utf-8'), re.MULTILINE) for (i, match) in enumerate(matches, start=1): (key, value) = match.groups() key = key.strip() value = value.strip() if (key == 'MACC / frame'): out['maccs_frame'] = int(value) pass elif (key == 'RAM size'): (ram, min) = value.split('(Minimum:') out['ram_usage_max'] = parse_with_unit(ram) out['ram_usage_min'] = parse_with_unit(min.rstrip(')')) pass elif (key == 'ROM size'): out['flash_usage'] = parse_with_unit(value) pass return out
def test_ram_use(): examples = [('\n AI_ARRAY_OBJ_DECLARE(\n input_1_output_array, AI_DATA_FORMAT_FLOAT, \n NULL, NULL, 1860,\n AI_STATIC)\n AI_ARRAY_OBJ_DECLARE(\n conv2d_1_output_array, AI_DATA_FORMAT_FLOAT, \n NULL, NULL, 29760,\n AI_STATIC)\n ', {'input_1_output_array': 1860, 'conv2d_1_output_array': 29760})] for (input, expected) in examples: out = extract_ram_use(input) assert (out == expected), out
def extract_ram_use(str): regex = 'AI_ARRAY_OBJ_DECLARE\\(([^)]*)\\)' matches = re.finditer(regex, str, re.MULTILINE) out = {} for (i, match) in enumerate(matches): (items,) = match.groups() items = [i.strip() for i in items.split(',')] (name, format, _, _, size, modifiers) = items out[name] = int(size) return out
def generatecode(model_path, out_path, name, model_type, compression): home = str(pathlib.Path.home()) version = os.environ.get('XCUBEAI_VERSION', '3.4.0') platform_name = platform.system().lower() if (platform_name == 'darwin'): platform_name = 'mac' p = 'STM32Cube/Repository/Packs/STMicroelectronics/X-CUBE-AI/{version}/Utilities/{os}/generatecode'.format(version=version, os=platform_name) default_path = os.path.join(home, p) cmd_path = os.environ.get('XCUBEAI_GENERATECODE', default_path) if (not os.path.exists(out_path)): os.makedirs(out_path) config = generate_config(model_path, out_path, name=name, model_type=model_type, compression=compression) config_path = os.path.join(out_path, 'config.ai') with open(config_path, 'w') as f: f.write(config) args = [cmd_path, '--auto', '-c', config_path] stdout = subprocess.check_output(args, stderr=subprocess.STDOUT) stats = extract_stats(stdout) assert len(stats.keys()), 'No model output. Stdout: {}'.format(stdout) with open(os.path.join(out_path, 'network.c'), 'r') as f: network_c = f.read() ram = extract_ram_use(network_c) stats['arrays'] = ram return stats
def parse(): parser = argparse.ArgumentParser(description='Process some integers.') a = parser.add_argument supported_types = '|'.join(model_options.keys()) a('model', metavar='PATH', type=str, help='The model to convert') a('out', metavar='DIR', type=str, help='Where to write generated output') a('--type', default='keras', help='Type of model. {}'.format(supported_types)) a('--name', default='network', help='Name of the generated network') a('--compression', default=None, type=int, help='Compression setting to use. Valid: 4|8') args = parser.parse_args() return args
def main(): args = parse() test_ram_use() stats = generatecode(args.model, args.out, name=args.name, model_type=args.type, compression=args.compression) print('Wrote model to', args.out) print('Model status: ', json.dumps(stats))
def load_model_info(jobs_dir, job_dir): (experiment, date, time, rnd, fold) = job_dir.split('-') hist_path = os.path.join(jobs_dir, job_dir, 'train.csv') df = pandas.read_csv(hist_path) df['epoch'] = (df.epoch + 1) df['fold'] = int(fold.lstrip('fold')) df['experiment'] = experiment df['run'] = '-'.join([date, time, rnd]) models = [] for fname in os.listdir(os.path.join(jobs_dir, job_dir)): if fname.endswith('model.hdf5'): models.append(fname) def get_epoch(s): e = s.split('-')[0].lstrip('e') e = int(e) return e models = sorted(models, key=get_epoch) assert models[0].startswith('e01-') last_model = models[(len(models) - 1)] expected_last = 'e{:02d}-'.format(len(models)) assert last_model.startswith(expected_last), (last_model, expected_last) df['model_path'] = [os.path.join(jobs_dir, job_dir, m) for m in models] return df
def load_train_history(jobs_dir, limit=None): jobs = os.listdir(jobs_dir) if limit: matching = [d for d in jobs if (limit in d)] else: matching = jobs dataframes = [] for job_dir in matching: try: df = load_model_info(jobs_dir, job_dir) except (FileNotFoundError, ValueError) as e: print('Failed to load job {}: {}'.format(job_dir, str(e))) continue dataframes.append(df) df = pandas.concat(dataframes) return df
def test_load_history(): jobs_dir = '../../jobs' job_id = 'sbcnn44k128aug-20190227-0220-48ba' df = load_history()
def pick_best(history, n_best=1): def best_by_loss(df): return df.sort_values('voted_val_acc', ascending=False).head(n_best) return history.groupby(['experiment', 'fold']).apply(best_by_loss)
def evaluate_model(predictor, model_path, val_data, test_data): def score(model, data): y_true = data.classID p = predictor(model, data) y_pred = numpy.argmax(p, axis=1) acc = sklearn.metrics.accuracy_score(y_true, y_pred) labels = list(range(len(urbansound8k.classnames))) confusion = sklearn.metrics.confusion_matrix(y_true, y_pred, labels=labels) return (acc, confusion) model = keras.models.load_model(model_path) salience_info = {'foreground': 1, 'background': 2} test_info = {'val': val_data, 'test': test_data} out = {} for (setname, data) in test_info.items(): for (variant, salience) in salience_info.items(): key = '{}_{}'.format(setname, variant) (acc, confusion) = score(model, data[(data.salience == salience)]) print('acc for ', key, acc) out[key] = confusion out['val'] = (out['val_foreground'] + out['val_background']) out['test'] = (out['test_foreground'] + out['test_background']) return out
def evaluate(models, folds_data, predictor, out_dir, dry_run=False): def eval_experiment(df): results = {} by_fold = df.sort_index(level='fold', ascending=True) for (idx, row) in by_fold.iterrows(): fold = row['fold'] assert (fold > 0), 'fold number should be 1 indexed' print('Testing model {} fold={}'.format(row['experiment'], fold)) model_path = row['model_path'] val = folds_data[(fold - 1)][1] test = folds_data[(fold - 1)][2] test_folds = test.fold.unique() assert (len(test_folds) == 1) assert (test_folds[0] == fold) val_folds = val.fold.unique() assert (len(val_folds) == 1) assert (val_folds[0] != fold) train_data = folds_data[(fold - 1)][0] train_files = set(train_data.slice_file_name.unique()) assert (len(train_files) > 6500), len(train_files) test_files = set(test.slice_file_name.unique()) assert (len(test_files) > 700) common_files = train_files.intersection(test_files) assert (len(common_files) == 0), common_files if dry_run: val = test[0:20] test = test[0:20] result = evaluate_model(predictor, model_path, val, test) for (k, v) in result.items(): if (results.get(k) is None): results[k] = [] results[k].append(v) exname = df['experiment'].unique()[0] results_path = os.path.join(out_dir, '{}.confusion.npz'.format(exname)) numpy.savez(results_path, **results) print('Wrote', results_path) return results_path out = models.groupby(level='experiment').apply(eval_experiment) return out
def parse(args): import argparse parser = argparse.ArgumentParser(description='Test trained models') a = parser.add_argument common.add_arguments(parser) a('--run', dest='run', default='', help='%(default)s') a('--check', action='store_true', default='', help='Run a check pass, not actually evaluating') a('--skip-stats', action='store_true', default='', help='Do not compute on-device stats') a('--out', dest='results_dir', default='./data/results', help='%(default)s') parsed = parser.parse_args(args) return parsed
def main(): args = parse(sys.argv[1:]) out_dir = os.path.join(args.results_dir, args.run) common.ensure_directories(out_dir) urbansound8k.maybe_download_dataset(args.datasets_dir) data = urbansound8k.load_dataset() folds = urbansound8k.folds(data) exsettings = common.load_settings_path(args.settings_path) frames = exsettings['frames'] voting = exsettings['voting'] overlap = exsettings['voting_overlap'] settings = features.settings(exsettings) def load_sample(sample): return features.load_sample(sample, settings, start_time=sample.start, window_frames=frames, feature_dir=args.features_dir, normalize=exsettings['normalize']) def predict(model, data): return features.predict_voted(exsettings, model, data, loader=load_sample, method=voting, overlap=overlap) history = load_train_history(args.models_dir, args.run) n_folds = len(history.fold.unique()) n_experiments = len(history.experiment.unique()) print('Found {} experiments across {} folds', n_folds, n_experiments) best = pick_best(history) print('Best models\n', best[['epoch', 'fold', 'voted_val_acc']]) print('Computing model info') def get_stats(row): ex = row.iloc[0] model = ex['model_path'] (model_stats, layer_info) = stats.model_info(model) layer_info_path = os.path.join(out_dir, '{}.layers.csv'.format(ex['experiment'])) layer_info.to_csv(layer_info_path) return pandas.Series(model_stats) if (not args.skip_stats): model_stats = best.groupby(level='experiment').apply(get_stats) print('Model stats\n', model_stats) model_stats.to_csv(os.path.join(out_dir, 'stm32stats.csv')) print('Testing models...') results = evaluate(best, folds, predictor=predict, out_dir=out_dir, dry_run=args.check)
def maybe_download_dataset(workdir): if (not os.path.exists(workdir)): os.makedirs(workdir) dir_path = os.path.join(workdir, 'UrbanSound8K') archive_path = (dir_path + '.tar.gz') last_progress = None def download_progress(count, blocksize, totalsize): nonlocal last_progress p = int((((count * blocksize) * 100) / totalsize)) if (p != last_progress): print('\r{}%'.format(p), end='\r') last_progress = p if (not os.path.exists(dir_path)): print('Could not find', dir_path) if (not os.path.exists(archive_path)): u = download_urls[0] print('Downloading...', u) urllib.request.urlretrieve(u, archive_path, reporthook=download_progress) print('Extracting...', archive_path) with tarfile.open(archive_path, 'r:gz') as archive: archive.extractall(workdir) return dir_path
def load_dataset(): metadata_path = os.path.join(here, 'datasets/UrbanSound8K.csv') samples = pandas.read_csv(metadata_path) return samples
def sample_path(sample, dataset_path=None): if (not dataset_path): dataset_path = default_path return os.path.join(dataset_path, 'audio', ('fold' + str(sample.fold)), sample.slice_file_name)
def folds(data): fold_idxs = folds_idx(n_folds=10) assert (len(fold_idxs) == 10) folds = [] for fold in fold_idxs: (train, val, test) = fold train = (numpy.array(train) + 1) val = (numpy.array(val) + 1) test = (numpy.array(test) + 1) fold_train = data[data.fold.isin(train)] fold_val = data[data.fold.isin(val)] fold_test = data[data.fold.isin(test)] train_folds = set(fold_train.fold.unique()) val_folds = set(fold_val.fold.unique()) test_folds = set(fold_test.fold.unique()) assert (len(train_folds) == 8), len(train_folds) assert (train_folds.intersection(val_folds) == set()) assert (train_folds.intersection(test_folds) == set()) assert (val_folds.intersection(test_folds) == set()) folds.append((fold_train, fold_val, fold_test)) return folds
def ensure_valid_fold(fold, n_folds=10): (train, val, test) = fold assert (len(train) == (n_folds - 2)), len(train) assert (0 <= train[0] < n_folds), train[0] assert (len(val) == 1), len(val) assert (0 <= val[0] < n_folds), val[0] assert (len(test) == 1), len(test) assert (0 <= test[0] < n_folds), test[0] assert (test[0] != val[0]) test_overlap = set(train).intersection(set(test)) val_overlap = set(train).intersection(set(val)) assert (test_overlap == set()), test_overlap assert (val_overlap == set()), val_overlap assert (sorted(((train + val) + test)) == list(range(0, n_folds))) return True
def folds_idx(n_folds): 'Generate fold indices for cross-validation.\n Each fold has 1 validation, 1 test set and the remaining train' test_fold = 10 folds = [] all_folds = list(range(0, n_folds)) for idx in range(0, n_folds): test = [all_folds[idx]] val = [all_folds[(idx - 1)]] train = list(set(all_folds).difference(set((test + val)))) fold = (train, val, test) ensure_valid_fold(fold) folds.append(fold) assert (len(folds) == n_folds), len(folds) return folds
def sbcnn_generator(n_iter=400, random_state=1): from sklearn.model_selection import ParameterSampler params = dict(kernel_t=range(3, 10, 2), kernel_f=range(3, 10, 2), pool_t=range(2, 5), pool_f=range(2, 5), kernels_start=range(16, 64), fully_connected=range(16, 128)) sampler = ParameterSampler(params, n_iter=n_iter, random_state=random_state) out_models = [] out_total_params = [] for p in sampler: s = {'model': 'sbcnn', 'frames': 31, 'n_mels': 60, 'samplerate': 22050} pool = (p['pool_f'], p['pool_t']) kernel = (p['kernel_f'], p['kernel_t']) for (k, v) in p.items(): s[k] = v s['pool'] = pool s['kernel'] = kernel (yield (p, s))
def generate_models(): gen = sbcnn_generator() data = {'model_path': [], 'gen_path': [], 'id': []} for out in iter(gen): model = None try: (params, settings) = out model = models.build(settings.copy()) except ValueError as e: print('Error:', e) continue for (k, v) in params.items(): if (data.get(k) is None): data[k] = [] data[k].append(v) model_id = str(uuid.uuid4()) out_dir = os.path.join('scan', model_id) os.makedirs(out_dir) model_path = os.path.join(out_dir, 'model.orig.hdf5') out_path = os.path.join(out_dir, 'gen') model.save(model_path) stats = stm32convert.generatecode(model_path, out_path, name='network', model_type='keras', compression=None) data['model_path'].append(model_path) data['gen_path'].append(out_path) data['id'].append(model_id) for (k, v) in stats.items(): if (data.get(k) is None): data[k] = [] data[k].append(v) df = pandas.DataFrame(data) return df
def main(): df = generate_models() df.to_csv('scan.csv')
def main(): settings = common.load_experiment('experiments', 'ldcnn20k60') def build(): return train.sb_cnn(settings) m = build() m.summary() m.save('model.wip.hdf5') s = settings shape = (s['n_mels'], s['frames'], 1) model_stats = stats.analyze_model(build, [shape], n_classes=10) (flops, params) = model_stats inference_flops = {name: v for (name, v) in flops.items() if (not stats.is_training_scope(name))} for (name, flop) in inference_flops.items(): print(name, flop)
def check_missing(df, field, name='name'): missing = df[df[field].isna()] if len(missing): print('WARN. Missing "{}" for {}'.format(field, list(missing[name])))
def logmel_models(data_path): df = pandas.read_csv(data_path) df = df[df['features'].str.contains('logmel')] df.index = df['name'] df['params'] = (df['kparams'] * 1000.0) df['window'] = ((df.frames * df.hop) / df.samplerate) df['t_step'] = (df.hop / df.samplerate) df['f_res'] = (df.samplerate / df.bands) df['macc_s'] = ((df['mmacc'] * 1000000.0) / df.window) return df
def model_table(data_path): df = logmel_models(data_path) table = pandas.DataFrame() table['Accuracy (%)'] = (df.accuracy * 100) table['MACC / second'] = ['{} M'.format(int((v / 1000000.0))) for v in df.macc_s] table['Model parameters'] = ['{} k'.format(int((v / 1000.0))) for v in df.params] table = table.sort_values('Accuracy (%)', ascending=False) return table.to_latex(column_format='lrrr')
def plot_models(data_path, figsize=(12, 4), max_params=128000.0, max_maccs=4500000.0): df = logmel_models(data_path) (fig, ax) = plt.subplots(1, figsize=figsize) check_missing(df, 'accuracy') check_missing(df, 'kparams') check_missing(df, 'mmacc') df.plot.scatter(x='params', y='macc_s', logx=True, logy=True, ax=ax) ax.set_xlabel('Model parameters') ax.set_ylabel('MACC / second') feasible_x = max_params feasible_y = max_maccs x = [0, feasible_x, feasible_x, 0] y = [0, 0, feasible_y, feasible_y] ax.fill(x, y, color='green', alpha=0.5) linestyle = dict(color='black', linewidth=0.5) ax.axvline(feasible_x, **linestyle) ax.axhline(feasible_y, **linestyle) def add_labels(row): xy = (row.params, row.macc_s) label = '{} {:.1f}%'.format(row['name'], (100 * row.accuracy)) ax.annotate(label, xy, xytext=(5, 40), textcoords='offset points', size=12, rotation=25, color='darkslategrey') df.apply(add_labels, axis=1) fig.tight_layout() return fig
def flatten(list): out = [] for x in list: for y in x: out.append(y) return out
def plot_spectrogram(f, ax=None, cmap=None): (y, sr) = librosa.load(f, sr=44100) fig = None if (not ax): (fig, ax) = plt.subplots(1, figsize=(16, 4)) S = numpy.abs(librosa.stft(y)) S = librosa.amplitude_to_db(S, ref=numpy.max) kwargs = dict(ax=ax, y_axis='log', x_axis='time', sr=sr) if (cmap is not None): kwargs['cmap'] = cmap librosa.display.specshow(S, **kwargs) return fig
def plot_spectrograms(files, titles, out=None): assert (len(files) == len(titles)) (fig, axs) = plt.subplots(2, (len(files) // 2), sharex=True, figsize=(16, 6)) axs = flatten(axs) for (i, (path, title, ax)) in enumerate(zip(files, titles, axs)): plot_spectrogram(path, ax=ax) ax.set_title(title) if ((i != 0) and (i != (len(files) / 2))): ax.set_ylabel('') ax.set_yticks([]) if (i < (len(files) / 2)): ax.set_xlabel('') if out: fig.savefig(out, bbox_inches='tight', pad_inches=0) return fig
def plot_examples(examples): examples = urbansound8k_examples here = os.path.dirname(__file__) base = os.path.join(here, '../microesc/../data/datasets/UrbanSound8K/audio/') paths = [os.path.join(base, e[0]) for e in examples.values()] fig = plot_spectrograms(paths, examples.keys()) return fig
def main(): plotname = os.path.basename(sys.argv[1]) here = os.path.dirname(__file__) plot_func = plots.get(plotname, None) if (not plot_func): sys.stderr.write('Plot {} not found. Supported: \n{}'.format(plotname, plots.keys())) return 1 out = plot_func() out_path = os.path.join(here, 'plots', plotname) if (not os.path.exists(os.path.dirname(out_path))): os.makedirs(os.path.dirname(out_path)) ext = os.path.splitext(plotname)[1] if (ext == '.png'): out.savefig(out_path, bbox_inches='tight') elif (ext == '.tex'): with open(out_path, 'w') as f: f.write(out) else: raise ValueError('Unknown extension {}'.format(ext))
def strformat(fmt, series): return [fmt.format(i) for i in series]
def downsample_from_name(name): if name.startswith('Stride'): return 'stride' else: return 'maxpool'
def accuracies(df, col): mean = (df[(col + '_mean')] * 100) std = (df[(col + '_std')] * 100) fmt = ['{:.1f}% +-{:.1f}'.format(*t) for t in zip(mean, std)] return fmt
def cpu_use(df): usage = (((df.utilization * 1000) * 1) / df.classifications_per_second).astype(int) return ['{:d} ms'.format(i).ljust(3) for i in usage]
def plot_augmentations(y, sr, time_shift=3000, pitch_shift=12, time_stretch=1.3): augmentations = {'Original': y, 'Timeshift left': y[time_shift:], 'Timeshift right': numpy.concatenate([numpy.zeros(time_shift), y[:(- time_shift)]]), 'Timestretch faster': librosa.effects.time_stretch(y, time_stretch), 'Timestretch slower': librosa.effects.time_stretch(y, (1 / time_stretch)), 'Pitchshift up': librosa.effects.pitch_shift(y, sr, pitch_shift), 'Pitchshift down': librosa.effects.pitch_shift(y, sr, (- pitch_shift))} layout = [['Original', 'Original', 'Original'], ['Timeshift right', 'Timestretch faster', 'Pitchshift up'], ['Timeshift left', 'Timestretch slower', 'Pitchshift down']] shape = numpy.array(layout).shape (fig, axs) = plt.subplots(shape[0], shape[1], figsize=(16, 6), sharex=True) for row in range(shape[0]): for col in range(shape[1]): description = layout[row][col] ax = axs[row][col] data = augmentations[description] S = numpy.abs(librosa.stft(data)) S = scipy.ndimage.filters.gaussian_filter(S, 0.7) S = librosa.amplitude_to_db(S, ref=numpy.max) S -= S.mean() librosa.display.specshow(S, ax=ax, sr=sr, y_axis='hz') ax.set_ylim(0, 5000) ax.set_title(description) return fig
def main(): path = '163459__littlebigsounds__lbs-fx-dog-small-alert-bark001.wav' (y, sr) = librosa.load(path, offset=0.1, duration=1.2) fig = plot_augmentations(y, sr) out = __file__.replace('.py', '.png') fig.savefig(out, bbox_inches='tight')
def bandpass_filter(lowcut, highcut, fs, order, output='sos'): assert ((order % 2) == 0), 'order must be multiple of 2' assert ((highcut * 0.95) < (fs / 2.0)), 'highcut {} above Nyquist for fs={}'.format(highcut, fs) assert (lowcut > 0.0), 'lowcut must be above 0' nyq = (0.5 * fs) low = (lowcut / nyq) high = min((highcut / nyq), 0.99) output = scipy.signal.butter((order / 2), [low, high], btype='band', output=output) return output
def filterbank(center, fraction, fs, order): reference = acoustics.octave.REFERENCE center = [f for f in center if (f < (fs / 2.0))] center = numpy.asarray(center) indices = acoustics.octave.index_of_frequency(center, fraction=fraction, ref=reference) center = acoustics.octave.exact_center_frequency(None, fraction=fraction, n=indices, ref=reference) lower = acoustics.octave.lower_frequency(center, fraction=fraction) upper = acoustics.octave.upper_frequency(center, fraction=fraction) nominal = acoustics.octave.nominal_center_frequency(None, fraction, indices) def f(low, high): return bandpass_filter(low, high, fs=fs, order=order) filterbank = [f(low, high) for (low, high) in zip(lower, upper)] return (nominal, filterbank)
def third_octave_filterbank(fs, order=8): from acoustics.standards import iec_61672_1_2013 as iec_61672 center = iec_61672.NOMINAL_THIRD_OCTAVE_CENTER_FREQUENCIES return filterbank(center, fraction=3, fs=fs, order=order)
def plot_filterbank_oct(ax, fs=44100): filterbank = third_octave_filterbank((fs / 2)) for (center, sos) in zip(filterbank[0], filterbank[1]): (w, h) = scipy.signal.sosfreqz(sos, worN=4096, fs=fs) db = (20 * numpy.log10((numpy.abs(h) + 1e-09))) ax.plot(w, db) ax.set_title('1/3 octave') ax.set_ylabel('Attenuation (dB)') ax.set_ylim((- 60), 5) ax.set_xlim(20.0, 20000.0)
def plot_filterbank_gammatone(ax, fs=44100): np = numpy from pyfilterbank import gammatone gfb = gammatone.GammatoneFilterbank(samplerate=44100, startband=(- 6), endband=26, density=1.5) def plotfun(x, y): xx = (x * fs) ax.plot(xx, (20 * np.log10((np.abs(y) + 1e-09)))) gfb.freqz(nfft=(2 * 4096), plotfun=plotfun) ax.set_title('Gammatone') ax.set_ylim([(- 80), 1]) ax.set_xlim([10, 20000.0]) return gfb
def plot_filterbank_mel(ax, n_mels=32, n_fft=4097, fmin=10, fmax=22050, fs=44100): from pyfilterbank import melbank (melmat, (melfreq, fftfreq)) = melbank.compute_melmat(n_mels, fmin, fmax, num_fft_bands=4097, sample_rate=fs) ax.plot(fftfreq, (20 * numpy.log10((melmat.T + 1e-09)))) ax.set_title('Mel-scale') ax.set_xlabel('Frequency (Hz)')
def main(): (fig, (gt_ax, oct_ax, mel_ax)) = plt.subplots(3, sharex=True, sharey=True, figsize=(12, 5)) axes = fig.gca() plot_filterbank_gammatone(gt_ax) plot_filterbank_mel(mel_ax) plot_filterbank_oct(oct_ax) axes.set_ylim([(- 40), 3]) axes.set_xlim([100, 20000]) fig.tight_layout() out = __file__.replace('.py', '.png') fig.savefig(out, bbox_inches='tight')
def plot_logloss(figsize=(6, 3)): (fig, ax) = plt.subplots(1, figsize=figsize) yhat = numpy.linspace(0.0, 1.0, 300) losses_0 = [log_loss([0], [x], labels=[0, 1]) for x in yhat] losses_1 = [log_loss([1], [x], labels=[0, 1]) for x in yhat] ax.plot(yhat, losses_0, label='true=0') ax.plot(yhat, losses_1, label='true=1') ax.legend() ax.set_ylim(0, 8) ax.set_xlim(0, 1) return fig
def main(): fig = plot_logloss() fig.tight_layout() out = __file__.replace('.py', '.png') fig.savefig(out, bbox_inches='tight')
def arglist(options): args = ['--{}={}'.format(k, v) for (k, v) in options.items()] return args
def command_for_job(options): args = ['python3', 'train.py'] args += arglist(options) return args
def generate_train_jobs(experiments, settings_path, folds, overrides): timestamp = datetime.datetime.now().strftime('%Y%m%d-%H%M') unique = str(uuid.uuid4())[0:4] def name(experiment, fold): name = '-'.join([experiment, timestamp, unique]) return (name + '-fold{}'.format(fold)) def job(exname, experiment): for fold in folds: n = name(exname, fold) options = {'name': n, 'fold': fold, 'settings': settings_path} for (k, v) in experiment.items(): options[k] = v for (k, v) in overrides.items(): options[k] = v cmd = command_for_job(options) return cmd jobs = [job(str(idx), ex) for (idx, ex) in experiments.iterrows()] return jobs
def parse(args): import argparse parser = argparse.ArgumentParser(description='Generate jobs') a = parser.add_argument a('--models', default='models.csv', help='%(default)s') a('--settings', default='experiments/ldcnn20k60.yaml', help='%(default)s') a('--jobs', dest='jobs_dir', default='./data/jobs', help='%(default)s') a('--check', action='store_true', help='Only run a pre-flight check') parsed = parser.parse_args(args) return parsed
def main(): args = parse(sys.argv[1:]) models = pandas.read_csv(args.models) settings = common.load_settings_path(args.settings) overrides = {} folds = list(range(0, 9)) if args.check: folds = (1,) overrides['train_samples'] = (settings['batch'] * 1) overrides['val_samples'] = (settings['batch'] * 1) cmds = generate_train_jobs(models, args.settings, folds, overrides) print('\n'.join((' '.join(cmd) for cmd in cmds)))
@pytest.mark.skip('fails right now') @pytest.mark.parametrize('family', FAMILIES) def test_models_basic(family): s = settings.load_settings({'model': family, 'frames': 31, 'n_mels': 60, 'samplerate': 22050}) if (family == 'sbcnn'): s['downsample_size'] = (3, 2) s['conv_size'] = (3, 3) if (family == 'strided'): s['downsample_size'] = (3, 3) s['conv_size'] = (3, 3) s['conv_block'] = 'conv' s['filters'] = 12 m = models.build(s) assert isinstance(m, keras.Model)
@pytest.mark.parametrize('conv_type', CONV_TYPES) def test_strided_variations(conv_type): s = settings.load_settings({'model': 'strided', 'frames': 31, 'n_mels': 60, 'samplerate': 22050, 'conv_block': conv_type, 'filters': 20}) s['conv_size'] = (3, 3) s['downsample_size'] = (2, 2) m = models.build(s) assert isinstance(m, keras.Model)
def test_conv_ds(): k = (5, 5) i = (60, 31, 16) ch = 16 conv = stats.compute_conv2d(*i, ch, *k) ds = stats.compute_conv2d_ds(*i, ch, *k) ratio = (conv / ds) assert (ratio > 9.0)
def test_conv_ds3x3(): k = (3, 3) i = (60, 31, 64) ch = 64 conv = stats.compute_conv2d(*i, ch, *k) ds = stats.compute_conv2d_ds(*i, ch, *k) ratio = (conv / ds) assert (ratio > 7.5)
@pytest.mark.skip('fails') def test_generator_fake_loader(): dataset_path = 'data/UrbanSound8K/' urbansound8k.default_path = dataset_path data = urbansound8k.load_dataset() (folds, test) = urbansound8k.folds(data) data_length = 16 batch_size = 8 frames = 72 bands = 32 n_classes = 10 def zero_loader(s): return numpy.zeros((bands, frames, 1)) fold = folds[0][0] X = fold[0:data_length] Y = fold.classID[0:data_length] g = train.dataframe_generator(X, Y, loader=zero_loader, batchsize=batch_size, n_classes=n_classes) n_batches = 3 batches = list(itertools.islice(g, n_batches)) assert (len(batches) == n_batches) assert (len(batches[0]) == 2) assert (batches[0][0].shape == (batch_size, bands, frames, 1)) assert (batches[0][1].shape == (batch_size, n_classes))
def test_windows_shorter_than_window(): frame_samples = 256 window_frames = 64 fs = 16000 length = (0.4 * fs) w = list(features.sample_windows(int(length), frame_samples, window_frames)) assert (len(w) == 1), len(w) assert (w[(- 1)][1] == length)
def test_window_typical(): frame_samples = 256 window_frames = 64 fs = 16000 length = (4.0 * fs) w = list(features.sample_windows(int(length), frame_samples, window_frames)) assert (len(w) == 8), len(w) assert (w[(- 1)][1] == length)
def _test_predict_windowed(): t = test[0:10] sbcnn16k32_settings = dict(feature='mels', samplerate=16000, n_mels=32, fmin=0, fmax=8000, n_fft=512, hop_length=256, augmentations=5) def load_sample32(sample): return features.load_sample(sample, sbcnn16k32_settings, window_frames=72, feature_dir='../../scratch/aug') mean_m = features.predict_voted(sbcnn16k32_settings, model, t, loader=load_sample32, method='mean')
def test_precompute(): settings = dict(feature='mels', samplerate=16000, n_mels=32, fmin=0, fmax=8000, n_fft=512, hop_length=256, augmentations=12) dir = './pre2' if os.path.exists(dir): shutil.rmtree(dir) workdir = os.path.abspath(os.path.join(os.path.dirname(__file__), '../data/')) data = urbansound8k.load_dataset() urbansound8k.maybe_download_dataset(workdir) d = os.path.join(dir, features.settings_id(settings)) expect_path = features.feature_path(data.iloc[0], d) assert (not os.path.exists(expect_path)), expect_path preprocess.precompute(data[0:4], settings, out_dir=d, verbose=0, force=True, n_jobs=2) assert os.path.exists(expect_path), expect_path
def test_grouped_confusion(): cm = numpy.array([[82, 0, 3, 0, 0, 10, 0, 4, 1, 0], [3, 29, 0, 0, 0, 0, 1, 0, 0, 0], [4, 3, 37, 14, 4, 4, 0, 0, 2, 32], [5, 2, 5, 78, 4, 0, 0, 0, 0, 6], [23, 2, 4, 1, 55, 4, 2, 6, 3, 0], [9, 0, 0, 4, 3, 70, 0, 5, 1, 1], [0, 0, 0, 5, 0, 0, 27, 0, 0, 0], [0, 0, 2, 0, 1, 1, 1, 91, 0, 0], [9, 11, 9, 4, 0, 1, 0, 0, 46, 3], [1, 7, 7, 0, 7, 0, 0, 0, 3, 75]]) (gcm, gnames) = report.grouped_confusion(cm, report.groups) assert (numpy.sum(cm) == numpy.sum(gcm)) assert (gnames[0] == 'social_activity') assert (gnames[3] == 'domestic_machines') expect_correct_social = (((((37 + 78) + 75) + (14 + 32)) + (5 + 6)) + (7 + 0)) assert (gcm[0][0] == expect_correct_social), (gcm[0][0], expect_correct_social) assert (gcm[3][3] == 82)
@pytest.mark.parametrize('example', CORRECT_FOLDS.keys()) def test_ensure_valid_fold_passes_correct(example): fold = CORRECT_FOLDS[example] folds.ensure_valid_fold(fold)
@pytest.mark.parametrize('example', WRONG_FOLDS.keys()) def test_ensure_valid_fold_detects_wrong(example): fold = WRONG_FOLDS[example] with pytest.raises(AssertionError) as e_info: folds.ensure_valid_fold(fold)
def test_folds_idx(): f = folds.folds_idx(10) print(('\n' + '\n'.join([str(i) for i in f]))) assert (f[0][2][0] == 0), 'first test fold should be 0' assert (f[(- 1)][2][0] == 9), 'last test fold should be 9'
def test_folds(): data = urbansound8k.load_dataset() f = urbansound8k.folds(data) assert (len(f) == 10)
class ConveRTModelConfig(NamedTuple): num_embed_hidden: int = 512 feed_forward1_hidden: int = 2048 feed_forward2_hidden: int = 1024 num_attention_project: int = 64 vocab_size: int = 25000 num_encoder_layers: int = 6 dropout_rate: float = 0.0 n: int = 121 relative_attns: list = [3, 5, 48, 48, 48, 48] num_attention_heads: int = 2 token_sequence_truncation: int = 60
class ConveRTTrainConfig(NamedTuple): sp_model_path: str = os.path.join(dirname, 'data/en.wiki.bpe.vs25000.model') dataset_path: str = os.path.join(dirname, 'data/sample-dataset.json') test_dataset_path: str = 'data/sample-dataset.json' model_save_dir: str = 'lightning_logs/checkpoints/' log_dir: str = 'lightning_logs' device: str = 'cpu' use_data_paraller: bool = True is_reddit: bool = True train_batch_size: int = 64 test_batch_size: int = 256 split_size: int = 8 learning_rate: float = 0.001 lr_warmup_start: float = 0.1 lr_warmup_end: float = 1.0 warmup_batch: float = 10000 final_batch: float = 100000000.0 learning_rate_end: float = 0.0001 epochs: int = 10 grad_norm_clip: float = 1.0 smoothing: float = 0.2 l2_weight_decay: float = 1e-05
class LossFunction(nn.Module): @staticmethod def cosine_similarity_matrix(context_embed: torch.Tensor, reply_embed: torch.Tensor) -> torch.Tensor: assert (context_embed.size(0) == reply_embed.size(0)) cosine_similarity = torch.matmul(context_embed, reply_embed.T) return cosine_similarity def forward(self, context_embed: torch.Tensor, reply_embed: torch.Tensor) -> torch.Tensor: cosine_similarity = self.cosine_similarity_matrix(context_embed, reply_embed) j = (- torch.sum(torch.diagonal(cosine_similarity))) cosine_similarity.diagonal().copy_(torch.zeros(cosine_similarity.size(0))) j = ((0.8 * j) + ((0.2 / (cosine_similarity.size(0) * (cosine_similarity.size(0) - 1))) * torch.sum(cosine_similarity))) j += torch.sum(torch.logsumexp(cosine_similarity, dim=0)) return j
@dataclass class EncoderInputFeature(): input_ids: torch.Tensor attention_mask: torch.Tensor position_ids: torch.Tensor input_lengths: torch.Tensor def pad_sequence(self, seq_len: int): self.input_ids = pad(self.input_ids, [0, (seq_len - self.input_ids.size(0))], 'constant', 0) self.attention_mask = pad(self.attention_mask, [0, (seq_len - self.attention_mask.size(0))], 'constant', 0) self.position_ids = pad(self.position_ids, [0, (seq_len - self.position_ids.size(0))], 'constant', 0)
@dataclass class EmbeddingPair(): context: EncoderInputFeature reply: EncoderInputFeature
class DataModule(pl.LightningDataModule): def __init__(self): super().__init__() self.input_attributes = ['input_ids', 'attention_mask', 'position_ids', 'input_lengths'] def batching_input_features(self, encoder_inputs: List[EncoderInputFeature]) -> EncoderInputFeature: max_seq_len = max([int(encoder_input.input_lengths.item()) for encoder_input in encoder_inputs]) for encoder_input in encoder_inputs: encoder_input.pad_sequence(max_seq_len) batch_features = {feature_name: torch.stack([getattr(encoder_input, feature_name) for encoder_input in encoder_inputs], dim=0) for feature_name in self.input_attributes} return EncoderInputFeature(**batch_features) def convert_collate_fn(self, features: List[EmbeddingPair]) -> EmbeddingPair: return EmbeddingPair(context=self.batching_input_features([feature.context for feature in features]), reply=self.batching_input_features([feature.reply for feature in features])) def train_dataloader(self, train_dataset): return DataLoader(train_dataset, config.train_batch_size, collate_fn=self.convert_collate_fn, drop_last=True) def val_dataloader(self): pass def test_dataloader(self): pass
class DatasetInstance(NamedTuple): context: List[str] response: str
def load_instances_from_reddit_json(dataset_path: str) -> List[DatasetInstance]: instances: List[DatasetInstance] = [] with open(dataset_path) as f: for line in f: x = json.loads(line) context_keys = sorted([key for key in x.keys() if ('context' in key)]) instance = DatasetInstance(context=[x[key] for key in context_keys], response=x['response']) instances.append(instance) return instances
class RedditData(torch.utils.data.Dataset): def __init__(self, instances: List[DatasetInstance], sp_processor: SentencePieceProcessor, truncation_length: int): self.sp_processor = sp_processor self.instances = instances self.truncation_length = truncation_length def __len__(self): return len(self.instances) def __getitem__(self, item): context_str = self.instances[item].context[0] context_embedding = self._convert_instance_to_embedding(context_str) reply_embedding = self._convert_instance_to_embedding(self.instances[item].response) return EmbeddingPair(context=context_embedding, reply=reply_embedding) def _convert_instance_to_embedding(self, input_str: str) -> EncoderInputFeature: input_ids = self.sp_processor.EncodeAsIds(input_str) if self.truncation_length: input_ids = input_ids[:self.truncation_length] attention_mask = [1 for _ in range(len(input_ids))] position_ids = [i for i in range(len(input_ids))] return EncoderInputFeature(input_ids=torch.tensor(input_ids).to(config.device), attention_mask=torch.tensor(attention_mask).to(config.device), position_ids=torch.tensor(position_ids).to(config.device), input_lengths=torch.tensor(len(input_ids)).to(config.device))
class LearningRateDecayCallback(pl.Callback): def __init__(self, config, lr_decay=True): super().__init__() self.lr_warmup_end = config.lr_warmup_end self.lr_warmup_start = config.lr_warmup_start self.learning_rate = config.learning_rate self.warmup_batch = config.warmup_batch self.final_batch = config.final_batch self.lr_decay = lr_decay def on_train_batch_end(self, trainer, pl_module, batch, batch_idx, dataloader_idx): '\n\n :param trainer:\n :type trainer:\n :param pl_module:\n :type pl_module:\n :param batch:\n :type batch:\n :param batch_idx:\n :type batch_idx:\n :param dataloader_idx:\n :type dataloader_idx:\n ' optimizer = trainer.optimizers[0] if self.lr_decay: if (batch_idx < self.warmup_batch): lr_mult = (float(batch_idx) / float(max(1, self.warmup_batch))) lr = (self.lr_warmup_start + (lr_mult * (self.lr_warmup_end - self.lr_warmup_start))) else: progress = (float((batch_idx - self.warmup_batch)) / float(max(1, (self.final_batch - self.warmup_batch)))) lr = max((self.learning_rate + ((0.5 * (1.0 + math.cos((math.pi * progress)))) * (self.lr_warmup_end - self.learning_rate))), self.learning_rate) for param_group in optimizer.param_groups: param_group['lr'] = lr
def set_seed(seed): random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed_all(seed)
def find_subword_params(model): 'Long winded helper fn to return Subword Embedding Params for clipping, as they are the only parameters that\n are gradient clipped in the paper, only calculated once after model instantiation, but before training' embeds = set() for (mn, m) in model.named_modules(): for (pn, p) in m.named_parameters(): if mn.startswith('transformer_layers.subword_embedding'): fpn = (('%s.%s' % (mn, pn)) if mn else pn) embeds.add(fpn) param_dict = {pn: p for (pn, p) in model.named_parameters()} return ([param_dict[pn] for pn in sorted(list(embeds))], embeds)
class SingleContextConvert(pl.LightningModule): def __init__(self, model_config: ConveRTModelConfig, train_config: ConveRTTrainConfig): super().__init__() self.model_config = model_config self.train_config = train_config self.transformer_layers = TransformerLayers(model_config) self.ff2_context = FeedForward2(model_config) self.ff2_reply = FeedForward2(model_config) self.loss_function = LossFunction() self.weight_decay = train_config.l2_weight_decay self.hparams = self.train_config._field_defaults self.hparams.update(self.model_config._field_defaults) self.subword_params = None logger.info('number of parameters: %e', sum((p.numel() for p in self.parameters()))) def register_subword_params(self): self.subword_params = find_subword_params(self)[0] def forward(self, x): return self.transformer_layers(x) def backward(self, trainer, loss, optimizer, optimizer_idx): 'override hook of lightning as want specific grad norm clip of only subword embedding parameters, after loss.backward()\n but before optimizer step' loss.backward() torch.nn.utils.clip_grad_norm_(self.subword_params, self.train_config.grad_norm_clip) def configure_optimizers(self): '\n here I did not implement weight decay on bias and Layernorm layers as is typical in modern NLP papers.\n I do not think the paper specified params to avoid weight decay on\n :return:\n :rtype:\n ' no_decay = ['bias', 'LayerNorm.weight'] params_decay = [p for (n, p) in self.named_parameters() if (not any(((nd in n) for nd in no_decay)))] params_nodecay = [p for (n, p) in self.named_parameters() if any(((nd in n) for nd in no_decay))] optim_groups = [{'params': params_decay, 'weight_decay': self.hparams.l2_weight_decay}, {'params': params_nodecay, 'weight_decay': 0.0}] optimizer = torch.optim.AdamW(optim_groups, lr=self.hparams.learning_rate) return optimizer def training_step(self, batch, batch_idx): batch_context = batch.context batch_reply = batch.reply rx = self(batch_context) ry = self(batch_reply) hx = self.ff2_context(rx, batch_context.attention_mask) hy = self.ff2_reply(ry, batch_reply.attention_mask) loss = self.loss_function(hx, hy) tqdm_dict = {'train_loss': loss} output = OrderedDict({'loss': loss, 'progress_bar': tqdm_dict, 'log': tqdm_dict}) return output def validation_step(self, batch, batch_idx): output = self.training_step(batch, batch_idx) val_output = {'val_loss': output['loss']} return val_output
def _parse_args(): 'Parse command-line arguments.' parser = argparse.ArgumentParser() parser.add_argument('--progress_bar_refresh_rate', type=int, default=1) parser.add_argument('--row_log_interval', type=int, default=1) args = parser.parse_args() return args
def main(**kwargs): set_seed(1) train_config = ConveRTTrainConfig() model_config = ConveRTModelConfig() tokenizer = SentencePieceProcessor() args = _parse_args() tokenizer.Load(train_config.sp_model_path) train_instances = load_instances_from_reddit_json(train_config.dataset_path) RD = RedditData(train_instances, tokenizer, 60) dm = DataModule() train_loader = dm.train_dataloader(RD) model = SingleContextConvert(model_config, train_config) lr_decay = LearningRateDecayCallback(train_config) model.register_subword_params() trainer = pl.Trainer.from_argparse_args(args, callbacks=[lr_decay], **kwargs) trainer.fit(model, train_dataloader=train_loader, val_dataloaders=train_loader)
@pytest.fixture def config(): return ConveRTTrainConfig()
@pytest.fixture def tokenizer() -> SentencePieceProcessor: tokenizer = SentencePieceProcessor() tokenizer.Load(config.sp_model_path) return tokenizer
def test_load_instances_from_reddit_json(config): instances = load_instances_from_reddit_json(config.dataset_path) assert (len(instances) == 1000)
class TestModelTraining(unittest.TestCase): 'Check can overfit small batch etc. without issues' def test_fast_dev_run(self): t = time() try: main(fast_dev_run=True) except: self.fail('Obvious Training Problem!') time_taken = (time() - t) self.assertLess(time_taken, 10)
@pytest.fixture def model_config(): return ConveRTModelConfig()
@pytest.fixture def train_config(): return ConveRTTrainConfig(train_batch_size=64, split_size=8, learning_rate=2e-05)
def test_circulant_t(): assert (circulant_mask(50, 47).sum().item() == 2494) try: circulant_mask(47, 50) circulant_mask(47, 47) circulant_mask(47, 45) except ExceptionType: self.fail('ciculant_t Failed')
def test_SubwordEmbedding(train_config, model_config): embedding = SubwordEmbedding(model_config) input_token_ids = torch.randint(high=model_config.vocab_size, size=(train_config.train_batch_size, SEQ_LEN)) positional_input = torch.randint(high=model_config.vocab_size, size=(train_config.train_batch_size, SEQ_LEN)) embedding_output = embedding(input_ids=input_token_ids, position_ids=positional_input) assert (embedding_output.size() == (train_config.train_batch_size, SEQ_LEN, model_config.num_embed_hidden))
def test_SelfAttention(model_config, train_config): attention = SelfAttention(model_config, relative_attention) query = torch.rand(train_config.train_batch_size, SEQ_LEN, model_config.num_embed_hidden) attn_mask = torch.ones(query.size()[:(- 1)], dtype=torch.float) output = attention(query, attn_mask) assert (output.size() == (train_config.train_batch_size, SEQ_LEN, model_config.num_embed_hidden))
def test_FeedForward1(train_config, model_config): ff1 = FeedForward1(model_config.num_embed_hidden, model_config.feed_forward1_hidden, model_config.dropout_rate) embed = torch.rand(train_config.train_batch_size, SEQ_LEN, model_config.num_embed_hidden) output = ff1(embed) assert (output.size() == embed.size())
def test_SharedInnerBlock(train_config, model_config): from random import randrange SIB = SharedInnerBlock(model_config, model_config.relative_attns[randrange(6)]) embed = torch.rand(train_config.train_batch_size, SEQ_LEN, model_config.num_embed_hidden) attn_mask = torch.ones(embed.size()[:(- 1)], dtype=torch.float) out1 = SIB(embed, attn_mask) assert (out1.size() == embed.size())
def test_MultiheadAttention(train_config, model_config): MHA = MultiheadAttention(model_config) embed = torch.rand(train_config.train_batch_size, SEQ_LEN, model_config.num_embed_hidden) attn_mask = torch.ones(embed.size()[:(- 1)], dtype=torch.float) assert ((model_config.num_embed_hidden % MHA.num_attention_heads) == 0) assert (MHA(embed, attn_mask).size() == (train_config.train_batch_size, SEQ_LEN, (model_config.num_embed_hidden * model_config.num_attention_heads)))
def test_TransformerLayers(model_config): TL = TransformerLayers(model_config) path = str(((Path(__file__).parents[1].resolve() / 'data') / 'batch_context.pickle')) with open(path, 'rb') as input_file: encoder_input = pickle.load(input_file) print(type(encoder_input)) embedding = SubwordEmbedding(model_config) emb_output = embedding(encoder_input.input_ids, encoder_input.position_ids) assert (TL(encoder_input).size() == (emb_output.size()[:(- 1)] + ((model_config.num_embed_hidden * model_config.num_attention_heads),)))
def test_FeedForward2(model_config, train_config): embed = torch.rand(train_config.train_batch_size, SEQ_LEN, (model_config.num_embed_hidden * model_config.num_attention_heads)) attn_mask = torch.ones(embed.size()[:(- 1)], dtype=torch.float) FF2 = FeedForward2(model_config) assert (FF2(embed, attn_mask).size() == (train_config.train_batch_size, model_config.num_embed_hidden))
def pytest_ignore_collect(path, config): 'Ignore paths that would otherwise be collceted by the doctest\n plugin and lead to ImportError due to missing dependencies.\n ' return any((path.fnmatch(ignore) for ignore in ignore_test_paths))
class Initializer(object): 'Base class for parameter tensor initializers.\n\n The :class:`Initializer` class represents a weight initializer used\n to initialize weight parameters in a neural network layer. It should be\n subclassed when implementing new types of weight initializers.\n\n ' def __call__(self, shape): '\n Makes :class:`Initializer` instances callable like a function, invoking\n their :meth:`sample()` method.\n ' return self.sample(shape) def sample(self, shape): '\n Sample should return a theano.tensor of size shape and data type\n theano.config.floatX.\n\n Parameters\n -----------\n shape : tuple or int\n Integer or tuple specifying the size of the returned\n matrix.\n returns : theano.tensor\n Matrix of size shape and dtype theano.config.floatX.\n ' raise NotImplementedError()
class Normal(Initializer): 'Sample initial weights from the Gaussian distribution.\n\n Initial weight parameters are sampled from N(mean, std).\n\n Parameters\n ----------\n std : float\n Std of initial parameters.\n mean : float\n Mean of initial parameters.\n ' def __init__(self, std=0.01, mean=0.0): self.std = std self.mean = mean def sample(self, shape): return floatX(get_rng().normal(self.mean, self.std, size=shape))
class Uniform(Initializer): 'Sample initial weights from the uniform distribution.\n\n Parameters are sampled from U(a, b).\n\n Parameters\n ----------\n range : float or tuple\n When std is None then range determines a, b. If range is a float the\n weights are sampled from U(-range, range). If range is a tuple the\n weights are sampled from U(range[0], range[1]).\n std : float or None\n If std is a float then the weights are sampled from\n U(mean - np.sqrt(3) * std, mean + np.sqrt(3) * std).\n mean : float\n see std for description.\n ' def __init__(self, range=0.01, std=None, mean=0.0): if (std is not None): a = (mean - (np.sqrt(3) * std)) b = (mean + (np.sqrt(3) * std)) else: try: (a, b) = range except TypeError: (a, b) = ((- range), range) self.range = (a, b) def sample(self, shape): return floatX(get_rng().uniform(low=self.range[0], high=self.range[1], size=shape))