How to use load_from_dir method in autotest

Best Python code snippet using autotest_python

train.py

Source:train.py

1import warnings2warnings.filterwarnings('ignore')3import os4import logging5import pickle6import importlib7import functools8from pathlib import Path9from argparse import Namespace10import torch11from texar.torch.run import Executor, metric, cond, action12from texar.torch.run.metric.summary import Average, RunningAverage13from torch.optim.adam import Adam14from torch.optim.lr_scheduler import LambdaLR15from typing import Any, Callable, List, Optional, Tuple16from .data import create_texar_dataset, load_tokenizer, TexarDataset17from .construct import create_model, load_hparams18from .metric import RawMetric19from .args import get_train_parser20from .model.lm import LM21device = torch.device("cuda" if torch.cuda.is_available() else "cpu")22torch.manual_seed(0)23logger = logging.getLogger(__name__)24def get_lr_scaler(step: int,25                  warmup_steps: int):26    """Noam learning rate schedule described in27    https://arxiv.org/pdf/1706.03762.pdf.28    """29    if step <= warmup_steps:30        lr_scaler = min(1.0, step / warmup_steps)  # Linear warmup31    else:32        lr_scaler = 0.9 ** (step / warmup_steps)  # Exponential decay33    return lr_scaler34def create_optimizer_and_scheduler(model: LM,35                                   learning_rate: float=0.001,36                                   dynamic_lr: bool=False,37                                   warmup_steps: int=400038                                   ) -> Tuple[Adam, LambdaLR]:39    # Define optimizer and scheduler40    if dynamic_lr:41        lr_scaler_fn = functools.partial(get_lr_scaler, warmup_steps=warmup_steps)42    else:43        lr_scaler_fn = lambda _: 1.044    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate,45                                 betas=(0.9, 0.998), eps=1e-9)46    lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_scaler_fn)47    return optimizer, lr_scheduler48def load_metrics(pg_metric_names: List[Any],49                 eval_metric_names: List[Any],50                 text_decode_fn: Callable51                 ) -> Tuple[List[Any], List[Any]]:52    metric_fns_map = {metric_name:53                      RawMetric.get_metric_from_name(metric_name)(decode_fn=text_decode_fn)54                      for metric_name in set(pg_metric_names + eval_metric_names)}55    pg_metric_fns = [metric_fns_map[metric_name] for metric_name in pg_metric_names]56    eval_metric_fns = [metric_fns_map[metric_name] for metric_name in eval_metric_names]57    return pg_metric_fns, eval_metric_fns58def create_texar_metrics(pg_metric_fns: List[Any],59                         eval_metric_fns: List[Any],60                         log_iterations: int=10061                         ) -> Tuple[List[RunningAverage], List[Average]]:62    train_metrics = [metric.RunningAverage(log_iterations),63                     metric.RunningAverage(log_iterations,64                                           pred_name='mle_loss'),65                     *(metric.RunningAverage(log_iterations,66                                             pred_name='{}_loss'.format(metric_fn.metric_name))67                       for metric_fn in pg_metric_fns)]68    if eval_metric_fns:69        eval_metrics = [*(metric.Average(pred_name=metric_fn.metric_name,70                                         higher_is_better=metric_fn.higher_is_better)71                          for metric_fn in eval_metric_fns),72                        metric.Average(pred_name='mle_loss')]73    else:74        eval_metrics = [metric.Average(pred_name='mle_loss')]75    return train_metrics, eval_metrics76def create_executor(model: LM,77                    save_dir: str,78                    train_data: TexarDataset,79                    eval_data: TexarDataset,80                    optimizer: Adam,81                    lr_scheduler: LambdaLR,82                    load_from_dir: Optional[str]=None,83                    train_metrics: List[RunningAverage]=[],84                    eval_metrics: List[Average]=[],85                    max_grad_norm: float=5.0,86                    accum_steps: int=1,87                    max_epochs: int=100,88                    patience: int=2,89                    valid_iterations: int=1000,90                    valid_epoch_end: bool=False,91                    log_iterations: int=10092                    ) -> Executor:93    # Use Executor API to train and validate94    executor = Executor(model=model,95                        train_data=train_data,96                        valid_data=eval_data,97                        test_data=eval_data,98                        optimizer=optimizer,99                        lr_scheduler=lr_scheduler,100                        grad_clip=max_grad_norm,101                        num_iters_per_update=accum_steps,102                        stop_training_on=cond.epoch(max_epochs),103                        log_every=cond.iteration(log_iterations),104                        validate_mode='eval',105                        test_mode='eval',106                        validate_every=([cond.epoch(1),107                                         cond.iteration(valid_iterations)] if valid_epoch_end108                                        else [cond.iteration(valid_iterations)]),109                        train_metrics=train_metrics + [metric.LR(optimizer)],110                        valid_metrics=eval_metrics,111                        save_every=cond.validation(better=True),112                        plateau_condition=[113                            cond.consecutive(cond.validation(better=False), 1)],114                        action_on_plateau=[115                            action.early_stop(patience=patience)],116                        max_to_keep=1,117                        checkpoint_dir=save_dir)118    if load_from_dir is not None:119         # Seems like a bug where checkpoint is only looked for in executor.checkpoint_dir120         # Get around this by temporarily changing checkpoint_dir to args.load_from_dir then changing it back to args.save_dir121        executor.checkpoint_dir = Path(load_from_dir)122        loaded_checkpoint = executor.load()123        logger.info("Initialized model from checkpoint {}".format(loaded_checkpoint))124        executor.checkpoint_dir = Path(save_dir)125    return executor126def train(args: Namespace,127          train_src_texts: List[str],128          train_tgt_texts: List[str],129          eval_src_texts: List[str],130          eval_tgt_texts: List[str],131          train_ref_texts: Optional[List[str]]=None,132          eval_ref_texts: Optional[List[str]]=None133          ) -> None:134    if args.load_from_dir:135        logger.info(("Loading model configuration from {}. " +136                     "All hyperparameter settings will be read from here " +137                     "and will override any settings provided as command-line arguments.").format(138            os.path.join(args.load_from_dir)))139        model_hparams = load_hparams(path=args.load_from_dir)140    else:141        assert args.config_file is not None142        model_hparams = load_hparams(path=args.config_file)143    tokenizer = load_tokenizer(model_hparams['tokenizer'])144    # Make dataset objects for Texar API145    texar_train_data, texar_eval_data = create_texar_dataset(146        tokenizer=tokenizer,147        train_src_texts=train_src_texts,148        train_tgt_texts=train_tgt_texts,149        eval_src_texts=eval_src_texts,150        eval_tgt_texts=eval_tgt_texts,151        train_ref_texts=train_ref_texts,152        eval_ref_texts=eval_ref_texts,153        max_src_length=args.max_src_length,154        max_tgt_length=args.max_tgt_length,155        batch_size=args.batch_size156    )157    # Initialize the policy gradient and evaluation metrics based on given metric names158    pg_metric_fns, eval_metric_fns = load_metrics(159        pg_metric_names=args.pg_metrics,160        eval_metric_names=args.eval_metrics,161        text_decode_fn=(lambda text:162                        tokenizer.map_id_to_text(text, skip_special_tokens=True))163    )164    # Initialize model with properties and policy gradient/evaluation functions165    model = create_model(hparams=model_hparams,166                         save_dir=args.save_dir)167    model.set_pg_and_eval_fns(pg_metric_fns, eval_metric_fns)168    # Create optimizer and LR scheduler169    optimizer, lr_scheduler = create_optimizer_and_scheduler(model,170                                                             learning_rate=args.learning_rate,171                                                             dynamic_lr=args.dynamic_lr,172                                                             warmup_steps=args.warmup_steps)173    # Wrap metrics inside Texar Metric API to be used by Executor174    (texar_train_metrics,175     texar_eval_metrics) = create_texar_metrics(pg_metric_fns=pg_metric_fns,176                                                eval_metric_fns=eval_metric_fns,177                                                log_iterations=args.log_iterations)178    # Create Texar Executor179    executor = create_executor(model=model,180                               save_dir=args.save_dir,181                               train_data=texar_train_data,182                               eval_data=texar_eval_data,183                               optimizer=optimizer,184                               lr_scheduler=lr_scheduler,185                               load_from_dir=args.load_from_dir,186                               train_metrics=texar_train_metrics,187                               eval_metrics=texar_eval_metrics,188                               max_grad_norm=args.max_grad_norm,189                               accum_steps=args.accum_steps,190                               max_epochs=args.max_epochs,191                               patience=args.patience,192                               valid_iterations=args.valid_iterations,193                               valid_epoch_end=args.valid_epoch_end,194                               log_iterations=args.log_iterations)195    logger.info("Validation result prior to training:")196    executor.test()...

evaluation.py

Source:evaluation.py

...6from torch_tools.visualization import to_image7from visualization import interpolate8from loading import load_from_dir9# %matplotlib inline10# deformator, G, shift_predictor = load_from_dir(11#     './models/pretrained/deformators/SN_MNIST/',12#     G_weights='./models/pretrained/generators/SN_MNIST/')13deformator, G, shift_predictor = load_from_dir(14    './models/pretrained/deformators/SN_Anime/',15    G_weights='./models/pretrained/generators/SN_Anime/')16# deformator, G, shift_predictor = load_from_dir(17#     './models/pretrained/deformators/BigGAN/',18#     G_weights='./models/pretrained/generators/BigGAN/G_ema.pth')19# deformator, G, shift_predictor = load_from_dir(20#     './models/pretrained/deformators/ProgGAN/',21#     G_weights='./models/pretrained/generators/ProgGAN/100_celeb_hq_network-snapshot-010403.pth')22# deformator, G, shift_predictor = load_from_dir(23#     './models/pretrained/deformators/StyleGAN2/',24#     G_weights='./models/pretrained/generators/StyleGAN2/stylegan2-ffhq-config-f.pt')25discovered_annotation = ''26for d in deformator.annotation.items():27    discovered_annotation += '{}: {}\n'.format(d[0], d[1])28print('human-annotated directions:\n' + discovered_annotation)29rows = 830plt.figure(figsize=(5, rows), dpi=250)31# set desired class for conditional GAN32# if is_conditional(G):33#     G.set_classes(12)34annotated = list(deformator.annotation.values())35inspection_dim = annotated[0]36# zs = torch.randn([rows, G.dim_z] if type(G.dim_z) == int else [rows] + G.dim_z, device='cuda')...

visualize.py

Source:visualize.py

1import time2import math3import os4import sys5if len(sys.argv) < 2:6    print('Usage: {} datasource'.format(sys.argv[0]))7    print('\tAvailable datasources: boxes, minipong, mediumpong...')8    exit(1)9import numpy as np10import torch11import torch.nn as nn12import torch.nn.functional as F13import torch.optim as optim14import imutil15from logutil import TimeSeries16from tqdm import tqdm17from spatial_recurrent import CSRN18from coordconv import CoordConv2d19from higgins import higgins_metric20from importlib import import_module21datasource = import_module('envs.' + sys.argv[1])22import models23def main():24    batch_size = 1625    latent_dim = 1226    true_latent_dim = 427    num_actions = 428    encoder = models.Encoder(latent_dim)29    decoder = models.Decoder(latent_dim)30    transition = models.Transition(latent_dim, num_actions)31    blur = models.GaussianSmoothing(channels=3, kernel_size=11, sigma=4.)32    higgins_scores = []33    #load_from_dir = '/mnt/nfs/experiments/demo_2018_12_12/scm-gan_81bd12cd'34    load_from_dir = '.'35    print('Loading models from directory {}'.format(load_from_dir))36    encoder.load_state_dict(torch.load(os.path.join(load_from_dir, 'model-encoder.pth')))37    decoder.load_state_dict(torch.load(os.path.join(load_from_dir, 'model-decoder.pth')))38    transition.load_state_dict(torch.load(os.path.join(load_from_dir, 'model-transition.pth')))39    encoder.eval()40    decoder.eval()41    transition.eval()42    for model in (encoder, decoder, transition):43        for child in model.children():44            if type(child) == nn.BatchNorm2d or type(child) == nn.BatchNorm1d:45                child.momentum = 046    states, rewards, dones, actions = datasource.get_trajectories(batch_size, timesteps=1)47    states = torch.Tensor(states).cuda()48    # Reconstruct the first timestep49    reconstructed = decoder(encoder(states[:, 0]))50    imutil.show(reconstructed)51if __name__ == '__main__':...

Automation Testing Tutorials

Learn to execute automation testing from scratch with LambdaTest Learning Hub. Right from setting up the prerequisites to run your first automation test, to following best practices and diving deeper into advanced test scenarios. LambdaTest Learning Hubs compile a list of step-by-step guides to help you be proficient with different test automation frameworks i.e. Selenium, Cypress, TestNG etc.