3dresnet_armory.py

from pathlib import Path
import json
import random
import os

import numpy as np
import torch
from torch.nn import CrossEntropyLoss
from torch.optim import SGD, lr_scheduler
import torch.nn.functional as F
import torch.multiprocessing as mp
import torch.distributed as dist
from torch.backends import cudnn
import torchvision

from PIL import Image

from opts import parse_opts
from model import generate_model, make_data_parallel
#from mean import get_mean_std

#from dataset import get_validation_data
#from utils import Logger, worker_init_fn, get_lr
# from validation import val_epoch

from datasets import preprocess_data

import pdb

from main import *
from utils import *

import logging
from typing import Union, Optional, Tuple

from art.estimators.classification import PyTorchClassifier
#from art.classifiers import PyTorchClassifier

   
logger = logging.getLogger(__name__)
    
opt = get_opt(['--attack_type','pgd_inf', '--n_classes', '101', '--model_depth', '101', '--model', 'resnext', 
          '--no_mean_norm', '--no_std_norm']) # '--use_ape'

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

MEAN = np.array([114.7748, 107.7354, 99.4750], dtype=np.float32)
STD = np.array([1, 1, 1], dtype=np.float32)


def preprocessing_fn_numpy(batch: np.ndarray):
    """
    batch is a batch of videos, (batch, frames, height, width, channels)
    Preprocessing resizes the height and width to 112 x 112 and reshapes
    each video to (n_stack, 3, 16, height, width), where n_stack = int(time/16).
    Outputs a list of videos, each of shape (n_stack, 3, 16, 112, 112)
    """
    sample_duration = 16  # expected number of consecutive frames as input to the model

    outputs = []
    for i, x in enumerate(batch):
        if x.ndim != 4:
            raise ValueError(f"sample {i} in batch has {x.ndim} dims, not 4 (FHWC)")
        if x.dtype in (float, np.float32):
            if x.max() > 1.0 or x.min() < 0.0:
                raise ValueError(f"sample {i} is a float but not in [0.0, 1.0] range")
            x = (255 * x).round().astype(np.uint8)
        if x.dtype != np.uint8:
            raise ValueError(f"sample {i} - unrecognized dtype {x.dtype}")

        # select a fixed number of consecutive frames
        total_frames = x.shape[0]
        if total_frames < sample_duration / 2:
            raise ValueError(
                f"video is too short; requires >= {sample_duration / 2} frames"
            )
        if total_frames <= sample_duration:  # cyclic pad if not enough frames
            x = np.vstack([x, x[: sample_duration - total_frames]])

        # apply MARS preprocessing: scaling, cropping, normalizing
#         opt = parse_opts(arguments=[])
        opt.modality = "RGB"
        opt.sample_size = 112
        x_Image = []  # convert each frame to PIL Image
        for frame in x:
            x_Image.append(Image.fromarray(frame))
        x_mars_preprocessed = preprocess_data.scale_crop(x_Image, 0, opt)

        # reshape
        x_reshaped = []
        for ns in range(int(total_frames / sample_duration)):
            np_frames = x_mars_preprocessed[
                :, ns * sample_duration : (ns + 1) * sample_duration, :, :
            ].numpy()
            x_reshaped.append(np_frames)
        outputs.append(np.array(x_reshaped, dtype=np.float32))
    return outputs


def preprocessing_fn_torch(
    batch: Union[torch.Tensor, np.ndarray],
    consecutive_frames: int = 16,
    scale_first: bool = True,
    align_corners: bool = False,
):
    """
    inputs - batch of videos each with shape (frames, height, width, channel)
    outputs - batch of videos each with shape (n_stack, channel, stack_frames, new_height, new_width)
        frames = n_stack * stack_frames (after padding)
        new_height = new_width = 112
    consecutive_frames - number of consecutive frames (stack_frames)
    After resizing, a center crop is performed to make the image square
    This is a differentiable alternative to MARS' PIL-based preprocessing.
        There are some
    """
    if not isinstance(batch, torch.Tensor):
        logger.warning(f"batch {type(batch)} is not a torch.Tensor. Casting")
        batch = torch.from_numpy(batch).to(DEVICE)
        # raise ValueError(f"batch {type(batch)} is not a torch.Tensor")
    if batch.dtype != torch.float32:
        raise ValueError(f"batch {batch.dtype} should be torch.float32")
    if batch.shape[0] != 1:
        raise ValueError(f"Batch size {batch.shape[0]} != 1")
    video = batch[0]

    if video.ndim != 4:
        raise ValueError(
            f"video dims {video.ndim} != 4 (frames, height, width, channel)"
        )
    if video.shape[0] < 1:
        raise ValueError("video must have at least one frame")
    if tuple(video.shape[1:]) == (240, 320, 3):
        standard_shape = True
    elif tuple(video.shape[1:]) == (226, 400, 3):
        logger.warning("Expected odd example shape (226, 400, 3)")
        standard_shape = False
    else:
        raise ValueError(f"frame shape {tuple(video.shape[1:])} not recognized")
    if video.max() > 1.0 or video.min() < 0.0:
        raise ValueError("input should be float32 in [0, 1] range")
    if not isinstance(consecutive_frames, int):
        raise ValueError(f"consecutive_frames {consecutive_frames} must be an int")
    if consecutive_frames < 1:
        raise ValueError(f"consecutive_frames {consecutive_frames} must be positive")

    # Select a integer multiple of consecutive frames
    while len(video) < consecutive_frames:
        # cyclic pad if insufficient for a single stack
        video = torch.cat([video, video[: consecutive_frames - len(video)]])
    if len(video) % consecutive_frames != 0:
        # cut trailing frames
        video = video[: len(video) - (len(video) % consecutive_frames)]

    if scale_first:
        # Attempts to directly follow MARS approach
        # (frames, height, width, channel) to (frames, channel, height, width)
        video = video.permute(0, 3, 1, 2)
        if standard_shape:  # 240 x 320
            sample_height, sample_width = 112, 149
        else:  # 226 x 400
            video = video[:, :, 1:-1, :]  # crop top/bottom pixels, reduce by 2
            sample_height, sample_width = 112, 200

        video = torch.nn.functional.interpolate(
            video,
            size=(sample_height, sample_width),
            mode="bilinear",
            align_corners=align_corners,
        )

        if standard_shape:
            crop_left = 18  # round((149 - 112)/2.0)
        else:
            crop_left = 40
        video = video[:, :, :, crop_left : crop_left + sample_height]

    else:
        # More efficient, but not MARS approach
        # Center crop
        sample_size = 112
        if standard_shape:
            crop_width = 40
        else:
            video = video[:, 1:-1, :, :]
            crop_width = 88
        video = video[:, :, crop_width:-crop_width, :]

        # Downsample to (112, 112) frame size
        # (frames, height, width, channel) to (frames, channel, height, width)
        video = video.permute(0, 3, 1, 2)
        video = torch.nn.functional.interpolate(
            video,
            size=(sample_size, sample_size),
            mode="bilinear",
            align_corners=align_corners,
        )

    if video.max() > 1.0:
        raise ValueError("Video exceeded max after interpolation")
    if video.min() < 0.0:
        raise ValueError("Video under min after interpolation")

    # reshape into stacks of frames
    video = torch.reshape(video, (-1, consecutive_frames) + video.shape[1:])

    # transpose to (stacks, channel, stack_frames, height, width)
    video = video.permute(0, 2, 1, 3, 4)
    # video = torch.transpose(video, axes=(0, 4, 1, 2, 3))

    # normalize before changing channel position?
    #video = torch.transpose(video, 1, 4)
    #video = ((video * 255) - torch.from_numpy(MEAN).to(DEVICE)) / torch.from_numpy(
    #    STD
    #).to(DEVICE)
    #video = torch.transpose(video, 4, 1)

    return video


def fit_preprocessing_fn_numpy(batch: np.ndarray):
    """
    Randomly sample a single stack from each video
    """
    x = preprocessing_fn_numpy(batch)
    x = np.stack([x_i[np.random.randint(x_i.shape[0])] for x_i in x])
    return x


preprocessing_fn = fit_preprocessing_fn_numpy


def make_model(
    model_status: str = "ucf101_trained", weights_path: Optional[str] = None, use_ape = False
) -> Tuple[torch.nn.DataParallel, torch.optim.SGD]:
    statuses = ("ucf101_trained", "kinetics_pretrained")
    if model_status not in statuses:
        raise ValueError(f"model_status {model_status} not in {statuses}")
    trained = model_status == "ucf101_trained"
    if not trained and weights_path is None:
        raise ValueError("weights_path cannot be None for 'kinetics_pretrained'")
    
    opt.device = torch.device('cpu' if opt.no_cuda else 'cuda')
    opt.use_ape = use_ape
    
    if not opt.no_cuda:
        cudnn.benchmark = True
    if opt.accimage:
        torchvision.set_image_backend('accimage')

    opt.resume_path = weights_path
	
    model = generate_model(opt.model, sample_duration=16)
    #model = generate_model('resnext_3bn_comb', sample_duration=16)  # Mine

    if opt.use_ape:
        in_ch = 3
        G = Generator(in_ch).to(opt.device)

    model = resume_model(opt.resume_path, opt.arch, model)
    #model = resume_model_my(opt.resume_path, opt.arch, model)  # Mine


    model = make_data_parallel(model, opt.distributed, opt.device)

    if opt.use_ape:
    #         G = make_data_parallel(G, opt.distributed, opt.device)
        if opt.ape_path is not None:
            #checkpoint = torch.load(opt.ape_path)
    #             G.load_state_dict(checkpoint['generator'])
            G.load_state_dict(torch.load(opt.resume_path)['generator'])
            G = make_data_parallel(G, opt.distributed, opt.device)
        model = torch.nn.Sequential(G, model)
        
    parameters = model.parameters()
    
    logger.info(f"Loading model... {opt.model} {opt.model_depth}")
    
    if opt.nesterov:
        dampening = 0
    else:
        dampening = opt.dampening
        
    optimizer = SGD(parameters,
                    lr=opt.learning_rate,
                    momentum=opt.momentum,
                    dampening=dampening,
                    weight_decay=opt.weight_decay,
                    nesterov=opt.nesterov)

    return model, optimizer

class OuterModel(torch.nn.Module):
    def __init__(
        self, weights_path: Optional[str], max_frames: int = 0, use_ape = False, **model_kwargs,
    ):
        """
        Max frames is the maximum number of input frames.
            If max_frames == 0, no clipping is done
            Else if max_frames > 0, frames are clipped to that number.
            This can be helpful for smaller memory cards.
        """
        super().__init__()
        max_frames = int(max_frames)
        if max_frames < 0:
            raise ValueError(f"max_frames {max_frames} cannot be negative")
        self.max_frames = max_frames
        self.model, self.optimizer = make_model(
            weights_path=weights_path, use_ape=use_ape, **model_kwargs
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        
        #print(x.shape)
        
        if self.max_frames:
            x = x[:, : self.max_frames]

        #print(x.shape)
        
        if self.training:
            # Use preprocessing_fn_numpy in dataset preprocessing
            return self.model(x)
        else:
            x = preprocessing_fn_torch(x)
            stack_outputs = self.model(x)
            output = stack_outputs.mean(axis=0, keepdims=True)

        return output

def get_art_model(
    model_kwargs: dict, wrapper_kwargs: dict, weights_path: Optional[str] = None
) -> PyTorchClassifier:
     
    model = OuterModel(weights_path=weights_path, **model_kwargs)

    model.to(opt.device)

    wrapped_model = PyTorchClassifier(
        model,
        loss=torch.nn.CrossEntropyLoss(),
        optimizer=model.optimizer,
        input_shape=(None, 240, 320, 3),
        nb_classes=101,
        clip_values=(0.0, 1.0),
        **wrapper_kwargs,
    )
    return wrapped_model