utils/augmentation.py

import numpy as np
from tqdm import tqdm

def jitter(x, sigma=0.03):
    # https://arxiv.org/pdf/1706.00527.pdf
    return x + np.random.normal(loc=0., scale=sigma, size=x.shape)


def scaling(x, sigma=0.1):
    # https://arxiv.org/pdf/1706.00527.pdf
    factor = np.random.normal(loc=1., scale=sigma, size=(x.shape[0],x.shape[2]))
    return np.multiply(x, factor[:,np.newaxis,:])

def rotation(x):
    x = np.array(x)
    flip = np.random.choice([-1, 1], size=(x.shape[0],x.shape[2]))
    rotate_axis = np.arange(x.shape[2])
    np.random.shuffle(rotate_axis)
    return flip[:,np.newaxis,:] * x[:,:,rotate_axis]

def permutation(x, max_segments=5, seg_mode="equal"):
    orig_steps = np.arange(x.shape[1])
    
    num_segs = np.random.randint(1, max_segments, size=(x.shape[0]))
    
    ret = np.zeros_like(x)
    for i, pat in enumerate(x):
        if num_segs[i] > 1:
            if seg_mode == "random":
                split_points = np.random.choice(x.shape[1]-2, num_segs[i]-1, replace=False)
                split_points.sort()
                splits = np.split(orig_steps, split_points)
            else:
                splits = np.array_split(orig_steps, num_segs[i])
            warp = np.concatenate(np.random.permutation(splits)).ravel()
            # ? Question: What is the point of making segments?
            # for i in range(len(splits)):
            #     permute = np.random.permutation(splits[i])


            ret[i] = pat[warp]
        else:
            ret[i] = pat
    return ret

def magnitude_warp(x, sigma=0.2, knot=4):
    from scipy.interpolate import CubicSpline
    orig_steps = np.arange(x.shape[1])
    
    random_warps = np.random.normal(loc=1.0, scale=sigma, size=(x.shape[0], knot+2, x.shape[2]))
    warp_steps = (np.ones((x.shape[2],1))*(np.linspace(0, x.shape[1]-1., num=knot+2))).T
    ret = np.zeros_like(x)
    for i, pat in enumerate(x):
        warper = np.array([CubicSpline(warp_steps[:,dim], random_warps[i,:,dim])(orig_steps) for dim in range(x.shape[2])]).T
        ret[i] = pat * warper

    return ret

def time_warp(x, sigma=0.2, knot=4):
    from scipy.interpolate import CubicSpline
    orig_steps = np.arange(x.shape[1])
    
    random_warps = np.random.normal(loc=1.0, scale=sigma, size=(x.shape[0], knot+2, x.shape[2]))
    warp_steps = (np.ones((x.shape[2],1))*(np.linspace(0, x.shape[1]-1., num=knot+2))).T
    
    ret = np.zeros_like(x)
    for i, pat in enumerate(x):
        for dim in range(x.shape[2]):
            time_warp = CubicSpline(warp_steps[:,dim], warp_steps[:,dim] * random_warps[i,:,dim])(orig_steps)
            scale = (x.shape[1]-1)/time_warp[-1]
            ret[i,:,dim] = np.interp(orig_steps, np.clip(scale*time_warp, 0, x.shape[1]-1), pat[:,dim]).T
    return ret

def window_slice(x, reduce_ratio=0.9):
    # https://halshs.archives-ouvertes.fr/halshs-01357973/document
    target_len = np.ceil(reduce_ratio*x.shape[1]).astype(int)
    if target_len >= x.shape[1]:
        return x
    starts = np.random.randint(low=0, high=x.shape[1]-target_len, size=(x.shape[0])).astype(int)
    ends = (target_len + starts).astype(int)
    
    ret = np.zeros_like(x)
    for i, pat in enumerate(x):
        for dim in range(x.shape[2]):
            ret[i,:,dim] = np.interp(np.linspace(0, target_len, num=x.shape[1]), np.arange(target_len), pat[starts[i]:ends[i],dim]).T
    return ret

def window_warp(x, window_ratio=0.1, scales=[0.5, 2.]):
    # https://halshs.archives-ouvertes.fr/halshs-01357973/document
    warp_scales = np.random.choice(scales, x.shape[0])
    warp_size = np.ceil(window_ratio*x.shape[1]).astype(int)
    window_steps = np.arange(warp_size)
        
    window_starts = np.random.randint(low=1, high=x.shape[1]-warp_size-1, size=(x.shape[0])).astype(int)
    window_ends = (window_starts + warp_size).astype(int)
            
    ret = np.zeros_like(x)
    for i, pat in enumerate(x):
        for dim in range(x.shape[2]):
            start_seg = pat[:window_starts[i],dim]
            window_seg = np.interp(np.linspace(0, warp_size-1, num=int(warp_size*warp_scales[i])), window_steps, pat[window_starts[i]:window_ends[i],dim])
            end_seg = pat[window_ends[i]:,dim]
            warped = np.concatenate((start_seg, window_seg, end_seg))                
            ret[i,:,dim] = np.interp(np.arange(x.shape[1]), np.linspace(0, x.shape[1]-1., num=warped.size), warped).T
    return ret

def spawner(x, labels, sigma=0.05, verbose=0):
    # https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6983028/
    # use verbose=-1 to turn off warnings
    # use verbose=1 to print out figures
    
    import utils.dtw as dtw
    random_points = np.random.randint(low=1, high=x.shape[1]-1, size=x.shape[0])
    window = np.ceil(x.shape[1] / 10.).astype(int)
    orig_steps = np.arange(x.shape[1])
    l = np.argmax(labels, axis=1) if labels.ndim > 1 else labels
    
    ret = np.zeros_like(x)
    # for i, pat in enumerate(tqdm(x)):
    for i, pat in enumerate(x):
        # guarentees that same one isnt selected
        choices = np.delete(np.arange(x.shape[0]), i)
        # remove ones of different classes
        choices = np.where(l[choices] == l[i])[0]
        if choices.size > 0:     
            random_sample = x[np.random.choice(choices)]
            # SPAWNER splits the path into two randomly
            path1 = dtw.dtw(pat[:random_points[i]], random_sample[:random_points[i]], dtw.RETURN_PATH, slope_constraint="symmetric", window=window)
            path2 = dtw.dtw(pat[random_points[i]:], random_sample[random_points[i]:], dtw.RETURN_PATH, slope_constraint="symmetric", window=window)
            combined = np.concatenate((np.vstack(path1), np.vstack(path2+random_points[i])), axis=1)
            if verbose:
                # print(random_points[i])
                dtw_value, cost, DTW_map, path = dtw.dtw(pat, random_sample, return_flag = dtw.RETURN_ALL, slope_constraint=slope_constraint, window=window)
                dtw.draw_graph1d(cost, DTW_map, path, pat, random_sample)
                dtw.draw_graph1d(cost, DTW_map, combined, pat, random_sample)
            mean = np.mean([pat[combined[0]], random_sample[combined[1]]], axis=0)
            for dim in range(x.shape[2]):
                ret[i,:,dim] = np.interp(orig_steps, np.linspace(0, x.shape[1]-1., num=mean.shape[0]), mean[:,dim]).T
        else:
            # if verbose > -1:
            #     print("There is only one pattern of class {}, skipping pattern average".format(l[i]))
            ret[i,:] = pat
    return jitter(ret, sigma=sigma)

def wdba(x, labels, batch_size=6, slope_constraint="symmetric", use_window=True, verbose=0):
    # https://ieeexplore.ieee.org/document/8215569
    # use verbose = -1 to turn off warnings    
    # slope_constraint is for DTW. "symmetric" or "asymmetric"
    x = np.array(x)
    import utils.dtw as dtw
    
    if use_window:
        window = np.ceil(x.shape[1] / 10.).astype(int)
    else:
        window = None
    orig_steps = np.arange(x.shape[1])
    l = np.argmax(labels, axis=1) if labels.ndim > 1 else labels
        
    ret = np.zeros_like(x)
    # for i in tqdm(range(ret.shape[0])):
    for i in range(ret.shape[0]):
        # get the same class as i
        choices = np.where(l == l[i])[0]
        if choices.size > 0:        
            # pick random intra-class pattern
            k = min(choices.size, batch_size)
            random_prototypes = x[np.random.choice(choices, k, replace=False)]
            
            # calculate dtw between all
            dtw_matrix = np.zeros((k, k))
            for p, prototype in enumerate(random_prototypes):
                for s, sample in enumerate(random_prototypes):
                    if p == s:
                        dtw_matrix[p, s] = 0.
                    else:
                        dtw_matrix[p, s] = dtw.dtw(prototype, sample, dtw.RETURN_VALUE, slope_constraint=slope_constraint, window=window)
                        
            # get medoid
            medoid_id = np.argsort(np.sum(dtw_matrix, axis=1))[0]
            nearest_order = np.argsort(dtw_matrix[medoid_id])
            medoid_pattern = random_prototypes[medoid_id]
            
            # start weighted DBA
            average_pattern = np.zeros_like(medoid_pattern)
            weighted_sums = np.zeros((medoid_pattern.shape[0]))
            for nid in nearest_order:
                if nid == medoid_id or dtw_matrix[medoid_id, nearest_order[1]] == 0.:
                    average_pattern += medoid_pattern
                    weighted_sums += np.ones_like(weighted_sums) 
                else:
                    path = dtw.dtw(medoid_pattern, random_prototypes[nid], dtw.RETURN_PATH, slope_constraint=slope_constraint, window=window)
                    dtw_value = dtw_matrix[medoid_id, nid]
                    warped = random_prototypes[nid, path[1]]
                    weight = np.exp(np.log(0.5)*dtw_value/dtw_matrix[medoid_id, nearest_order[1]])
                    average_pattern[path[0]] += weight * warped
                    weighted_sums[path[0]] += weight 
            
            ret[i,:] = average_pattern / weighted_sums[:,np.newaxis]
        else:
            # if verbose > -1:
            #     print("There is only one pattern of class {}, skipping pattern average".format(l[i]))
            ret[i,:] = x[i]
    return ret

# Proposed

def random_guided_warp(x, labels, slope_constraint="symmetric", use_window=True, dtw_type="normal", verbose=0):
    # use verbose = -1 to turn off warnings
    # slope_constraint is for DTW. "symmetric" or "asymmetric"
    # dtw_type is for shapeDTW or DTW. "normal" or "shape"
    
    import utils.dtw as dtw
    
    if use_window:
        window = np.ceil(x.shape[1] / 10.).astype(int)
    else:
        window = None
    orig_steps = np.arange(x.shape[1])
    l = np.argmax(labels, axis=1) if labels.ndim > 1 else labels
    
    ret = np.zeros_like(x)
    # for i, pat in enumerate(tqdm(x)):
    for i, pat in enumerate(x):
        # guarentees that same one isnt selected
        choices = np.delete(np.arange(x.shape[0]), i)
        # remove ones of different classes
        choices = np.where(l[choices] == l[i])[0]
        if choices.size > 0:        
            # pick random intra-class pattern
            random_prototype = x[np.random.choice(choices)]
            
            if dtw_type == "shape":
                path = dtw.shape_dtw(random_prototype, pat, dtw.RETURN_PATH, slope_constraint=slope_constraint, window=window)
            else:
                path = dtw.dtw(random_prototype, pat, dtw.RETURN_PATH, slope_constraint=slope_constraint, window=window)
                            
            # Time warp
            warped = pat[path[1]]
            for dim in range(x.shape[2]):
                ret[i,:,dim] = np.interp(orig_steps, np.linspace(0, x.shape[1]-1., num=warped.shape[0]), warped[:,dim]).T
        else:
            # if verbose > -1:
            #     print("There is only one pattern of class {}, skipping timewarping".format(l[i]))
            ret[i,:] = pat
    return ret

def random_guided_warp_shape(x, labels, slope_constraint="symmetric", use_window=True):
    return random_guided_warp(x, labels, slope_constraint, use_window, dtw_type="shape")

def discriminative_guided_warp(x, labels, batch_size=6, slope_constraint="symmetric", use_window=True, dtw_type="normal", use_variable_slice=True, verbose=0):
    # use verbose = -1 to turn off warnings
    # slope_constraint is for DTW. "symmetric" or "asymmetric"
    # dtw_type is for shapeDTW or DTW. "normal" or "shape"
    
    import utils.dtw as dtw
    
    if use_window:
        window = np.ceil(x.shape[1] / 10.).astype(int)
    else:
        window = None
    orig_steps = np.arange(x.shape[1])
    l = np.argmax(labels, axis=1) if labels.ndim > 1 else labels
    
    positive_batch = np.ceil(batch_size / 2).astype(int)
    negative_batch = np.floor(batch_size / 2).astype(int)
        
    ret = np.zeros_like(x)
    warp_amount = np.zeros(x.shape[0])
    # for i, pat in enumerate(tqdm(x)):
    for i, pat in enumerate(x):
        # guarentees that same one isnt selected
        choices = np.delete(np.arange(x.shape[0]), i)
        
        # remove ones of different classes
        positive = np.where(l[choices] == l[i])[0]
        negative = np.where(l[choices] != l[i])[0]
        
        if positive.size > 0 and negative.size > 0:
            pos_k = min(positive.size, positive_batch)
            neg_k = min(negative.size, negative_batch)
            positive_prototypes = x[np.random.choice(positive, pos_k, replace=False)]
            negative_prototypes = x[np.random.choice(negative, neg_k, replace=False)]
                        
            # vector embedding and nearest prototype in one
            pos_aves = np.zeros((pos_k))
            neg_aves = np.zeros((pos_k))
            if dtw_type == "shape":
                for p, pos_prot in enumerate(positive_prototypes):
                    for ps, pos_samp in enumerate(positive_prototypes):
                        if p != ps:
                            pos_aves[p] += (1./(pos_k-1.))*dtw.shape_dtw(pos_prot, pos_samp, dtw.RETURN_VALUE, slope_constraint=slope_constraint, window=window)
                    for ns, neg_samp in enumerate(negative_prototypes):
                        neg_aves[p] += (1./neg_k)*dtw.shape_dtw(pos_prot, neg_samp, dtw.RETURN_VALUE, slope_constraint=slope_constraint, window=window)
                selected_id = np.argmax(neg_aves - pos_aves)
                path = dtw.shape_dtw(positive_prototypes[selected_id], pat, dtw.RETURN_PATH, slope_constraint=slope_constraint, window=window)
            else:
                for p, pos_prot in enumerate(positive_prototypes):
                    for ps, pos_samp in enumerate(positive_prototypes):
                        if p != ps:
                            pos_aves[p] += (1./(pos_k-1.))*dtw.dtw(pos_prot, pos_samp, dtw.RETURN_VALUE, slope_constraint=slope_constraint, window=window)
                    for ns, neg_samp in enumerate(negative_prototypes):
                        neg_aves[p] += (1./neg_k)*dtw.dtw(pos_prot, neg_samp, dtw.RETURN_VALUE, slope_constraint=slope_constraint, window=window)
                selected_id = np.argmax(neg_aves - pos_aves)
                path = dtw.dtw(positive_prototypes[selected_id], pat, dtw.RETURN_PATH, slope_constraint=slope_constraint, window=window)
                   
            # Time warp
            warped = pat[path[1]]
            warp_path_interp = np.interp(orig_steps, np.linspace(0, x.shape[1]-1., num=warped.shape[0]), path[1])
            warp_amount[i] = np.sum(np.abs(orig_steps-warp_path_interp))
            for dim in range(x.shape[2]):
                ret[i,:,dim] = np.interp(orig_steps, np.linspace(0, x.shape[1]-1., num=warped.shape[0]), warped[:,dim]).T
        else:
            # if verbose > -1:
            #     print("There is only one pattern of class {}".format(l[i]))
            ret[i,:] = pat
            warp_amount[i] = 0.
    if use_variable_slice:
        max_warp = np.max(warp_amount)
        if max_warp == 0:
            # unchanged
            ret = window_slice(ret, reduce_ratio=0.9)
        else:
            for i, pat in enumerate(ret):
                # Variable Sllicing
                ret[i] = window_slice(pat[np.newaxis,:,:], reduce_ratio=0.9+0.1*warp_amount[i]/max_warp)[0]
    return ret

def discriminative_guided_warp_shape(x, labels, batch_size=6, slope_constraint="symmetric", use_window=True):
    return discriminative_guided_warp(x, labels, batch_size, slope_constraint, use_window, dtw_type="shape")


def run_augmentation(x, y, args):
    print("Augmenting %s"%args.data)
    np.random.seed(args.seed)
    x_aug = x
    y_aug = y
    if args.augmentation_ratio > 0:
        augmentation_tags = "%d"%args.augmentation_ratio
        for n in range(args.augmentation_ratio):
            x_temp, augmentation_tags = augment(x, y, args)
            x_aug = np.append(x_aug, x_temp, axis=0)
            y_aug = np.append(y_aug, y, axis=0)
            print("Round %d: %s done"%(n, augmentation_tags))
        if args.extra_tag:
            augmentation_tags += "_"+args.extra_tag
    else:
        augmentation_tags = args.extra_tag
    return x_aug, y_aug, augmentation_tags

def run_augmentation_single(x, y, args):
    # print("Augmenting %s"%args.data)
    np.random.seed(args.seed)

    x_aug = x
    y_aug = y


    if len(x.shape)<3:
        # Augmenting on the entire series: using the input data as "One Big Batch"
        #   Before  -   (sequence_length, num_channels)
        #   After   -   (1, sequence_length, num_channels)
        # Note: the 'sequence_length' here is actually the length of the entire series
        x_input = x[np.newaxis,:]
    elif len(x.shape)==3:
        # Augmenting on the batch series: keep current dimension (batch_size, sequence_length, num_channels)
        x_input = x
    else:
        raise ValueError("Input must be (batch_size, sequence_length, num_channels) dimensional")

    if args.augmentation_ratio > 0:
        augmentation_tags = "%d"%args.augmentation_ratio
        for n in range(args.augmentation_ratio):
            x_aug, augmentation_tags = augment(x_input, y, args)
            # print("Round %d: %s done"%(n, augmentation_tags))
        if args.extra_tag:
            augmentation_tags += "_"+args.extra_tag
    else:
        augmentation_tags = args.extra_tag

    if(len(x.shape)<3):
        # Reverse to two-dimensional in whole series augmentation scenario
        x_aug = x_aug.squeeze(0)
    return x_aug, y_aug, augmentation_tags


def augment(x, y, args):
    import utils.augmentation as aug
    augmentation_tags = ""
    if args.jitter:
        x = aug.jitter(x)
        augmentation_tags += "_jitter"
    if args.scaling:
        x = aug.scaling(x)
        augmentation_tags += "_scaling"
    if args.rotation:
        x = aug.rotation(x)
        augmentation_tags += "_rotation"
    if args.permutation:
        x = aug.permutation(x)
        augmentation_tags += "_permutation"
    if args.randompermutation:
        x = aug.permutation(x, seg_mode="random")
        augmentation_tags += "_randomperm"
    if args.magwarp:
        x = aug.magnitude_warp(x)
        augmentation_tags += "_magwarp"
    if args.timewarp:
        x = aug.time_warp(x)
        augmentation_tags += "_timewarp"
    if args.windowslice:
        x = aug.window_slice(x)
        augmentation_tags += "_windowslice"
    if args.windowwarp:
        x = aug.window_warp(x)
        augmentation_tags += "_windowwarp"
    if args.spawner:
        x = aug.spawner(x, y)
        augmentation_tags += "_spawner"
    if args.dtwwarp:
        x = aug.random_guided_warp(x, y)
        augmentation_tags += "_rgw"
    if args.shapedtwwarp:
        x = aug.random_guided_warp_shape(x, y)
        augmentation_tags += "_rgws"
    if args.wdba:
        x = aug.wdba(x, y)
        augmentation_tags += "_wdba"
    if args.discdtw:
        x = aug.discriminative_guided_warp(x, y)
        augmentation_tags += "_dgw"
    if args.discsdtw:
        x = aug.discriminative_guided_warp_shape(x, y)
        augmentation_tags += "_dgws"
    return x, augmentation_tags