imagenet_compress.py

from utils.torch.rand import *
from utils.torch.modules import ImageNet
from model.imagenet_train import Model
from torch.utils.data import *
from discretization import *
from torchvision import datasets, transforms
import random
import time
import argparse
from tqdm import tqdm
import pickle

class ANS:
    def __init__(self, pmfs, bits=31, quantbits=8):
        self.device = pmfs.device
        self.bits = bits
        self.quantbits = quantbits

        # mask of 2**bits - 1 bits
        self.mask = (1 << bits) - 1

        # normalization constants
        self.lbound = 1 << 32
        self.tail_bits = (1 << 32) - 1

        self.seq_len, self.support = pmfs.shape

        # compute pmf's and cdf's scaled up by 2**n
        multiplier = (1 << self.bits) - (1 << self.quantbits)
        self.pmfs = (pmfs * multiplier).long()

        # add ones to counter zero probabilities
        self.pmfs += torch.ones_like(self.pmfs)

        # add remnant to the maximum value of the probabilites
        self.pmfs[torch.arange(0, self.seq_len),torch.argmax(self.pmfs, dim=1)] += ((1 << self.bits) - self.pmfs.sum(1))

        # compute cdf's
        self.cdfs = torch.cumsum(self.pmfs, dim=1) # compute CDF (scaled up to 2**n)
        self.cdfs = torch.cat([torch.zeros([self.cdfs.shape[0], 1], dtype=torch.long, device=self.device), self.cdfs], dim=1) # pad with 0 at the beginning

        # move cdf's and pmf's the cpu for faster encoding and decoding
        self.cdfs = self.cdfs.cpu().numpy()
        self.pmfs = self.pmfs.cpu().numpy()

        assert self.cdfs.shape == (self.seq_len, self.support + 1)
        assert np.all(self.cdfs[:,-1] == (1 << bits))

    def encode(self, x, symbols):
        for i, s in enumerate(symbols):
            pmf = int(self.pmfs[i,s])
            if x[-1] >= ((self.lbound >> self.bits) << 32) * pmf:
                x.append(x[-1] >> 32)
                x[-2] = x[-2] & self.tail_bits
            x[-1] = ((x[-1] // pmf) << self.bits) + (x[-1] % pmf) + int(self.cdfs[i, s])
        return x

    def decode(self, x):
        sequence = np.zeros((self.seq_len,), dtype=np.int64)
        for i in reversed(range(self.seq_len)):
            masked_x = x[-1] & self.mask
            s = np.searchsorted(self.cdfs[i,:-1], masked_x, 'right') - 1
            sequence[i] = s
            x[-1] = int(self.pmfs[i,s]) * (x[-1] >> self.bits) + masked_x - int(self.cdfs[i, s])
            if x[-1] < self.lbound:
                x[-1] = (x[-1] << 32) | x.pop(-2)
        sequence = torch.from_numpy(sequence).to(self.device)
        return x, sequence

def compress(quantbits, nz, bitswap, gpu):
    # model and compression params
    zdim = 8*16*16
    zrange = torch.arange(zdim)
    xdim = 32**2 * 3
    xrange = torch.arange(xdim)
    ansbits = 31 # ANS precision
    type = torch.float64 # datatype throughout compression
    device = f"cuda:{gpu}" # gpu

    # set up the different channel dimension for different latent depths
    if nz == 8:
        reswidth = 252
    elif nz == 4:
        reswidth = 254
    elif nz == 2:
        reswidth = 255
    else:
        reswidth = 256
    assert nz > 0

    print(f"{'Bit-Swap' if bitswap else 'BB-ANS'} - ImageNet - {nz} latent layers - {quantbits} bits quantization")

    # seed for replicating experiment and stability
    np.random.seed(100)
    random.seed(50)
    torch.manual_seed(50)
    torch.cuda.manual_seed(50)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = True

    # compression experiment params
    experiments = 100
    ndatapoints = 100
    decompress = False

    # <=== MODEL ===>
    model = Model(xs = (3, 32, 32), nz=nz, zchannels=8, nprocessing=4, kernel_size=3, resdepth=8, reswidth=reswidth).to(device)
    model.load_state_dict(
        torch.load(f'model/params/imagenet/nz{nz}',
                   map_location=lambda storage, location: storage
                   )
    )
    model.eval()

    print("Discretizing")
    # get discretization bins for latent variables
    zendpoints, zcentres = discretize(nz, quantbits, type, device, model, "imagenet")

    # get discretization bins for discretized logistic
    xbins = ImageBins(type, device, xdim)
    xendpoints = xbins.endpoints()
    xcentres = xbins.centres()

    print("Load data..")
    # <=== DATA ===>
    class ToInt:
        def __call__(self, pic):
            return pic * 255
    transform_ops = transforms.Compose([transforms.ToTensor(), ToInt()])
    test_set = modules.ImageNet(root='model/data/imagenet/test', file='test.npy', transform=transform_ops)

    # sample (experiments, ndatapoints) from test set with replacement
    if not os.path.exists("bitstreams/imagenet/indices"):
        randindices = np.random.choice(len(test_set.dataset), size=(experiments, ndatapoints), replace=False)
        np.save("bitstreams/imagenet/indices", randindices)
    else:
        randindices = np.load("bitstreams/imagenet/indices")

    print("Setting up metrics..")
    # metrics for the results
    nets = np.zeros((experiments, ndatapoints), dtype=np.float)
    elbos = np.zeros((experiments, ndatapoints), dtype=np.float)
    cma = np.zeros((experiments, ndatapoints), dtype=np.float)
    total = np.zeros((experiments, ndatapoints), dtype=np.float)

    print("Compression..")
    for ei in range(experiments):
        print(f"Experiment {ei + 1}")
        subset = Subset(test_set, randindices[ei])
        test_loader = DataLoader(
            dataset=subset,
            batch_size=1, shuffle=False, drop_last=True)
        datapoints = list(test_loader)

        # < ===== COMPRESSION ===>
        # initialize compression
        model.compress()
        state = list(map(int, np.random.randint(low=1 << 16, high=(1 << 32) - 1, size=10000, dtype=np.uint32))) # fill state list with 'random' bits
        state[-1] = state[-1] << 32
        initialstate = state.copy()
        restbits = None

        # <===== SENDER =====>
        iterator = tqdm(range(len(datapoints)), desc="Sender")
        for xi in iterator:
            x = datapoints[xi]
            x = x.to(device).view(xdim)

            # calculate ELBO
            with torch.no_grad():
                model.compress(False)
                logrecon, logdec, logenc, _ = model.loss(x.view((-1,) + model.xs))
                elbo = -logrecon + torch.sum(-logdec + logenc)
                model.compress(True)

            if bitswap:
                # < ===== Bit-Swap ====>
                # inference and generative model
                for zi in range(nz):
                    # inference model
                    input = zcentres[zi - 1, zrange, zsym] if zi > 0 else xcentres[xrange, x.long()]
                    mu, scale = model.infer(zi)(given=input)
                    cdfs = logistic_cdf(zendpoints[zi].t(), mu, scale).t()# most expensive calculation?
                    pmfs = cdfs[:, 1:] - cdfs[:, :-1]
                    pmfs = torch.cat((cdfs[:,0].unsqueeze(1), pmfs, 1. - cdfs[:,-1].unsqueeze(1)), dim=1)

                    # decode z
                    state, zsymtop = ANS(pmfs, bits=ansbits, quantbits=quantbits).decode(state)

                    # save excess bits for calculations
                    if xi == zi == 0:
                        restbits = state.copy()
                        assert len(restbits) > 1, "too few initial bits" # otherwise initial state consists of too few bits

                    # generative model
                    z = zcentres[zi, zrange, zsymtop]
                    mu, scale = model.generate(zi)(given=z)
                    cdfs = logistic_cdf((zendpoints[zi - 1] if zi > 0 else xendpoints).t(), mu, scale).t() # most expensive calculation?
                    pmfs = cdfs[:, 1:] - cdfs[:, :-1]
                    pmfs = torch.cat((cdfs[:,0].unsqueeze(1), pmfs, 1. - cdfs[:,-1].unsqueeze(1)), dim=1)

                    # encode z or x
                    state = ANS(pmfs, bits=ansbits, quantbits=(quantbits if zi > 0 else 8)).encode(state, zsym if zi > 0 else x.long())

                    zsym = zsymtop
            else:
                # < ===== BB-ANS ====>
                # inference and generative model
                zs = []
                for zi in range(nz):
                    # inference model
                    input = zcentres[zi - 1, zrange, zsym] if zi > 0 else xcentres[xrange, x.long()]
                    mu, scale = model.infer(zi)(given=input)
                    cdfs = logistic_cdf(zendpoints[zi].t(), mu, scale).t() # most expensive calculation?
                    pmfs = cdfs[:, 1:] - cdfs[:, :-1]
                    pmfs = torch.cat((cdfs[:, 0].unsqueeze(1), pmfs, 1. - cdfs[:, -1].unsqueeze(1)), dim=1)

                    # decode z
                    state, zsymtop = ANS(pmfs, bits=ansbits, quantbits=quantbits).decode(state)
                    zs.append(zsymtop)

                    zsym = zsymtop

                # save excess bits for calculations
                if xi == 0:
                    restbits = state.copy()
                    assert len(restbits) > 1  # otherwise initial state consists of too few bits

                for zi in range(nz):
                    # generative model
                    zsymtop = zs.pop(0)
                    z = zcentres[zi, zrange, zsymtop]
                    mu, scale = model.generate(zi)(given=z)
                    cdfs = logistic_cdf((zendpoints[zi - 1] if zi > 0 else xendpoints).t(), mu, scale).t() # most expensive calculation?
                    pmfs = cdfs[:, 1:] - cdfs[:, :-1]
                    pmfs = torch.cat((cdfs[:, 0].unsqueeze(1), pmfs, 1. - cdfs[:, -1].unsqueeze(1)), dim=1)

                    # encode z or x
                    state = ANS(pmfs, bits=ansbits, quantbits=(quantbits if zi > 0 else 8)).encode(state, zsym if zi > 0 else x.long())

                    zsym = zsymtop

                assert zs == []

            # prior
            cdfs = logistic_cdf(zendpoints[-1].t(), torch.zeros(1, device=device, dtype=type), torch.ones(1, device=device, dtype=type)).t()
            pmfs = cdfs[:, 1:] - cdfs[:, :-1]
            pmfs = torch.cat((cdfs[:, 0].unsqueeze(1), pmfs, 1. - cdfs[:, -1].unsqueeze(1)), dim=1)

            # encode prior
            state = ANS(pmfs, bits=ansbits, quantbits=quantbits).encode(state, zsymtop)

            # calculating bits
            totaladdedbits = (len(state) - len(initialstate)) * 32
            totalbits = (len(state) - (len(restbits) - 1)) * 32

            # logging
            nets[ei, xi] = (totaladdedbits / xdim) - nets[ei, :xi].sum()
            elbos[ei, xi] = elbo.item() / xdim
            cma[ei, xi] = totalbits / (xdim * (xi + 1))
            total[ei, xi] = totalbits

            iterator.set_postfix_str(s=f"N:{nets[ei,:xi+1].mean():.2f}±{nets[ei,:xi+1].std():.2f}, D:{nets[ei,:xi+1].mean()-elbos[ei,:xi+1].mean():.4f}, C: {cma[ei,:xi+1].mean():.2f}, T: {totalbits:.0f}", refresh=False)

        # write state to file
        with open(f"bitstreams/imagenet/nz{nz}/{'Bit-Swap' if bitswap else 'BB-ANS'}/{quantbits}bits_nz{nz}_experiment{ei + 1}", "wb") as fp:
            pickle.dump(state, fp)

        state = None
        # open state file
        with open(f"bitstreams/imagenet/nz{nz}/{'Bit-Swap' if bitswap else 'BB-ANS'}/{quantbits}bits_nz{nz}_experiment{ei + 1}", "rb") as fp:
            state = pickle.load(fp)

        if not decompress:
            continue

        # <===== RECEIVER =====>
        datapoints.reverse()
        iterator = tqdm(range(len(datapoints)), desc="Receiver", postfix=f"decoded {None}")
        for xi in iterator:
            x = datapoints[xi]
            x = x.to(device).view(xdim)

            # prior
            cdfs = logistic_cdf(zendpoints[-1].t(), torch.zeros(1, device=device, dtype=type), torch.ones(1, device=device, dtype=type)).t()
            pmfs = cdfs[:, 1:] - cdfs[:, :-1]
            pmfs = torch.cat((cdfs[:, 0].unsqueeze(1), pmfs, 1. - cdfs[:, -1].unsqueeze(1)), dim=1)

            # decode z
            state, zsymtop = ANS(pmfs, bits=ansbits, quantbits=quantbits).decode(state)

            if bitswap:
                # < ===== Bit-Swap ====>
                # inference and generative model
                for zi in reversed(range(nz)):
                    # generative model
                    z = zcentres[zi, zrange, zsymtop]
                    mu, scale = model.generate(zi)(given=z)
                    cdfs = logistic_cdf((zendpoints[zi - 1] if zi > 0 else xendpoints).t(), mu, scale).t() # most expensive calculation?
                    pmfs = cdfs[:, 1:] - cdfs[:, :-1]
                    pmfs = torch.cat((cdfs[:, 0].unsqueeze(1), pmfs, 1. - cdfs[:, -1].unsqueeze(1)), dim=1)

                    # decode z or x
                    state, sym = ANS(pmfs, bits=ansbits, quantbits=quantbits if zi > 0 else 8).decode(state)

                    # inference model
                    input = zcentres[zi - 1, zrange, sym] if zi > 0 else xcentres[xrange, sym]
                    mu, scale = model.infer(zi)(given=input)
                    cdfs = logistic_cdf(zendpoints[zi].t(), mu, scale).t() # most expensive calculation?
                    pmfs = cdfs[:, 1:] - cdfs[:, :-1]
                    pmfs = torch.cat((cdfs[:, 0].unsqueeze(1), pmfs, 1. - cdfs[:, -1].unsqueeze(1)), dim=1)

                    # encode z
                    state = ANS(pmfs, bits=ansbits, quantbits=quantbits).encode(state, zsymtop)

                    zsymtop = sym

                assert torch.all(x.long() == zsymtop), f"decoded datapoint does not match {xi + 1}"

            else:
                # < ===== BB-ANS ====>
                # inference and generative model
                zs = [zsymtop]
                for zi in reversed(range(nz)):
                    # generative model
                    z = zcentres[zi, zrange, zsymtop]
                    mu, scale = model.generate(zi)(given=z)
                    cdfs = logistic_cdf((zendpoints[zi - 1] if zi > 0 else xendpoints).t(), mu, scale).t() # most expensive calculation?
                    pmfs = cdfs[:, 1:] - cdfs[:, :-1]
                    pmfs = torch.cat((cdfs[:, 0].unsqueeze(1), pmfs, 1. - cdfs[:, -1].unsqueeze(1)), dim=1)

                    # decode z or x
                    state, sym = ANS(pmfs, bits=ansbits, quantbits=quantbits if zi > 0 else 8).decode(state)
                    zs.append(sym)
                    zsymtop = sym

                zsymtop = zs.pop(0)
                for zi in reversed(range(nz)):
                    # inference model
                    sym = zs.pop(0) if zi > 0 else zs[0]

                    input = zcentres[zi - 1, zrange, sym] if zi > 0 else xcentres[xrange, sym]
                    mu, scale = model.infer(zi)(given=input)
                    cdfs = logistic_cdf(zendpoints[zi].t(), mu, scale).t() # most expensive calculation?
                    pmfs = cdfs[:, 1:] - cdfs[:, :-1]
                    pmfs = torch.cat((cdfs[:, 0].unsqueeze(1), pmfs, 1. - cdfs[:, -1].unsqueeze(1)), dim=1)

                    # encode z
                    state = ANS(pmfs, bits=ansbits, quantbits=quantbits).encode(state, zsymtop)

                    zsymtop = sym
                # check if decoded datapoint matches the real datapoint
                assert torch.all(x.long() == zs[0]), f"decoded datapoint does not match {xi + 1}"
            iterator.set_postfix_str(s=f"decoded {len(datapoints) - xi}")

        # check if the initial state matches the output state
        assert initialstate == state

    print(f"N:{nets.mean():.4f}±{nets.std():.2f}, E:{elbos.mean():.4f}±{elbos.std():.2f}, D:{nets.mean() - elbos.mean():.6f}")

    # save experiments
    np.save(f"plots/mnist{nz}/{'bitswap' if bitswap else 'bbans'}_{quantbits}bits_nets",nets)
    np.save(f"plots/mnist{nz}/{'bitswap' if bitswap else 'bbans'}_{quantbits}bits_elbos", elbos)
    np.save(f"plots/mnist{nz}/{'bitswap' if bitswap else 'bbans'}_{quantbits}bits_cmas",cma)
    np.save(f"plots/mnist{nz}/{'bitswap' if bitswap else 'bbans'}_{quantbits}bits_total", total)

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--gpu', default=0, type=int) # assign to gpu
    parser.add_argument('--nz', default=2, type=int) # choose number of latent variables
    parser.add_argument('--quantbits', default=10, type=int) # choose discretization precision
    parser.add_argument('--bitswap', default=1, type=int) # choose whether to use Bit-Swap or not

    args = parser.parse_args()
    print(args)

    gpu = args.gpu
    nz = args.nz
    quantbits = args.quantbits
    bitswap = args.bitswap

    for nz in [2, 4]:
        for bits in [quantbits]:
            for bitswap in [bitswap]:
                compress(bits, nz, bitswap, gpu)