qmlp.py

   1 #!/usr/bin/env python
   2
   3 # @XREMOTE_HOST: elk.fleuret.org
   4 # @XREMOTE_EXEC: python
   5 # @XREMOTE_PRE: source ${HOME}/misc/venv/pytorch/bin/activate
   6 # @XREMOTE_PRE: killall -u ${USER} -q -9 python || true
   7 # @XREMOTE_PRE: ln -sf ${HOME}/data/pytorch ./data
   8 # @XREMOTE_SEND: *.py *.sh
   9
  10 # Any copyright is dedicated to the Public Domain.
  11 # https://creativecommons.org/publicdomain/zero/1.0/
  12
  13 # Written by Francois Fleuret <francois@fleuret.org>
  14
  15 import math, sys
  16
  17 import torch, torchvision
  18
  19 from torch import nn
  20 from torch.nn import functional as F
  21
  22 ######################################################################
  23
  24 nb_quantization_levels = 101
  25
  26
  27 def quantize(x, xmin, xmax):
  28     return (
  29         ((x - xmin) / (xmax - xmin) * nb_quantization_levels)
  30         .long()
  31         .clamp(min=0, max=nb_quantization_levels - 1)
  32     )
  33
  34
  35 def dequantize(q, xmin, xmax):
  36     return q / nb_quantization_levels * (xmax - xmin) + xmin
  37
  38
  39 ######################################################################
  40
  41
  42 def generate_sets_and_params(
  43     batch_nb_mlps,
  44     nb_samples,
  45     batch_size,
  46     nb_epochs,
  47     device=torch.device("cpu"),
  48     print_log=False,
  49     save_as_examples=False,
  50 ):
  51     data_input = torch.zeros(batch_nb_mlps, 2 * nb_samples, 2, device=device)
  52     data_targets = torch.zeros(
  53         batch_nb_mlps, 2 * nb_samples, dtype=torch.int64, device=device
  54     )
  55
  56     while (data_targets.float().mean(-1) - 0.5).abs().max() > 0.1:
  57         i = (data_targets.float().mean(-1) - 0.5).abs() > 0.1
  58         nb = i.sum()
  59
  60         nb_rec = 8
  61         nb_values = 2  # more increases the min-max gap
  62         support = torch.rand(nb, nb_rec, 2, nb_values, device=device) * 2 - 1
  63         support = support.sort(-1).values
  64         support = support[:, :, :, torch.tensor([0, nb_values - 1])].view(nb, nb_rec, 4)
  65
  66         x = torch.rand(nb, 2 * nb_samples, 2, device=device) * 2 - 1
  67         y = (
  68             (
  69                 (x[:, None, :, 0] >= support[:, :, None, 0]).long()
  70                 * (x[:, None, :, 0] <= support[:, :, None, 1]).long()
  71                 * (x[:, None, :, 1] >= support[:, :, None, 2]).long()
  72                 * (x[:, None, :, 1] <= support[:, :, None, 3]).long()
  73             )
  74             .max(dim=1)
  75             .values
  76         )
  77
  78         data_input[i], data_targets[i] = x, y
  79
  80     train_input, train_targets = (
  81         data_input[:, :nb_samples],
  82         data_targets[:, :nb_samples],
  83     )
  84     test_input, test_targets = data_input[:, nb_samples:], data_targets[:, nb_samples:]
  85
  86     q_train_input = quantize(train_input, -1, 1)
  87     train_input = dequantize(q_train_input, -1, 1)
  88     train_targets = train_targets
  89
  90     q_test_input = quantize(test_input, -1, 1)
  91     test_input = dequantize(q_test_input, -1, 1)
  92     test_targets = test_targets
  93
  94     if save_as_examples:
  95         for k in range(q_train_input.size(0)):
  96             with open(f"example_{k:04d}.dat", "w") as f:
  97                 for u, c in zip(train_input[k], train_targets[k]):
  98                     f.write(f"{c} {u[0].item()} {u[1].item()}\n")
  99
 100     hidden_dim = 32
 101     w1 = torch.randn(batch_nb_mlps, hidden_dim, 2, device=device) / math.sqrt(2)
 102     b1 = torch.zeros(batch_nb_mlps, hidden_dim, device=device)
 103     w2 = torch.randn(batch_nb_mlps, 2, hidden_dim, device=device) / math.sqrt(
 104         hidden_dim
 105     )
 106     b2 = torch.zeros(batch_nb_mlps, 2, device=device)
 107
 108     w1.requires_grad_()
 109     b1.requires_grad_()
 110     w2.requires_grad_()
 111     b2.requires_grad_()
 112     optimizer = torch.optim.Adam([w1, b1, w2, b2], lr=1e-2)
 113
 114     criterion = nn.CrossEntropyLoss()
 115     criterion.to(device)
 116
 117     for k in range(nb_epochs):
 118         acc_train_loss = 0.0
 119         nb_train_errors = 0
 120
 121         for input, targets in zip(
 122             train_input.split(batch_size, dim=1), train_targets.split(batch_size, dim=1)
 123         ):
 124             h = torch.einsum("mij,mnj->mni", w1, input) + b1[:, None, :]
 125             h = F.relu(h)
 126             output = torch.einsum("mij,mnj->mni", w2, h) + b2[:, None, :]
 127             loss = F.cross_entropy(
 128                 output.reshape(-1, output.size(-1)), targets.reshape(-1)
 129             )
 130             acc_train_loss += loss.item() * input.size(0)
 131
 132             wta = output.argmax(-1)
 133             nb_train_errors += (wta != targets).long().sum(-1)
 134
 135             optimizer.zero_grad()
 136             loss.backward()
 137             optimizer.step()
 138
 139         with torch.no_grad():
 140             for p in [w1, b1, w2, b2]:
 141                 m = (
 142                     torch.rand(p.size(), device=p.device) <= k / (nb_epochs - 1)
 143                 ).long()
 144                 pq = quantize(p, -2, 2)
 145                 p[...] = (1 - m) * p + m * dequantize(pq, -2, 2)
 146
 147         train_error = nb_train_errors / train_input.size(1)
 148         acc_train_loss = acc_train_loss / train_input.size(1)
 149
 150         # print(f"{k=} {acc_train_loss=} {train_error=}")
 151
 152     acc_test_loss = 0
 153     nb_test_errors = 0
 154
 155     for input, targets in zip(
 156         test_input.split(batch_size, dim=1), test_targets.split(batch_size, dim=1)
 157     ):
 158         h = torch.einsum("mij,mnj->mni", w1, input) + b1[:, None, :]
 159         h = F.relu(h)
 160         output = torch.einsum("mij,mnj->mni", w2, h) + b2[:, None, :]
 161         loss = F.cross_entropy(output.reshape(-1, output.size(-1)), targets.reshape(-1))
 162         acc_test_loss += loss.item() * input.size(0)
 163
 164         wta = output.argmax(-1)
 165         nb_test_errors += (wta != targets).long().sum(-1)
 166
 167     test_error = nb_test_errors / test_input.size(1)
 168     q_params = torch.cat(
 169         [quantize(p.view(batch_nb_mlps, -1), -2, 2) for p in [w1, b1, w2, b2]], dim=1
 170     )
 171     q_train_set = torch.cat([q_train_input, train_targets[:, :, None]], -1).reshape(
 172         batch_nb_mlps, -1
 173     )
 174     q_test_set = torch.cat([q_test_input, test_targets[:, :, None]], -1).reshape(
 175         batch_nb_mlps, -1
 176     )
 177
 178     return q_train_set, q_test_set, q_params, test_error
 179
 180
 181 ######################################################################
 182
 183
 184 def evaluate_q_params(
 185         q_params, q_set, batch_size=25, device=torch.device("cpu"), nb_mlps_per_batch=1024,
 186         save_as_examples=False,
 187 ):
 188     errors = []
 189     nb_mlps = q_params.size(0)
 190
 191     for n in range(0, nb_mlps, nb_mlps_per_batch):
 192         batch_nb_mlps = min(nb_mlps_per_batch, nb_mlps - n)
 193         batch_q_params = q_params[n : n + batch_nb_mlps]
 194         batch_q_set = q_set[n : n + batch_nb_mlps]
 195         hidden_dim = 32
 196         w1 = torch.empty(batch_nb_mlps, hidden_dim, 2, device=device)
 197         b1 = torch.empty(batch_nb_mlps, hidden_dim, device=device)
 198         w2 = torch.empty(batch_nb_mlps, 2, hidden_dim, device=device)
 199         b2 = torch.empty(batch_nb_mlps, 2, device=device)
 200
 201         with torch.no_grad():
 202             k = 0
 203             for p in [w1, b1, w2, b2]:
 204                 print(f"{p.size()=}")
 205                 x = dequantize(
 206                     batch_q_params[:, k : k + p.numel() // batch_nb_mlps], -2, 2
 207                 ).view(p.size())
 208                 p.copy_(x)
 209                 k += p.numel() // batch_nb_mlps
 210
 211         batch_q_set = batch_q_set.view(batch_nb_mlps, -1, 3)
 212         data_input = dequantize(batch_q_set[:, :, :2], -1, 1).to(device)
 213         data_targets = batch_q_set[:, :, 2].to(device)
 214
 215         print(f"{data_input.size()=} {data_targets.size()=}")
 216
 217         criterion = nn.CrossEntropyLoss()
 218         criterion.to(device)
 219
 220         acc_loss = 0.0
 221         nb_errors = 0
 222
 223         for input, targets in zip(
 224             data_input.split(batch_size, dim=1), data_targets.split(batch_size, dim=1)
 225         ):
 226             h = torch.einsum("mij,mnj->mni", w1, input) + b1[:, None, :]
 227             h = F.relu(h)
 228             output = torch.einsum("mij,mnj->mni", w2, h) + b2[:, None, :]
 229             loss = F.cross_entropy(
 230                 output.reshape(-1, output.size(-1)), targets.reshape(-1)
 231             )
 232             acc_loss += loss.item() * input.size(0)
 233             wta = output.argmax(-1)
 234             nb_errors += (wta != targets).long().sum(-1)
 235
 236         errors.append(nb_errors / data_input.size(1))
 237         acc_loss = acc_loss / data_input.size(1)
 238
 239     return torch.cat(errors)
 240
 241
 242 ######################################################################
 243
 244
 245 def generate_sequence_and_test_set(
 246     nb_mlps,
 247     nb_samples,
 248     batch_size,
 249     nb_epochs,
 250     device,
 251     nb_mlps_per_batch=1024,
 252 ):
 253     seqs, q_test_sets, test_errors = [], [], []
 254
 255     for n in range(0, nb_mlps, nb_mlps_per_batch):
 256         q_train_set, q_test_set, q_params, test_error = generate_sets_and_params(
 257             batch_nb_mlps=min(nb_mlps_per_batch, nb_mlps - n),
 258             nb_samples=nb_samples,
 259             batch_size=batch_size,
 260             nb_epochs=nb_epochs,
 261             device=device,
 262         )
 263
 264         seqs.append(
 265             torch.cat(
 266                 [
 267                     q_train_set,
 268                     q_train_set.new_full(
 269                         (
 270                             q_train_set.size(0),
 271                             1,
 272                         ),
 273                         nb_quantization_levels,
 274                     ),
 275                     q_params,
 276                 ],
 277                 dim=-1,
 278             )
 279         )
 280
 281         q_test_sets.append(q_test_set)
 282         test_errors.append(test_error)
 283
 284     seq = torch.cat(seqs)
 285     q_test_set = torch.cat(q_test_sets)
 286     test_error = torch.cat(test_errors)
 287
 288     return seq, q_test_set, test_error
 289
 290
 291 ######################################################################
 292
 293 if __name__ == "__main__":
 294     import time
 295
 296     batch_nb_mlps, nb_samples = 128, 2500
 297
 298     generate_sets_and_params(
 299         batch_nb_mlps=10,
 300         nb_samples=nb_samples,
 301         batch_size=25,
 302         nb_epochs=100,
 303         device=torch.device("cpu"),
 304         print_log=False,
 305         save_as_examples=True,
 306     )
 307
 308     exit(0)
 309
 310     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 311
 312     start_time = time.perf_counter()
 313
 314     data = []
 315
 316     seq, q_test_set, test_error = generate_sequence_and_test_set(
 317         nb_mlps=batch_nb_mlps,
 318         nb_samples=nb_samples,
 319         device=device,
 320         batch_size=25,
 321         nb_epochs=250,
 322         nb_mlps_per_batch=17,
 323     )
 324
 325     end_time = time.perf_counter()
 326     print(f"{seq.size(0) / (end_time - start_time):.02f} samples per second")
 327
 328     q_train_set = seq[:, : nb_samples * 3]
 329     q_params = seq[:, nb_samples * 3 + 1 :]
 330     print(f"SANITY #2 {q_train_set.size()=} {q_params.size()=} {seq.size()=}")
 331     error_train = evaluate_q_params(q_params, q_train_set, nb_mlps_per_batch=17)
 332     print(f"train {error_train*100}%")
 333     error_test = evaluate_q_params(q_params, q_test_set, nb_mlps_per_batch=17)
 334     print(f"test {error_test*100}%")