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
10 # Any copyright is dedicated to the Public Domain.
11 # https://creativecommons.org/publicdomain/zero/1.0/
13 # Written by Francois Fleuret <francois@fleuret.org>
17 import torch, torchvision
20 from torch.nn import functional as F
22 ######################################################################
24 nb_quantization_levels = 101
27 def quantize(x, xmin, xmax):
29 ((x - xmin) / (xmax - xmin) * nb_quantization_levels)
31 .clamp(min=0, max=nb_quantization_levels - 1)
35 def dequantize(q, xmin, xmax):
36 return q / nb_quantization_levels * (xmax - xmin) + xmin
39 ######################################################################
44 def generate_sets_and_params(
49 device=torch.device("cpu"),
52 data_input = torch.zeros(batch_nb_mlps, 2 * nb_samples, 2, device=device)
53 data_targets = torch.zeros(
54 batch_nb_mlps, 2 * nb_samples, dtype=torch.int64, device=device
57 while (data_targets.float().mean(-1) - 0.5).abs().max() > 0.1:
58 i = (data_targets.float().mean(-1) - 0.5).abs() > 0.1
62 support = torch.rand(nb, nb_rec, 2, 3, device=device) * 2 - 1
63 support = support.sort(-1).values
64 support = support[:, :, :, torch.tensor([0, 2])].view(nb, nb_rec, 4)
66 x = torch.rand(nb, 2 * nb_samples, 2, device=device) * 2 - 1
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()
78 data_input[i], data_targets[i] = x, y
80 train_input, train_targets = (
81 data_input[:, :nb_samples],
82 data_targets[:, :nb_samples],
84 test_input, test_targets = data_input[:, nb_samples:], data_targets[:, nb_samples:]
86 q_train_input = quantize(train_input, -1, 1)
87 train_input = dequantize(q_train_input, -1, 1)
88 train_targets = train_targets
90 q_test_input = quantize(test_input, -1, 1)
91 test_input = dequantize(q_test_input, -1, 1)
92 test_targets = test_targets
95 w1 = torch.randn(batch_nb_mlps, hidden_dim, 2, device=device) / math.sqrt(2)
96 b1 = torch.zeros(batch_nb_mlps, hidden_dim, device=device)
97 w2 = torch.randn(batch_nb_mlps, 2, hidden_dim, device=device) / math.sqrt(hidden_dim)
98 b2 = torch.zeros(batch_nb_mlps, 2, device=device)
104 optimizer = torch.optim.Adam([w1, b1, w2, b2], lr=1e-2)
106 criterion = nn.CrossEntropyLoss()
109 for k in range(nb_epochs):
113 for input, targets in zip(
114 train_input.split(batch_size, dim=1), train_targets.split(batch_size, dim=1)
116 h = torch.einsum("mij,mnj->mni", w1, input) + b1[:, None, :]
118 output = torch.einsum("mij,mnj->mni", w2, h) + b2[:, None, :]
119 loss = F.cross_entropy(
120 output.reshape(-1, output.size(-1)), targets.reshape(-1)
122 acc_train_loss += loss.item() * input.size(0)
124 wta = output.argmax(-1)
125 nb_train_errors += (wta != targets).long().sum(-1)
127 optimizer.zero_grad()
131 with torch.no_grad():
132 for p in [w1, b1, w2, b2]:
134 torch.rand(p.size(), device=p.device) <= k / (nb_epochs - 1)
136 pq = quantize(p, -2, 2)
137 p[...] = (1 - m) * p + m * dequantize(pq, -2, 2)
139 train_error = nb_train_errors / train_input.size(1)
140 acc_train_loss = acc_train_loss / train_input.size(1)
142 # print(f"{k=} {acc_train_loss=} {train_error=}")
144 q_params = torch.cat(
145 [quantize(p.view(batch_nb_mlps, -1), -2, 2) for p in [w1, b1, w2, b2]], dim=1
147 q_train_set = torch.cat([q_train_input, train_targets[:, :, None]], -1).reshape(
150 q_test_set = torch.cat([q_test_input, test_targets[:, :, None]], -1).reshape(
154 return q_train_set, q_test_set, q_params
157 ######################################################################
160 def evaluate_q_params(q_params, q_set, batch_size=25, device=torch.device("cpu"), nb_mlps_per_batch=1024):
163 nb_mlps = q_params.size(0)
165 for n in range(0,nb_mlps,nb_mlps_per_batch):
166 batch_nb_mlps = min(nb_mlps_per_batch,nb_mlps-n)
167 batch_q_params = q_params[n:n+batch_nb_mlps]
168 batch_q_set = q_set[n:n+batch_nb_mlps]
170 w1 = torch.empty(batch_nb_mlps, hidden_dim, 2, device=device)
171 b1 = torch.empty(batch_nb_mlps, hidden_dim, device=device)
172 w2 = torch.empty(batch_nb_mlps, 2, hidden_dim, device=device)
173 b2 = torch.empty(batch_nb_mlps, 2, device=device)
175 with torch.no_grad():
177 for p in [w1, b1, w2, b2]:
178 print(f"{p.size()=}")
179 x = dequantize(batch_q_params[:, k : k + p.numel() // batch_nb_mlps], -2, 2).view(
183 k += p.numel() // batch_nb_mlps
185 batch_q_set = batch_q_set.view(batch_nb_mlps, -1, 3)
186 data_input = dequantize(batch_q_set[:, :, :2], -1, 1).to(device)
187 data_targets = batch_q_set[:, :, 2].to(device)
189 print(f"{data_input.size()=} {data_targets.size()=}")
191 criterion = nn.CrossEntropyLoss()
197 for input, targets in zip(
198 data_input.split(batch_size, dim=1), data_targets.split(batch_size, dim=1)
200 h = torch.einsum("mij,mnj->mni", w1, input) + b1[:, None, :]
202 output = torch.einsum("mij,mnj->mni", w2, h) + b2[:, None, :]
203 loss = F.cross_entropy(output.reshape(-1, output.size(-1)), targets.reshape(-1))
204 acc_loss += loss.item() * input.size(0)
205 wta = output.argmax(-1)
206 nb_errors += (wta != targets).long().sum(-1)
208 errors.append(nb_errors / data_input.size(1))
209 acc_loss = acc_loss / data_input.size(1)
212 return torch.cat(errors)
215 ######################################################################
218 def generate_sequence_and_test_set(
224 nb_mlps_per_batch=1024,
227 seqs, q_test_sets = [],[]
229 for n in range(0,nb_mlps,nb_mlps_per_batch):
230 q_train_set, q_test_set, q_params = generate_sets_and_params(
231 batch_nb_mlps = min(nb_mlps_per_batch, nb_mlps - n),
232 nb_samples=nb_samples,
233 batch_size=batch_size,
238 seqs.append(torch.cat(
241 q_train_set.new_full(
246 nb_quantization_levels,
253 q_test_sets.append(q_test_set)
255 seq = torch.cat(seqs)
256 q_test_set = torch.cat(q_test_sets)
258 return seq, q_test_set
261 ######################################################################
263 if __name__ == "__main__":
266 batch_nb_mlps, nb_samples = 128, 500
268 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
270 start_time = time.perf_counter()
274 seq, q_test_set = generate_sequence_and_test_set(
275 nb_mlps=batch_nb_mlps,
276 nb_samples=nb_samples,
283 end_time = time.perf_counter()
284 print(f"{seq.size(0) / (end_time - start_time):.02f} samples per second")
286 q_train_set = seq[:, : nb_samples * 3]
287 q_params = seq[:, nb_samples * 3 + 1 :]
288 print(f"SANITY #2 {q_train_set.size()=} {q_params.size()=} {seq.size()=}")
289 error_train = evaluate_q_params(q_params, q_train_set, nb_mlps_per_batch=17)
290 print(f"train {error_train*100}%")
291 error_test = evaluate_q_params(q_params, q_test_set, nb_mlps_per_batch=17)
292 print(f"test {error_test*100}%")