File size: 10,433 Bytes
a8639ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
import torch

# from architecture import DecoderTransformer
from builtin_architecture import make_model
import os
import sys
import time
from dataset import dataset, get_train_dataset, get_dataloader
import torch.nn.functional as F
from tqdm import tqdm, trange
import heapq

EXPERIMENT_DIRECTORY = "runs/code-decoder-v23-mega"  # "runs/code-decoder-v22-bigset-tuner"  # "runs/code-decoder-v21-alltrains-tuner"#"runs/code-decoder-v19-bigset-5k"#"runs/code-decoder-v18-allTrains-customTokenizer"#"runs/code-decoder-v17-bpe-upscale"#"runs/code-decoder-v16-upscale"#"runs/code-decoder-v13-rescaling-smaller-retrained"  # "runs/code-decoder-v12-dummy"  # "runs/code-decoder-v11-vanilla-alphabet"#"runs/code-decoder-v10-vanilla-smaller-batchfirst"#"runs/code-decoder-v9-vanilla-smaller"#"runs/code-decoder-v8-smaller"  # "runs/code-decoder-v4-improved"  # shakespeare-test, run1-python

device = "mps" if torch.backends.mps.is_available() else "cpu"

device = "cpu"


def evaluate_topk(model, start_sequence, amt=10, k=20, temperature=1.0, device="cpu"):
    generated_sequence = start_sequence.clone().to(device)

    model.eval()
    with torch.no_grad():
        for _ in trange(amt, leave=False, dynamic_ncols=True, desc="topk"):
            seq = generated_sequence
            results = model(seq, transpose=True)
            results = results.transpose(0, 1)

            logits = results.reshape(-1, results.size(-1))[-1]

            logits = logits / temperature

            top_k_values, top_k_indices = torch.topk(logits, k)
            top_k_probs = F.softmax(top_k_values, dim=-1)

            sampled_index = torch.multinomial(top_k_probs, 1).item()
            next_token = top_k_indices[sampled_index].unsqueeze(0)

            generated_sequence = torch.cat(
                (generated_sequence, next_token.unsqueeze(0)), dim=1
            )

    return generated_sequence


def evaluate_topp(model, start_sequence, amt=10, p=0.9, temperature=1.0, device="cpu"):
    generated_sequence = start_sequence.clone().to(device)

    model.eval()
    with torch.no_grad():
        for _ in trange(amt, leave=False, dynamic_ncols=True, desc="topp"):
            seq = generated_sequence
            results = model(seq, transpose=True)
            results = results.transpose(0, 1)

            logits = results.reshape(-1, results.size(-1))[-1]
            logits = logits / temperature

            probs = F.softmax(logits, dim=-1)

            sorted_probs, sorted_indices = torch.sort(probs, descending=True)
            cumulative_probs = torch.cumsum(sorted_probs, dim=-1)

            cutoff_idx = torch.where(cumulative_probs > p)[0][0] + 1
            top_p_probs = sorted_probs[:cutoff_idx]
            top_p_indices = sorted_indices[:cutoff_idx]

            # Normalize selected probabilities
            top_p_probs /= top_p_probs.sum()

            # Sample from the top-p tokens
            sampled_index = torch.multinomial(top_p_probs, 1).item()
            next_token = top_p_indices[sampled_index].unsqueeze(0)

            generated_sequence = torch.cat(
                (generated_sequence, next_token.unsqueeze(0)), dim=1
            )

    return generated_sequence


def evaluate_beam(model, start_sequence, k=2, amt=10, temperature=0.8, device="cpu"):
    generated_sequence = start_sequence.clone().to(device)

    model.eval()

    # Initialize beam candidates (shape: [k, seq_len])
    current_beams = generated_sequence.expand(k, -1)
    current_beam_scores = torch.zeros(k, device=device)

    with torch.no_grad():
        for _ in trange(amt, leave=False, dynamic_ncols=True, desc="beam"):
            all_candidates = []

            # Process each beam
            for i in range(k):
                seq = current_beams[i].unsqueeze(0)  # Shape: [1, seq_len]
                results = model(seq, transpose=True)
                results = results.transpose(0, 1)  # Ensure batch-first shape

                logits = results[:, -1, :] / temperature  # Last token logits
                topk_values, topk_indices = torch.topk(logits, k)  # Shape: [1, k]

                # Expand beam by top-k choices
                for j in range(k):
                    candidate = torch.cat((seq, topk_indices[:, j].unsqueeze(0)), dim=1)
                    score = current_beam_scores[i] + topk_values[:, j]
                    all_candidates.append((candidate, score))

            # Select top-k sequences
            all_candidates.sort(key=lambda x: x[1], reverse=True)  # Sort by score
            top_candidates = all_candidates[:k]  # Keep top-k

            current_beams = torch.cat([candidate for candidate, _ in top_candidates])
            current_beam_scores = torch.tensor(
                [score.item() for _, score in top_candidates], device=device
            )

    return current_beams[0]  # Return the best beam sequence


def evaluate(
    model,
    start_sequence,
    amt=10,
):
    generated_sequence = start_sequence.clone()
    generated_sequence = generated_sequence.to(device)

    model.eval()
    with torch.no_grad():
        for _ in trange(amt, leave=False):
            seq = generated_sequence
            results = model(seq, transpose=True)
            results = results.transpose(0, 1)

            next_token = torch.argmax(results.reshape(-1, results.size(-1)), dim=1)[
                -1
            ].unsqueeze(0)

            generated_sequence = torch.cat(
                (generated_sequence, next_token.unsqueeze(0)), dim=1
            )

    return generated_sequence


def tester_exactly_like_trainingmanager_please_please_work(model, rawbatch):
    labels = rawbatch[:, 1:].contiguous()
    batch = rawbatch[:, :-1].contiguous()
    results = model(batch, transpose=True)
    results = results.transpose(0, 1)
    print(
        torch.sum(
            torch.argmax(results.reshape(-1, results.size(-1)), dim=1)
            == labels.reshape(-1)
        )
        / len(labels.reshape(-1))
    )
    return torch.argmax(results.reshape(-1, results.size(-1)), dim=1), labels.reshape(
        -1
    )


def tester_exactly_like_trainingmanager_only_last_please_work(model, rawbatch):
    labels = rawbatch[:, 1:].contiguous()
    batch = rawbatch[:, :-1].contiguous()

    batch = batch[-1].unsqueeze(0)
    labels = labels[-1].unsqueeze(0)  # works bc my data is initially batch-first

    results = model(batch, transpose=True)
    results = results.transpose(0, 1)
    print(
        torch.sum(
            torch.argmax(results.reshape(-1, results.size(-1)), dim=1)
            == labels.reshape(-1)
        )
        / len(labels.reshape(-1))
    )
    return torch.argmax(results.reshape(-1, results.size(-1)), dim=1), labels.reshape(
        -1
    )

    # def tester_exactly_like_trainingmanager_just_next_given_seq_pls(model, seq):
    #     seq = seq.unsqueeze(0)

    #     results = model(batch, transpose=True)
    #     results = results.transpose(0, 1)

    return torch.argmax(results.reshape(-1, results.size(-1)), dim=1)[-1]


def compute_entropy(logits):

    probs = F.softmax(logits, dim=-1)
    entropy = -(probs * probs.log()).sum(dim=-1)  # Entropy, I guess
    return entropy.mean().item()


def main():
    # net = DecoderTransformer(vocab_size=199, num_blocks=1)
    net = make_model()
    net.to(device)
    print(os.path.join(EXPERIMENT_DIRECTORY, "ckpt", "latest.pt"))
    net.load_state_dict(
        torch.load(
            os.path.join(EXPERIMENT_DIRECTORY, "ckpt", "latest.pt"), weights_only=True
        )
    )

    for name, param in net.named_parameters():
        if torch.isnan(param).any():
            print(f"NaN found in {name}")
    for name, param in net.named_parameters():
        if param.grad is not None and torch.isnan(param.grad).any():
            print(f"NaN found in gradients of {name}")
    loader = get_dataloader(get_train_dataset())
    torch.random.manual_seed(
        sum([ord(i) for i in input("seed? ")])
    )  # so people can write whatever there
    for data in loader:
        batch, attn_mask = data

        print(
            tester_exactly_like_trainingmanager_please_please_work(net, rawbatch=batch)
        )
        print("pretty please")

        print(
            tester_exactly_like_trainingmanager_only_last_please_work(
                net, rawbatch=batch
            )
        )
        print("please please please")

        # print(
        #     tester_exactly_like_trainingmanager_just_next_given_seq_pls(
        #         net, seq=batch[:, :-1].contiguous()[-1]
        #     )
        # )
        # print(f"Answer was {batch[:,1:].contiguous()[-1][-1]}")
        # print("please please please")

        # print(
        #     tester_exactly_like_trainingmanager_just_next_given_seq_pls(
        #         net, seq=batch[:, :-1].contiguous()[-1][:10]
        #     )
        # )
        # print(f"Answer was {batch[:,1:].contiguous()[-1][10]}")
        # print("please please please")

        labels = batch[:, 1:].contiguous()
        batch = batch[:, :-1].contiguous()

        batch = batch[0]
        labels = labels[0]

        batch = batch[:100]
        labels = labels[:100]
        print("Getting first 100 tokens for batch and labels")

        # inp, mask = dataset[0]

        # inp = inp[:-1]
        print(batch)
        print(dataset.manager.decode(batch))
        print("batch ^ labels v")
        print(dataset.manager.decode(labels))
        print("that's inp I guess ^^")
        with torch.no_grad():
            logits = net(batch.unsqueeze(0))  # Pass batch through model
            entropy = compute_entropy(
                logits[:, -1, :]
            )  # Compute entropy at last token position

        print(f"Entropy of last token: {entropy:.4f}")
        # print("USING TOPK")
        # result = evaluate_topk(net, batch.unsqueeze(0), amt=100)
        # print(result)
        # print(
        #     dataset.manager.decode(result[0]),
        #     " | PREFIX FROM TRAIN DSET:",
        #     dataset.manager.decode(batch),
        # )

        print("USING BEAM")
        result = evaluate_beam(net, batch.unsqueeze(0), amt=100, k=3)

        result = dataset.manager.decode(result)
        batch_str = dataset.manager.decode(batch)

        result = f"<data>\n{batch_str}</data>\n{result[len(batch_str):]}"

        print(result)

        # print(dataset.manager.raw_decode(81))

        break


if __name__ == "__main__":
    main()