FAIR/engine_self_training.py at main · Externalhappy/FAIR · GitHub

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
import math
import sys
from typing import Iterable
import numpy as np
import torch
import torch.nn.functional as F
import torch.nn as nn
import utils.utils as utils
from timm.utils import accuracy
from tqdm import tqdm
from utils.utils import *
from utils.weighted_text_similarity import *
import gc
gc.collect()
torch.cuda.empty_cache()

def train_one_epoch(args, model: torch.nn.Module,
                    data_loader_u: Iterable,
                    optimizer: torch.optim.Optimizer,
                    amp_autocast, device: torch.device,
                    epoch: int,
                    loss_scaler,
                    lr_schedule_values,
                    train_config,
                    start_steps=None,
                    stats_dict=None,
                    dataset_params=None
                    ):
    model.train()
    metric_logger = utils.MetricLogger(delimiter="  ")
    header = f'Epoch: [{epoch + 1}/{args.epochs}] Iter:'
    print_freq = 10
    start_time = time.time()
    # -----------------------------------------------------------------------------------------------
    for data_iter_step, (inputs, true_labels) in enumerate(metric_logger.log_every(data_loader_u, print_freq, header)):
        img_train_weak = inputs[:, 0].to(device, non_blocking=True)
        img_train_strong = inputs[:, 1].to(device, non_blocking=True)
        true_labels = true_labels.to(device, non_blocking=True)
        # ----------------------- assign learning rate for each step ---------------------------------
        it = start_steps + data_iter_step  # global training iteration
        if lr_schedule_values is not None:
            for i, param_group in enumerate(optimizer.param_groups):
                if lr_schedule_values is not None:
                    param_group["lr"] = lr_schedule_values[it]
        # ----------------------------------- Generate pseudo labels based on the neighbors/diff ---------------------------------------------------------
        with torch.no_grad():
            feat_weak, cls_token =  model(img_train_weak, return_cls_token=True) # (batch_size, embed_dim)
            # --------------------------------------------------------------------------------------------------------------------------------------------
            img_train_crops = inputs[:, 2:].to(device, non_blocking=True)
            crop_feats, crop_cls_token = model(img_train_crops.flatten(0, 1), return_cls_token=True)
            crop_feats = crop_feats.view(img_train_crops.shape[0], img_train_crops.shape[1], -1) # (batch_size, num_crops, embed_dim)
            crop_cls_token = crop_cls_token.view(img_train_crops.shape[0], img_train_crops.shape[1], -1) # (batch_size, num_crops, embed_dim)
            # Compute crop weights by comparing each crop to the weakly augmented image.
            if train_config['use_cls_token']:
                crop_weights = torch.einsum('bd,bcd->bc', cls_token, crop_cls_token)
            else:
                crop_weights = torch.einsum('bd,bcd->bc', feat_weak, crop_feats)
            top_k_indices = torch.topk(crop_weights, k=args.n_crops, dim=-1).indices
            crop_feats = torch.gather(crop_feats, 1, top_k_indices.unsqueeze(-1).expand(-1, -1, crop_feats.shape[-1]))
            crop_weights = torch.gather(crop_weights, 1, top_k_indices)
            crop_weights /= crop_weights.max(dim=-1, keepdim=True).values
            # --------------------------------------------------------------------------------------------------------------------------------------------
            w_i = (train_config['image_scale'] * crop_weights).unsqueeze(-1)
            sampled_crop_embeds = (crop_feats * w_i).sum(dim=1) / w_i.sum(dim=1)
            sampled_crop_embeds = F.normalize(sampled_crop_embeds, dim=-1)
            ##------------------- ours -------------------------------------------------
            logits_classifier = 100 * sampled_crop_embeds @ model.get_classifier().T
            pseudo_labels = torch.argmax(logits_classifier, dim=-1)
            pl_certainty = compute_text_similarity_weights(logits_classifier, model.get_classifier(), dataset_params)
        # ----------------------------------- Training ---------------------------------------------------------
        with amp_autocast():
            feat_strong =  model(img_train_strong)
            output = 100. * feat_strong @ model.get_classifier().t()
            loss = (pl_certainty * F.cross_entropy(output, pseudo_labels, reduction='none')).mean()
            ##---------------------------- Consistency regularization ------------------------------------------------------
            if args.reg_batch:
                probs = F.softmax(output, dim=-1)
                probs_batch_avg = probs.mean(0)
                if data_iter_step == 0:
                    probs_avg = probs_batch_avg
                else:
                    probs_avg = 0.5 * (probs_avg.detach() + probs_batch_avg)
                loss += -train_config['consistency_weight'] * torch.log(probs_avg).mean() / 0.5
        #------------------------------------------------------ For the record -------------------------------
        metric_logger.update(loss=loss.item())
        metric_logger.update(acc_PL=(pseudo_labels == true_labels).float().mean().item() * 100)
        # # ---------------------------------------------------------------------
        if not math.isfinite(loss.item()):
            print(f"Loss is {loss.item()}, stopping training")
            breakpoint()
            sys.exit(1)
        # ## ---------------------- for grad -------------------------------------
        optimizer.zero_grad()
        if loss.requires_grad:
            if loss_scaler is not None:
                grad_norm = loss_scaler(loss, optimizer, clip_grad=1.0, parameters=model.parameters(), create_graph=False)
                metric_logger.update(grad_norm=grad_norm)
            else:
                loss.backward(create_graph=True)
                optimizer.step()
        ##------------------------------------------------------------------------------------
        torch.cuda.synchronize()
        torch.cuda.empty_cache()
        ##------------------------------------------------------------------------------------
    epoch_time = time.time() - start_time
    print('-----------------------------------------------------------------------')
    print(f"Averaged stats: {epoch} : {metric_logger}, Time: {epoch_time:.2f}s")
    print('-----------------------------------------------------------------------')
    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}, stats_dict

def ours_eval(model, inputs):
    feat_test = model(inputs)
    output = 100. * feat_test @ model.get_classifier().t()
    return output

@torch.no_grad()
def evaluate(data_loader, model, device, eval_func=ours_eval, classnames=None,
             show_per_class=False, show_harmonic_mean=False, prev_logits=None, n_iter=0):
    metric_logger = utils.MetricLogger(delimiter="  ")
    header = 'Test:'
    class_correct = defaultdict(int)
    class_total = defaultdict(int)
    model.eval()
    all_logits, all_labels = [], [] if prev_logits is not None else (None, None)
    for batch in metric_logger.log_every(data_loader, 10, header):
        inputs, target = batch[0].to(device, non_blocking=True), batch[1].to(device, non_blocking=True)
        output = eval_func(model, inputs)
        if prev_logits is not None:
            all_logits.append(output)
            all_labels.append(target)
        _, preds = torch.max(output, 1)
        correct_preds = preds.eq(target)
        for label, is_correct in zip(target, correct_preds):
            class_correct[label.item()] += is_correct.item()
            class_total[label.item()] += 1
        acc = accuracy(output, target)[0]
        metric_logger.meters['acc'].update(acc.item(), n=inputs.size(0))
    return_results = {k: meter.global_avg for k, meter in metric_logger.meters.items()}
    if prev_logits is not None:
        all_logits = torch.cat(all_logits, dim=0) + prev_logits
        all_labels = torch.cat(all_labels, dim=0)
        return_results['logits'] = all_logits
        overall_acc = accuracy(all_logits / (n_iter + 1), all_labels)[0].item()
    else:
        overall_acc = metric_logger.acc.global_avg
    print(f"\n=> Overall Acc@1: {overall_acc:.2f}\n")
    return return_results