training sweeper checkin

codes/sweep.py

@@ -0,0 +1,52 @@
+import functools
+import os
+from multiprocessing.pool import ThreadPool
+
+import torch
+
+from train import Trainer
+from utils import options as option
+
+def launch_trainer(opt, opt_path=''):
+    rank = opt['gpu_ids'][0]
+    os.environ['CUDA_VISIBLE_DEVICES'] = [rank]
+    print('export CUDA_VISIBLE_DEVICES=' + rank)
+    trainer = Trainer()
+    opt['dist'] = False
+    trainer.rank = -1
+    torch.cuda.set_device(rank)
+    trainer.init(opt_path, opt, 'none')
+    trainer.do_training()
+
+if __name__ == '__main__':
+    """
+    Ad-hoc script (hard coded; no command-line parameters) that spawns multiple separate trainers from a single options
+    file, with a hard-coded set of modifications.
+    """
+    base_opt = '../options/train_diffusion_tts.yml'
+    modifications = {
+        'baseline': {},
+        'only_conv': {'networks': {'generator': {'kwargs': {'cond_transformer_depth': 4, 'mid_transformer_depth': 1}}}},
+        'intermediary_attention': {'networks': {'generator': {'kwargs': {'attention_resolutions': [32,64], 'num_res_blocks': [2, 2, 2, 2, 2, 2, 2]}}}},
+        'more_resblocks': {'networks': {'generator': {'kwargs': {'num_res_blocks': [3, 3, 3, 3, 3, 3, 2]}}}},
+        'less_resblocks': {'networks': {'generator': {'kwargs': {'num_res_blocks': [1, 1, 1, 1, 1, 1, 1]}}}},
+        'wider': {'networks': {'generator': {'kwargs': {'channel_mult': [1,2,4,6,8,8,8]}}}},
+        'inject_every_layer': {'networks': {'generator': {'kwargs': {'token_conditioning_resolutions': [1,2,4,8,16,32,64]}}}},
+        'deep_conditioning': {'networks': {'generator': {'kwargs': {'cond_transformer_depth': 12}}}},
+    }
+    opt = option.parse(base_opt, is_train=True)
+    all_opts = []
+    for i, (mod, mod_dict) in enumerate(modifications.items()):
+        nd = opt.copy()
+        nd.update(mod_dict)
+        opt['gpu_ids'] = [i]
+        nd['name'] = f'{nd["name"]}_{mod}'
+        nd['wandb_run_name'] = mod
+        base_path = nd['path']['log']
+        for k, p in nd['path'].items():
+            if isinstance(p, str) and base_path in p:
+                nd['path'][k] = p.replace(base_path, f'{base_path}\\{mod}')
+        all_opts.append(nd)
+
+    with ThreadPool(len(modifications)) as pool:
+        list(pool.imap(functools.partial(launch_trainer, opt_path=base_opt), all_opts))

codes/trainer/steps.py

@@ -320,7 +320,7 @@ class ConfigurableStep(Module):
                 else:
                     self.nan_counter = 0
 
-            if self.clip_grad_eps is not None:
+            if self.clip_grad_eps is not None and self.clip_grad_eps != 0:
                 for pgn, pg in zip(opt._group_names, opt.param_groups):
                     grad_norm = clip_grad_norm(pg['params'], pgn, self.clip_grad_eps)
                     if torch.isnan(grad_norm):