davanstrien HF Staff commited on
Commit
e8231ab
·
verified ·
1 Parent(s): 6e90523

Upload train_detr.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. train_detr.py +5 -5
train_detr.py CHANGED
@@ -218,7 +218,7 @@ t_start = time.time()
218
  log("starting training; epochs", ARGS.epochs, "batch", ARGS.batch_size)
219
  for epoch in range(1, ARGS.epochs+1):
220
  model.train()
221
- ep_loss = 0.0; ep_ce = 0.0; ep_bbox = 0.0; ep_giou = 0.0; nb = 0
222
  for images, anns, sizes, idx in tqdm(train_dl, desc="epoch %d"%epoch):
223
  anns_pp = [{"image_id": int(idx[k]), "annotations": anns[k]} for k in range(len(images))]
224
  enc = processor(images=images, annotations=anns_pp, return_tensors="pt")
@@ -230,14 +230,14 @@ for epoch in range(1, ARGS.epochs+1):
230
  loss.backward()
231
  optim.step()
232
  ep_loss += loss.item()
233
- ep_ce += out.loss_ce.item()
234
- ep_bbox += out.loss_bbox.item()
235
- ep_giou += out.loss_giou.item()
236
  nb += 1
237
  global_step += 1
238
  if ARGS.max_steps and global_step >= ARGS.max_steps:
239
  break
240
- log("epoch %d loss %.4f ce %.4f bbox %.4f giou %.4f" % (epoch, ep_loss/nb, ep_ce/nb, ep_bbox/nb, ep_giou/nb))
 
241
  m = evaluate()
242
  log("VAL", json.dumps(m))
243
  if m and m["AP"] > best["AP"]:
 
218
  log("starting training; epochs", ARGS.epochs, "batch", ARGS.batch_size)
219
  for epoch in range(1, ARGS.epochs+1):
220
  model.train()
221
+ ep_loss = 0.0; ep_losssum = {}; nb = 0
222
  for images, anns, sizes, idx in tqdm(train_dl, desc="epoch %d"%epoch):
223
  anns_pp = [{"image_id": int(idx[k]), "annotations": anns[k]} for k in range(len(images))]
224
  enc = processor(images=images, annotations=anns_pp, return_tensors="pt")
 
230
  loss.backward()
231
  optim.step()
232
  ep_loss += loss.item()
233
+ for kk, vv in out.loss_dict.items():
234
+ ep_losssum[kk] = ep_losssum.get(kk, 0.0) + vv.detach().item()
 
235
  nb += 1
236
  global_step += 1
237
  if ARGS.max_steps and global_step >= ARGS.max_steps:
238
  break
239
+ log("epoch %d loss %.4f %s" % (epoch, ep_loss/nb,
240
+ " ".join("%s %.4f"%(k, v/nb) for k,v in ep_losssum.items())))
241
  m = evaluate()
242
  log("VAL", json.dumps(m))
243
  if m and m["AP"] > best["AP"]: