# DAY 2 EXERCISES 2022-02

Select a free account from the table below and write your name to the "User" column. If you have trouble editing this page, you can ask for an account in Zoom chat.

Passwords will be provided in Zoom chat.

| Training account | User |
| ---------------- | --------------------- |
| training080 | Stefan Rua |
| training081 | Pierpaolo Pierpaolo |
| training082 | César |
| training083 | Patrik Lauha |
| training084 | Till Sawala |
| training085 | Derui Zhu |
| training086 | Mika Jalava |
| training087 | Katarina TD |
| training088 | Ana Gonzalez |
| training089 | Alvaro Tolosa-Delgado |
| training090 | Yassine Kaddouri |
| training091 | saurav kumar |
| training092 | Oleksandr Karasov |
| training093 | E. Honkavaara |
| training094 | Vaclav Hrbek |
| training095 | Panu Somervuo |
| training096 | Mauro |
| training097 | Outi |
| training098 | Nitin Shukla |
| training099 | Rémi Pétuya |
| training100 | Timo Weckman |
| training101 | Parasuram |
| training102 | |
| training103 | Akash James |
| training104 | |
| training105 | Daniel Caviedes |
| training106 | |
| training107 | Enda Carroll |
| training108 | Malik |
| training109 | |
| training110 | Yorgos Sof |
| training111 | Carlos Sánchez |
| training112 | Pablo Sanz Mercado |
| training113 | Ossi Bister |
| training114 | |
| training115 | Andrej S. |
| training116 | Raquel Oliveira |
| training117 | Andrea Kulakov |
| training118 | |
| training119 | Johannes Kruse |
| training120 | Amhar Jaher |
| training121 | |
| training122 | |
| training123 | Edixon Parraga |
| training124 | Betzabeth Leon |

## Exercise 5

### dvc (dogs vs cats)

| Submitter | Model description | Test accuracy |
| --------- | ----------------- | ------------- |
| Markus | original dvc-cnn-simple.h5 | 73.03% |
| Stefan | original dvc-cnn-simple.h5 | 62.52% |
| Stefan | original dvc-bit.h5 | 98.48% |
| Pierpaolo | original `dvc-cnn-simple.h5` | 73.01% |
| Pierpaolo | original `dvc-vgg16-reuse.h5` | 88.68% |
| Pierpaolo | original `dvc-vgg16-finetune.h5` | 92.18% |
| Rémi | original dvc=cnn=simple.h5 | 96.42% |
| Rémi | pretrained | 79.37 % |
| Ana Gonza | dvc-cnn-simple.h5 | 69.36% |
| Ana Gonza | dvc-vgg16-finetune.h5 | 91.70% |
| Ana Gonza | dvc-vgg16-reuse.h5 | 89.07% |
| Outi| dvc original| 72.94%|
|Outi| finetuned | 92.68%|
|saurav kumar| dvc-vgg16-pretrained| 88.58%|
|saurav kumar| dvc-vgg16-pretrained(finetuned)| 91.29%|
|saurav kumar| dvc-cnn-simple| 69.73%|
|Outi | reuse |86.55%|
| Andrea Kulakov | dvc-vgg16-reuse.h5 | 83.78% |
| Andrea Kulakov | dvc-vgg16-finetune.h5 | 91.69% |
| Andrea Kulakov | dvc-cnn-simple.h5 | 72.40% |
| Panu | dvc_tfr-cnn-simple.h5 | 70.33% |
| Panu | dvc-vgg16-reuse.h5 | 89.15% |
| Panu | dvc-vgg16-finetune.h5 | 92.53% |
| Panu | dvc-bit.h5 | 98.47% |
| Panu | tf2 vit | 98.60% |
| Vaclav | dvc-bit.h5 | 98.32% |
| Patrik | dvc-cnn-simple | 70.08% |
| Patrik | dvc-vgg16-finetune| 92.25%|
| Patrik | dvc-vgg16-reuse |88.60%|
| Patrik |dvc-bit|98.35%|
| Amhar| dvc-vgg16-reuse.h5 | 87.86% |
| Amhar| dvc_tfr-cnn-simple.h5 | 70.44% |
| Amhar| dvc-bit.h5 | 98.53% |
| Amhar| dvc-vgg16-finetune.h5 | 92.53% |
| Mika | dvc-vgg16-finetune.h5 | 91.17% |
| Mika | dvc-vgg16-reuse.h5 | 85.86% |
| Mika | dvc-cnn-simple.h5 | 64.65% |
| Till | tf2-dvc-cnn-simple | 72.68% |
| Till | tf2-dvc-cnn-pretrained (VGG16) | 88.61% |
| Yorgos | tf2-dvc-cnn-simple | 67.95% |
| Yorgos | tf2-dvc-cnn-pretrained | 92.40% |
| Weckman | pytorch-finetune | 95.86% |
| Pablo Sanz | dvc-cnn-simple |58.95% |
| Pablo Sanz | dvc-vgg16-reuse |88.95%|
| Pablo Sanz | dvc-vgg16-finetune| 92.70%|
| Alvaro | original `dvc-cnn-simple.h5` | 72% |
| Alvaro | original `dvc-vgg16-reuse.h5` | 88% |
| Alvaro | original `dvc-vgg16-finetune.h5` | 92% |

### gtsrb (traffic signs)

| Submitter | Model description | Test accuracy |
| --------- | ----------------- | ------------- |
| Markus | original gtsrb-cnn-simple.h5 | 72.58% |
| Andrej S. | pyTorch cnn_pretrained | 67.10% |
| Andrej S. | Task 2: pyTorch cnn_pretrained, Transform rotate 5 degr| 65.76% |
| Stefan | gtsrb-cnn-simple | 84.03% |
| Stefan | gtsrb-bit | 57.82% |
| Stefan | gtsrb-vgg16-finetune | 61.03% |
| Pierpaolo | original `gtsrb-cnn-simple.h5` | 83.75% |
| Pierpaolo | original `ggtsrb-vgg16-reuse.h5` |55.80% |
| Pierpaolo | original `gtsrb-vgg16-finetune.h5` | 62.88% |
|Outi| simple| 81.80%|
|Outi|simple, 25 epochs, dropout 0.4| 82.64%|
|Outi | pretrained/reuse|56.29%|
|Outi | bit | 53.50%|
|Outi |pretrained/finetune| 64.80%|
| Saurav kumar | gtsrb-cnn-simple |81.70% |
| Saurav kumar | gtsrb-cnn-pretrained |63.27% |
| Saurav kumar | gtsrb-cnn-gtrisb-bit |57.13% |
| Panu | gtsrb-cnn-simple.h5 | 81.19% |
| Panu | gtsrb-vgg16-reuse.h5 | 55.41% |
| Panu | gtsrb-vgg16-finetune.h5 | 62.30% |
| Panu | gtsrb-bit.h5 | 55.41% |
| Vaclav | gtsrb_cnn_pretrained | 67.51% |
| Ana Gonzalez | gtsrb-bit.h5 | 55.44% |
| Ana Gonzalez | gtsrb-vgg16-finetune.h5 | 63.58% |
| Ana Gonzalez | gtsrb-cnn-simple.h5 | 81.44% |
| Ana Gonzalez | gtsrb-vgg16-reuse.h5 | 55.47% |
| Pablo Sanz |simple| 72.03%|
| Pablo Sanz | pretrained|97.58%|
| Rémi | simple | 96.11% |
| Rémi | pretrained | 79.68% |
| Rémi | BiT | 69.16% |
| Andrea Kulakov | gtsrb-vgg16-finetune.h5 | 62.49% |

## Exercise 6

| Submitter | Model description | Test accuracy |
| --------- | ----------------- | ------------- |
| Stefan | tf2-20ng-rnn | 84.40% |
| Stefan | tf2-20ng-cnn | 91.22% |
| Stefan | tf2-20ng-cnn (500 epochs) | 95.95% |
| Stefan | tf2-20ng-cnn (500 epochs, batch size of 32) | 12.83% |
| Stefan | tf2-20ng-cnn (100 epochs, batch size of 1024) | 97.45% |
| Stefan | tf2-20ng-bert | 82.82% |
| Pierpaolo | original `tf2-20ng-rnn` | 85.22% |
| Pierpaolo | original `tf2-20ng-cnn` | 95.57% |
| Pierpaolo | original `tf2-20ng-bert` | 82.57% |
| Pierpaolo | `tf2-20ng-cnn` + extra dense layer (200 epochs) | 96.17% |
| Pierpaolo | `tf2-20ng-cnn` + extra dense layer (200 epochs, batch size 256) | 97.05% |
| Pierpaolo | `tf2-20ng-cnn` + extra dense layer (200 epochs, batch size 256, adam) | 96.05% |
| Andrea Kulakov | tf2-20ng-rnn.py | 83.85% |
| Andrea Kulakov | tf2-20ng-cnn.py - much faster than rnn | 96.53% |
| Andrea Kulakov | tf2-20ng-bert.py - way much slower than rnn (00:11:32) | 82.53% |
|Outi | bert|84.30%|
|Outi |cnn| 96.22%|
|Outi |rnn| 87.50%|
| Rémi | RNN walltime 00:02:32 | 85.80%|
| Rémi | CNN walltime 00:00:52 | 95.85% |
| Rémi | BERT walltime 00:11:39 | 83.05%|
| Rémi | BERT LR=4e-5 BATCH_SIZE 16 | 83.65% |
| Ana G | RNN wall time 00:02:33 | 85.55% |
| Ana G | CNN wall time 00:00:53 | 96.35% |
| Ana G | Bert wall time 00:11:32 | 82.98% |
| Panu | tf2-20ng-rnn 00:02:33 | 83.17% |
| Panu | tf2-20ng-cnn 00:00:54 | 93.85% |
| Panu | tf2-20ng-bert 00:11:25 | 83.55% |
| Mika | tf2-20ng-rnn cpu 00:02:36 | 81.65% |
| Mika | tf2-20ng-cnn cpu 00:00:57 | 96.35% |
| Mika | tf2-20ng-bert cpu 00:09:09 | 82.82% |
| Saurav | tf2-20ng-rnn | 83.80% |
| Saurav | tf2-20ng-cnn | 95.65% |
| Saurav | tf2-20ng-bert | 83.55% |
| Pablo Sanz | tf2-20ng-rnn | 84.68% |
| Pablo Sanz | tf2-20ng-cnn | 94.67% |
| Pablo Sanz | tf2-20ng-bert | 82.90% |
| Andrej S. | pytorch_20ng_bert | 81.2% |

## Exercise 7

You can check time with `seff SLURMID` and check "Job Wall-clock time".

| Submitter | Model description | Time | Test accuracy |
| --------- | ----------------- | ---- | ------------- |
| Mats | tf2-dvc-cnn-simple, 1 GPU | 00:00:54 | 65.31% |
| Stefan | tf2-dvc-cnn-simple, 4 GPUs | 00:00:34 (wall) | 59.80% |
| Pablo Sanz | dvc-ccn-simple, 1 GPUs | 0m44.292s | |
| Pablo Sanz | dvc-ccn-simple, 4 GPUs | 0m29.767s | 58.95% |
| Pablo Sanz | dvc-vgg16-reuse, 1 GPUs | 0m33.834s | |
| Pablo Sanz | dvc-vgg16-reuse, 4 GPUs | 0m32.074s | 88.95% |
| Pablo Sanz | dvc-vgg16-finetune, 1 GPUs | 0m36.021s | |
| Pablo Sanz | dvc-vgg16-finetune, 4 GPUs | 0m35.944s | 92.70% |
| Mika | dvc-ccn-simple 1 GPU | cpu 00:02:34 wall 00:01:24 | 64.65% |
| Mika | dvc-ccn-simple 4 GPU | cpu 00:00:57 wall 00:00:35 | 71.41% |
| Pierpaolo | `dvc-ccn-simple` 2 GPU | real 0m59.765s | 60.59% |
| Pierpaolo | `dvc-ccn-simple` 3 GPU | real 1m6.243s | 66.46% |
| Pierpaolo | `dvc-ccn-simple` 4 GPU | real 1m0.664s | 65.56% |
| saurav| dvc-ccn-simple 1 GPU | % |
| saurav | dvc-ccn-simple 4 GPU| 68.1% |