# DAY 1 EXERCISES 2022-02

## Exercise 2 (02-tf2-mnist-mlp)

| Submitter | Model description | Test accuracy |
| --------- | ----------------- | ------------- |
| Markus | original "mlp_model" | 95.54% |
| Markus | Task 1 example answer | 97.12% |
| Stefan | 784-1000-500-250-250-250-10 model used as a baseline for fully connected MLP's in https://arxiv.org/pdf/1507.02672.pdf, each layer followed by a dropout layer with rate=0.2 | 98.04% |
|Outi |Task 1 answer |97.09% |
| Markus | Task 1 example answer with 20 epochs | 97.49% |
|Carlos S. | Task 1 | 97.53% |
|César | Task 1 | 97.13% |
|Rémi | Task 1 |97.03% |
|Pierpaolo | Task 1 | 98.02% |
|Shaheryar | Task 1 - 64 units, ReLU activation, Dropout layer rate of 0.1 | 97.84% |
| Andrea Kulakov | | 97.14% |

## Exercise 3 (03-tf2-mnist-cnn)

| Submitter | Model description | Test accuracy |
| --------- | ----------------- | ------------- |
| Markus | original "cnn_model" | 98.46% |
| Stefan | Task 1: "better_cnn_model" | 98.95% |
| Pierpaolo | Task 1: "cnn_model" | 99.01% |
| Pierpaolo | Task2: "cnn_model" with (`adam`, `batch=200`, `epochs=15`) | 99.24% |
| Rémi | Task 1: more complex cnn | 98.95 % |
| Outi | Task 1 | 98.88% |
| César | Task 1 | 98.99% |
| Carlos S. | Task 1 | 99.01% |
| Shaheryar | Task 1 - "A more complex CNN model" | 98.89% |
| Patrik | Task 1 model with 10 epochs | 98.98% |
| Markus | "better cnn_model", 5 epochs | 99.06% |
| Markus | "better cnn_model", 20 epochs | 99.32% |
| Andrej | Task 1: 5 epochs, batch_size=16 | 99.06% |
| Outi | Task 2 - 10 epochs | 99.21% |
| Weckman | Task 2 | 99.31%
| Panu | Task 2: adam, 5 epochs, batch_size 32 | 99.04% |
| Enda | Task 2: adam, 5 epochs, batch_size 128 | 99.04% |
| Shaheryar | Task 2: Nadam, 5 epochs, shuffle=True | 99.06% |
| Stefan | Task 2: "better_cnn_model" with `adagrad` | 99.19% |
| saurav | Task 2: "better_cnn_model" with adam (test_set)| 99.00% |
| Yorgos | Task 1: adam, 5 epochs, batch_size 128 | 98.93% |
| Andrea Kulakov | optimizer = SGD(lr=0.1, momentum=0.9) | 98.98% |

State of the art in MNIST classification: 99.91% (https://paperswithcode.com/sota/image-classification-on-mnist)

## Exercise 4 (04-tf2-imdb-rnn)

| Submitter | Model description | Test accuracy |
| --------- | ----------------- | ------------- |
| Mats | Original "rnn_model" | 82.01% |
| Mats | Task 1, model answer, 5 epochs | 83.49% |
| Mats | Task 1, bidirectional model answer, 5 epochs | 83.32% |
| Shaheryar | Task 1 - Two LSTM layers (Using Dropout=0.1, RepeatVector(1) before second LSTM layer) | 83.02% |
| Shaheryar | Task 2 - Model tuning (Using callbacks, dropout=0.2, recurrent_dropout=0.3 in model from Task 1) | 84.22% | Rémi | task 1, bidir | 81.08 % |
| Mats | CNN | 85.02% |
| Shaheryar | CNN | 88.71% |
| Stefan | Task 2: model answer, all suggested modifications (dropout, regularization, early stopping) | 84.18% |
| Andrea Kulakov | dropout=0.2, recurrent_dropout=0.3 | 83.28%