Week3 tutorials update by cindyhfls · Pull Request #1031 · NeuromatchAcademy/course-content-dl

cindyhfls · 2026-05-18T18:37:15Z

No description provided.

…iscussion only Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

W3D4 T1: trimmed from 7 exercises to 5 (GridWorld, MDP, Q steps-to-go, Value Iteration, Policy Iteration); added scaffolding markdown before each exercise; Q-learning and ε-greedy moved to W3D5 T1 W3D5: new T1 created (Q-learning + ε-greedy, moved from W3D4); former T1 (DL Discussion 3) renamed to T2 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

W3D4 T1: trimmed from 7 exercises to 5 (GridWorld, MDP, Q steps-to-go, Value Iteration, Policy Iteration); added scaffolding markdown before each exercise; Q-learning and ε-greedy moved to W3D5 T1 W3D5: - New T1: Q-learning + ε-greedy (moved from W3D4), feedback_prefix W3D5_T1 - T2: DL Discussion 3 (renamed from old T1); title corrected to Tutorial 2 - T3 (Bonus): Reinforcement Learning For Games pulled from upstream W3D5_Tutorial1; title changed to Bonus Tutorial, day line and feedback_prefix (W3D5_T3_Bonus) updated - Folder renamed W3D5_DeepLearningDiscussion3 → W3D5_AdvancedReinforcementLearningAndDeepLearningDiscussion3 - materials.yml: name updated, tutorials count 1 → 3, slides added for T1 (ztgws), T2 (2f7ay), T3 Bonus (3zn9w) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

W3D2: update Tutorial1 content and README to match DeepLearningDiscussion2 naming W3D3: apply coding style and dependency inlining to Tutorial1 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ean up, fill most of the code and kept code completion to minimum

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

iamzoltan · 2026-06-09T13:11:43Z

   },
   "source": [
-    "# Tutorial 1: Deep Learning Case Study 2: Architectures and Multimodal DL \n",
+    "# Tutorial 1: Deep Learning Thinking 2: Architectures and Multimodal DL thinking\n",


iamzoltan · 2026-06-09T13:11:57Z

   "source": [
    "---\n",
-    "# Section 1: Intro Deep Learning Case Study 2\n",
+    "# Section 1: Intro Deep Learning Thinking 2\n",


Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

iamzoltan

This ones is also almost ready. some minor comments

iamzoltan · 2026-06-12T03:12:38Z

+    "cellView": "form"
+   },
+   "outputs": [],
+   "source": "# @title Module: Load utilities\n\nimport os\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport torchvision\n\n\n\ndef load_encoder(save_direc, model_type=\"simclr\", dataset_type=\"full\", \n                 neg_pairs=\"all\", verbose=True):\n    \"\"\"\n    load_encoder(save_direc)\n\n    Loads encoder (VAE or SimCLR) with pre-trained parameters.\n\n    Required args:\n    - save_direc (str): directory where pre-trained encoder parameters are \n        saved\n\n    Optional args:\n    - model_type (str): type of pre-trained encoder model (default: \"simclr\")\n    - dataset_type (str): type of dataset used in the pre-training \n        (default: \"full\") \n    - neg_pairs (str or num): Number of negeative pairs used in loss \n        calculation, if loading a SimCLR model. (default: \"all\")\n    - verbose (bool): If True, details of the encoder being loaded are \n        printed. (default: True)\n\n    Returns:\n    - encoder (models.EncoderCore): encoder loaded with pre-trained parameters\n    \"\"\"\n\n\n    if dataset_type == \"full\":\n        dataset_type_str = \"\"\n    elif dataset_type in [\"biased\", \"bias_ctrl\"]:\n        dataset_type_str = f\"_{dataset_type}\"\n    else:\n        raise ValueError(\"dataset_type can only be 'full', 'biased' or \"\n          f\"'bias_ctrl', but found '{dataset_type}'.\")\n\n    vae = False\n    simclr_transforms_str, simclr_transforms_str_pr = \"\", \"\"\n    neg_str, neg_str_pr = \"\", \"\"\n    \n    seed = 2021\n    if model_type == \"vae\":\n        batch_size = 500\n        vae = True\n        model_name = \"VAE\"\n        if dataset_type == \"full\":\n            num_epochs = 300\n        elif dataset_type in [\"biased\", \"bias_ctrl\"]:\n            num_epochs = 450\n        \n    elif model_type == \"simclr\":\n        batch_size = 1000\n        model_name = \"SimCLR\"\n        simclr_transforms_str_pr = (\"\\nwith the following random affine \"\n            \"transforms:\\n\\tdegrees=90\\n\\ttranslation=(0.2, 0.2)\"\n            \"\\n\\tscale=(0.8, 1.2)\")\n        simclr_transforms_str = \"_deg90_trans0-2_scale0-8to1-2\"\n        if dataset_type == \"full\":\n            num_epochs = 60\n            if neg_pairs != \"all\":\n                neg_pairs = int(neg_pairs)\n                if neg_pairs != 2:\n                    raise ValueError(\"If not 'all'', neg_pairs must be set \"\n                        \"to 2, as that is the only value that was used in \"\n                        \"the saved models.\")\n                neg_str_pr = (f\"\\nwith {neg_pairs} negative pairs per image \"\n                    \"used in the contrastive loss, and\")\n                neg_str = f\"_{neg_pairs}neg\"\n        elif dataset_type in [\"biased\", \"bias_ctrl\"]:\n            if neg_pairs != \"all\":\n                raise ValueError(\n                    \"No saved model for SimCLR with few negative pairs using \"\n                    \"the biased or bias_ctrl datasets.\"\n                    )\n            num_epochs = 150\n    \n    elif model_type == \"supervised\":\n        model_name = model_type\n        batch_size = 1000\n        num_epochs = 10\n\n    elif model_type == \"random\":\n        model_name = model_type\n        batch_size = 0\n        num_epochs = 0\n\n    else:\n        raise ValueError(\"Recorded model types only include 'supervised', \"\n            f\"'random', 'vae', 'simclr', but not '{model_type}'.\")\n        \n    encoder_path = (\n        f\"{model_type}_encoder{dataset_type_str}{neg_str}_{num_epochs}ep_\"\n        f\"bs{batch_size}{simclr_transforms_str}_seed{seed}.pth\"\n    )\n    full_path = os.path.join(save_direc, \"checkpoints\", encoder_path)\n    \n    if verbose:\n        model_details = (f\"    => trained for {num_epochs} epochs \"\n            f\"(batch_size of {batch_size}) on the {dataset_type} dSprites \"\n            f\"subset dataset{neg_str_pr}{simclr_transforms_str_pr}.\")\n        print(f\"Loading {model_name} encoder from '{full_path}'.\\n\"\n            f\"{model_details}\")\n\n    encoder = models.EncoderCore(vae=vae)\n    encoder.load_state_dict(torch.load(full_path))\n\n    return encoder\n\n\ndef load_vae_decoder(save_direc, verbose=True):\n    \"\"\"\n    load_vae_decoder(save_direc)\n\n    Loads VAE decoder with pre-trained parameters.\n\n    Required args:\n    - save_direc (str): directory where pre-trained decoder parameters are \n        saved\n\n    Optional args:\n    - verbose (bool): If True, details of the decoder being loaded are \n        printed. (default: True)\n\n    Returns:\n    - encoder (models.VAE_decoder): decoder loaded with pre-trained parameters\n    \"\"\"\n    \n    batch_size = 500\n    seed = 2021\n    model_name = \"VAE\"\n    num_epochs = 300\n    decoder_path =f\"vae_decoder_{num_epochs}ep_bs{batch_size}_seed{seed}.pth\"\n\n    full_path = os.path.join(save_direc, \"checkpoints\", decoder_path)\n    \n    if verbose:\n        model_details = (f\"   => trained for {num_epochs} epochs \"\n            f\"(batch_size of {batch_size}) on the full dSprites subset \"\n            \"dataset.\")\n        print(f\"Loading {model_name} decoder from '{full_path}'.\\n\"\n            f\"{model_details}\")\n\n    decoder = models.VAE_decoder()\n    decoder.load_state_dict(torch.load(full_path))\n\n    return decoder\n\n\nclass ResNet18_with_encoder(torchvision.models.resnet.ResNet):\n    \"\"\"\n    ResNet18_with_encoder()\n\n    torchvision ResNet18 with explicitly defined encoder attribute, and \n    get_features() method.\n\n    Optional args:\n    - pretrained (bool): If True, the model is pretrained. (default: True)\n    \"\"\"\n    \n    def __init__(self, pretrained=True):\n\n        self._untrained = not(pretrained)\n\n        weights = torchvision.models.ResNet18_Weights.DEFAULT if pretrained else None\n        resnet18 = torchvision.models.resnet18(\n            weights=weights, progress=False\n            )\n        self.__dict__.update(resnet18.__dict__)\n        \n        self.pretrained = pretrained\n        self.input_dim = (3, 224, 224)\n\n        self._define_encoder()\n        self.feat_size = self._get_feat_extr_output_size(self.input_dim)\n\n    @property\n    def untrained(self):\n        return self._untrained\n\n    @property\n    def vae(self):\n        return False\n\n    def _define_encoder(self):\n        # first 8\n        self.encoder = nn.Sequential(\n            self.conv1,\n            self.bn1,\n            self.relu,\n            self.maxpool,\n            self.layer1,\n            self.layer2,\n            self.layer3,\n            self.layer4,\n        )\n\n    def _get_feat_extr_output_size(self, input_dim):\n        dummy_tensor = torch.ones(1, *input_dim)\n        reset_training = self.training\n        self.eval()\n        with torch.no_grad():   \n            output_dim = self.encoder(dummy_tensor).shape\n        if reset_training:\n            self.train()\n        return np.prod(output_dim)\n\n    def get_features(self, X):\n        with torch.no_grad():\n            feats = self.encoder(X)\n        return feats\n\n    def forward(self, *args, **kwargs):\n        self._untrained = False\n        super().forward(*args, **kwargs)\n\n\nclass VGG16_with_encoder(torchvision.models.vgg.VGG):\n    \"\"\"\n    VGG16_with_encoder()\n\n    torchvision VGG16 with explicitly defined encoder attribute, and \n    get_features() method.\n\n    Optional args:\n    - pretrained (bool): If True, the model is pretrained. (default: True)\n    \"\"\"\n\n    def __init__(self, pretrained=True):\n\n        self._untrained = not(pretrained)\n\n        weights = torchvision.models.VGG16_Weights.DEFAULT if pretrained else None\n        vgg16 = torchvision.models.vgg16(\n            weights=weights, progress=False\n            )\n        self.__dict__.update(vgg16.__dict__)\n        \n        self.pretrained = pretrained\n\n        self._define_encoder()\n        self.input_dim = (3, 64, 64)\n        self.feat_size = self._get_feat_extr_output_size(self.input_dim)\n\n    @property\n    def untrained(self):\n        return self._untrained\n\n    @property\n    def vae(self):\n        return False\n\n    def _define_encoder(self):\n        self.encoder = self.features # alias\n\n    def _get_feat_extr_output_size(self, input_dim):\n        dummy_tensor = torch.ones(1, *input_dim)\n        reset_training = self.training\n        self.eval()\n        with torch.no_grad():   \n            output_dim = self.encoder(dummy_tensor).shape\n        if reset_training:\n            self.train()\n        return np.prod(output_dim)\n\n    def get_features(self, X):\n        with torch.no_grad():\n            feats = self.encoder(X)\n        return feats\n\n    def forward(self, *args, **kwargs):\n        self._untrained = False\n        super().forward(*args, **kwargs)\n\n\nclass SimCLR_spijk_with_encoder(nn.Module):\n    \"\"\"\n    SimCLR_spijk_with_encoder()\n\n    SimCLR implementation from https://github.com/Spijkervet/SimCLR, with \n    explicitly defined get_features() method.\n\n    Optional args:\n    - pretrained (bool): If True, the model is pretrained. (default: True)\n    \"\"\"\n\n    def __init__(self, pretrained=True):\n        \n        self.projection_dim = 64\n        self._untrained = not(pretrained)\n\n        import simclr\n\n        encoder = simclr.modules.get_resnet(\"resnet18\", pretrained=pretrained)\n        simclr_model = simclr.SimCLR(\n            encoder, self.projection_dim, encoder.fc.in_features\n            )\n        self.__dict__.update(simclr_model.__dict__)\n\n        self.pretrained = pretrained\n\n        if self.pretrained:\n            src = (\"https://github.com/Spijkervet/SimCLR/releases/download/\"\n                \"1.1/checkpoint_100.tar\")\n\n            state_dict = torch.hub.load_state_dict_from_url(\n                src, progress=False, map_location=\"cpu\"\n                )\n\n            self.load_state_dict(state_dict)\n\n        self.input_dim = (3, 224, 224)\n        self.feat_size = self._get_feat_extr_output_size(self.input_dim)\n\n\n    @property\n    def untrained(self):\n        return self._untrained\n\n    @property\n    def vae(self):\n        return False\n\n    def _get_feat_extr_output_size(self, input_dim):\n        dummy_tensor = torch.ones(1, *input_dim)\n        reset_training = self.training\n        self.eval()\n        with torch.no_grad():   \n            output_dim = self.encoder(dummy_tensor).shape\n        if reset_training:\n            self.train()\n        return np.prod(output_dim)\n\n    def get_features(self, X):\n        with torch.no_grad():\n            feats = self.encoder(X)\n        return feats\n\n    def forward(self, *args, **kwargs):\n        self._untrained = False\n        super().forward(*args, **kwargs)\n"


Can we separate this into a list of strings?

iamzoltan · 2026-06-12T03:22:07Z

-    "    #################################################\n",
-    "    # Set the proper actions\n",
-    "    raise NotImplementedError(\"Implement `plan` function!\")\n",
-    "    #################################################\n",
    "    while goal_queue:\n",
-    "      goal = goal_queue.pop(0)  # pop from front of list\n",


was removing this student question intentional?

iamzoltan · 2026-06-12T03:23:18Z

    "\n",
    "The following code:\n",
-    "* Loads the parameters of a full Variational AutoEncoder (VAE) network (encoder and decoder) pre-trained on the generative task of reconstructing the input images, under the Kullback–Leibler Divergence (KLD) minimization constraint over the latent space that characterizes VAEs, using `load.load_encoder()` and `load.load_decoder()`."
+    "* Loads the parameters of a full Variational AutoEncoder (VAE) network (encoder and decoder) pre-trained on the generative task of reconstructing the input images, under the Kullback\u2013Leibler Divergence (KLD) minimization constraint over the latent space that characterizes VAEs, using `load.load_encoder()` and `load.load_decoder()`."


can we convert these back to their chars?

cindyhfls and others added 8 commits May 18, 2026 14:28

W3D2/W3D5: rename case study → discussion; trim W3D5 to DeepLearningD…

e39ba54

…iscussion only Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

W3D4 T1: remove BFS references not present in upstream

9719269

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

W3D5: update day names in title cells to match folder rename

9c3d210

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

W3D2, W3D3: sync content with playground

4d05472

W3D2: update Tutorial1 content and README to match DeepLearningDiscussion2 naming W3D3: apply coding style and dependency inlining to Tutorial1 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

adding back some docstring and helper functions deleted by Claude, cl…

f303535

…ean up, fill most of the code and kept code completion to minimum

add back docstring deleted by Claude and bug fix

dd75164

cindyhfls marked this pull request as ready for review May 26, 2026 13:33

cindyhfls marked this pull request as draft May 29, 2026 02:20

cindyhfls changed the base branch from main to staging May 29, 2026 02:20

cindyhfls marked this pull request as ready for review May 29, 2026 02:20

cindyhfls force-pushed the week3-tutorials-update branch from c552757 to dd75164 Compare June 9, 2026 04:46

W3D2-W3D5: add Jiaxin Cindy Tu to content reviewers/editors

83de14f

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

iamzoltan reviewed Jun 9, 2026

View reviewed changes

cindyhfls and others added 11 commits June 9, 2026 10:25

remove outdated instructor and student tutorial files

d910129

update colab and kaggle links

a98a94f

add missing package dependencies to requirements.txt and environment.yml

8950d84

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

update requirements

14933ee

pretrained is outdated in pytorch

c048a7b

add week3 changes to materials.yml

941059f

named DL Discussion 2 for W3D2 materials.yml

93306de

Bonus_Tutorial1: fix pretrained=True → DenseNet121_Weights.DEFAULT

0c2641e

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

remove W3D2 readme

d378e13

remove the legacy files from renaming folder

d8c0e71

fixed static reference in week3 and deployment tutorials

eb7f937

iamzoltan reviewed Jun 12, 2026

View reviewed changes

cindyhfls added 2 commits June 11, 2026 23:59

fix the json.dump modification of characters

0814137

fix the accidental deletion of bonus day

6747a33

cindyhfls added 4 commits June 12, 2026 03:10

move Q-Learning and epsilon greedy back to W3D4

e1c1aac

fix W3D4 day and tutorial titles

0b8119c

update requirements

97a3c6c

remove import check for W3D2-W3D5

2b036e7

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Week3 tutorials update#1031

Week3 tutorials update#1031
cindyhfls wants to merge 26 commits into
NeuromatchAcademy:stagingfrom
cindyhfls:week3-tutorials-update

cindyhfls commented May 18, 2026

Uh oh!

iamzoltan Jun 9, 2026

Uh oh!

iamzoltan Jun 9, 2026

Uh oh!

iamzoltan left a comment

Uh oh!

iamzoltan Jun 12, 2026

Uh oh!

iamzoltan Jun 12, 2026

Uh oh!

cindyhfls Jun 12, 2026

Uh oh!

iamzoltan Jun 12, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

cindyhfls commented May 18, 2026

Uh oh!

iamzoltan Jun 9, 2026

Choose a reason for hiding this comment

Uh oh!

iamzoltan Jun 9, 2026

Choose a reason for hiding this comment

Uh oh!

iamzoltan left a comment

Choose a reason for hiding this comment

Uh oh!

iamzoltan Jun 12, 2026

Choose a reason for hiding this comment

Uh oh!

iamzoltan Jun 12, 2026

Choose a reason for hiding this comment

Uh oh!

cindyhfls Jun 12, 2026

Choose a reason for hiding this comment

Uh oh!

iamzoltan Jun 12, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants