feat: single model, single optimizer template (#34)

* feat: add single model, single optimizer template

* Update README.md

* feat: archive as python package system

* feat: add hubconf, tests for single, TODOs tip

* tip improvement

* readme improve, script for find and replace

* kwargs to Checkpoint handler

* mimic tree

* feat: more utils functions and tests

* feat: handlers usage in main.py, more comments

* feat: add custom additional events, fix #30

* fix: all in 1 deps install, amp option fix #32

* fix: model, opt, loss, lr_s from initialize, fix #32

* chore: up README about distributed launching

* feat: add resume_from command line argument to resume from checkpoint, fix #31

* fix: comments and docstring

.github/run_code_style.sh

@@ -2,13 +2,13 @@
 
 set -xeu
 
-if [ $1 = "lint" ]; then
+if [ $1 == "lint" ]; then
     flake8 app templates tests --config .flake8
     isort app templates tests --check --settings pyproject.toml
     black app templates tests --check --config pyproject.toml
-elif [ $1 = "fmt" ]; then
+elif [ $1 == "fmt" ]; then
     isort app templates tests --color --settings pyproject.toml
     black app templates tests --config pyproject.toml
-elif [ $1 = "install" ]; then
+elif [ $1 == "install" ]; then
     pip install flake8 "black==20.8b1" "isort==5.7.0"
 fi

.github/run_test.sh

@@ -2,11 +2,11 @@
 
 set -xeu
 
-if [ $1 = "generate" ]; then
+if [ $1 == "generate" ]; then
     python ./tests/generate.py
-elif [ $1 = "unittest" ]; then
+elif [ $1 == "unittest" ]; then
     pytest ./tests/unittest -vvv -ra --color=yes --durations=0
-elif [ $1 = "integration" ]; then
+elif [ $1 == "integration" ]; then
     for file in $(find ./tests/integration -iname "*.sh")
     do
         bash $file

app/codegen.py

@@ -12,7 +12,7 @@ def __init__(self, templates_dir: str = "./templates", dist_dir: str = "./dist")
         self.dist_dir = Path(dist_dir)
         self.template_list = [p.stem for p in self.templates_dir.iterdir() if p.is_dir() and not p.stem.startswith("_")]
         self.rendered_code = {t: {} for t in self.template_list}
-        self.available_archive_formats = sorted(map(lambda x: x[0], shutil.get_archive_formats()), reverse=True)
+        self.available_archive_formats = [x[0] for x in shutil.get_archive_formats()[::-1]]
 
     def render_templates(self, template_name: str, config: dict):
         """Renders all the templates files from template folder for the given config."""
@@ -31,27 +31,31 @@ def render_templates(self, template_name: str, config: dict):
             self.rendered_code[template_name][fname] = code
             yield fname, code
 
-    def mk_dist_template_dir(self, template_name: str):
-        self.dist_template_dir = Path(f"{self.dist_dir}/{template_name}")
-        self.dist_template_dir.mkdir(parents=True, exist_ok=True)
+    def make_and_write(self, template_name: str):
+        """Make the directories first and write to the files"""
+        for p in (self.templates_dir / template_name).rglob("*"):
+            if not p.stem.startswith("_") and p.is_dir():
+                # p is templates/template_name/...
+                # remove "templates" from p.parts and join with "/", so we'll have
+                # template_name/...
+                p = "/".join(p.parts[1:])
+            else:
+                p = template_name
 
-    def write_file(self, fname: str, code: str) -> None:
-        """Creates `fname` with content `code` in `dist_dir/template_name`."""
-        (self.dist_template_dir / fname).write_text(code)
+            if not (self.dist_dir / p).is_dir():
+                (self.dist_dir / p).mkdir(parents=True, exist_ok=True)
 
-    def write_files(self, template_name):
-        """Writes all rendered code for the specified template."""
-        # Save files with rendered code to the disk
         for fname, code in self.rendered_code[template_name].items():
-            self.write_file(fname, code)
+            (self.dist_dir / template_name / fname).write_text(code)
 
     def make_archive(self, template_name, archive_format):
         """Creates dist dir with generated code, then makes the archive."""
-        self.mk_dist_template_dir(template_name)
-        self.write_files(template_name)
+
+        self.make_and_write(template_name)
         archive_fname = shutil.make_archive(
-            base_name=str(self.dist_template_dir),
+            base_name=template_name,
+            root_dir=self.dist_dir,
             format=archive_format,
-            base_dir=self.dist_template_dir,
+            base_dir=template_name,
         )
-        return archive_fname
+        return shutil.move(archive_fname, self.dist_dir / archive_fname.split("/")[-1])

app/streamlit_app.py

@@ -1,7 +1,5 @@
-import os
 import shutil
 from pathlib import Path
-from subprocess import check_output
 
 import streamlit as st
 from codegen import CodeGenerator
@@ -10,6 +8,22 @@
 __version__ = "0.1.0"
 
 
+FOLDER_TO_TEMPLATE_NAME = {
+    "Single Model, Single Optimizer": "single",
+    "Generative Adversarial Network": "gan",
+    "Image Classification": "image_classification",
+}
+
+TIP = """
+> **💡 TIP**
+>
+> To quickly adapt to the generated code structure, there are TODOs in the files that are needed to be edited.
+> [PyCharm TODO comments](https://www.jetbrains.com/help/pycharm/using-todo.html) or
+> [VSCode Todo Tree](https://marketplace.visualstudio.com/items?itemName=Gruntfuggly.todo-tree)
+> can help you find them easily.
+"""
+
+
 class App:
     page_title = "Code Generator"
     page_icon = "https://raw.githubusercontent.com/pytorch/ignite/master/assets/logo/ignite_logomark.svg"
@@ -39,18 +53,19 @@ def sidebar(self, template_list=None, config=None):
         template_list = template_list or []
         st.markdown("### Choose a Template")
         self.template_name = st.selectbox("Available Templates are:", options=template_list)
+        self.template_name = FOLDER_TO_TEMPLATE_NAME[self.template_name]
         with st.sidebar:
             if self.template_name:
                 config = config(self.template_name)
                 self.config = config.get_configs()
             else:
                 self.config = {}
 
-    def render_code(self, fname="", code=""):
+    def render_code(self, fname: str = "", code: str = ""):
         """Main content with the code."""
-        with st.beta_expander(f"View rendered {fname}"):
+        with st.beta_expander(f"View rendered {fname}", expanded=fname.endswith(".md")):
             if fname.endswith(".md"):
-                st.markdown(code)
+                st.markdown(code, unsafe_allow_html=True)
             else:
                 col1, col2 = st.beta_columns([1, 20])
                 with col1:
@@ -59,22 +74,57 @@ def render_code(self, fname="", code=""):
                     st.code(code)
 
     def render_directory(self, dir):
-        output = check_output(["tree", dir], encoding="utf-8")
+        """tree command is not available in all systems."""
+        output = f"{dir}\n"
+        # https://stackoverflow.com/questions/9727673/list-directory-tree-structure-in-python
+        # prefix components:
+        space = "    "
+        branch = "│   "
+        # pointers:
+        tee = "├── "
+        last = "└── "
+        file_count = 0
+        dir_count = 0
+
+        def tree(dir_path: Path, prefix: str = ""):
+            """A recursive generator, given a directory Path object
+            will yield a visual tree structure line by line
+            with each line prefixed by the same characters
+            """
+            nonlocal file_count
+            nonlocal dir_count
+            contents = sorted(dir_path.iterdir())
+            # contents each get pointers that are ├── with a final └── :
+            pointers = [tee] * (len(contents) - 1) + [last]
+            for pointer, path in zip(pointers, contents):
+                if path.is_file():
+                    file_count += 1
+                yield prefix + pointer + path.name
+                if path.is_dir():  # extend the prefix and recurse:
+                    dir_count += 1
+                    extension = branch if pointer == tee else space
+                    # i.e. space because last, └── , above so no more |
+                    yield from tree(path, prefix=prefix + extension)
+
+        for line in tree(dir):
+            output += line + "\n"
+        output += f"\n{dir_count} directories, {file_count} files"
         st.markdown("Generated files and directory structure")
         st.code(output)
 
     def add_sidebar(self):
         def config(template_name):
             return import_from_file("template_config", f"./templates/{template_name}/_sidebar.py")
 
-        self.sidebar(self.codegen.template_list, config)
+        self.sidebar([*FOLDER_TO_TEMPLATE_NAME], config)
 
     def add_content(self):
         """Get generated/rendered code from the codegen."""
         content = [*self.codegen.render_templates(self.template_name, self.config)]
-        if st.checkbox("View rendered code ?"):
+        if st.checkbox("View rendered code ?", value=True):
             for fname, code in content:
-                self.render_code(fname, code)
+                if len(code):  # don't show files which don't have content in them
+                    self.render_code(fname, code)
 
     def add_download(self):
         st.markdown("")
@@ -94,12 +144,13 @@ def add_download(self):
                 shutil.copy(archive_fname, dist_path)
                 st.success(f"Download link : [{archive_fname}](./static/{archive_fname})")
                 with col2:
-                    self.render_directory(os.path.join(self.codegen.dist_dir, self.template_name))
+                    self.render_directory(Path(self.codegen.dist_dir, self.template_name))
 
     def run(self):
         self.add_sidebar()
         self.add_content()
         self.add_download()
+        st.info(TIP)
 
 
 def main():

requirements-dev.txt

@@ -1,3 +1,5 @@
+-r requirements.txt
+
 # dev
 pytorch-ignite
 torch

templates/_base/_argparse.pyi

@@ -0,0 +1,141 @@
+{% block imports %}
+from argparse import ArgumentParser
+{% endblock %}
+
+{% block defaults %}
+DEFAULTS = {
+    "use_amp": {
+        "action": "store_true",
+        "help": "use torch.cuda.amp for automatic mixed precision"
+    },
+    "resume_from": {
+        "default": None,
+        "type": str,
+        "help": "path to the checkpoint file to resume, can also url starting with https (None)"
+    },
+    "seed": {
+        "default": 666,
+        "type": int,
+        "help": "seed to use in ignite.utils.manual_seed() (666)"
+    },
+    "verbose": {
+        "action": "store_true",
+        "help": "use logging.INFO in ignite.utils.setup_logger",
+    },
+
+    # distributed training options
+    "backend": {
+        "default": None,
+        "type": str,
+        "help": "backend to use for distributed training (None)",
+    },
+    "nproc_per_node": {
+        "default": {{nproc_per_node}},
+        "type": int,
+        "help": """number of processes to launch on each node, for GPU training
+                this is recommended to be set to the number of GPUs in your system
+                so that each process can be bound to a single GPU ({{ nproc_per_node }})""",
+    },
+    "nnodes": {
+        "default": {{nnodes}},
+        "type": int,
+        "help": "number of nodes to use for distributed training ({{ nnodes }})",
+    },
+    "node_rank": {
+        "default": {{node_rank}},
+        "type": int,
+        "help": "rank of the node for multi-node distributed training ({{ node_rank }})",
+    },
+    "master_addr": {
+        "default": {{master_addr}},
+        "type": str,
+        "help": "master node TCP/IP address for torch native backends ({{ master_addr }})",
+    },
+    "master_port": {
+        "default": {{master_port}},
+        "type": int,
+        "help": "master node port for torch native backends ({{ master_port }})",
+    },
+
+    # ignite handlers options
+    "output_path": {
+        "default": "{{output_path}}",
+        "type": str,
+        "help": "output path to indicate where to_save objects are stored ({{output_path}})",
+    },
+    "save_every_iters": {
+        "default": {{save_every_iters}},
+        "type": int,
+        "help": "Saving iteration interval ({{save_every_iters}})",
+    },
+    "n_saved": {
+        "default": {{n_saved}},
+        "type": int,
+        "help": "number of best models to store ({{ n_saved }})",
+    },
+    "log_every_iters": {
+        "default": {{log_every_iters}},
+        "type": int,
+        "help": "logging interval for iteration progress bar ({{log_every_iters}})",
+    },
+    "with_pbars": {
+        "default": {{with_pbars}},
+        "type": bool,
+        "help": "show epoch-wise and iteration-wise progress bars ({{with_pbars}})",
+    },
+    "with_pbar_on_iters": {
+        "default": {{with_pbar_on_iters}},
+        "type": bool,
+        "help": "show iteration progress bar or not ({{with_pbar_on_iters}})",
+    },
+    "stop_on_nan": {
+        "default": {{stop_on_nan}},
+        "type": bool,
+        "help": "stop the training if engine output contains NaN/inf values (stop_on_nan)",
+    },
+    "clear_cuda_cache": {
+        "default": {{clear_cuda_cache}},
+        "type": bool,
+        "help": "clear cuda cache every end of epoch ({{clear_cuda_cache}})",
+    },
+    "with_gpu_stats": {
+        "default": {{with_gpu_stats}},
+        "type": bool,
+        "help": "show gpu information, requires pynvml ({{with_gpu_stats}})",
+    },
+    "patience": {
+        "default": {{patience}},
+        "type": int,
+        "help": "number of events to wait if no improvement and then stop the training ({{patience}})"
+    },
+    "limit_sec": {
+        "default": {{limit_sec}},
+        "type": int,
+        "help": "maximum time before training terminates in seconds ({{limit_sec}})"
+    },
+
+    # ignite logger options
+    "filepath": {
+        "default": "{{ filepath }}",
+        "type": str,
+        "help": "logging file path ({{ filepath }})",
+    },
+    "logger_log_every_iters": {
+        "default": {{logger_log_every_iters}},
+        "type": int,
+        "help": "logging interval for experiment tracking system ({{logger_log_every_iters}})",
+    },
+}
+{% endblock %}
+
+
+{% block get_default_parser %}
+def get_default_parser() -> ArgumentParser:
+    """Get the default configs for training."""
+    parser = ArgumentParser(add_help=False)
+
+    for key, value in DEFAULTS.items():
+        parser.add_argument(f"--{key}", **value)
+
+    return parser
+{% endblock %}

templates/_base/_events.pyi

@@ -0,0 +1,28 @@
+"""
+Additional Events to inspect the training at custom events.
+"""
+
+from ignite.engine.events import EventEnum
+
+
+class TrainEvents(EventEnum):
+    """Additional Training Events. includes
+
+    - BACKWARD_COMPLETED : trigger after calling loss.backward()
+    - OPTIM_STEP_COMPLETED : trigger after calling optimizer.step()
+    """
+
+    BACKWARD_COMPLETED = "backward_completed"
+    OPTIM_STEP_COMPLETED = "optim_step_completed"
+
+
+# define events and attribute mapping
+# so that we can trigger them with custom filter function
+train_events_to_attr = {
+    TrainEvents.BACKWARD_COMPLETED: "backward_completed",
+    TrainEvents.OPTIM_STEP_COMPLETED: "optim_step_completed",
+}
+
+# Any custom events can go below
+# fire them in process_function of the respective engine and
+# register them with the respective engine in the `engines.py`